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         refcount_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 (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 300                 if (!list_empty(&lo->plh_segs))
 301                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 302                 pnfs_detach_layout_hdr(lo);
 303                 spin_unlock(&inode->i_lock);
 304                 pnfs_free_layout_hdr(lo);
 305         }
 306 }
 307
 308 static void
 309 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 310                          u32 seq)
 311 {
 312         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 313                 iomode = IOMODE_ANY;
 314         lo->plh_return_iomode = iomode;
 315         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 316         if (seq != 0) {
 317                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 318                 lo->plh_return_seq = seq;
 319         }
 320 }
 321
 322 static void
 323 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 324 {
 325         struct pnfs_layout_segment *lseg;
 326         lo->plh_return_iomode = 0;
 327         lo->plh_return_seq = 0;
 328         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 329         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 330                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 331                         continue;
 332                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 333         }
 334 }
 335
 336 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 337 {
 338         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 339         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 340         smp_mb__after_atomic();
 341         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 342         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 343 }
 344
 345 static void
 346 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 347                 struct list_head *free_me)
 348 {
 349         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 350         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 351         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 352                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 353         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 354                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 355 }
 356
 357 /*
 358  * Update the seqid of a layout stateid
 359  */
 360 bool nfs4_refresh_layout_stateid(nfs4_stateid *dst, struct inode *inode)
 361 {
 362         struct pnfs_layout_hdr *lo;
 363         bool ret = false;
 364
 365         spin_lock(&inode->i_lock);
 366         lo = NFS_I(inode)->layout;
 367         if (lo && nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 368                 dst->seqid = lo->plh_stateid.seqid;
 369                 ret = true;
 370         }
 371         spin_unlock(&inode->i_lock);
 372         return ret;
 373 }
 374
 375 /*
 376  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 377  *
 378  * In order to continue using the pnfs_layout_hdr, a full recovery
 379  * is required.
 380  * Note that caller must hold inode->i_lock.
 381  */
 382 int
 383 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 384                 struct list_head *lseg_list)
 385 {
 386         struct pnfs_layout_range range = {
 387                 .iomode = IOMODE_ANY,
 388                 .offset = 0,
 389                 .length = NFS4_MAX_UINT64,
 390         };
 391         struct pnfs_layout_segment *lseg, *next;
 392
 393         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 394         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 395                 pnfs_clear_lseg_state(lseg, lseg_list);
 396         pnfs_clear_layoutreturn_info(lo);
 397         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 398         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 399             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 400                 pnfs_clear_layoutreturn_waitbit(lo);
 401         return !list_empty(&lo->plh_segs);
 402 }
 403
 404 static int
 405 pnfs_iomode_to_fail_bit(u32 iomode)
 406 {
 407         return iomode == IOMODE_RW ?
 408                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 409 }
 410
 411 static void
 412 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 413 {
 414         lo->plh_retry_timestamp = jiffies;
 415         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 416                 refcount_inc(&lo->plh_refcount);
 417 }
 418
 419 static void
 420 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 421 {
 422         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 423                 refcount_dec(&lo->plh_refcount);
 424 }
 425
 426 static void
 427 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 428 {
 429         struct inode *inode = lo->plh_inode;
 430         struct pnfs_layout_range range = {
 431                 .iomode = iomode,
 432                 .offset = 0,
 433                 .length = NFS4_MAX_UINT64,
 434         };
 435         LIST_HEAD(head);
 436
 437         spin_lock(&inode->i_lock);
 438         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 439         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 440         spin_unlock(&inode->i_lock);
 441         pnfs_free_lseg_list(&head);
 442         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 443                         iomode == IOMODE_RW ?  "RW" : "READ");
 444 }
 445
 446 static bool
 447 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 448 {
 449         unsigned long start, end;
 450         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 451
 452         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 453                 return false;
 454         end = jiffies;
 455         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 456         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 457                 /* It is time to retry the failed layoutgets */
 458                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 459                 return false;
 460         }
 461         return true;
 462 }
 463
 464 static void
 465 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 466                 const struct pnfs_layout_range *range,
 467                 const nfs4_stateid *stateid)
 468 {
 469         INIT_LIST_HEAD(&lseg->pls_list);
 470         INIT_LIST_HEAD(&lseg->pls_lc_list);
 471         refcount_set(&lseg->pls_refcount, 1);
 472         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 473         lseg->pls_layout = lo;
 474         lseg->pls_range = *range;
 475         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 476 }
 477
 478 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 479 {
 480         if (lseg != NULL) {
 481                 struct inode *inode = lseg->pls_layout->plh_inode;
 482                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 483         }
 484 }
 485
 486 static void
 487 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 488                 struct pnfs_layout_segment *lseg)
 489 {
 490         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 491         list_del_init(&lseg->pls_list);
 492         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 493         refcount_dec(&lo->plh_refcount);
 494         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 495                 return;
 496         if (list_empty(&lo->plh_segs) &&
 497             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 498             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 499                 if (atomic_read(&lo->plh_outstanding) == 0)
 500                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 501                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 502         }
 503 }
 504
 505 static bool
 506 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 507                 struct pnfs_layout_segment *lseg)
 508 {
 509         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 510             pnfs_layout_is_valid(lo)) {
 511                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 512                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 513                 return true;
 514         }
 515         return false;
 516 }
 517
 518 void
 519 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 520 {
 521         struct pnfs_layout_hdr *lo;
 522         struct inode *inode;
 523
 524         if (!lseg)
 525                 return;
 526
 527         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 528                 refcount_read(&lseg->pls_refcount),
 529                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 530
 531         lo = lseg->pls_layout;
 532         inode = lo->plh_inode;
 533
 534         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 535                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 536                         spin_unlock(&inode->i_lock);
 537                         return;
 538                 }
 539                 pnfs_get_layout_hdr(lo);
 540                 pnfs_layout_remove_lseg(lo, lseg);
 541                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 542                         lseg = NULL;
 543                 spin_unlock(&inode->i_lock);
 544                 pnfs_free_lseg(lseg);
 545                 pnfs_put_layout_hdr(lo);
 546         }
 547 }
 548 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 549
 550 /*
 551  * is l2 fully contained in l1?
 552  *   start1                             end1
 553  *   [----------------------------------)
 554  *           start2           end2
 555  *           [----------------)
 556  */
 557 static bool
 558 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 559                  const struct pnfs_layout_range *l2)
 560 {
 561         u64 start1 = l1->offset;
 562         u64 end1 = pnfs_end_offset(start1, l1->length);
 563         u64 start2 = l2->offset;
 564         u64 end2 = pnfs_end_offset(start2, l2->length);
 565
 566         return (start1 <= start2) && (end1 >= end2);
 567 }
 568
 569 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 570                 struct list_head *tmp_list)
 571 {
 572         if (!refcount_dec_and_test(&lseg->pls_refcount))
 573                 return false;
 574         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 575         list_add(&lseg->pls_list, tmp_list);
 576         return true;
 577 }
 578
 579 /* Returns 1 if lseg is removed from list, 0 otherwise */
 580 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 581                              struct list_head *tmp_list)
 582 {
 583         int rv = 0;
 584
 585         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 586                 /* Remove the reference keeping the lseg in the
 587                  * list.  It will now be removed when all
 588                  * outstanding io is finished.
 589                  */
 590                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 591                         refcount_read(&lseg->pls_refcount));
 592                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 593                         rv = 1;
 594         }
 595         return rv;
 596 }
 597
 598 /*
 599  * Compare 2 layout stateid sequence ids, to see which is newer,
 600  * taking into account wraparound issues.
 601  */
 602 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 603 {
 604         return (s32)(s1 - s2) > 0;
 605 }
 606
 607 static bool
 608 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 609                  const struct pnfs_layout_range *recall_range)
 610 {
 611         return (recall_range->iomode == IOMODE_ANY ||
 612                 lseg_range->iomode == recall_range->iomode) &&
 613                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 614 }
 615
 616 static bool
 617 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 618                 const struct pnfs_layout_range *recall_range,
 619                 u32 seq)
 620 {
 621         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 622                 return false;
 623         if (recall_range == NULL)
 624                 return true;
 625         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 626 }
 627
 628 /**
 629  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 630  * @lo: layout header containing the lsegs
 631  * @tmp_list: list head where doomed lsegs should go
 632  * @recall_range: optional recall range argument to match (may be NULL)
 633  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 634  *
 635  * Walk the list of lsegs in the layout header, and tear down any that should
 636  * be destroyed. If "recall_range" is specified then the segment must match
 637  * that range. If "seq" is non-zero, then only match segments that were handed
 638  * out at or before that sequence.
 639  *
 640  * Returns number of matching invalid lsegs remaining in list after scanning
 641  * it and purging them.
 642  */
 643 int
 644 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 645                             struct list_head *tmp_list,
 646                             const struct pnfs_layout_range *recall_range,
 647                             u32 seq)
 648 {
 649         struct pnfs_layout_segment *lseg, *next;
 650         int remaining = 0;
 651
 652         dprintk("%s:Begin lo %p\n", __func__, lo);
 653
 654         if (list_empty(&lo->plh_segs))
 655                 return 0;
 656         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 657                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 658                         dprintk("%s: freeing lseg %p iomode %d seq %u"
 659                                 "offset %llu length %llu\n", __func__,
 660                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 661                                 lseg->pls_range.offset, lseg->pls_range.length);
 662                         if (!mark_lseg_invalid(lseg, tmp_list))
 663                                 remaining++;
 664                 }
 665         dprintk("%s:Return %i\n", __func__, remaining);
 666         return remaining;
 667 }
 668
 669 static void
 670 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 671                 struct list_head *free_me,
 672                 const struct pnfs_layout_range *range,
 673                 u32 seq)
 674 {
 675         struct pnfs_layout_segment *lseg, *next;
 676
 677         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 678                 if (pnfs_match_lseg_recall(lseg, range, seq))
 679                         list_move_tail(&lseg->pls_list, free_me);
 680         }
 681 }
 682
 683 /* note free_me must contain lsegs from a single layout_hdr */
 684 void
 685 pnfs_free_lseg_list(struct list_head *free_me)
 686 {
 687         struct pnfs_layout_segment *lseg, *tmp;
 688
 689         if (list_empty(free_me))
 690                 return;
 691
 692         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 693                 list_del(&lseg->pls_list);
 694                 pnfs_free_lseg(lseg);
 695         }
 696 }
 697
 698 void
 699 pnfs_destroy_layout(struct nfs_inode *nfsi)
 700 {
 701         struct pnfs_layout_hdr *lo;
 702         LIST_HEAD(tmp_list);
 703
 704         spin_lock(&nfsi->vfs_inode.i_lock);
 705         lo = nfsi->layout;
 706         if (lo) {
 707                 pnfs_get_layout_hdr(lo);
 708                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 709                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 710                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 711                 spin_unlock(&nfsi->vfs_inode.i_lock);
 712                 pnfs_free_lseg_list(&tmp_list);
 713                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 714                 pnfs_put_layout_hdr(lo);
 715         } else
 716                 spin_unlock(&nfsi->vfs_inode.i_lock);
 717 }
 718 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 719
 720 static bool
 721 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 722                 struct list_head *layout_list)
 723 {
 724         struct pnfs_layout_hdr *lo;
 725         bool ret = false;
 726
 727         spin_lock(&inode->i_lock);
 728         lo = NFS_I(inode)->layout;
 729         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 730                 pnfs_get_layout_hdr(lo);
 731                 list_add(&lo->plh_bulk_destroy, layout_list);
 732                 ret = true;
 733         }
 734         spin_unlock(&inode->i_lock);
 735         return ret;
 736 }
 737
 738 /* Caller must hold rcu_read_lock and clp->cl_lock */
 739 static int
 740 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 741                 struct nfs_server *server,
 742                 struct list_head *layout_list)
 743 {
 744         struct pnfs_layout_hdr *lo, *next;
 745         struct inode *inode;
 746
 747         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 748                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
 749                         continue;
 750                 inode = igrab(lo->plh_inode);
 751                 if (inode == NULL)
 752                         continue;
 753                 list_del_init(&lo->plh_layouts);
 754                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 755                         continue;
 756                 rcu_read_unlock();
 757                 spin_unlock(&clp->cl_lock);
 758                 iput(inode);
 759                 spin_lock(&clp->cl_lock);
 760                 rcu_read_lock();
 761                 return -EAGAIN;
 762         }
 763         return 0;
 764 }
 765
 766 static int
 767 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 768                 bool is_bulk_recall)
 769 {
 770         struct pnfs_layout_hdr *lo;
 771         struct inode *inode;
 772         LIST_HEAD(lseg_list);
 773         int ret = 0;
 774
 775         while (!list_empty(layout_list)) {
 776                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 777                                 plh_bulk_destroy);
 778                 dprintk("%s freeing layout for inode %lu\n", __func__,
 779                         lo->plh_inode->i_ino);
 780                 inode = lo->plh_inode;
 781
 782                 pnfs_layoutcommit_inode(inode, false);
 783
 784                 spin_lock(&inode->i_lock);
 785                 list_del_init(&lo->plh_bulk_destroy);
 786                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 787                         if (is_bulk_recall)
 788                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 789                         ret = -EAGAIN;
 790                 }
 791                 spin_unlock(&inode->i_lock);
 792                 pnfs_free_lseg_list(&lseg_list);
 793                 /* Free all lsegs that are attached to commit buckets */
 794                 nfs_commit_inode(inode, 0);
 795                 pnfs_put_layout_hdr(lo);
 796                 iput(inode);
 797         }
 798         return ret;
 799 }
 800
 801 int
 802 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 803                 struct nfs_fsid *fsid,
 804                 bool is_recall)
 805 {
 806         struct nfs_server *server;
 807         LIST_HEAD(layout_list);
 808
 809         spin_lock(&clp->cl_lock);
 810         rcu_read_lock();
 811 restart:
 812         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 813                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 814                         continue;
 815                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 816                                 server,
 817                                 &layout_list) != 0)
 818                         goto restart;
 819         }
 820         rcu_read_unlock();
 821         spin_unlock(&clp->cl_lock);
 822
 823         if (list_empty(&layout_list))
 824                 return 0;
 825         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 826 }
 827
 828 int
 829 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 830                 bool is_recall)
 831 {
 832         struct nfs_server *server;
 833         LIST_HEAD(layout_list);
 834
 835         spin_lock(&clp->cl_lock);
 836         rcu_read_lock();
 837 restart:
 838         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 839                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 840                                         server,
 841                                         &layout_list) != 0)
 842                         goto restart;
 843         }
 844         rcu_read_unlock();
 845         spin_unlock(&clp->cl_lock);
 846
 847         if (list_empty(&layout_list))
 848                 return 0;
 849         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 850 }
 851
 852 /*
 853  * Called by the state manger to remove all layouts established under an
 854  * expired lease.
 855  */
 856 void
 857 pnfs_destroy_all_layouts(struct nfs_client *clp)
 858 {
 859         nfs4_deviceid_mark_client_invalid(clp);
 860         nfs4_deviceid_purge_client(clp);
 861
 862         pnfs_destroy_layouts_byclid(clp, false);
 863 }
 864
 865 /* update lo->plh_stateid with new if is more recent */
 866 void
 867 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 868                         bool update_barrier)
 869 {
 870         u32 oldseq, newseq, new_barrier = 0;
 871
 872         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 873         newseq = be32_to_cpu(new->seqid);
 874
 875         if (!pnfs_layout_is_valid(lo)) {
 876                 nfs4_stateid_copy(&lo->plh_stateid, new);
 877                 lo->plh_barrier = newseq;
 878                 pnfs_clear_layoutreturn_info(lo);
 879                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 880                 return;
 881         }
 882         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 883                 nfs4_stateid_copy(&lo->plh_stateid, new);
 884                 /*
 885                  * Because of wraparound, we want to keep the barrier
 886                  * "close" to the current seqids.
 887                  */
 888                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 889         }
 890         if (update_barrier)
 891                 new_barrier = be32_to_cpu(new->seqid);
 892         else if (new_barrier == 0)
 893                 return;
 894         if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 895                 lo->plh_barrier = new_barrier;
 896 }
 897
 898 static bool
 899 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 900                 const nfs4_stateid *stateid)
 901 {
 902         u32 seqid = be32_to_cpu(stateid->seqid);
 903
 904         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 905 }
 906
 907 /* lget is set to 1 if called from inside send_layoutget call chain */
 908 static bool
 909 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 910 {
 911         return lo->plh_block_lgets ||
 912                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 913 }
 914
 915 /*
 916  * Get layout from server.
 917  *    for now, assume that whole file layouts are requested.
 918  *    arg->offset: 0
 919  *    arg->length: all ones
 920  */
 921 static struct pnfs_layout_segment *
 922 send_layoutget(struct pnfs_layout_hdr *lo,
 923            struct nfs_open_context *ctx,
 924            nfs4_stateid *stateid,
 925            const struct pnfs_layout_range *range,
 926            long *timeout, gfp_t gfp_flags)
 927 {
 928         struct inode *ino = lo->plh_inode;
 929         struct nfs_server *server = NFS_SERVER(ino);
 930         struct nfs4_layoutget *lgp;
 931         loff_t i_size;
 932
 933         dprintk("--> %s\n", __func__);
 934
 935         /*
 936          * Synchronously retrieve layout information from server and
 937          * store in lseg. If we race with a concurrent seqid morphing
 938          * op, then re-send the LAYOUTGET.
 939          */
 940         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 941         if (lgp == NULL)
 942                 return ERR_PTR(-ENOMEM);
 943
 944         i_size = i_size_read(ino);
 945
 946         lgp->args.minlength = PAGE_SIZE;
 947         if (lgp->args.minlength > range->length)
 948                 lgp->args.minlength = range->length;
 949         if (range->iomode == IOMODE_READ) {
 950                 if (range->offset >= i_size)
 951                         lgp->args.minlength = 0;
 952                 else if (i_size - range->offset < lgp->args.minlength)
 953                         lgp->args.minlength = i_size - range->offset;
 954         }
 955         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 956         pnfs_copy_range(&lgp->args.range, range);
 957         lgp->args.type = server->pnfs_curr_ld->id;
 958         lgp->args.inode = ino;
 959         lgp->args.ctx = get_nfs_open_context(ctx);
 960         nfs4_stateid_copy(&lgp->args.stateid, stateid);
 961         lgp->gfp_flags = gfp_flags;
 962         lgp->cred = lo->plh_lc_cred;
 963
 964         return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
 965 }
 966
 967 static void pnfs_clear_layoutcommit(struct inode *inode,
 968                 struct list_head *head)
 969 {
 970         struct nfs_inode *nfsi = NFS_I(inode);
 971         struct pnfs_layout_segment *lseg, *tmp;
 972
 973         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 974                 return;
 975         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 976                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 977                         continue;
 978                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 979         }
 980 }
 981
 982 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
 983                 const nfs4_stateid *arg_stateid,
 984                 const struct pnfs_layout_range *range,
 985                 const nfs4_stateid *stateid)
 986 {
 987         struct inode *inode = lo->plh_inode;
 988         LIST_HEAD(freeme);
 989
 990         spin_lock(&inode->i_lock);
 991         if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
 992             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
 993                 goto out_unlock;
 994         if (stateid) {
 995                 u32 seq = be32_to_cpu(arg_stateid->seqid);
 996
 997                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
 998                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
 999                 pnfs_set_layout_stateid(lo, stateid, true);
1000         } else
1001                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1002 out_unlock:
1003         pnfs_clear_layoutreturn_waitbit(lo);
1004         spin_unlock(&inode->i_lock);
1005         pnfs_free_lseg_list(&freeme);
1006
1007 }
1008
1009 static bool
1010 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1011                 nfs4_stateid *stateid,
1012                 enum pnfs_iomode *iomode)
1013 {
1014         /* Serialise LAYOUTGET/LAYOUTRETURN */
1015         if (atomic_read(&lo->plh_outstanding) != 0)
1016                 return false;
1017         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1018                 return false;
1019         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1020         pnfs_get_layout_hdr(lo);
1021         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1022                 if (stateid != NULL) {
1023                         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1024                         if (lo->plh_return_seq != 0)
1025                                 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1026                 }
1027                 if (iomode != NULL)
1028                         *iomode = lo->plh_return_iomode;
1029                 pnfs_clear_layoutreturn_info(lo);
1030                 return true;
1031         }
1032         if (stateid != NULL)
1033                 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1034         if (iomode != NULL)
1035                 *iomode = IOMODE_ANY;
1036         return true;
1037 }
1038
1039 static void
1040 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1041                 struct pnfs_layout_hdr *lo,
1042                 const nfs4_stateid *stateid,
1043                 enum pnfs_iomode iomode)
1044 {
1045         struct inode *inode = lo->plh_inode;
1046
1047         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1048         args->inode = inode;
1049         args->range.iomode = iomode;
1050         args->range.offset = 0;
1051         args->range.length = NFS4_MAX_UINT64;
1052         args->layout = lo;
1053         nfs4_stateid_copy(&args->stateid, stateid);
1054 }
1055
1056 static int
1057 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1058                        enum pnfs_iomode iomode, bool sync)
1059 {
1060         struct inode *ino = lo->plh_inode;
1061         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1062         struct nfs4_layoutreturn *lrp;
1063         int status = 0;
1064
1065         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1066         if (unlikely(lrp == NULL)) {
1067                 status = -ENOMEM;
1068                 spin_lock(&ino->i_lock);
1069                 pnfs_clear_layoutreturn_waitbit(lo);
1070                 spin_unlock(&ino->i_lock);
1071                 pnfs_put_layout_hdr(lo);
1072                 goto out;
1073         }
1074
1075         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1076         lrp->args.ld_private = &lrp->ld_private;
1077         lrp->clp = NFS_SERVER(ino)->nfs_client;
1078         lrp->cred = lo->plh_lc_cred;
1079         if (ld->prepare_layoutreturn)
1080                 ld->prepare_layoutreturn(&lrp->args);
1081
1082         status = nfs4_proc_layoutreturn(lrp, sync);
1083 out:
1084         dprintk("<-- %s status: %d\n", __func__, status);
1085         return status;
1086 }
1087
1088 /* Return true if layoutreturn is needed */
1089 static bool
1090 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1091 {
1092         struct pnfs_layout_segment *s;
1093
1094         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1095                 return false;
1096
1097         /* Defer layoutreturn until all lsegs are done */
1098         list_for_each_entry(s, &lo->plh_segs, pls_list) {
1099                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1100                         return false;
1101         }
1102
1103         return true;
1104 }
1105
1106 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1107 {
1108         struct inode *inode= lo->plh_inode;
1109
1110         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1111                 return;
1112         spin_lock(&inode->i_lock);
1113         if (pnfs_layout_need_return(lo)) {
1114                 nfs4_stateid stateid;
1115                 enum pnfs_iomode iomode;
1116                 bool send;
1117
1118                 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1119                 spin_unlock(&inode->i_lock);
1120                 if (send) {
1121                         /* Send an async layoutreturn so we dont deadlock */
1122                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1123                 }
1124         } else
1125                 spin_unlock(&inode->i_lock);
1126 }
1127
1128 /*
1129  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1130  * when the layout segment list is empty.
1131  *
1132  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1133  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1134  * deviceid is marked invalid.
1135  */
1136 int
1137 _pnfs_return_layout(struct inode *ino)
1138 {
1139         struct pnfs_layout_hdr *lo = NULL;
1140         struct nfs_inode *nfsi = NFS_I(ino);
1141         LIST_HEAD(tmp_list);
1142         nfs4_stateid stateid;
1143         int status = 0;
1144         bool send;
1145
1146         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1147
1148         spin_lock(&ino->i_lock);
1149         lo = nfsi->layout;
1150         if (!lo) {
1151                 spin_unlock(&ino->i_lock);
1152                 dprintk("NFS: %s no layout to return\n", __func__);
1153                 goto out;
1154         }
1155         /* Reference matched in nfs4_layoutreturn_release */
1156         pnfs_get_layout_hdr(lo);
1157         /* Is there an outstanding layoutreturn ? */
1158         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1159                 spin_unlock(&ino->i_lock);
1160                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1161                                         TASK_UNINTERRUPTIBLE))
1162                         goto out_put_layout_hdr;
1163                 spin_lock(&ino->i_lock);
1164         }
1165         pnfs_clear_layoutcommit(ino, &tmp_list);
1166         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1167
1168         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1169                 struct pnfs_layout_range range = {
1170                         .iomode         = IOMODE_ANY,
1171                         .offset         = 0,
1172                         .length         = NFS4_MAX_UINT64,
1173                 };
1174                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1175         }
1176
1177         /* Don't send a LAYOUTRETURN if list was initially empty */
1178         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1179                 spin_unlock(&ino->i_lock);
1180                 dprintk("NFS: %s no layout segments to return\n", __func__);
1181                 goto out_put_layout_hdr;
1182         }
1183
1184         send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1185         spin_unlock(&ino->i_lock);
1186         if (send)
1187                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1188 out_put_layout_hdr:
1189         pnfs_free_lseg_list(&tmp_list);
1190         pnfs_put_layout_hdr(lo);
1191 out:
1192         dprintk("<-- %s status: %d\n", __func__, status);
1193         return status;
1194 }
1195
1196 int
1197 pnfs_commit_and_return_layout(struct inode *inode)
1198 {
1199         struct pnfs_layout_hdr *lo;
1200         int ret;
1201
1202         spin_lock(&inode->i_lock);
1203         lo = NFS_I(inode)->layout;
1204         if (lo == NULL) {
1205                 spin_unlock(&inode->i_lock);
1206                 return 0;
1207         }
1208         pnfs_get_layout_hdr(lo);
1209         /* Block new layoutgets and read/write to ds */
1210         lo->plh_block_lgets++;
1211         spin_unlock(&inode->i_lock);
1212         filemap_fdatawait(inode->i_mapping);
1213         ret = pnfs_layoutcommit_inode(inode, true);
1214         if (ret == 0)
1215                 ret = _pnfs_return_layout(inode);
1216         spin_lock(&inode->i_lock);
1217         lo->plh_block_lgets--;
1218         spin_unlock(&inode->i_lock);
1219         pnfs_put_layout_hdr(lo);
1220         return ret;
1221 }
1222
1223 bool pnfs_roc(struct inode *ino,
1224                 struct nfs4_layoutreturn_args *args,
1225                 struct nfs4_layoutreturn_res *res,
1226                 const struct rpc_cred *cred)
1227 {
1228         struct nfs_inode *nfsi = NFS_I(ino);
1229         struct nfs_open_context *ctx;
1230         struct nfs4_state *state;
1231         struct pnfs_layout_hdr *lo;
1232         struct pnfs_layout_segment *lseg, *next;
1233         nfs4_stateid stateid;
1234         enum pnfs_iomode iomode = 0;
1235         bool layoutreturn = false, roc = false;
1236         bool skip_read = false;
1237
1238         if (!nfs_have_layout(ino))
1239                 return false;
1240 retry:
1241         spin_lock(&ino->i_lock);
1242         lo = nfsi->layout;
1243         if (!lo || !pnfs_layout_is_valid(lo) ||
1244             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1245                 goto out_noroc;
1246         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1247                 pnfs_get_layout_hdr(lo);
1248                 spin_unlock(&ino->i_lock);
1249                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1250                                 TASK_UNINTERRUPTIBLE);
1251                 pnfs_put_layout_hdr(lo);
1252                 goto retry;
1253         }
1254
1255         /* no roc if we hold a delegation */
1256         if (nfs4_check_delegation(ino, FMODE_READ)) {
1257                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1258                         goto out_noroc;
1259                 skip_read = true;
1260         }
1261
1262         list_for_each_entry(ctx, &nfsi->open_files, list) {
1263                 state = ctx->state;
1264                 if (state == NULL)
1265                         continue;
1266                 /* Don't return layout if there is open file state */
1267                 if (state->state & FMODE_WRITE)
1268                         goto out_noroc;
1269                 if (state->state & FMODE_READ)
1270                         skip_read = true;
1271         }
1272
1273
1274         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1275                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1276                         continue;
1277                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1278                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1279                         continue;
1280                 /*
1281                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1282                  * doesn't invalidate the layout for us.
1283                  */
1284                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1285                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1286                         continue;
1287                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1288         }
1289
1290         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1291                 goto out_noroc;
1292
1293         /* ROC in two conditions:
1294          * 1. there are ROC lsegs
1295          * 2. we don't send layoutreturn
1296          */
1297         /* lo ref dropped in pnfs_roc_release() */
1298         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1299         /* If the creds don't match, we can't compound the layoutreturn */
1300         if (!layoutreturn || cred != lo->plh_lc_cred)
1301                 goto out_noroc;
1302
1303         roc = layoutreturn;
1304         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1305         res->lrs_present = 0;
1306         layoutreturn = false;
1307
1308 out_noroc:
1309         spin_unlock(&ino->i_lock);
1310         pnfs_layoutcommit_inode(ino, true);
1311         if (roc) {
1312                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1313                 if (ld->prepare_layoutreturn)
1314                         ld->prepare_layoutreturn(args);
1315                 return true;
1316         }
1317         if (layoutreturn)
1318                 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1319         return false;
1320 }
1321
1322 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1323                 struct nfs4_layoutreturn_res *res,
1324                 int ret)
1325 {
1326         struct pnfs_layout_hdr *lo = args->layout;
1327         const nfs4_stateid *arg_stateid = NULL;
1328         const nfs4_stateid *res_stateid = NULL;
1329         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1330
1331         if (ret == 0) {
1332                 arg_stateid = &args->stateid;
1333                 if (res->lrs_present)
1334                         res_stateid = &res->stateid;
1335         }
1336         pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1337                         res_stateid);
1338         if (ld_private && ld_private->ops && ld_private->ops->free)
1339                 ld_private->ops->free(ld_private);
1340         pnfs_put_layout_hdr(lo);
1341         trace_nfs4_layoutreturn_on_close(args->inode, 0);
1342 }
1343
1344 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1345 {
1346         struct nfs_inode *nfsi = NFS_I(ino);
1347         struct pnfs_layout_hdr *lo;
1348         bool sleep = false;
1349
1350         /* we might not have grabbed lo reference. so need to check under
1351          * i_lock */
1352         spin_lock(&ino->i_lock);
1353         lo = nfsi->layout;
1354         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1355                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1356                 sleep = true;
1357         }
1358         spin_unlock(&ino->i_lock);
1359         return sleep;
1360 }
1361
1362 /*
1363  * Compare two layout segments for sorting into layout cache.
1364  * We want to preferentially return RW over RO layouts, so ensure those
1365  * are seen first.
1366  */
1367 static s64
1368 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1369            const struct pnfs_layout_range *l2)
1370 {
1371         s64 d;
1372
1373         /* high offset > low offset */
1374         d = l1->offset - l2->offset;
1375         if (d)
1376                 return d;
1377
1378         /* short length > long length */
1379         d = l2->length - l1->length;
1380         if (d)
1381                 return d;
1382
1383         /* read > read/write */
1384         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1385 }
1386
1387 static bool
1388 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1389                 const struct pnfs_layout_range *l2)
1390 {
1391         return pnfs_lseg_range_cmp(l1, l2) > 0;
1392 }
1393
1394 static bool
1395 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1396                 struct pnfs_layout_segment *old)
1397 {
1398         return false;
1399 }
1400
1401 void
1402 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1403                    struct pnfs_layout_segment *lseg,
1404                    bool (*is_after)(const struct pnfs_layout_range *,
1405                            const struct pnfs_layout_range *),
1406                    bool (*do_merge)(struct pnfs_layout_segment *,
1407                            struct pnfs_layout_segment *),
1408                    struct list_head *free_me)
1409 {
1410         struct pnfs_layout_segment *lp, *tmp;
1411
1412         dprintk("%s:Begin\n", __func__);
1413
1414         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1415                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1416                         continue;
1417                 if (do_merge(lseg, lp)) {
1418                         mark_lseg_invalid(lp, free_me);
1419                         continue;
1420                 }
1421                 if (is_after(&lseg->pls_range, &lp->pls_range))
1422                         continue;
1423                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1424                 dprintk("%s: inserted lseg %p "
1425                         "iomode %d offset %llu length %llu before "
1426                         "lp %p iomode %d offset %llu length %llu\n",
1427                         __func__, lseg, lseg->pls_range.iomode,
1428                         lseg->pls_range.offset, lseg->pls_range.length,
1429                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1430                         lp->pls_range.length);
1431                 goto out;
1432         }
1433         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1434         dprintk("%s: inserted lseg %p "
1435                 "iomode %d offset %llu length %llu at tail\n",
1436                 __func__, lseg, lseg->pls_range.iomode,
1437                 lseg->pls_range.offset, lseg->pls_range.length);
1438 out:
1439         pnfs_get_layout_hdr(lo);
1440
1441         dprintk("%s:Return\n", __func__);
1442 }
1443 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1444
1445 static void
1446 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1447                    struct pnfs_layout_segment *lseg,
1448                    struct list_head *free_me)
1449 {
1450         struct inode *inode = lo->plh_inode;
1451         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1452
1453         if (ld->add_lseg != NULL)
1454                 ld->add_lseg(lo, lseg, free_me);
1455         else
1456                 pnfs_generic_layout_insert_lseg(lo, lseg,
1457                                 pnfs_lseg_range_is_after,
1458                                 pnfs_lseg_no_merge,
1459                                 free_me);
1460 }
1461
1462 static struct pnfs_layout_hdr *
1463 alloc_init_layout_hdr(struct inode *ino,
1464                       struct nfs_open_context *ctx,
1465                       gfp_t gfp_flags)
1466 {
1467         struct pnfs_layout_hdr *lo;
1468
1469         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1470         if (!lo)
1471                 return NULL;
1472         refcount_set(&lo->plh_refcount, 1);
1473         INIT_LIST_HEAD(&lo->plh_layouts);
1474         INIT_LIST_HEAD(&lo->plh_segs);
1475         INIT_LIST_HEAD(&lo->plh_return_segs);
1476         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1477         lo->plh_inode = ino;
1478         lo->plh_lc_cred = get_rpccred(ctx->cred);
1479         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1480         return lo;
1481 }
1482
1483 static struct pnfs_layout_hdr *
1484 pnfs_find_alloc_layout(struct inode *ino,
1485                        struct nfs_open_context *ctx,
1486                        gfp_t gfp_flags)
1487         __releases(&ino->i_lock)
1488         __acquires(&ino->i_lock)
1489 {
1490         struct nfs_inode *nfsi = NFS_I(ino);
1491         struct pnfs_layout_hdr *new = NULL;
1492
1493         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1494
1495         if (nfsi->layout != NULL)
1496                 goto out_existing;
1497         spin_unlock(&ino->i_lock);
1498         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1499         spin_lock(&ino->i_lock);
1500
1501         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1502                 nfsi->layout = new;
1503                 return new;
1504         } else if (new != NULL)
1505                 pnfs_free_layout_hdr(new);
1506 out_existing:
1507         pnfs_get_layout_hdr(nfsi->layout);
1508         return nfsi->layout;
1509 }
1510
1511 /*
1512  * iomode matching rules:
1513  * iomode       lseg    strict match
1514  *                      iomode
1515  * -----        -----   ------ -----
1516  * ANY          READ    N/A    true
1517  * ANY          RW      N/A    true
1518  * RW           READ    N/A    false
1519  * RW           RW      N/A    true
1520  * READ         READ    N/A    true
1521  * READ         RW      true   false
1522  * READ         RW      false  true
1523  */
1524 static bool
1525 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1526                  const struct pnfs_layout_range *range,
1527                  bool strict_iomode)
1528 {
1529         struct pnfs_layout_range range1;
1530
1531         if ((range->iomode == IOMODE_RW &&
1532              ls_range->iomode != IOMODE_RW) ||
1533             (range->iomode != ls_range->iomode &&
1534              strict_iomode) ||
1535             !pnfs_lseg_range_intersecting(ls_range, range))
1536                 return 0;
1537
1538         /* range1 covers only the first byte in the range */
1539         range1 = *range;
1540         range1.length = 1;
1541         return pnfs_lseg_range_contained(ls_range, &range1);
1542 }
1543
1544 /*
1545  * lookup range in layout
1546  */
1547 static struct pnfs_layout_segment *
1548 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1549                 struct pnfs_layout_range *range,
1550                 bool strict_iomode)
1551 {
1552         struct pnfs_layout_segment *lseg, *ret = NULL;
1553
1554         dprintk("%s:Begin\n", __func__);
1555
1556         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1557                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1558                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1559                     pnfs_lseg_range_match(&lseg->pls_range, range,
1560                                           strict_iomode)) {
1561                         ret = pnfs_get_lseg(lseg);
1562                         break;
1563                 }
1564         }
1565
1566         dprintk("%s:Return lseg %p ref %d\n",
1567                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1568         return ret;
1569 }
1570
1571 /*
1572  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1573  * to the MDS or over pNFS
1574  *
1575  * The nfs_inode read_io and write_io fields are cumulative counters reset
1576  * when there are no layout segments. Note that in pnfs_update_layout iomode
1577  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1578  * WRITE request.
1579  *
1580  * A return of true means use MDS I/O.
1581  *
1582  * From rfc 5661:
1583  * If a file's size is smaller than the file size threshold, data accesses
1584  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1585  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1586  * server.  If both file size and I/O size are provided, the client SHOULD
1587  * reach or exceed  both thresholds before sending its read or write
1588  * requests to the data server.
1589  */
1590 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1591                                      struct inode *ino, int iomode)
1592 {
1593         struct nfs4_threshold *t = ctx->mdsthreshold;
1594         struct nfs_inode *nfsi = NFS_I(ino);
1595         loff_t fsize = i_size_read(ino);
1596         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1597
1598         if (t == NULL)
1599                 return ret;
1600
1601         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1602                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1603
1604         switch (iomode) {
1605         case IOMODE_READ:
1606                 if (t->bm & THRESHOLD_RD) {
1607                         dprintk("%s fsize %llu\n", __func__, fsize);
1608                         size_set = true;
1609                         if (fsize < t->rd_sz)
1610                                 size = true;
1611                 }
1612                 if (t->bm & THRESHOLD_RD_IO) {
1613                         dprintk("%s nfsi->read_io %llu\n", __func__,
1614                                 nfsi->read_io);
1615                         io_set = true;
1616                         if (nfsi->read_io < t->rd_io_sz)
1617                                 io = true;
1618                 }
1619                 break;
1620         case IOMODE_RW:
1621                 if (t->bm & THRESHOLD_WR) {
1622                         dprintk("%s fsize %llu\n", __func__, fsize);
1623                         size_set = true;
1624                         if (fsize < t->wr_sz)
1625                                 size = true;
1626                 }
1627                 if (t->bm & THRESHOLD_WR_IO) {
1628                         dprintk("%s nfsi->write_io %llu\n", __func__,
1629                                 nfsi->write_io);
1630                         io_set = true;
1631                         if (nfsi->write_io < t->wr_io_sz)
1632                                 io = true;
1633                 }
1634                 break;
1635         }
1636         if (size_set && io_set) {
1637                 if (size && io)
1638                         ret = true;
1639         } else if (size || io)
1640                 ret = true;
1641
1642         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1643         return ret;
1644 }
1645
1646 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1647 {
1648         /*
1649          * send layoutcommit as it can hold up layoutreturn due to lseg
1650          * reference
1651          */
1652         pnfs_layoutcommit_inode(lo->plh_inode, false);
1653         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1654                                    nfs_wait_bit_killable,
1655                                    TASK_UNINTERRUPTIBLE);
1656 }
1657
1658 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1659 {
1660         unsigned long *bitlock = &lo->plh_flags;
1661
1662         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1663         smp_mb__after_atomic();
1664         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1665 }
1666
1667 /*
1668  * Layout segment is retreived from the server if not cached.
1669  * The appropriate layout segment is referenced and returned to the caller.
1670  */
1671 struct pnfs_layout_segment *
1672 pnfs_update_layout(struct inode *ino,
1673                    struct nfs_open_context *ctx,
1674                    loff_t pos,
1675                    u64 count,
1676                    enum pnfs_iomode iomode,
1677                    bool strict_iomode,
1678                    gfp_t gfp_flags)
1679 {
1680         struct pnfs_layout_range arg = {
1681                 .iomode = iomode,
1682                 .offset = pos,
1683                 .length = count,
1684         };
1685         unsigned pg_offset;
1686         struct nfs_server *server = NFS_SERVER(ino);
1687         struct nfs_client *clp = server->nfs_client;
1688         struct pnfs_layout_hdr *lo = NULL;
1689         struct pnfs_layout_segment *lseg = NULL;
1690         nfs4_stateid stateid;
1691         long timeout = 0;
1692         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1693         bool first;
1694
1695         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1696                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1697                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1698                 goto out;
1699         }
1700
1701         if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1702                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1703                                  PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1704                 goto out;
1705         }
1706
1707         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1708                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1709                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1710                 goto out;
1711         }
1712
1713 lookup_again:
1714         nfs4_client_recover_expired_lease(clp);
1715         first = false;
1716         spin_lock(&ino->i_lock);
1717         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1718         if (lo == NULL) {
1719                 spin_unlock(&ino->i_lock);
1720                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1721                                  PNFS_UPDATE_LAYOUT_NOMEM);
1722                 goto out;
1723         }
1724
1725         /* Do we even need to bother with this? */
1726         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1727                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1728                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1729                 dprintk("%s matches recall, use MDS\n", __func__);
1730                 goto out_unlock;
1731         }
1732
1733         /* if LAYOUTGET already failed once we don't try again */
1734         if (pnfs_layout_io_test_failed(lo, iomode)) {
1735                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1736                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1737                 goto out_unlock;
1738         }
1739
1740         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1741         if (lseg) {
1742                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1743                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1744                 goto out_unlock;
1745         }
1746
1747         if (!nfs4_valid_open_stateid(ctx->state)) {
1748                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1749                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1750                 goto out_unlock;
1751         }
1752
1753         /*
1754          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1755          * stateid, or it has been invalidated, then we must use the open
1756          * stateid.
1757          */
1758         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1759
1760                 /*
1761                  * The first layoutget for the file. Need to serialize per
1762                  * RFC 5661 Errata 3208.
1763                  */
1764                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1765                                      &lo->plh_flags)) {
1766                         spin_unlock(&ino->i_lock);
1767                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1768                                     TASK_UNINTERRUPTIBLE);
1769                         pnfs_put_layout_hdr(lo);
1770                         dprintk("%s retrying\n", __func__);
1771                         goto lookup_again;
1772                 }
1773
1774                 first = true;
1775                 if (nfs4_select_rw_stateid(ctx->state,
1776                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1777                                         NULL, &stateid, NULL) != 0) {
1778                         trace_pnfs_update_layout(ino, pos, count,
1779                                         iomode, lo, lseg,
1780                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1781                         goto out_unlock;
1782                 }
1783         } else {
1784                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1785         }
1786
1787         /*
1788          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1789          * for LAYOUTRETURN even if first is true.
1790          */
1791         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1792                 spin_unlock(&ino->i_lock);
1793                 dprintk("%s wait for layoutreturn\n", __func__);
1794                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1795                         if (first)
1796                                 pnfs_clear_first_layoutget(lo);
1797                         pnfs_put_layout_hdr(lo);
1798                         dprintk("%s retrying\n", __func__);
1799                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1800                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
1801                         goto lookup_again;
1802                 }
1803                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1804                                 PNFS_UPDATE_LAYOUT_RETURN);
1805                 goto out_put_layout_hdr;
1806         }
1807
1808         if (pnfs_layoutgets_blocked(lo)) {
1809                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1810                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1811                 goto out_unlock;
1812         }
1813         atomic_inc(&lo->plh_outstanding);
1814         spin_unlock(&ino->i_lock);
1815
1816         if (list_empty(&lo->plh_layouts)) {
1817                 /* The lo must be on the clp list if there is any
1818                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1819                  */
1820                 spin_lock(&clp->cl_lock);
1821                 if (list_empty(&lo->plh_layouts))
1822                         list_add_tail(&lo->plh_layouts, &server->layouts);
1823                 spin_unlock(&clp->cl_lock);
1824         }
1825
1826         pg_offset = arg.offset & ~PAGE_MASK;
1827         if (pg_offset) {
1828                 arg.offset -= pg_offset;
1829                 arg.length += pg_offset;
1830         }
1831         if (arg.length != NFS4_MAX_UINT64)
1832                 arg.length = PAGE_ALIGN(arg.length);
1833
1834         lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1835         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1836                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1837         atomic_dec(&lo->plh_outstanding);
1838         if (IS_ERR(lseg)) {
1839                 switch(PTR_ERR(lseg)) {
1840                 case -EBUSY:
1841                         if (time_after(jiffies, giveup))
1842                                 lseg = NULL;
1843                         break;
1844                 case -ERECALLCONFLICT:
1845                         /* Huh? We hold no layouts, how is there a recall? */
1846                         if (first) {
1847                                 lseg = NULL;
1848                                 break;
1849                         }
1850                         /* Destroy the existing layout and start over */
1851                         if (time_after(jiffies, giveup))
1852                                 pnfs_destroy_layout(NFS_I(ino));
1853                         /* Fallthrough */
1854                 case -EAGAIN:
1855                         break;
1856                 default:
1857                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1858                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1859                                 lseg = NULL;
1860                         }
1861                         goto out_put_layout_hdr;
1862                 }
1863                 if (lseg) {
1864                         if (first)
1865                                 pnfs_clear_first_layoutget(lo);
1866                         trace_pnfs_update_layout(ino, pos, count,
1867                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1868                         pnfs_put_layout_hdr(lo);
1869                         goto lookup_again;
1870                 }
1871         } else {
1872                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1873         }
1874
1875 out_put_layout_hdr:
1876         if (first)
1877                 pnfs_clear_first_layoutget(lo);
1878         pnfs_put_layout_hdr(lo);
1879 out:
1880         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1881                         "(%s, offset: %llu, length: %llu)\n",
1882                         __func__, ino->i_sb->s_id,
1883                         (unsigned long long)NFS_FILEID(ino),
1884                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1885                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1886                         (unsigned long long)pos,
1887                         (unsigned long long)count);
1888         return lseg;
1889 out_unlock:
1890         spin_unlock(&ino->i_lock);
1891         goto out_put_layout_hdr;
1892 }
1893 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1894
1895 static bool
1896 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1897 {
1898         switch (range->iomode) {
1899         case IOMODE_READ:
1900         case IOMODE_RW:
1901                 break;
1902         default:
1903                 return false;
1904         }
1905         if (range->offset == NFS4_MAX_UINT64)
1906                 return false;
1907         if (range->length == 0)
1908                 return false;
1909         if (range->length != NFS4_MAX_UINT64 &&
1910             range->length > NFS4_MAX_UINT64 - range->offset)
1911                 return false;
1912         return true;
1913 }
1914
1915 struct pnfs_layout_segment *
1916 pnfs_layout_process(struct nfs4_layoutget *lgp)
1917 {
1918         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1919         struct nfs4_layoutget_res *res = &lgp->res;
1920         struct pnfs_layout_segment *lseg;
1921         struct inode *ino = lo->plh_inode;
1922         LIST_HEAD(free_me);
1923
1924         if (!pnfs_sanity_check_layout_range(&res->range))
1925                 return ERR_PTR(-EINVAL);
1926
1927         /* Inject layout blob into I/O device driver */
1928         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1929         if (IS_ERR_OR_NULL(lseg)) {
1930                 if (!lseg)
1931                         lseg = ERR_PTR(-ENOMEM);
1932
1933                 dprintk("%s: Could not allocate layout: error %ld\n",
1934                        __func__, PTR_ERR(lseg));
1935                 return lseg;
1936         }
1937
1938         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1939
1940         spin_lock(&ino->i_lock);
1941         if (pnfs_layoutgets_blocked(lo)) {
1942                 dprintk("%s forget reply due to state\n", __func__);
1943                 goto out_forget;
1944         }
1945
1946         if (!pnfs_layout_is_valid(lo)) {
1947                 /* We have a completely new layout */
1948                 pnfs_set_layout_stateid(lo, &res->stateid, true);
1949         } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1950                 /* existing state ID, make sure the sequence number matches. */
1951                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1952                         dprintk("%s forget reply due to sequence\n", __func__);
1953                         goto out_forget;
1954                 }
1955                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1956         } else {
1957                 /*
1958                  * We got an entirely new state ID.  Mark all segments for the
1959                  * inode invalid, and retry the layoutget
1960                  */
1961                 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1962                 goto out_forget;
1963         }
1964
1965         pnfs_get_lseg(lseg);
1966         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1967
1968
1969         if (res->return_on_close)
1970                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1971
1972         spin_unlock(&ino->i_lock);
1973         pnfs_free_lseg_list(&free_me);
1974         return lseg;
1975
1976 out_forget:
1977         spin_unlock(&ino->i_lock);
1978         lseg->pls_layout = lo;
1979         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1980         if (!pnfs_layout_is_valid(lo))
1981                 nfs_commit_inode(ino, 0);
1982         return ERR_PTR(-EAGAIN);
1983 }
1984
1985 /**
1986  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1987  * @lo: pointer to layout header
1988  * @tmp_list: list header to be used with pnfs_free_lseg_list()
1989  * @return_range: describe layout segment ranges to be returned
1990  *
1991  * This function is mainly intended for use by layoutrecall. It attempts
1992  * to free the layout segment immediately, or else to mark it for return
1993  * as soon as its reference count drops to zero.
1994  */
1995 int
1996 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1997                                 struct list_head *tmp_list,
1998                                 const struct pnfs_layout_range *return_range,
1999                                 u32 seq)
2000 {
2001         struct pnfs_layout_segment *lseg, *next;
2002         int remaining = 0;
2003
2004         dprintk("%s:Begin lo %p\n", __func__, lo);
2005
2006         if (list_empty(&lo->plh_segs))
2007                 return 0;
2008
2009         assert_spin_locked(&lo->plh_inode->i_lock);
2010
2011         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2012                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2013                         dprintk("%s: marking lseg %p iomode %d "
2014                                 "offset %llu length %llu\n", __func__,
2015                                 lseg, lseg->pls_range.iomode,
2016                                 lseg->pls_range.offset,
2017                                 lseg->pls_range.length);
2018                         if (mark_lseg_invalid(lseg, tmp_list))
2019                                 continue;
2020                         remaining++;
2021                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2022                 }
2023
2024         if (remaining)
2025                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2026
2027         return remaining;
2028 }
2029
2030 void pnfs_error_mark_layout_for_return(struct inode *inode,
2031                                        struct pnfs_layout_segment *lseg)
2032 {
2033         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2034         struct pnfs_layout_range range = {
2035                 .iomode = lseg->pls_range.iomode,
2036                 .offset = 0,
2037                 .length = NFS4_MAX_UINT64,
2038         };
2039         bool return_now = false;
2040
2041         spin_lock(&inode->i_lock);
2042         if (!pnfs_layout_is_valid(lo)) {
2043                 spin_unlock(&inode->i_lock);
2044                 return;
2045         }
2046         pnfs_set_plh_return_info(lo, range.iomode, 0);
2047         /*
2048          * mark all matching lsegs so that we are sure to have no live
2049          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2050          * for how it works.
2051          */
2052         if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2053                 nfs4_stateid stateid;
2054                 enum pnfs_iomode iomode;
2055
2056                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2057                 spin_unlock(&inode->i_lock);
2058                 if (return_now)
2059                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2060         } else {
2061                 spin_unlock(&inode->i_lock);
2062                 nfs_commit_inode(inode, 0);
2063         }
2064 }
2065 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2066
2067 void
2068 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2069 {
2070         if (pgio->pg_lseg == NULL ||
2071             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2072                 return;
2073         pnfs_put_lseg(pgio->pg_lseg);
2074         pgio->pg_lseg = NULL;
2075 }
2076 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2077
2078 /*
2079  * Check for any intersection between the request and the pgio->pg_lseg,
2080  * and if none, put this pgio->pg_lseg away.
2081  */
2082 static void
2083 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2084 {
2085         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2086                 pnfs_put_lseg(pgio->pg_lseg);
2087                 pgio->pg_lseg = NULL;
2088         }
2089 }
2090
2091 void
2092 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2093 {
2094         u64 rd_size = req->wb_bytes;
2095
2096         pnfs_generic_pg_check_layout(pgio);
2097         pnfs_generic_pg_check_range(pgio, req);
2098         if (pgio->pg_lseg == NULL) {
2099                 if (pgio->pg_dreq == NULL)
2100                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2101                 else
2102                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2103
2104                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2105                                                    req->wb_context,
2106                                                    req_offset(req),
2107                                                    rd_size,
2108                                                    IOMODE_READ,
2109                                                    false,
2110                                                    GFP_KERNEL);
2111                 if (IS_ERR(pgio->pg_lseg)) {
2112                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2113                         pgio->pg_lseg = NULL;
2114                         return;
2115                 }
2116         }
2117         /* If no lseg, fall back to read through mds */
2118         if (pgio->pg_lseg == NULL)
2119                 nfs_pageio_reset_read_mds(pgio);
2120
2121 }
2122 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2123
2124 void
2125 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2126                            struct nfs_page *req, u64 wb_size)
2127 {
2128         pnfs_generic_pg_check_layout(pgio);
2129         pnfs_generic_pg_check_range(pgio, req);
2130         if (pgio->pg_lseg == NULL) {
2131                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2132                                                    req->wb_context,
2133                                                    req_offset(req),
2134                                                    wb_size,
2135                                                    IOMODE_RW,
2136                                                    false,
2137                                                    GFP_NOFS);
2138                 if (IS_ERR(pgio->pg_lseg)) {
2139                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2140                         pgio->pg_lseg = NULL;
2141                         return;
2142                 }
2143         }
2144         /* If no lseg, fall back to write through mds */
2145         if (pgio->pg_lseg == NULL)
2146                 nfs_pageio_reset_write_mds(pgio);
2147 }
2148 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2149
2150 void
2151 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2152 {
2153         if (desc->pg_lseg) {
2154                 pnfs_put_lseg(desc->pg_lseg);
2155                 desc->pg_lseg = NULL;
2156         }
2157 }
2158 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2159
2160 /*
2161  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2162  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2163  */
2164 size_t
2165 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2166                      struct nfs_page *prev, struct nfs_page *req)
2167 {
2168         unsigned int size;
2169         u64 seg_end, req_start, seg_left;
2170
2171         size = nfs_generic_pg_test(pgio, prev, req);
2172         if (!size)
2173                 return 0;
2174
2175         /*
2176          * 'size' contains the number of bytes left in the current page (up
2177          * to the original size asked for in @req->wb_bytes).
2178          *
2179          * Calculate how many bytes are left in the layout segment
2180          * and if there are less bytes than 'size', return that instead.
2181          *
2182          * Please also note that 'end_offset' is actually the offset of the
2183          * first byte that lies outside the pnfs_layout_range. FIXME?
2184          *
2185          */
2186         if (pgio->pg_lseg) {
2187                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2188                                      pgio->pg_lseg->pls_range.length);
2189                 req_start = req_offset(req);
2190
2191                 /* start of request is past the last byte of this segment */
2192                 if (req_start >= seg_end)
2193                         return 0;
2194
2195                 /* adjust 'size' iff there are fewer bytes left in the
2196                  * segment than what nfs_generic_pg_test returned */
2197                 seg_left = seg_end - req_start;
2198                 if (seg_left < size)
2199                         size = (unsigned int)seg_left;
2200         }
2201
2202         return size;
2203 }
2204 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2205
2206 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2207 {
2208         struct nfs_pageio_descriptor pgio;
2209
2210         /* Resend all requests through the MDS */
2211         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2212                               hdr->completion_ops);
2213         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2214         return nfs_pageio_resend(&pgio, hdr);
2215 }
2216 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2217
2218 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2219 {
2220
2221         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2222         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2223             PNFS_LAYOUTRET_ON_ERROR) {
2224                 pnfs_return_layout(hdr->inode);
2225         }
2226         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2227                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2228 }
2229
2230 /*
2231  * Called by non rpc-based layout drivers
2232  */
2233 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2234 {
2235         if (likely(!hdr->pnfs_error)) {
2236                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2237                                 hdr->mds_offset + hdr->res.count);
2238                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2239         }
2240         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2241         if (unlikely(hdr->pnfs_error))
2242                 pnfs_ld_handle_write_error(hdr);
2243         hdr->mds_ops->rpc_release(hdr);
2244 }
2245 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2246
2247 static void
2248 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2249                 struct nfs_pgio_header *hdr)
2250 {
2251         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2252
2253         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2254                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2255                 nfs_pageio_reset_write_mds(desc);
2256                 mirror->pg_recoalesce = 1;
2257         }
2258         hdr->release(hdr);
2259 }
2260
2261 static enum pnfs_try_status
2262 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2263                         const struct rpc_call_ops *call_ops,
2264                         struct pnfs_layout_segment *lseg,
2265                         int how)
2266 {
2267         struct inode *inode = hdr->inode;
2268         enum pnfs_try_status trypnfs;
2269         struct nfs_server *nfss = NFS_SERVER(inode);
2270
2271         hdr->mds_ops = call_ops;
2272
2273         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2274                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2275         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2276         if (trypnfs != PNFS_NOT_ATTEMPTED)
2277                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2278         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2279         return trypnfs;
2280 }
2281
2282 static void
2283 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2284               struct nfs_pgio_header *hdr, int how)
2285 {
2286         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2287         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2288         enum pnfs_try_status trypnfs;
2289
2290         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2291         switch (trypnfs) {
2292         case PNFS_NOT_ATTEMPTED:
2293                 pnfs_write_through_mds(desc, hdr);
2294         case PNFS_ATTEMPTED:
2295                 break;
2296         case PNFS_TRY_AGAIN:
2297                 /* cleanup hdr and prepare to redo pnfs */
2298                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2299                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2300                         list_splice_init(&hdr->pages, &mirror->pg_list);
2301                         mirror->pg_recoalesce = 1;
2302                 }
2303                 hdr->mds_ops->rpc_release(hdr);
2304         }
2305 }
2306
2307 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2308 {
2309         pnfs_put_lseg(hdr->lseg);
2310         nfs_pgio_header_free(hdr);
2311 }
2312
2313 int
2314 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2315 {
2316         struct nfs_pgio_header *hdr;
2317         int ret;
2318
2319         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2320         if (!hdr) {
2321                 desc->pg_error = -ENOMEM;
2322                 return desc->pg_error;
2323         }
2324         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2325
2326         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2327         ret = nfs_generic_pgio(desc, hdr);
2328         if (!ret)
2329                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2330
2331         return ret;
2332 }
2333 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2334
2335 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2336 {
2337         struct nfs_pageio_descriptor pgio;
2338
2339         /* Resend all requests through the MDS */
2340         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2341         return nfs_pageio_resend(&pgio, hdr);
2342 }
2343 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2344
2345 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2346 {
2347         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2348         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2349             PNFS_LAYOUTRET_ON_ERROR) {
2350                 pnfs_return_layout(hdr->inode);
2351         }
2352         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2353                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2354 }
2355
2356 /*
2357  * Called by non rpc-based layout drivers
2358  */
2359 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2360 {
2361         if (likely(!hdr->pnfs_error))
2362                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2363         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2364         if (unlikely(hdr->pnfs_error))
2365                 pnfs_ld_handle_read_error(hdr);
2366         hdr->mds_ops->rpc_release(hdr);
2367 }
2368 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2369
2370 static void
2371 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2372                 struct nfs_pgio_header *hdr)
2373 {
2374         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2375
2376         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2377                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2378                 nfs_pageio_reset_read_mds(desc);
2379                 mirror->pg_recoalesce = 1;
2380         }
2381         hdr->release(hdr);
2382 }
2383
2384 /*
2385  * Call the appropriate parallel I/O subsystem read function.
2386  */
2387 static enum pnfs_try_status
2388 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2389                        const struct rpc_call_ops *call_ops,
2390                        struct pnfs_layout_segment *lseg)
2391 {
2392         struct inode *inode = hdr->inode;
2393         struct nfs_server *nfss = NFS_SERVER(inode);
2394         enum pnfs_try_status trypnfs;
2395
2396         hdr->mds_ops = call_ops;
2397
2398         dprintk("%s: Reading ino:%lu %u@%llu\n",
2399                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2400
2401         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2402         if (trypnfs != PNFS_NOT_ATTEMPTED)
2403                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2404         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2405         return trypnfs;
2406 }
2407
2408 /* Resend all requests through pnfs. */
2409 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2410 {
2411         struct nfs_pageio_descriptor pgio;
2412
2413         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2414                 /* Prevent deadlocks with layoutreturn! */
2415                 pnfs_put_lseg(hdr->lseg);
2416                 hdr->lseg = NULL;
2417
2418                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2419                                         hdr->completion_ops);
2420                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2421         }
2422 }
2423 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2424
2425 static void
2426 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2427 {
2428         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2429         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2430         enum pnfs_try_status trypnfs;
2431
2432         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2433         switch (trypnfs) {
2434         case PNFS_NOT_ATTEMPTED:
2435                 pnfs_read_through_mds(desc, hdr);
2436         case PNFS_ATTEMPTED:
2437                 break;
2438         case PNFS_TRY_AGAIN:
2439                 /* cleanup hdr and prepare to redo pnfs */
2440                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2441                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2442                         list_splice_init(&hdr->pages, &mirror->pg_list);
2443                         mirror->pg_recoalesce = 1;
2444                 }
2445                 hdr->mds_ops->rpc_release(hdr);
2446         }
2447 }
2448
2449 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2450 {
2451         pnfs_put_lseg(hdr->lseg);
2452         nfs_pgio_header_free(hdr);
2453 }
2454
2455 int
2456 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2457 {
2458         struct nfs_pgio_header *hdr;
2459         int ret;
2460
2461         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2462         if (!hdr) {
2463                 desc->pg_error = -ENOMEM;
2464                 return desc->pg_error;
2465         }
2466         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2467         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2468         ret = nfs_generic_pgio(desc, hdr);
2469         if (!ret)
2470                 pnfs_do_read(desc, hdr);
2471         return ret;
2472 }
2473 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2474
2475 static void pnfs_clear_layoutcommitting(struct inode *inode)
2476 {
2477         unsigned long *bitlock = &NFS_I(inode)->flags;
2478
2479         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2480         smp_mb__after_atomic();
2481         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2482 }
2483
2484 /*
2485  * There can be multiple RW segments.
2486  */
2487 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2488 {
2489         struct pnfs_layout_segment *lseg;
2490
2491         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2492                 if (lseg->pls_range.iomode == IOMODE_RW &&
2493                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2494                         list_add(&lseg->pls_lc_list, listp);
2495         }
2496 }
2497
2498 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2499 {
2500         struct pnfs_layout_segment *lseg, *tmp;
2501
2502         /* Matched by references in pnfs_set_layoutcommit */
2503         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2504                 list_del_init(&lseg->pls_lc_list);
2505                 pnfs_put_lseg(lseg);
2506         }
2507
2508         pnfs_clear_layoutcommitting(inode);
2509 }
2510
2511 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2512 {
2513         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2514 }
2515 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2516
2517 void
2518 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2519                 loff_t end_pos)
2520 {
2521         struct nfs_inode *nfsi = NFS_I(inode);
2522         bool mark_as_dirty = false;
2523
2524         spin_lock(&inode->i_lock);
2525         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2526                 nfsi->layout->plh_lwb = end_pos;
2527                 mark_as_dirty = true;
2528                 dprintk("%s: Set layoutcommit for inode %lu ",
2529                         __func__, inode->i_ino);
2530         } else if (end_pos > nfsi->layout->plh_lwb)
2531                 nfsi->layout->plh_lwb = end_pos;
2532         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2533                 /* references matched in nfs4_layoutcommit_release */
2534                 pnfs_get_lseg(lseg);
2535         }
2536         spin_unlock(&inode->i_lock);
2537         dprintk("%s: lseg %p end_pos %llu\n",
2538                 __func__, lseg, nfsi->layout->plh_lwb);
2539
2540         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2541          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2542         if (mark_as_dirty)
2543                 mark_inode_dirty_sync(inode);
2544 }
2545 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2546
2547 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2548 {
2549         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2550
2551         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2552                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2553         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2554 }
2555
2556 /*
2557  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2558  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2559  * data to disk to allow the server to recover the data if it crashes.
2560  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2561  * is off, and a COMMIT is sent to a data server, or
2562  * if WRITEs to a data server return NFS_DATA_SYNC.
2563  */
2564 int
2565 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2566 {
2567         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2568         struct nfs4_layoutcommit_data *data;
2569         struct nfs_inode *nfsi = NFS_I(inode);
2570         loff_t end_pos;
2571         int status;
2572
2573         if (!pnfs_layoutcommit_outstanding(inode))
2574                 return 0;
2575
2576         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2577
2578         status = -EAGAIN;
2579         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2580                 if (!sync)
2581                         goto out;
2582                 status = wait_on_bit_lock_action(&nfsi->flags,
2583                                 NFS_INO_LAYOUTCOMMITTING,
2584                                 nfs_wait_bit_killable,
2585                                 TASK_KILLABLE);
2586                 if (status)
2587                         goto out;
2588         }
2589
2590         status = -ENOMEM;
2591         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2592         data = kzalloc(sizeof(*data), GFP_NOFS);
2593         if (!data)
2594                 goto clear_layoutcommitting;
2595
2596         status = 0;
2597         spin_lock(&inode->i_lock);
2598         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2599                 goto out_unlock;
2600
2601         INIT_LIST_HEAD(&data->lseg_list);
2602         pnfs_list_write_lseg(inode, &data->lseg_list);
2603
2604         end_pos = nfsi->layout->plh_lwb;
2605
2606         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2607         spin_unlock(&inode->i_lock);
2608
2609         data->args.inode = inode;
2610         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2611         nfs_fattr_init(&data->fattr);
2612         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2613         data->res.fattr = &data->fattr;
2614         if (end_pos != 0)
2615                 data->args.lastbytewritten = end_pos - 1;
2616         else
2617                 data->args.lastbytewritten = U64_MAX;
2618         data->res.server = NFS_SERVER(inode);
2619
2620         if (ld->prepare_layoutcommit) {
2621                 status = ld->prepare_layoutcommit(&data->args);
2622                 if (status) {
2623                         put_rpccred(data->cred);
2624                         spin_lock(&inode->i_lock);
2625                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2626                         if (end_pos > nfsi->layout->plh_lwb)
2627                                 nfsi->layout->plh_lwb = end_pos;
2628                         goto out_unlock;
2629                 }
2630         }
2631
2632
2633         status = nfs4_proc_layoutcommit(data, sync);
2634 out:
2635         if (status)
2636                 mark_inode_dirty_sync(inode);
2637         dprintk("<-- %s status %d\n", __func__, status);
2638         return status;
2639 out_unlock:
2640         spin_unlock(&inode->i_lock);
2641         kfree(data);
2642 clear_layoutcommitting:
2643         pnfs_clear_layoutcommitting(inode);
2644         goto out;
2645 }
2646 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2647
2648 int
2649 pnfs_generic_sync(struct inode *inode, bool datasync)
2650 {
2651         return pnfs_layoutcommit_inode(inode, true);
2652 }
2653 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2654
2655 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2656 {
2657         struct nfs4_threshold *thp;
2658
2659         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2660         if (!thp) {
2661                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2662                 return NULL;
2663         }
2664         return thp;
2665 }
2666
2667 #if IS_ENABLED(CONFIG_NFS_V4_2)
2668 int
2669 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2670 {
2671         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2672         struct nfs_server *server = NFS_SERVER(inode);
2673         struct nfs_inode *nfsi = NFS_I(inode);
2674         struct nfs42_layoutstat_data *data;
2675         struct pnfs_layout_hdr *hdr;
2676         int status = 0;
2677
2678         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2679                 goto out;
2680
2681         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2682                 goto out;
2683
2684         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2685                 goto out;
2686
2687         spin_lock(&inode->i_lock);
2688         if (!NFS_I(inode)->layout) {
2689                 spin_unlock(&inode->i_lock);
2690                 goto out_clear_layoutstats;
2691         }
2692         hdr = NFS_I(inode)->layout;
2693         pnfs_get_layout_hdr(hdr);
2694         spin_unlock(&inode->i_lock);
2695
2696         data = kzalloc(sizeof(*data), gfp_flags);
2697         if (!data) {
2698                 status = -ENOMEM;
2699                 goto out_put;
2700         }
2701
2702         data->args.fh = NFS_FH(inode);
2703         data->args.inode = inode;
2704         status = ld->prepare_layoutstats(&data->args);
2705         if (status)
2706                 goto out_free;
2707
2708         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2709
2710 out:
2711         dprintk("%s returns %d\n", __func__, status);
2712         return status;
2713
2714 out_free:
2715         kfree(data);
2716 out_put:
2717         pnfs_put_layout_hdr(hdr);
2718 out_clear_layoutstats:
2719         smp_mb__before_atomic();
2720         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2721         smp_mb__after_atomic();
2722         goto out;
2723 }
2724 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2725 #endif
2726
2727 unsigned int layoutstats_timer;
2728 module_param(layoutstats_timer, uint, 0644);
2729 EXPORT_SYMBOL_GPL(layoutstats_timer);