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 <[email protected]>
   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 #include "nfs4_fs.h"
  41
  42 #define NFSDBG_FACILITY         NFSDBG_PNFS
  43 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  44
  45 /* Locking:
  46  *
  47  * pnfs_spinlock:
  48  *      protects pnfs_modules_tbl.
  49  */
  50 static DEFINE_SPINLOCK(pnfs_spinlock);
  51
  52 /*
  53  * pnfs_modules_tbl holds all pnfs modules
  54  */
  55 static LIST_HEAD(pnfs_modules_tbl);
  56
  57 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  58 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  59                 struct list_head *free_me,
  60                 const struct pnfs_layout_range *range,
  61                 u32 seq);
  62 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  63                                 struct list_head *tmp_list);
  64
  65 /* Return the registered pnfs layout driver module matching given id */
  66 static struct pnfs_layoutdriver_type *
  67 find_pnfs_driver_locked(u32 id)
  68 {
  69         struct pnfs_layoutdriver_type *local;
  70
  71         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  72                 if (local->id == id)
  73                         goto out;
  74         local = NULL;
  75 out:
  76         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  77         return local;
  78 }
  79
  80 static struct pnfs_layoutdriver_type *
  81 find_pnfs_driver(u32 id)
  82 {
  83         struct pnfs_layoutdriver_type *local;
  84
  85         spin_lock(&pnfs_spinlock);
  86         local = find_pnfs_driver_locked(id);
  87         if (local != NULL && !try_module_get(local->owner)) {
  88                 dprintk("%s: Could not grab reference on module\n", __func__);
  89                 local = NULL;
  90         }
  91         spin_unlock(&pnfs_spinlock);
  92         return local;
  93 }
  94
  95 void
  96 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  97 {
  98         if (nfss->pnfs_curr_ld) {
  99                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
 100                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 101                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 102                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 103                         nfs4_deviceid_purge_client(nfss->nfs_client);
 104                 module_put(nfss->pnfs_curr_ld->owner);
 105         }
 106         nfss->pnfs_curr_ld = NULL;
 107 }
 108
 109 /*
 110  * When the server sends a list of layout types, we choose one in the order
 111  * given in the list below.
 112  *
 113  * FIXME: should this list be configurable in some fashion? module param?
 114  *        mount option? something else?
 115  */
 116 static const u32 ld_prefs[] = {
 117         LAYOUT_SCSI,
 118         LAYOUT_BLOCK_VOLUME,
 119         LAYOUT_OSD2_OBJECTS,
 120         LAYOUT_FLEX_FILES,
 121         LAYOUT_NFSV4_1_FILES,
 122         0
 123 };
 124
 125 static int
 126 ld_cmp(const void *e1, const void *e2)
 127 {
 128         u32 ld1 = *((u32 *)e1);
 129         u32 ld2 = *((u32 *)e2);
 130         int i;
 131
 132         for (i = 0; ld_prefs[i] != 0; i++) {
 133                 if (ld1 == ld_prefs[i])
 134                         return -1;
 135
 136                 if (ld2 == ld_prefs[i])
 137                         return 1;
 138         }
 139         return 0;
 140 }
 141
 142 /*
 143  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 144  * Currently only one pNFS layout driver per filesystem is supported.
 145  *
 146  * @ids array of layout types supported by MDS.
 147  */
 148 void
 149 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 150                       struct nfs_fsinfo *fsinfo)
 151 {
 152         struct pnfs_layoutdriver_type *ld_type = NULL;
 153         u32 id;
 154         int i;
 155
 156         if (fsinfo->nlayouttypes == 0)
 157                 goto out_no_driver;
 158         if (!(server->nfs_client->cl_exchange_flags &
 159                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 160                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 161                         __func__, server->nfs_client->cl_exchange_flags);
 162                 goto out_no_driver;
 163         }
 164
 165         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 166                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 167
 168         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 169                 id = fsinfo->layouttype[i];
 170                 ld_type = find_pnfs_driver(id);
 171                 if (!ld_type) {
 172                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 173                                         id);
 174                         ld_type = find_pnfs_driver(id);
 175                 }
 176                 if (ld_type)
 177                         break;
 178         }
 179
 180         if (!ld_type) {
 181                 dprintk("%s: No pNFS module found!\n", __func__);
 182                 goto out_no_driver;
 183         }
 184
 185         server->pnfs_curr_ld = ld_type;
 186         if (ld_type->set_layoutdriver
 187             && ld_type->set_layoutdriver(server, mntfh)) {
 188                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 189                         "driver %u.\n", __func__, id);
 190                 module_put(ld_type->owner);
 191                 goto out_no_driver;
 192         }
 193         /* Bump the MDS count */
 194         atomic_inc(&server->nfs_client->cl_mds_count);
 195
 196         dprintk("%s: pNFS module for %u set\n", __func__, id);
 197         return;
 198
 199 out_no_driver:
 200         dprintk("%s: Using NFSv4 I/O\n", __func__);
 201         server->pnfs_curr_ld = NULL;
 202 }
 203
 204 int
 205 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 206 {
 207         int status = -EINVAL;
 208         struct pnfs_layoutdriver_type *tmp;
 209
 210         if (ld_type->id == 0) {
 211                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 212                 return status;
 213         }
 214         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 215                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 216                        "alloc_lseg and free_lseg.\n", __func__);
 217                 return status;
 218         }
 219
 220         spin_lock(&pnfs_spinlock);
 221         tmp = find_pnfs_driver_locked(ld_type->id);
 222         if (!tmp) {
 223                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 224                 status = 0;
 225                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 226                         ld_type->name);
 227         } else {
 228                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 229                         __func__, ld_type->id);
 230         }
 231         spin_unlock(&pnfs_spinlock);
 232
 233         return status;
 234 }
 235 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 236
 237 void
 238 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 239 {
 240         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 241         spin_lock(&pnfs_spinlock);
 242         list_del(&ld_type->pnfs_tblid);
 243         spin_unlock(&pnfs_spinlock);
 244 }
 245 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 246
 247 /*
 248  * pNFS client layout cache
 249  */
 250
 251 /* Need to hold i_lock if caller does not already hold reference */
 252 void
 253 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 254 {
 255         refcount_inc(&lo->plh_refcount);
 256 }
 257
 258 static struct pnfs_layout_hdr *
 259 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 260 {
 261         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 262         return ld->alloc_layout_hdr(ino, gfp_flags);
 263 }
 264
 265 static void
 266 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 267 {
 268         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 269         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 270
 271         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
 272                 struct nfs_client *clp = server->nfs_client;
 273
 274                 spin_lock(&clp->cl_lock);
 275                 list_del_rcu(&lo->plh_layouts);
 276                 spin_unlock(&clp->cl_lock);
 277         }
 278         put_cred(lo->plh_lc_cred);
 279         return ld->free_layout_hdr(lo);
 280 }
 281
 282 static void
 283 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 284 {
 285         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 286         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 287         nfsi->layout = NULL;
 288         /* Reset MDS Threshold I/O counters */
 289         nfsi->write_io = 0;
 290         nfsi->read_io = 0;
 291 }
 292
 293 void
 294 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 295 {
 296         struct inode *inode;
 297         unsigned long i_state;
 298
 299         if (!lo)
 300                 return;
 301         inode = lo->plh_inode;
 302         pnfs_layoutreturn_before_put_layout_hdr(lo);
 303
 304         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 305                 if (!list_empty(&lo->plh_segs))
 306                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 307                 pnfs_detach_layout_hdr(lo);
 308                 i_state = inode->i_state;
 309                 spin_unlock(&inode->i_lock);
 310                 pnfs_free_layout_hdr(lo);
 311                 /* Notify pnfs_destroy_layout_final() that we're done */
 312                 if (i_state & (I_FREEING | I_CLEAR))
 313                         wake_up_var(lo);
 314         }
 315 }
 316
 317 static struct inode *
 318 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
 319 {
 320         struct inode *inode = igrab(lo->plh_inode);
 321         if (inode)
 322                 return inode;
 323         set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 324         return NULL;
 325 }
 326
 327 static void
 328 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 329                          u32 seq)
 330 {
 331         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 332                 iomode = IOMODE_ANY;
 333         lo->plh_return_iomode = iomode;
 334         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 335         if (seq != 0) {
 336                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 337                 lo->plh_return_seq = seq;
 338         }
 339 }
 340
 341 static void
 342 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 343 {
 344         struct pnfs_layout_segment *lseg;
 345         lo->plh_return_iomode = 0;
 346         lo->plh_return_seq = 0;
 347         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 348         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 349                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 350                         continue;
 351                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 352         }
 353 }
 354
 355 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 356 {
 357         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 358         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 359         smp_mb__after_atomic();
 360         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 361         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 362 }
 363
 364 static void
 365 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 366                 struct list_head *free_me)
 367 {
 368         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 369         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 370         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 371                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 372         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 373                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 374 }
 375
 376 /*
 377  * Update the seqid of a layout stateid after receiving
 378  * NFS4ERR_OLD_STATEID
 379  */
 380 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
 381                 struct pnfs_layout_range *dst_range,
 382                 struct inode *inode)
 383 {
 384         struct pnfs_layout_hdr *lo;
 385         struct pnfs_layout_range range = {
 386                 .iomode = IOMODE_ANY,
 387                 .offset = 0,
 388                 .length = NFS4_MAX_UINT64,
 389         };
 390         bool ret = false;
 391         LIST_HEAD(head);
 392         int err;
 393
 394         spin_lock(&inode->i_lock);
 395         lo = NFS_I(inode)->layout;
 396         if (lo &&  pnfs_layout_is_valid(lo) &&
 397             nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 398                 /* Is our call using the most recent seqid? If so, bump it */
 399                 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
 400                         nfs4_stateid_seqid_inc(dst);
 401                         ret = true;
 402                         goto out;
 403                 }
 404                 /* Try to update the seqid to the most recent */
 405                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 406                 if (err != -EBUSY) {
 407                         dst->seqid = lo->plh_stateid.seqid;
 408                         *dst_range = range;
 409                         ret = true;
 410                 }
 411         }
 412 out:
 413         spin_unlock(&inode->i_lock);
 414         pnfs_free_lseg_list(&head);
 415         return ret;
 416 }
 417
 418 /*
 419  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 420  *
 421  * In order to continue using the pnfs_layout_hdr, a full recovery
 422  * is required.
 423  * Note that caller must hold inode->i_lock.
 424  */
 425 int
 426 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 427                 struct list_head *lseg_list)
 428 {
 429         struct pnfs_layout_range range = {
 430                 .iomode = IOMODE_ANY,
 431                 .offset = 0,
 432                 .length = NFS4_MAX_UINT64,
 433         };
 434         struct pnfs_layout_segment *lseg, *next;
 435
 436         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 437         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 438                 pnfs_clear_lseg_state(lseg, lseg_list);
 439         pnfs_clear_layoutreturn_info(lo);
 440         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 441         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 442             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 443                 pnfs_clear_layoutreturn_waitbit(lo);
 444         return !list_empty(&lo->plh_segs);
 445 }
 446
 447 static int
 448 pnfs_iomode_to_fail_bit(u32 iomode)
 449 {
 450         return iomode == IOMODE_RW ?
 451                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 452 }
 453
 454 static void
 455 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 456 {
 457         lo->plh_retry_timestamp = jiffies;
 458         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 459                 refcount_inc(&lo->plh_refcount);
 460 }
 461
 462 static void
 463 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 464 {
 465         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 466                 refcount_dec(&lo->plh_refcount);
 467 }
 468
 469 static void
 470 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 471 {
 472         struct inode *inode = lo->plh_inode;
 473         struct pnfs_layout_range range = {
 474                 .iomode = iomode,
 475                 .offset = 0,
 476                 .length = NFS4_MAX_UINT64,
 477         };
 478         LIST_HEAD(head);
 479
 480         spin_lock(&inode->i_lock);
 481         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 482         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 483         spin_unlock(&inode->i_lock);
 484         pnfs_free_lseg_list(&head);
 485         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 486                         iomode == IOMODE_RW ?  "RW" : "READ");
 487 }
 488
 489 static bool
 490 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 491 {
 492         unsigned long start, end;
 493         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 494
 495         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 496                 return false;
 497         end = jiffies;
 498         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 499         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 500                 /* It is time to retry the failed layoutgets */
 501                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 502                 return false;
 503         }
 504         return true;
 505 }
 506
 507 static void
 508 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 509                 const struct pnfs_layout_range *range,
 510                 const nfs4_stateid *stateid)
 511 {
 512         INIT_LIST_HEAD(&lseg->pls_list);
 513         INIT_LIST_HEAD(&lseg->pls_lc_list);
 514         INIT_LIST_HEAD(&lseg->pls_commits);
 515         refcount_set(&lseg->pls_refcount, 1);
 516         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 517         lseg->pls_layout = lo;
 518         lseg->pls_range = *range;
 519         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 520 }
 521
 522 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 523 {
 524         if (lseg != NULL) {
 525                 struct inode *inode = lseg->pls_layout->plh_inode;
 526                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 527         }
 528 }
 529
 530 static void
 531 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 532                 struct pnfs_layout_segment *lseg)
 533 {
 534         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 535         list_del_init(&lseg->pls_list);
 536         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 537         refcount_dec(&lo->plh_refcount);
 538         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 539                 return;
 540         if (list_empty(&lo->plh_segs) &&
 541             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 542             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 543                 if (atomic_read(&lo->plh_outstanding) == 0)
 544                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 545                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 546         }
 547 }
 548
 549 static bool
 550 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 551                 struct pnfs_layout_segment *lseg)
 552 {
 553         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 554             pnfs_layout_is_valid(lo)) {
 555                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 556                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 557                 return true;
 558         }
 559         return false;
 560 }
 561
 562 void
 563 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 564 {
 565         struct pnfs_layout_hdr *lo;
 566         struct inode *inode;
 567
 568         if (!lseg)
 569                 return;
 570
 571         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 572                 refcount_read(&lseg->pls_refcount),
 573                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 574
 575         lo = lseg->pls_layout;
 576         inode = lo->plh_inode;
 577
 578         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 579                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 580                         spin_unlock(&inode->i_lock);
 581                         return;
 582                 }
 583                 pnfs_get_layout_hdr(lo);
 584                 pnfs_layout_remove_lseg(lo, lseg);
 585                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 586                         lseg = NULL;
 587                 spin_unlock(&inode->i_lock);
 588                 pnfs_free_lseg(lseg);
 589                 pnfs_put_layout_hdr(lo);
 590         }
 591 }
 592 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 593
 594 /*
 595  * is l2 fully contained in l1?
 596  *   start1                             end1
 597  *   [----------------------------------)
 598  *           start2           end2
 599  *           [----------------)
 600  */
 601 static bool
 602 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 603                  const struct pnfs_layout_range *l2)
 604 {
 605         u64 start1 = l1->offset;
 606         u64 end1 = pnfs_end_offset(start1, l1->length);
 607         u64 start2 = l2->offset;
 608         u64 end2 = pnfs_end_offset(start2, l2->length);
 609
 610         return (start1 <= start2) && (end1 >= end2);
 611 }
 612
 613 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 614                 struct list_head *tmp_list)
 615 {
 616         if (!refcount_dec_and_test(&lseg->pls_refcount))
 617                 return false;
 618         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 619         list_add(&lseg->pls_list, tmp_list);
 620         return true;
 621 }
 622
 623 /* Returns 1 if lseg is removed from list, 0 otherwise */
 624 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 625                              struct list_head *tmp_list)
 626 {
 627         int rv = 0;
 628
 629         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 630                 /* Remove the reference keeping the lseg in the
 631                  * list.  It will now be removed when all
 632                  * outstanding io is finished.
 633                  */
 634                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 635                         refcount_read(&lseg->pls_refcount));
 636                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 637                         rv = 1;
 638         }
 639         return rv;
 640 }
 641
 642 /*
 643  * Compare 2 layout stateid sequence ids, to see which is newer,
 644  * taking into account wraparound issues.
 645  */
 646 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 647 {
 648         return (s32)(s1 - s2) > 0;
 649 }
 650
 651 static bool
 652 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 653                  const struct pnfs_layout_range *recall_range)
 654 {
 655         return (recall_range->iomode == IOMODE_ANY ||
 656                 lseg_range->iomode == recall_range->iomode) &&
 657                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 658 }
 659
 660 static bool
 661 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 662                 const struct pnfs_layout_range *recall_range,
 663                 u32 seq)
 664 {
 665         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 666                 return false;
 667         if (recall_range == NULL)
 668                 return true;
 669         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 670 }
 671
 672 /**
 673  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 674  * @lo: layout header containing the lsegs
 675  * @tmp_list: list head where doomed lsegs should go
 676  * @recall_range: optional recall range argument to match (may be NULL)
 677  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 678  *
 679  * Walk the list of lsegs in the layout header, and tear down any that should
 680  * be destroyed. If "recall_range" is specified then the segment must match
 681  * that range. If "seq" is non-zero, then only match segments that were handed
 682  * out at or before that sequence.
 683  *
 684  * Returns number of matching invalid lsegs remaining in list after scanning
 685  * it and purging them.
 686  */
 687 int
 688 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 689                             struct list_head *tmp_list,
 690                             const struct pnfs_layout_range *recall_range,
 691                             u32 seq)
 692 {
 693         struct pnfs_layout_segment *lseg, *next;
 694         int remaining = 0;
 695
 696         dprintk("%s:Begin lo %p\n", __func__, lo);
 697
 698         if (list_empty(&lo->plh_segs))
 699                 return 0;
 700         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 701                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 702                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 703                                 "offset %llu length %llu\n", __func__,
 704                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 705                                 lseg->pls_range.offset, lseg->pls_range.length);
 706                         if (!mark_lseg_invalid(lseg, tmp_list))
 707                                 remaining++;
 708                 }
 709         dprintk("%s:Return %i\n", __func__, remaining);
 710         return remaining;
 711 }
 712
 713 static void
 714 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 715                 struct list_head *free_me,
 716                 const struct pnfs_layout_range *range,
 717                 u32 seq)
 718 {
 719         struct pnfs_layout_segment *lseg, *next;
 720
 721         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 722                 if (pnfs_match_lseg_recall(lseg, range, seq))
 723                         list_move_tail(&lseg->pls_list, free_me);
 724         }
 725 }
 726
 727 /* note free_me must contain lsegs from a single layout_hdr */
 728 void
 729 pnfs_free_lseg_list(struct list_head *free_me)
 730 {
 731         struct pnfs_layout_segment *lseg, *tmp;
 732
 733         if (list_empty(free_me))
 734                 return;
 735
 736         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 737                 list_del(&lseg->pls_list);
 738                 pnfs_free_lseg(lseg);
 739         }
 740 }
 741
 742 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 743 {
 744         struct pnfs_layout_hdr *lo;
 745         LIST_HEAD(tmp_list);
 746
 747         spin_lock(&nfsi->vfs_inode.i_lock);
 748         lo = nfsi->layout;
 749         if (lo) {
 750                 pnfs_get_layout_hdr(lo);
 751                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 752                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 753                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 754                 spin_unlock(&nfsi->vfs_inode.i_lock);
 755                 pnfs_free_lseg_list(&tmp_list);
 756                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 757                 pnfs_put_layout_hdr(lo);
 758         } else
 759                 spin_unlock(&nfsi->vfs_inode.i_lock);
 760         return lo;
 761 }
 762
 763 void pnfs_destroy_layout(struct nfs_inode *nfsi)
 764 {
 765         __pnfs_destroy_layout(nfsi);
 766 }
 767 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 768
 769 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
 770                                 struct pnfs_layout_hdr *lo)
 771 {
 772         bool ret;
 773
 774         spin_lock(&nfsi->vfs_inode.i_lock);
 775         ret = nfsi->layout != lo;
 776         spin_unlock(&nfsi->vfs_inode.i_lock);
 777         return ret;
 778 }
 779
 780 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 781 {
 782         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 783
 784         if (lo)
 785                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
 786 }
 787
 788 static bool
 789 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 790                 struct list_head *layout_list)
 791 {
 792         struct pnfs_layout_hdr *lo;
 793         bool ret = false;
 794
 795         spin_lock(&inode->i_lock);
 796         lo = NFS_I(inode)->layout;
 797         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 798                 pnfs_get_layout_hdr(lo);
 799                 list_add(&lo->plh_bulk_destroy, layout_list);
 800                 ret = true;
 801         }
 802         spin_unlock(&inode->i_lock);
 803         return ret;
 804 }
 805
 806 /* Caller must hold rcu_read_lock and clp->cl_lock */
 807 static int
 808 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 809                 struct nfs_server *server,
 810                 struct list_head *layout_list)
 811         __must_hold(&clp->cl_lock)
 812         __must_hold(RCU)
 813 {
 814         struct pnfs_layout_hdr *lo, *next;
 815         struct inode *inode;
 816
 817         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 818                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 819                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 820                     !list_empty(&lo->plh_bulk_destroy))
 821                         continue;
 822                 /* If the sb is being destroyed, just bail */
 823                 if (!nfs_sb_active(server->super))
 824                         break;
 825                 inode = pnfs_grab_inode_layout_hdr(lo);
 826                 if (inode != NULL) {
 827                         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
 828                                 list_del_rcu(&lo->plh_layouts);
 829                         if (pnfs_layout_add_bulk_destroy_list(inode,
 830                                                 layout_list))
 831                                 continue;
 832                         rcu_read_unlock();
 833                         spin_unlock(&clp->cl_lock);
 834                         iput(inode);
 835                 } else {
 836                         rcu_read_unlock();
 837                         spin_unlock(&clp->cl_lock);
 838                 }
 839                 nfs_sb_deactive(server->super);
 840                 spin_lock(&clp->cl_lock);
 841                 rcu_read_lock();
 842                 return -EAGAIN;
 843         }
 844         return 0;
 845 }
 846
 847 static int
 848 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 849                 bool is_bulk_recall)
 850 {
 851         struct pnfs_layout_hdr *lo;
 852         struct inode *inode;
 853         LIST_HEAD(lseg_list);
 854         int ret = 0;
 855
 856         while (!list_empty(layout_list)) {
 857                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 858                                 plh_bulk_destroy);
 859                 dprintk("%s freeing layout for inode %lu\n", __func__,
 860                         lo->plh_inode->i_ino);
 861                 inode = lo->plh_inode;
 862
 863                 pnfs_layoutcommit_inode(inode, false);
 864
 865                 spin_lock(&inode->i_lock);
 866                 list_del_init(&lo->plh_bulk_destroy);
 867                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 868                         if (is_bulk_recall)
 869                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 870                         ret = -EAGAIN;
 871                 }
 872                 spin_unlock(&inode->i_lock);
 873                 pnfs_free_lseg_list(&lseg_list);
 874                 /* Free all lsegs that are attached to commit buckets */
 875                 nfs_commit_inode(inode, 0);
 876                 pnfs_put_layout_hdr(lo);
 877                 nfs_iput_and_deactive(inode);
 878         }
 879         return ret;
 880 }
 881
 882 int
 883 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 884                 struct nfs_fsid *fsid,
 885                 bool is_recall)
 886 {
 887         struct nfs_server *server;
 888         LIST_HEAD(layout_list);
 889
 890         spin_lock(&clp->cl_lock);
 891         rcu_read_lock();
 892 restart:
 893         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 894                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 895                         continue;
 896                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 897                                 server,
 898                                 &layout_list) != 0)
 899                         goto restart;
 900         }
 901         rcu_read_unlock();
 902         spin_unlock(&clp->cl_lock);
 903
 904         if (list_empty(&layout_list))
 905                 return 0;
 906         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 907 }
 908
 909 int
 910 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 911                 bool is_recall)
 912 {
 913         struct nfs_server *server;
 914         LIST_HEAD(layout_list);
 915
 916         spin_lock(&clp->cl_lock);
 917         rcu_read_lock();
 918 restart:
 919         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 920                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 921                                         server,
 922                                         &layout_list) != 0)
 923                         goto restart;
 924         }
 925         rcu_read_unlock();
 926         spin_unlock(&clp->cl_lock);
 927
 928         if (list_empty(&layout_list))
 929                 return 0;
 930         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 931 }
 932
 933 /*
 934  * Called by the state manager to remove all layouts established under an
 935  * expired lease.
 936  */
 937 void
 938 pnfs_destroy_all_layouts(struct nfs_client *clp)
 939 {
 940         nfs4_deviceid_mark_client_invalid(clp);
 941         nfs4_deviceid_purge_client(clp);
 942
 943         pnfs_destroy_layouts_byclid(clp, false);
 944 }
 945
 946 static void
 947 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
 948 {
 949         const struct cred *old;
 950
 951         if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
 952                 old = xchg(&lo->plh_lc_cred, get_cred(cred));
 953                 put_cred(old);
 954         }
 955 }
 956
 957 /* update lo->plh_stateid with new if is more recent */
 958 void
 959 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 960                         const struct cred *cred, bool update_barrier)
 961 {
 962         u32 oldseq, newseq, new_barrier = 0;
 963
 964         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 965         newseq = be32_to_cpu(new->seqid);
 966
 967         if (!pnfs_layout_is_valid(lo)) {
 968                 pnfs_set_layout_cred(lo, cred);
 969                 nfs4_stateid_copy(&lo->plh_stateid, new);
 970                 lo->plh_barrier = newseq;
 971                 pnfs_clear_layoutreturn_info(lo);
 972                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 973                 return;
 974         }
 975         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 976                 nfs4_stateid_copy(&lo->plh_stateid, new);
 977                 /*
 978                  * Because of wraparound, we want to keep the barrier
 979                  * "close" to the current seqids.
 980                  */
 981                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 982         }
 983         if (update_barrier)
 984                 new_barrier = be32_to_cpu(new->seqid);
 985         else if (new_barrier == 0)
 986                 return;
 987         if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 988                 lo->plh_barrier = new_barrier;
 989 }
 990
 991 static bool
 992 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 993                 const nfs4_stateid *stateid)
 994 {
 995         u32 seqid = be32_to_cpu(stateid->seqid);
 996
 997         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 998 }
 999
1000 /* lget is set to 1 if called from inside send_layoutget call chain */
1001 static bool
1002 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1003 {
1004         return lo->plh_block_lgets ||
1005                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1006 }
1007
1008 static struct nfs_server *
1009 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1010 {
1011         struct nfs_server *server;
1012
1013         if (inode) {
1014                 server = NFS_SERVER(inode);
1015         } else {
1016                 struct dentry *parent_dir = dget_parent(ctx->dentry);
1017                 server = NFS_SERVER(parent_dir->d_inode);
1018                 dput(parent_dir);
1019         }
1020         return server;
1021 }
1022
1023 static void nfs4_free_pages(struct page **pages, size_t size)
1024 {
1025         int i;
1026
1027         if (!pages)
1028                 return;
1029
1030         for (i = 0; i < size; i++) {
1031                 if (!pages[i])
1032                         break;
1033                 __free_page(pages[i]);
1034         }
1035         kfree(pages);
1036 }
1037
1038 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1039 {
1040         struct page **pages;
1041         int i;
1042
1043         pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1044         if (!pages) {
1045                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1046                 return NULL;
1047         }
1048
1049         for (i = 0; i < size; i++) {
1050                 pages[i] = alloc_page(gfp_flags);
1051                 if (!pages[i]) {
1052                         dprintk("%s: failed to allocate page\n", __func__);
1053                         nfs4_free_pages(pages, i);
1054                         return NULL;
1055                 }
1056         }
1057
1058         return pages;
1059 }
1060
1061 static struct nfs4_layoutget *
1062 pnfs_alloc_init_layoutget_args(struct inode *ino,
1063            struct nfs_open_context *ctx,
1064            const nfs4_stateid *stateid,
1065            const struct pnfs_layout_range *range,
1066            gfp_t gfp_flags)
1067 {
1068         struct nfs_server *server = pnfs_find_server(ino, ctx);
1069         size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1070         size_t max_pages = max_response_pages(server);
1071         struct nfs4_layoutget *lgp;
1072
1073         dprintk("--> %s\n", __func__);
1074
1075         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1076         if (lgp == NULL)
1077                 return NULL;
1078
1079         if (max_reply_sz) {
1080                 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1081                 if (npages < max_pages)
1082                         max_pages = npages;
1083         }
1084
1085         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1086         if (!lgp->args.layout.pages) {
1087                 kfree(lgp);
1088                 return NULL;
1089         }
1090         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1091         lgp->res.layoutp = &lgp->args.layout;
1092
1093         /* Don't confuse uninitialised result and success */
1094         lgp->res.status = -NFS4ERR_DELAY;
1095
1096         lgp->args.minlength = PAGE_SIZE;
1097         if (lgp->args.minlength > range->length)
1098                 lgp->args.minlength = range->length;
1099         if (ino) {
1100                 loff_t i_size = i_size_read(ino);
1101
1102                 if (range->iomode == IOMODE_READ) {
1103                         if (range->offset >= i_size)
1104                                 lgp->args.minlength = 0;
1105                         else if (i_size - range->offset < lgp->args.minlength)
1106                                 lgp->args.minlength = i_size - range->offset;
1107                 }
1108         }
1109         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1110         pnfs_copy_range(&lgp->args.range, range);
1111         lgp->args.type = server->pnfs_curr_ld->id;
1112         lgp->args.inode = ino;
1113         lgp->args.ctx = get_nfs_open_context(ctx);
1114         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1115         lgp->gfp_flags = gfp_flags;
1116         lgp->cred = ctx->cred;
1117         return lgp;
1118 }
1119
1120 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1121 {
1122         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1123
1124         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1125         if (lgp->args.inode)
1126                 pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1127         put_nfs_open_context(lgp->args.ctx);
1128         kfree(lgp);
1129 }
1130
1131 static void pnfs_clear_layoutcommit(struct inode *inode,
1132                 struct list_head *head)
1133 {
1134         struct nfs_inode *nfsi = NFS_I(inode);
1135         struct pnfs_layout_segment *lseg, *tmp;
1136
1137         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1138                 return;
1139         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1140                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1141                         continue;
1142                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1143         }
1144 }
1145
1146 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1147                 const nfs4_stateid *arg_stateid,
1148                 const struct pnfs_layout_range *range,
1149                 const nfs4_stateid *stateid)
1150 {
1151         struct inode *inode = lo->plh_inode;
1152         LIST_HEAD(freeme);
1153
1154         spin_lock(&inode->i_lock);
1155         if (!pnfs_layout_is_valid(lo) ||
1156             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1157                 goto out_unlock;
1158         if (stateid) {
1159                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1160
1161                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1162                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1163                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1164         } else
1165                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1166 out_unlock:
1167         pnfs_clear_layoutreturn_waitbit(lo);
1168         spin_unlock(&inode->i_lock);
1169         pnfs_free_lseg_list(&freeme);
1170
1171 }
1172
1173 static bool
1174 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1175                 nfs4_stateid *stateid,
1176                 const struct cred **cred,
1177                 enum pnfs_iomode *iomode)
1178 {
1179         /* Serialise LAYOUTGET/LAYOUTRETURN */
1180         if (atomic_read(&lo->plh_outstanding) != 0)
1181                 return false;
1182         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1183                 return false;
1184         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1185         pnfs_get_layout_hdr(lo);
1186         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1187                 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1188                 *cred = get_cred(lo->plh_lc_cred);
1189                 if (lo->plh_return_seq != 0)
1190                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1191                 if (iomode != NULL)
1192                         *iomode = lo->plh_return_iomode;
1193                 pnfs_clear_layoutreturn_info(lo);
1194                 return true;
1195         }
1196         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1197         *cred = get_cred(lo->plh_lc_cred);
1198         if (iomode != NULL)
1199                 *iomode = IOMODE_ANY;
1200         return true;
1201 }
1202
1203 static void
1204 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1205                 struct pnfs_layout_hdr *lo,
1206                 const nfs4_stateid *stateid,
1207                 enum pnfs_iomode iomode)
1208 {
1209         struct inode *inode = lo->plh_inode;
1210
1211         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1212         args->inode = inode;
1213         args->range.iomode = iomode;
1214         args->range.offset = 0;
1215         args->range.length = NFS4_MAX_UINT64;
1216         args->layout = lo;
1217         nfs4_stateid_copy(&args->stateid, stateid);
1218 }
1219
1220 static int
1221 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1222                        const nfs4_stateid *stateid,
1223                        const struct cred **pcred,
1224                        enum pnfs_iomode iomode,
1225                        bool sync)
1226 {
1227         struct inode *ino = lo->plh_inode;
1228         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1229         struct nfs4_layoutreturn *lrp;
1230         const struct cred *cred = *pcred;
1231         int status = 0;
1232
1233         *pcred = NULL;
1234         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1235         if (unlikely(lrp == NULL)) {
1236                 status = -ENOMEM;
1237                 spin_lock(&ino->i_lock);
1238                 pnfs_clear_layoutreturn_waitbit(lo);
1239                 spin_unlock(&ino->i_lock);
1240                 put_cred(cred);
1241                 pnfs_put_layout_hdr(lo);
1242                 goto out;
1243         }
1244
1245         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1246         lrp->args.ld_private = &lrp->ld_private;
1247         lrp->clp = NFS_SERVER(ino)->nfs_client;
1248         lrp->cred = cred;
1249         if (ld->prepare_layoutreturn)
1250                 ld->prepare_layoutreturn(&lrp->args);
1251
1252         status = nfs4_proc_layoutreturn(lrp, sync);
1253 out:
1254         dprintk("<-- %s status: %d\n", __func__, status);
1255         return status;
1256 }
1257
1258 static bool
1259 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1260                                 enum pnfs_iomode iomode,
1261                                 u32 seq)
1262 {
1263         struct pnfs_layout_range recall_range = {
1264                 .length = NFS4_MAX_UINT64,
1265                 .iomode = iomode,
1266         };
1267         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1268                                                &recall_range, seq) != -EBUSY;
1269 }
1270
1271 /* Return true if layoutreturn is needed */
1272 static bool
1273 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1274 {
1275         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1276                 return false;
1277         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1278                                                lo->plh_return_seq);
1279 }
1280
1281 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1282 {
1283         struct inode *inode= lo->plh_inode;
1284
1285         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1286                 return;
1287         spin_lock(&inode->i_lock);
1288         if (pnfs_layout_need_return(lo)) {
1289                 const struct cred *cred;
1290                 nfs4_stateid stateid;
1291                 enum pnfs_iomode iomode;
1292                 bool send;
1293
1294                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1295                 spin_unlock(&inode->i_lock);
1296                 if (send) {
1297                         /* Send an async layoutreturn so we dont deadlock */
1298                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1299                 }
1300         } else
1301                 spin_unlock(&inode->i_lock);
1302 }
1303
1304 /*
1305  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1306  * when the layout segment list is empty.
1307  *
1308  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1309  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1310  * deviceid is marked invalid.
1311  */
1312 int
1313 _pnfs_return_layout(struct inode *ino)
1314 {
1315         struct pnfs_layout_hdr *lo = NULL;
1316         struct nfs_inode *nfsi = NFS_I(ino);
1317         LIST_HEAD(tmp_list);
1318         const struct cred *cred;
1319         nfs4_stateid stateid;
1320         int status = 0;
1321         bool send, valid_layout;
1322
1323         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1324
1325         spin_lock(&ino->i_lock);
1326         lo = nfsi->layout;
1327         if (!lo) {
1328                 spin_unlock(&ino->i_lock);
1329                 dprintk("NFS: %s no layout to return\n", __func__);
1330                 goto out;
1331         }
1332         /* Reference matched in nfs4_layoutreturn_release */
1333         pnfs_get_layout_hdr(lo);
1334         /* Is there an outstanding layoutreturn ? */
1335         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1336                 spin_unlock(&ino->i_lock);
1337                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1338                                         TASK_UNINTERRUPTIBLE))
1339                         goto out_put_layout_hdr;
1340                 spin_lock(&ino->i_lock);
1341         }
1342         valid_layout = pnfs_layout_is_valid(lo);
1343         pnfs_clear_layoutcommit(ino, &tmp_list);
1344         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1345
1346         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1347                 struct pnfs_layout_range range = {
1348                         .iomode         = IOMODE_ANY,
1349                         .offset         = 0,
1350                         .length         = NFS4_MAX_UINT64,
1351                 };
1352                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1353         }
1354
1355         /* Don't send a LAYOUTRETURN if list was initially empty */
1356         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1357                         !valid_layout) {
1358                 spin_unlock(&ino->i_lock);
1359                 dprintk("NFS: %s no layout segments to return\n", __func__);
1360                 goto out_wait_layoutreturn;
1361         }
1362
1363         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1364         spin_unlock(&ino->i_lock);
1365         if (send)
1366                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1367 out_wait_layoutreturn:
1368         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1369 out_put_layout_hdr:
1370         pnfs_free_lseg_list(&tmp_list);
1371         pnfs_put_layout_hdr(lo);
1372 out:
1373         dprintk("<-- %s status: %d\n", __func__, status);
1374         return status;
1375 }
1376
1377 int
1378 pnfs_commit_and_return_layout(struct inode *inode)
1379 {
1380         struct pnfs_layout_hdr *lo;
1381         int ret;
1382
1383         spin_lock(&inode->i_lock);
1384         lo = NFS_I(inode)->layout;
1385         if (lo == NULL) {
1386                 spin_unlock(&inode->i_lock);
1387                 return 0;
1388         }
1389         pnfs_get_layout_hdr(lo);
1390         /* Block new layoutgets and read/write to ds */
1391         lo->plh_block_lgets++;
1392         spin_unlock(&inode->i_lock);
1393         filemap_fdatawait(inode->i_mapping);
1394         ret = pnfs_layoutcommit_inode(inode, true);
1395         if (ret == 0)
1396                 ret = _pnfs_return_layout(inode);
1397         spin_lock(&inode->i_lock);
1398         lo->plh_block_lgets--;
1399         spin_unlock(&inode->i_lock);
1400         pnfs_put_layout_hdr(lo);
1401         return ret;
1402 }
1403
1404 bool pnfs_roc(struct inode *ino,
1405                 struct nfs4_layoutreturn_args *args,
1406                 struct nfs4_layoutreturn_res *res,
1407                 const struct cred *cred)
1408 {
1409         struct nfs_inode *nfsi = NFS_I(ino);
1410         struct nfs_open_context *ctx;
1411         struct nfs4_state *state;
1412         struct pnfs_layout_hdr *lo;
1413         struct pnfs_layout_segment *lseg, *next;
1414         const struct cred *lc_cred;
1415         nfs4_stateid stateid;
1416         enum pnfs_iomode iomode = 0;
1417         bool layoutreturn = false, roc = false;
1418         bool skip_read = false;
1419
1420         if (!nfs_have_layout(ino))
1421                 return false;
1422 retry:
1423         rcu_read_lock();
1424         spin_lock(&ino->i_lock);
1425         lo = nfsi->layout;
1426         if (!lo || !pnfs_layout_is_valid(lo) ||
1427             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1428                 lo = NULL;
1429                 goto out_noroc;
1430         }
1431         pnfs_get_layout_hdr(lo);
1432         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1433                 spin_unlock(&ino->i_lock);
1434                 rcu_read_unlock();
1435                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1436                                 TASK_UNINTERRUPTIBLE);
1437                 pnfs_put_layout_hdr(lo);
1438                 goto retry;
1439         }
1440
1441         /* no roc if we hold a delegation */
1442         if (nfs4_check_delegation(ino, FMODE_READ)) {
1443                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1444                         goto out_noroc;
1445                 skip_read = true;
1446         }
1447
1448         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1449                 state = ctx->state;
1450                 if (state == NULL)
1451                         continue;
1452                 /* Don't return layout if there is open file state */
1453                 if (state->state & FMODE_WRITE)
1454                         goto out_noroc;
1455                 if (state->state & FMODE_READ)
1456                         skip_read = true;
1457         }
1458
1459
1460         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1461                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1462                         continue;
1463                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1464                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1465                         continue;
1466                 /*
1467                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1468                  * doesn't invalidate the layout for us.
1469                  */
1470                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1471                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1472                         continue;
1473                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1474         }
1475
1476         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1477                 goto out_noroc;
1478
1479         /* ROC in two conditions:
1480          * 1. there are ROC lsegs
1481          * 2. we don't send layoutreturn
1482          */
1483         /* lo ref dropped in pnfs_roc_release() */
1484         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1485         /* If the creds don't match, we can't compound the layoutreturn */
1486         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1487                 goto out_noroc;
1488
1489         roc = layoutreturn;
1490         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1491         res->lrs_present = 0;
1492         layoutreturn = false;
1493         put_cred(lc_cred);
1494
1495 out_noroc:
1496         spin_unlock(&ino->i_lock);
1497         rcu_read_unlock();
1498         pnfs_layoutcommit_inode(ino, true);
1499         if (roc) {
1500                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1501                 if (ld->prepare_layoutreturn)
1502                         ld->prepare_layoutreturn(args);
1503                 pnfs_put_layout_hdr(lo);
1504                 return true;
1505         }
1506         if (layoutreturn)
1507                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1508         pnfs_put_layout_hdr(lo);
1509         return false;
1510 }
1511
1512 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1513                   struct nfs4_layoutreturn_res **respp, int *ret)
1514 {
1515         struct nfs4_layoutreturn_args *arg = *argpp;
1516         int retval = -EAGAIN;
1517
1518         if (!arg)
1519                 return 0;
1520         /* Handle Layoutreturn errors */
1521         switch (*ret) {
1522         case 0:
1523                 retval = 0;
1524                 break;
1525         case -NFS4ERR_NOMATCHING_LAYOUT:
1526                 /* Was there an RPC level error? If not, retry */
1527                 if (task->tk_rpc_status == 0)
1528                         break;
1529                 /* If the call was not sent, let caller handle it */
1530                 if (!RPC_WAS_SENT(task))
1531                         return 0;
1532                 /*
1533                  * Otherwise, assume the call succeeded and
1534                  * that we need to release the layout
1535                  */
1536                 *ret = 0;
1537                 (*respp)->lrs_present = 0;
1538                 retval = 0;
1539                 break;
1540         case -NFS4ERR_DELAY:
1541                 /* Let the caller handle the retry */
1542                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1543                 return 0;
1544         case -NFS4ERR_OLD_STATEID:
1545                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1546                                                      &arg->range, arg->inode))
1547                         break;
1548                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1549                 return -EAGAIN;
1550         }
1551         *argpp = NULL;
1552         *respp = NULL;
1553         return retval;
1554 }
1555
1556 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1557                 struct nfs4_layoutreturn_res *res,
1558                 int ret)
1559 {
1560         struct pnfs_layout_hdr *lo = args->layout;
1561         struct inode *inode = args->inode;
1562         const nfs4_stateid *res_stateid = NULL;
1563         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1564
1565         switch (ret) {
1566         case -NFS4ERR_NOMATCHING_LAYOUT:
1567                 spin_lock(&inode->i_lock);
1568                 if (pnfs_layout_is_valid(lo) &&
1569                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1570                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1571                 pnfs_clear_layoutreturn_waitbit(lo);
1572                 spin_unlock(&inode->i_lock);
1573                 break;
1574         case 0:
1575                 if (res->lrs_present)
1576                         res_stateid = &res->stateid;
1577                 fallthrough;
1578         default:
1579                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1580                                              res_stateid);
1581         }
1582         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1583         if (ld_private && ld_private->ops && ld_private->ops->free)
1584                 ld_private->ops->free(ld_private);
1585         pnfs_put_layout_hdr(lo);
1586 }
1587
1588 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1589 {
1590         struct nfs_inode *nfsi = NFS_I(ino);
1591         struct pnfs_layout_hdr *lo;
1592         bool sleep = false;
1593
1594         /* we might not have grabbed lo reference. so need to check under
1595          * i_lock */
1596         spin_lock(&ino->i_lock);
1597         lo = nfsi->layout;
1598         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1599                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1600                 sleep = true;
1601         }
1602         spin_unlock(&ino->i_lock);
1603         return sleep;
1604 }
1605
1606 /*
1607  * Compare two layout segments for sorting into layout cache.
1608  * We want to preferentially return RW over RO layouts, so ensure those
1609  * are seen first.
1610  */
1611 static s64
1612 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1613            const struct pnfs_layout_range *l2)
1614 {
1615         s64 d;
1616
1617         /* high offset > low offset */
1618         d = l1->offset - l2->offset;
1619         if (d)
1620                 return d;
1621
1622         /* short length > long length */
1623         d = l2->length - l1->length;
1624         if (d)
1625                 return d;
1626
1627         /* read > read/write */
1628         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1629 }
1630
1631 static bool
1632 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1633                 const struct pnfs_layout_range *l2)
1634 {
1635         return pnfs_lseg_range_cmp(l1, l2) > 0;
1636 }
1637
1638 static bool
1639 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1640                 struct pnfs_layout_segment *old)
1641 {
1642         return false;
1643 }
1644
1645 void
1646 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1647                    struct pnfs_layout_segment *lseg,
1648                    bool (*is_after)(const struct pnfs_layout_range *,
1649                            const struct pnfs_layout_range *),
1650                    bool (*do_merge)(struct pnfs_layout_segment *,
1651                            struct pnfs_layout_segment *),
1652                    struct list_head *free_me)
1653 {
1654         struct pnfs_layout_segment *lp, *tmp;
1655
1656         dprintk("%s:Begin\n", __func__);
1657
1658         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1659                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1660                         continue;
1661                 if (do_merge(lseg, lp)) {
1662                         mark_lseg_invalid(lp, free_me);
1663                         continue;
1664                 }
1665                 if (is_after(&lseg->pls_range, &lp->pls_range))
1666                         continue;
1667                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1668                 dprintk("%s: inserted lseg %p "
1669                         "iomode %d offset %llu length %llu before "
1670                         "lp %p iomode %d offset %llu length %llu\n",
1671                         __func__, lseg, lseg->pls_range.iomode,
1672                         lseg->pls_range.offset, lseg->pls_range.length,
1673                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1674                         lp->pls_range.length);
1675                 goto out;
1676         }
1677         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1678         dprintk("%s: inserted lseg %p "
1679                 "iomode %d offset %llu length %llu at tail\n",
1680                 __func__, lseg, lseg->pls_range.iomode,
1681                 lseg->pls_range.offset, lseg->pls_range.length);
1682 out:
1683         pnfs_get_layout_hdr(lo);
1684
1685         dprintk("%s:Return\n", __func__);
1686 }
1687 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1688
1689 static void
1690 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1691                    struct pnfs_layout_segment *lseg,
1692                    struct list_head *free_me)
1693 {
1694         struct inode *inode = lo->plh_inode;
1695         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1696
1697         if (ld->add_lseg != NULL)
1698                 ld->add_lseg(lo, lseg, free_me);
1699         else
1700                 pnfs_generic_layout_insert_lseg(lo, lseg,
1701                                 pnfs_lseg_range_is_after,
1702                                 pnfs_lseg_no_merge,
1703                                 free_me);
1704 }
1705
1706 static struct pnfs_layout_hdr *
1707 alloc_init_layout_hdr(struct inode *ino,
1708                       struct nfs_open_context *ctx,
1709                       gfp_t gfp_flags)
1710 {
1711         struct pnfs_layout_hdr *lo;
1712
1713         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1714         if (!lo)
1715                 return NULL;
1716         refcount_set(&lo->plh_refcount, 1);
1717         INIT_LIST_HEAD(&lo->plh_layouts);
1718         INIT_LIST_HEAD(&lo->plh_segs);
1719         INIT_LIST_HEAD(&lo->plh_return_segs);
1720         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1721         lo->plh_inode = ino;
1722         lo->plh_lc_cred = get_cred(ctx->cred);
1723         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1724         return lo;
1725 }
1726
1727 static struct pnfs_layout_hdr *
1728 pnfs_find_alloc_layout(struct inode *ino,
1729                        struct nfs_open_context *ctx,
1730                        gfp_t gfp_flags)
1731         __releases(&ino->i_lock)
1732         __acquires(&ino->i_lock)
1733 {
1734         struct nfs_inode *nfsi = NFS_I(ino);
1735         struct pnfs_layout_hdr *new = NULL;
1736
1737         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1738
1739         if (nfsi->layout != NULL)
1740                 goto out_existing;
1741         spin_unlock(&ino->i_lock);
1742         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1743         spin_lock(&ino->i_lock);
1744
1745         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1746                 nfsi->layout = new;
1747                 return new;
1748         } else if (new != NULL)
1749                 pnfs_free_layout_hdr(new);
1750 out_existing:
1751         pnfs_get_layout_hdr(nfsi->layout);
1752         return nfsi->layout;
1753 }
1754
1755 /*
1756  * iomode matching rules:
1757  * iomode       lseg    strict match
1758  *                      iomode
1759  * -----        -----   ------ -----
1760  * ANY          READ    N/A    true
1761  * ANY          RW      N/A    true
1762  * RW           READ    N/A    false
1763  * RW           RW      N/A    true
1764  * READ         READ    N/A    true
1765  * READ         RW      true   false
1766  * READ         RW      false  true
1767  */
1768 static bool
1769 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1770                  const struct pnfs_layout_range *range,
1771                  bool strict_iomode)
1772 {
1773         struct pnfs_layout_range range1;
1774
1775         if ((range->iomode == IOMODE_RW &&
1776              ls_range->iomode != IOMODE_RW) ||
1777             (range->iomode != ls_range->iomode &&
1778              strict_iomode) ||
1779             !pnfs_lseg_range_intersecting(ls_range, range))
1780                 return false;
1781
1782         /* range1 covers only the first byte in the range */
1783         range1 = *range;
1784         range1.length = 1;
1785         return pnfs_lseg_range_contained(ls_range, &range1);
1786 }
1787
1788 /*
1789  * lookup range in layout
1790  */
1791 static struct pnfs_layout_segment *
1792 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1793                 struct pnfs_layout_range *range,
1794                 bool strict_iomode)
1795 {
1796         struct pnfs_layout_segment *lseg, *ret = NULL;
1797
1798         dprintk("%s:Begin\n", __func__);
1799
1800         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1801                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1802                     pnfs_lseg_range_match(&lseg->pls_range, range,
1803                                           strict_iomode)) {
1804                         ret = pnfs_get_lseg(lseg);
1805                         break;
1806                 }
1807         }
1808
1809         dprintk("%s:Return lseg %p ref %d\n",
1810                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1811         return ret;
1812 }
1813
1814 /*
1815  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1816  * to the MDS or over pNFS
1817  *
1818  * The nfs_inode read_io and write_io fields are cumulative counters reset
1819  * when there are no layout segments. Note that in pnfs_update_layout iomode
1820  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1821  * WRITE request.
1822  *
1823  * A return of true means use MDS I/O.
1824  *
1825  * From rfc 5661:
1826  * If a file's size is smaller than the file size threshold, data accesses
1827  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1828  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1829  * server.  If both file size and I/O size are provided, the client SHOULD
1830  * reach or exceed  both thresholds before sending its read or write
1831  * requests to the data server.
1832  */
1833 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1834                                      struct inode *ino, int iomode)
1835 {
1836         struct nfs4_threshold *t = ctx->mdsthreshold;
1837         struct nfs_inode *nfsi = NFS_I(ino);
1838         loff_t fsize = i_size_read(ino);
1839         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1840
1841         if (t == NULL)
1842                 return ret;
1843
1844         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1845                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1846
1847         switch (iomode) {
1848         case IOMODE_READ:
1849                 if (t->bm & THRESHOLD_RD) {
1850                         dprintk("%s fsize %llu\n", __func__, fsize);
1851                         size_set = true;
1852                         if (fsize < t->rd_sz)
1853                                 size = true;
1854                 }
1855                 if (t->bm & THRESHOLD_RD_IO) {
1856                         dprintk("%s nfsi->read_io %llu\n", __func__,
1857                                 nfsi->read_io);
1858                         io_set = true;
1859                         if (nfsi->read_io < t->rd_io_sz)
1860                                 io = true;
1861                 }
1862                 break;
1863         case IOMODE_RW:
1864                 if (t->bm & THRESHOLD_WR) {
1865                         dprintk("%s fsize %llu\n", __func__, fsize);
1866                         size_set = true;
1867                         if (fsize < t->wr_sz)
1868                                 size = true;
1869                 }
1870                 if (t->bm & THRESHOLD_WR_IO) {
1871                         dprintk("%s nfsi->write_io %llu\n", __func__,
1872                                 nfsi->write_io);
1873                         io_set = true;
1874                         if (nfsi->write_io < t->wr_io_sz)
1875                                 io = true;
1876                 }
1877                 break;
1878         }
1879         if (size_set && io_set) {
1880                 if (size && io)
1881                         ret = true;
1882         } else if (size || io)
1883                 ret = true;
1884
1885         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1886         return ret;
1887 }
1888
1889 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1890 {
1891         /*
1892          * send layoutcommit as it can hold up layoutreturn due to lseg
1893          * reference
1894          */
1895         pnfs_layoutcommit_inode(lo->plh_inode, false);
1896         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1897                                    nfs_wait_bit_killable,
1898                                    TASK_KILLABLE);
1899 }
1900
1901 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1902 {
1903         atomic_inc(&lo->plh_outstanding);
1904 }
1905
1906 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1907 {
1908         if (atomic_dec_and_test(&lo->plh_outstanding))
1909                 wake_up_var(&lo->plh_outstanding);
1910 }
1911
1912 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1913 {
1914         unsigned long *bitlock = &lo->plh_flags;
1915
1916         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1917         smp_mb__after_atomic();
1918         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1919 }
1920
1921 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1922                                 struct nfs_server *server)
1923 {
1924         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1925                 struct nfs_client *clp = server->nfs_client;
1926
1927                 /* The lo must be on the clp list if there is any
1928                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1929                  */
1930                 spin_lock(&clp->cl_lock);
1931                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1932                 spin_unlock(&clp->cl_lock);
1933         }
1934 }
1935
1936 /*
1937  * Layout segment is retreived from the server if not cached.
1938  * The appropriate layout segment is referenced and returned to the caller.
1939  */
1940 struct pnfs_layout_segment *
1941 pnfs_update_layout(struct inode *ino,
1942                    struct nfs_open_context *ctx,
1943                    loff_t pos,
1944                    u64 count,
1945                    enum pnfs_iomode iomode,
1946                    bool strict_iomode,
1947                    gfp_t gfp_flags)
1948 {
1949         struct pnfs_layout_range arg = {
1950                 .iomode = iomode,
1951                 .offset = pos,
1952                 .length = count,
1953         };
1954         unsigned pg_offset;
1955         struct nfs_server *server = NFS_SERVER(ino);
1956         struct nfs_client *clp = server->nfs_client;
1957         struct pnfs_layout_hdr *lo = NULL;
1958         struct pnfs_layout_segment *lseg = NULL;
1959         struct nfs4_layoutget *lgp;
1960         nfs4_stateid stateid;
1961         long timeout = 0;
1962         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1963         bool first;
1964
1965         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1966                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1967                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1968                 goto out;
1969         }
1970
1971         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1972                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1973                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1974                 goto out;
1975         }
1976
1977 lookup_again:
1978         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1979         if (IS_ERR(lseg))
1980                 goto out;
1981         first = false;
1982         spin_lock(&ino->i_lock);
1983         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1984         if (lo == NULL) {
1985                 spin_unlock(&ino->i_lock);
1986                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1987                                  PNFS_UPDATE_LAYOUT_NOMEM);
1988                 goto out;
1989         }
1990
1991         /* Do we even need to bother with this? */
1992         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1993                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1994                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1995                 dprintk("%s matches recall, use MDS\n", __func__);
1996                 goto out_unlock;
1997         }
1998
1999         /* if LAYOUTGET already failed once we don't try again */
2000         if (pnfs_layout_io_test_failed(lo, iomode)) {
2001                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2002                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2003                 goto out_unlock;
2004         }
2005
2006         /*
2007          * If the layout segment list is empty, but there are outstanding
2008          * layoutget calls, then they might be subject to a layoutrecall.
2009          */
2010         if (list_empty(&lo->plh_segs) &&
2011             atomic_read(&lo->plh_outstanding) != 0) {
2012                 spin_unlock(&ino->i_lock);
2013                 lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
2014                                         !atomic_read(&lo->plh_outstanding)));
2015                 if (IS_ERR(lseg))
2016                         goto out_put_layout_hdr;
2017                 pnfs_put_layout_hdr(lo);
2018                 goto lookup_again;
2019         }
2020
2021         /*
2022          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2023          * for LAYOUTRETURN.
2024          */
2025         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2026                 spin_unlock(&ino->i_lock);
2027                 dprintk("%s wait for layoutreturn\n", __func__);
2028                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2029                 if (!IS_ERR(lseg)) {
2030                         pnfs_put_layout_hdr(lo);
2031                         dprintk("%s retrying\n", __func__);
2032                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2033                                                  lseg,
2034                                                  PNFS_UPDATE_LAYOUT_RETRY);
2035                         goto lookup_again;
2036                 }
2037                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2038                                          PNFS_UPDATE_LAYOUT_RETURN);
2039                 goto out_put_layout_hdr;
2040         }
2041
2042         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2043         if (lseg) {
2044                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2045                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2046                 goto out_unlock;
2047         }
2048
2049         /*
2050          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2051          * stateid, or it has been invalidated, then we must use the open
2052          * stateid.
2053          */
2054         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2055                 int status;
2056
2057                 /*
2058                  * The first layoutget for the file. Need to serialize per
2059                  * RFC 5661 Errata 3208.
2060                  */
2061                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2062                                      &lo->plh_flags)) {
2063                         spin_unlock(&ino->i_lock);
2064                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2065                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2066                                                 TASK_KILLABLE));
2067                         if (IS_ERR(lseg))
2068                                 goto out_put_layout_hdr;
2069                         pnfs_put_layout_hdr(lo);
2070                         dprintk("%s retrying\n", __func__);
2071                         goto lookup_again;
2072                 }
2073
2074                 spin_unlock(&ino->i_lock);
2075                 first = true;
2076                 status = nfs4_select_rw_stateid(ctx->state,
2077                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2078                                         NULL, &stateid, NULL);
2079                 if (status != 0) {
2080                         lseg = ERR_PTR(status);
2081                         trace_pnfs_update_layout(ino, pos, count,
2082                                         iomode, lo, lseg,
2083                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2084                         nfs4_schedule_stateid_recovery(server, ctx->state);
2085                         pnfs_clear_first_layoutget(lo);
2086                         pnfs_put_layout_hdr(lo);
2087                         goto lookup_again;
2088                 }
2089                 spin_lock(&ino->i_lock);
2090         } else {
2091                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2092         }
2093
2094         if (pnfs_layoutgets_blocked(lo)) {
2095                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2096                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2097                 goto out_unlock;
2098         }
2099         nfs_layoutget_begin(lo);
2100         spin_unlock(&ino->i_lock);
2101
2102         _add_to_server_list(lo, server);
2103
2104         pg_offset = arg.offset & ~PAGE_MASK;
2105         if (pg_offset) {
2106                 arg.offset -= pg_offset;
2107                 arg.length += pg_offset;
2108         }
2109         if (arg.length != NFS4_MAX_UINT64)
2110                 arg.length = PAGE_ALIGN(arg.length);
2111
2112         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2113         if (!lgp) {
2114                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2115                                          PNFS_UPDATE_LAYOUT_NOMEM);
2116                 nfs_layoutget_end(lo);
2117                 goto out_put_layout_hdr;
2118         }
2119
2120         lseg = nfs4_proc_layoutget(lgp, &timeout);
2121         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2122                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2123         nfs_layoutget_end(lo);
2124         if (IS_ERR(lseg)) {
2125                 switch(PTR_ERR(lseg)) {
2126                 case -EBUSY:
2127                         if (time_after(jiffies, giveup))
2128                                 lseg = NULL;
2129                         break;
2130                 case -ERECALLCONFLICT:
2131                 case -EAGAIN:
2132                         break;
2133                 default:
2134                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2135                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2136                                 lseg = NULL;
2137                         }
2138                         goto out_put_layout_hdr;
2139                 }
2140                 if (lseg) {
2141                         if (first)
2142                                 pnfs_clear_first_layoutget(lo);
2143                         trace_pnfs_update_layout(ino, pos, count,
2144                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2145                         pnfs_put_layout_hdr(lo);
2146                         goto lookup_again;
2147                 }
2148         } else {
2149                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2150         }
2151
2152 out_put_layout_hdr:
2153         if (first)
2154                 pnfs_clear_first_layoutget(lo);
2155         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2156                                  PNFS_UPDATE_LAYOUT_EXIT);
2157         pnfs_put_layout_hdr(lo);
2158 out:
2159         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2160                         "(%s, offset: %llu, length: %llu)\n",
2161                         __func__, ino->i_sb->s_id,
2162                         (unsigned long long)NFS_FILEID(ino),
2163                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2164                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2165                         (unsigned long long)pos,
2166                         (unsigned long long)count);
2167         return lseg;
2168 out_unlock:
2169         spin_unlock(&ino->i_lock);
2170         goto out_put_layout_hdr;
2171 }
2172 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2173
2174 static bool
2175 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2176 {
2177         switch (range->iomode) {
2178         case IOMODE_READ:
2179         case IOMODE_RW:
2180                 break;
2181         default:
2182                 return false;
2183         }
2184         if (range->offset == NFS4_MAX_UINT64)
2185                 return false;
2186         if (range->length == 0)
2187                 return false;
2188         if (range->length != NFS4_MAX_UINT64 &&
2189             range->length > NFS4_MAX_UINT64 - range->offset)
2190                 return false;
2191         return true;
2192 }
2193
2194 static struct pnfs_layout_hdr *
2195 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2196 {
2197         struct pnfs_layout_hdr *lo;
2198
2199         spin_lock(&ino->i_lock);
2200         lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2201         if (!lo)
2202                 goto out_unlock;
2203         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2204                 goto out_unlock;
2205         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2206                 goto out_unlock;
2207         if (pnfs_layoutgets_blocked(lo))
2208                 goto out_unlock;
2209         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2210                 goto out_unlock;
2211         nfs_layoutget_begin(lo);
2212         spin_unlock(&ino->i_lock);
2213         _add_to_server_list(lo, NFS_SERVER(ino));
2214         return lo;
2215
2216 out_unlock:
2217         spin_unlock(&ino->i_lock);
2218         pnfs_put_layout_hdr(lo);
2219         return NULL;
2220 }
2221
2222 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2223                                      struct nfs_open_context *ctx)
2224 {
2225         struct inode *ino = data->dentry->d_inode;
2226         struct pnfs_layout_range rng = {
2227                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2228                           IOMODE_RW: IOMODE_READ,
2229                 .offset = 0,
2230                 .length = NFS4_MAX_UINT64,
2231         };
2232         struct nfs4_layoutget *lgp;
2233         struct pnfs_layout_hdr *lo;
2234
2235         /* Heuristic: don't send layoutget if we have cached data */
2236         if (rng.iomode == IOMODE_READ &&
2237            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2238                 return;
2239
2240         lo = _pnfs_grab_empty_layout(ino, ctx);
2241         if (!lo)
2242                 return;
2243         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2244                                              &rng, GFP_KERNEL);
2245         if (!lgp) {
2246                 pnfs_clear_first_layoutget(lo);
2247                 nfs_layoutget_end(lo);
2248                 pnfs_put_layout_hdr(lo);
2249                 return;
2250         }
2251         data->lgp = lgp;
2252         data->o_arg.lg_args = &lgp->args;
2253         data->o_res.lg_res = &lgp->res;
2254 }
2255
2256 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2257                                      struct nfs_open_context *ctx)
2258 {
2259         struct pnfs_layout_range rng = {
2260                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2261                           IOMODE_RW: IOMODE_READ,
2262                 .offset = 0,
2263                 .length = NFS4_MAX_UINT64,
2264         };
2265         struct nfs4_layoutget *lgp;
2266
2267         lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2268                                              &rng, GFP_KERNEL);
2269         if (!lgp)
2270                 return;
2271         data->lgp = lgp;
2272         data->o_arg.lg_args = &lgp->args;
2273         data->o_res.lg_res = &lgp->res;
2274 }
2275
2276 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2277                          struct nfs_open_context *ctx)
2278 {
2279         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2280
2281         if (!(pnfs_enabled_sb(server) &&
2282               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2283                 return;
2284         /* Could check on max_ops, but currently hardcoded high enough */
2285         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2286                 return;
2287         if (data->state)
2288                 _lgopen_prepare_attached(data, ctx);
2289         else
2290                 _lgopen_prepare_floating(data, ctx);
2291 }
2292
2293 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2294                        struct nfs_open_context *ctx)
2295 {
2296         struct pnfs_layout_hdr *lo;
2297         struct pnfs_layout_segment *lseg;
2298         struct nfs_server *srv = NFS_SERVER(ino);
2299         u32 iomode;
2300
2301         if (!lgp)
2302                 return;
2303         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2304         if (lgp->res.status) {
2305                 switch (lgp->res.status) {
2306                 default:
2307                         break;
2308                 /*
2309                  * Halt lgopen attempts if the server doesn't recognise
2310                  * the "current stateid" value, the layout type, or the
2311                  * layoutget operation as being valid.
2312                  * Also if it complains about too many ops in the compound
2313                  * or of the request/reply being too big.
2314                  */
2315                 case -NFS4ERR_BAD_STATEID:
2316                 case -NFS4ERR_NOTSUPP:
2317                 case -NFS4ERR_REP_TOO_BIG:
2318                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2319                 case -NFS4ERR_REQ_TOO_BIG:
2320                 case -NFS4ERR_TOO_MANY_OPS:
2321                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2322                         srv->caps &= ~NFS_CAP_LGOPEN;
2323                 }
2324                 return;
2325         }
2326         if (!lgp->args.inode) {
2327                 lo = _pnfs_grab_empty_layout(ino, ctx);
2328                 if (!lo)
2329                         return;
2330                 lgp->args.inode = ino;
2331         } else
2332                 lo = NFS_I(lgp->args.inode)->layout;
2333
2334         lseg = pnfs_layout_process(lgp);
2335         if (!IS_ERR(lseg)) {
2336                 iomode = lgp->args.range.iomode;
2337                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2338                 pnfs_put_lseg(lseg);
2339         }
2340 }
2341
2342 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2343 {
2344         if (lgp != NULL) {
2345                 struct inode *inode = lgp->args.inode;
2346                 if (inode) {
2347                         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2348                         pnfs_clear_first_layoutget(lo);
2349                         nfs_layoutget_end(lo);
2350                 }
2351                 pnfs_layoutget_free(lgp);
2352         }
2353 }
2354
2355 struct pnfs_layout_segment *
2356 pnfs_layout_process(struct nfs4_layoutget *lgp)
2357 {
2358         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2359         struct nfs4_layoutget_res *res = &lgp->res;
2360         struct pnfs_layout_segment *lseg;
2361         struct inode *ino = lo->plh_inode;
2362         LIST_HEAD(free_me);
2363
2364         if (!pnfs_sanity_check_layout_range(&res->range))
2365                 return ERR_PTR(-EINVAL);
2366
2367         /* Inject layout blob into I/O device driver */
2368         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2369         if (IS_ERR_OR_NULL(lseg)) {
2370                 if (!lseg)
2371                         lseg = ERR_PTR(-ENOMEM);
2372
2373                 dprintk("%s: Could not allocate layout: error %ld\n",
2374                        __func__, PTR_ERR(lseg));
2375                 return lseg;
2376         }
2377
2378         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2379
2380         spin_lock(&ino->i_lock);
2381         if (pnfs_layoutgets_blocked(lo)) {
2382                 dprintk("%s forget reply due to state\n", __func__);
2383                 goto out_forget;
2384         }
2385
2386         if (!pnfs_layout_is_valid(lo)) {
2387                 /* We have a completely new layout */
2388                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2389         } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2390                 /* existing state ID, make sure the sequence number matches. */
2391                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2392                         dprintk("%s forget reply due to sequence\n", __func__);
2393                         goto out_forget;
2394                 }
2395                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2396         } else {
2397                 /*
2398                  * We got an entirely new state ID.  Mark all segments for the
2399                  * inode invalid, and retry the layoutget
2400                  */
2401                 pnfs_mark_layout_stateid_invalid(lo, &free_me);
2402                 goto out_forget;
2403         }
2404
2405         pnfs_get_lseg(lseg);
2406         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2407
2408
2409         if (res->return_on_close)
2410                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2411
2412         spin_unlock(&ino->i_lock);
2413         pnfs_free_lseg_list(&free_me);
2414         return lseg;
2415
2416 out_forget:
2417         spin_unlock(&ino->i_lock);
2418         lseg->pls_layout = lo;
2419         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2420         return ERR_PTR(-EAGAIN);
2421 }
2422
2423 /**
2424  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2425  * @lo: pointer to layout header
2426  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2427  * @return_range: describe layout segment ranges to be returned
2428  * @seq: stateid seqid to match
2429  *
2430  * This function is mainly intended for use by layoutrecall. It attempts
2431  * to free the layout segment immediately, or else to mark it for return
2432  * as soon as its reference count drops to zero.
2433  *
2434  * Returns
2435  * - 0: a layoutreturn needs to be scheduled.
2436  * - EBUSY: there are layout segment that are still in use.
2437  * - ENOENT: there are no layout segments that need to be returned.
2438  */
2439 int
2440 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2441                                 struct list_head *tmp_list,
2442                                 const struct pnfs_layout_range *return_range,
2443                                 u32 seq)
2444 {
2445         struct pnfs_layout_segment *lseg, *next;
2446         int remaining = 0;
2447
2448         dprintk("%s:Begin lo %p\n", __func__, lo);
2449
2450         assert_spin_locked(&lo->plh_inode->i_lock);
2451
2452         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2453                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2454                         dprintk("%s: marking lseg %p iomode %d "
2455                                 "offset %llu length %llu\n", __func__,
2456                                 lseg, lseg->pls_range.iomode,
2457                                 lseg->pls_range.offset,
2458                                 lseg->pls_range.length);
2459                         if (mark_lseg_invalid(lseg, tmp_list))
2460                                 continue;
2461                         remaining++;
2462                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2463                 }
2464
2465         if (remaining) {
2466                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2467                 return -EBUSY;
2468         }
2469
2470         if (!list_empty(&lo->plh_return_segs)) {
2471                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2472                 return 0;
2473         }
2474
2475         return -ENOENT;
2476 }
2477
2478 static void
2479 pnfs_mark_layout_for_return(struct inode *inode,
2480                             const struct pnfs_layout_range *range)
2481 {
2482         struct pnfs_layout_hdr *lo;
2483         bool return_now = false;
2484
2485         spin_lock(&inode->i_lock);
2486         lo = NFS_I(inode)->layout;
2487         if (!pnfs_layout_is_valid(lo)) {
2488                 spin_unlock(&inode->i_lock);
2489                 return;
2490         }
2491         pnfs_set_plh_return_info(lo, range->iomode, 0);
2492         /*
2493          * mark all matching lsegs so that we are sure to have no live
2494          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2495          * for how it works.
2496          */
2497         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2498                 const struct cred *cred;
2499                 nfs4_stateid stateid;
2500                 enum pnfs_iomode iomode;
2501
2502                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2503                 spin_unlock(&inode->i_lock);
2504                 if (return_now)
2505                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2506         } else {
2507                 spin_unlock(&inode->i_lock);
2508                 nfs_commit_inode(inode, 0);
2509         }
2510 }
2511
2512 void pnfs_error_mark_layout_for_return(struct inode *inode,
2513                                        struct pnfs_layout_segment *lseg)
2514 {
2515         struct pnfs_layout_range range = {
2516                 .iomode = lseg->pls_range.iomode,
2517                 .offset = 0,
2518                 .length = NFS4_MAX_UINT64,
2519         };
2520
2521         pnfs_mark_layout_for_return(inode, &range);
2522 }
2523 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2524
2525 static bool
2526 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2527 {
2528         return pnfs_layout_is_valid(lo) &&
2529                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2530                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2531 }
2532
2533 static struct pnfs_layout_segment *
2534 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2535                      const struct pnfs_layout_range *range,
2536                      enum pnfs_iomode iomode)
2537 {
2538         struct pnfs_layout_segment *lseg;
2539
2540         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2541                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2542                         continue;
2543                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2544                         continue;
2545                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2546                         continue;
2547                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2548                         return lseg;
2549         }
2550         return NULL;
2551 }
2552
2553 /* Find open file states whose mode matches that of the range */
2554 static bool
2555 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2556                                  const struct pnfs_layout_range *range)
2557 {
2558         struct list_head *head;
2559         struct nfs_open_context *ctx;
2560         fmode_t mode = 0;
2561
2562         if (!pnfs_layout_can_be_returned(lo) ||
2563             !pnfs_find_first_lseg(lo, range, range->iomode))
2564                 return false;
2565
2566         head = &NFS_I(lo->plh_inode)->open_files;
2567         list_for_each_entry_rcu(ctx, head, list) {
2568                 if (ctx->state)
2569                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2570         }
2571
2572         switch (range->iomode) {
2573         default:
2574                 break;
2575         case IOMODE_READ:
2576                 mode &= ~FMODE_WRITE;
2577                 break;
2578         case IOMODE_RW:
2579                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2580                         mode &= ~FMODE_READ;
2581         }
2582         return mode == 0;
2583 }
2584
2585 static int
2586 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2587 {
2588         const struct pnfs_layout_range *range = data;
2589         struct pnfs_layout_hdr *lo;
2590         struct inode *inode;
2591 restart:
2592         rcu_read_lock();
2593         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2594                 if (!pnfs_layout_can_be_returned(lo) ||
2595                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2596                         continue;
2597                 inode = lo->plh_inode;
2598                 spin_lock(&inode->i_lock);
2599                 if (!pnfs_should_return_unused_layout(lo, range)) {
2600                         spin_unlock(&inode->i_lock);
2601                         continue;
2602                 }
2603                 spin_unlock(&inode->i_lock);
2604                 inode = pnfs_grab_inode_layout_hdr(lo);
2605                 if (!inode)
2606                         continue;
2607                 rcu_read_unlock();
2608                 pnfs_mark_layout_for_return(inode, range);
2609                 iput(inode);
2610                 cond_resched();
2611                 goto restart;
2612         }
2613         rcu_read_unlock();
2614         return 0;
2615 }
2616
2617 void
2618 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2619                                  enum pnfs_iomode iomode)
2620 {
2621         struct pnfs_layout_range range = {
2622                 .iomode = iomode,
2623                 .offset = 0,
2624                 .length = NFS4_MAX_UINT64,
2625         };
2626
2627         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2628                         &range);
2629 }
2630
2631 void
2632 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2633 {
2634         if (pgio->pg_lseg == NULL ||
2635             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2636                 return;
2637         pnfs_put_lseg(pgio->pg_lseg);
2638         pgio->pg_lseg = NULL;
2639 }
2640 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2641
2642 /*
2643  * Check for any intersection between the request and the pgio->pg_lseg,
2644  * and if none, put this pgio->pg_lseg away.
2645  */
2646 void
2647 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2648 {
2649         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2650                 pnfs_put_lseg(pgio->pg_lseg);
2651                 pgio->pg_lseg = NULL;
2652         }
2653 }
2654 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2655
2656 void
2657 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2658 {
2659         u64 rd_size = req->wb_bytes;
2660
2661         pnfs_generic_pg_check_layout(pgio);
2662         pnfs_generic_pg_check_range(pgio, req);
2663         if (pgio->pg_lseg == NULL) {
2664                 if (pgio->pg_dreq == NULL)
2665                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2666                 else
2667                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2668
2669                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2670                                                    nfs_req_openctx(req),
2671                                                    req_offset(req),
2672                                                    rd_size,
2673                                                    IOMODE_READ,
2674                                                    false,
2675                                                    GFP_KERNEL);
2676                 if (IS_ERR(pgio->pg_lseg)) {
2677                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2678                         pgio->pg_lseg = NULL;
2679                         return;
2680                 }
2681         }
2682         /* If no lseg, fall back to read through mds */
2683         if (pgio->pg_lseg == NULL)
2684                 nfs_pageio_reset_read_mds(pgio);
2685
2686 }
2687 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2688
2689 void
2690 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2691                            struct nfs_page *req, u64 wb_size)
2692 {
2693         pnfs_generic_pg_check_layout(pgio);
2694         pnfs_generic_pg_check_range(pgio, req);
2695         if (pgio->pg_lseg == NULL) {
2696                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2697                                                    nfs_req_openctx(req),
2698                                                    req_offset(req),
2699                                                    wb_size,
2700                                                    IOMODE_RW,
2701                                                    false,
2702                                                    GFP_KERNEL);
2703                 if (IS_ERR(pgio->pg_lseg)) {
2704                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2705                         pgio->pg_lseg = NULL;
2706                         return;
2707                 }
2708         }
2709         /* If no lseg, fall back to write through mds */
2710         if (pgio->pg_lseg == NULL)
2711                 nfs_pageio_reset_write_mds(pgio);
2712 }
2713 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2714
2715 void
2716 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2717 {
2718         if (desc->pg_lseg) {
2719                 pnfs_put_lseg(desc->pg_lseg);
2720                 desc->pg_lseg = NULL;
2721         }
2722 }
2723 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2724
2725 /*
2726  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2727  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2728  */
2729 size_t
2730 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2731                      struct nfs_page *prev, struct nfs_page *req)
2732 {
2733         unsigned int size;
2734         u64 seg_end, req_start, seg_left;
2735
2736         size = nfs_generic_pg_test(pgio, prev, req);
2737         if (!size)
2738                 return 0;
2739
2740         /*
2741          * 'size' contains the number of bytes left in the current page (up
2742          * to the original size asked for in @req->wb_bytes).
2743          *
2744          * Calculate how many bytes are left in the layout segment
2745          * and if there are less bytes than 'size', return that instead.
2746          *
2747          * Please also note that 'end_offset' is actually the offset of the
2748          * first byte that lies outside the pnfs_layout_range. FIXME?
2749          *
2750          */
2751         if (pgio->pg_lseg) {
2752                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2753                                      pgio->pg_lseg->pls_range.length);
2754                 req_start = req_offset(req);
2755
2756                 /* start of request is past the last byte of this segment */
2757                 if (req_start >= seg_end)
2758                         return 0;
2759
2760                 /* adjust 'size' iff there are fewer bytes left in the
2761                  * segment than what nfs_generic_pg_test returned */
2762                 seg_left = seg_end - req_start;
2763                 if (seg_left < size)
2764                         size = (unsigned int)seg_left;
2765         }
2766
2767         return size;
2768 }
2769 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2770
2771 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2772 {
2773         struct nfs_pageio_descriptor pgio;
2774
2775         /* Resend all requests through the MDS */
2776         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2777                               hdr->completion_ops);
2778         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2779         return nfs_pageio_resend(&pgio, hdr);
2780 }
2781 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2782
2783 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2784 {
2785
2786         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2787         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2788             PNFS_LAYOUTRET_ON_ERROR) {
2789                 pnfs_return_layout(hdr->inode);
2790         }
2791         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2792                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2793 }
2794
2795 /*
2796  * Called by non rpc-based layout drivers
2797  */
2798 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2799 {
2800         if (likely(!hdr->pnfs_error)) {
2801                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2802                                 hdr->mds_offset + hdr->res.count);
2803                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2804         }
2805         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2806         if (unlikely(hdr->pnfs_error))
2807                 pnfs_ld_handle_write_error(hdr);
2808         hdr->mds_ops->rpc_release(hdr);
2809 }
2810 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2811
2812 static void
2813 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2814                 struct nfs_pgio_header *hdr)
2815 {
2816         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2817
2818         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2819                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2820                 nfs_pageio_reset_write_mds(desc);
2821                 mirror->pg_recoalesce = 1;
2822         }
2823         hdr->completion_ops->completion(hdr);
2824 }
2825
2826 static enum pnfs_try_status
2827 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2828                         const struct rpc_call_ops *call_ops,
2829                         struct pnfs_layout_segment *lseg,
2830                         int how)
2831 {
2832         struct inode *inode = hdr->inode;
2833         enum pnfs_try_status trypnfs;
2834         struct nfs_server *nfss = NFS_SERVER(inode);
2835
2836         hdr->mds_ops = call_ops;
2837
2838         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2839                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2840         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2841         if (trypnfs != PNFS_NOT_ATTEMPTED)
2842                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2843         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2844         return trypnfs;
2845 }
2846
2847 static void
2848 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2849               struct nfs_pgio_header *hdr, int how)
2850 {
2851         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2852         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2853         enum pnfs_try_status trypnfs;
2854
2855         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2856         switch (trypnfs) {
2857         case PNFS_NOT_ATTEMPTED:
2858                 pnfs_write_through_mds(desc, hdr);
2859         case PNFS_ATTEMPTED:
2860                 break;
2861         case PNFS_TRY_AGAIN:
2862                 /* cleanup hdr and prepare to redo pnfs */
2863                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2864                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2865                         list_splice_init(&hdr->pages, &mirror->pg_list);
2866                         mirror->pg_recoalesce = 1;
2867                 }
2868                 hdr->mds_ops->rpc_release(hdr);
2869         }
2870 }
2871
2872 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2873 {
2874         pnfs_put_lseg(hdr->lseg);
2875         nfs_pgio_header_free(hdr);
2876 }
2877
2878 int
2879 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2880 {
2881         struct nfs_pgio_header *hdr;
2882         int ret;
2883
2884         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2885         if (!hdr) {
2886                 desc->pg_error = -ENOMEM;
2887                 return desc->pg_error;
2888         }
2889         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2890
2891         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2892         ret = nfs_generic_pgio(desc, hdr);
2893         if (!ret)
2894                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2895
2896         return ret;
2897 }
2898 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2899
2900 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2901 {
2902         struct nfs_pageio_descriptor pgio;
2903
2904         /* Resend all requests through the MDS */
2905         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2906         return nfs_pageio_resend(&pgio, hdr);
2907 }
2908 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2909
2910 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2911 {
2912         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2913         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2914             PNFS_LAYOUTRET_ON_ERROR) {
2915                 pnfs_return_layout(hdr->inode);
2916         }
2917         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2918                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2919 }
2920
2921 /*
2922  * Called by non rpc-based layout drivers
2923  */
2924 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2925 {
2926         if (likely(!hdr->pnfs_error))
2927                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2928         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2929         if (unlikely(hdr->pnfs_error))
2930                 pnfs_ld_handle_read_error(hdr);
2931         hdr->mds_ops->rpc_release(hdr);
2932 }
2933 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2934
2935 static void
2936 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2937                 struct nfs_pgio_header *hdr)
2938 {
2939         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2940
2941         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2942                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2943                 nfs_pageio_reset_read_mds(desc);
2944                 mirror->pg_recoalesce = 1;
2945         }
2946         hdr->completion_ops->completion(hdr);
2947 }
2948
2949 /*
2950  * Call the appropriate parallel I/O subsystem read function.
2951  */
2952 static enum pnfs_try_status
2953 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2954                        const struct rpc_call_ops *call_ops,
2955                        struct pnfs_layout_segment *lseg)
2956 {
2957         struct inode *inode = hdr->inode;
2958         struct nfs_server *nfss = NFS_SERVER(inode);
2959         enum pnfs_try_status trypnfs;
2960
2961         hdr->mds_ops = call_ops;
2962
2963         dprintk("%s: Reading ino:%lu %u@%llu\n",
2964                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2965
2966         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2967         if (trypnfs != PNFS_NOT_ATTEMPTED)
2968                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2969         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2970         return trypnfs;
2971 }
2972
2973 /* Resend all requests through pnfs. */
2974 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
2975                            unsigned int mirror_idx)
2976 {
2977         struct nfs_pageio_descriptor pgio;
2978
2979         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2980                 /* Prevent deadlocks with layoutreturn! */
2981                 pnfs_put_lseg(hdr->lseg);
2982                 hdr->lseg = NULL;
2983
2984                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2985                                         hdr->completion_ops);
2986                 pgio.pg_mirror_idx = mirror_idx;
2987                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2988         }
2989 }
2990 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2991
2992 static void
2993 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2994 {
2995         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2996         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2997         enum pnfs_try_status trypnfs;
2998
2999         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3000         switch (trypnfs) {
3001         case PNFS_NOT_ATTEMPTED:
3002                 pnfs_read_through_mds(desc, hdr);
3003         case PNFS_ATTEMPTED:
3004                 break;
3005         case PNFS_TRY_AGAIN:
3006                 /* cleanup hdr and prepare to redo pnfs */
3007                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3008                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3009                         list_splice_init(&hdr->pages, &mirror->pg_list);
3010                         mirror->pg_recoalesce = 1;
3011                 }
3012                 hdr->mds_ops->rpc_release(hdr);
3013         }
3014 }
3015
3016 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3017 {
3018         pnfs_put_lseg(hdr->lseg);
3019         nfs_pgio_header_free(hdr);
3020 }
3021
3022 int
3023 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3024 {
3025         struct nfs_pgio_header *hdr;
3026         int ret;
3027
3028         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3029         if (!hdr) {
3030                 desc->pg_error = -ENOMEM;
3031                 return desc->pg_error;
3032         }
3033         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3034         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3035         ret = nfs_generic_pgio(desc, hdr);
3036         if (!ret)
3037                 pnfs_do_read(desc, hdr);
3038         return ret;
3039 }
3040 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3041
3042 static void pnfs_clear_layoutcommitting(struct inode *inode)
3043 {
3044         unsigned long *bitlock = &NFS_I(inode)->flags;
3045
3046         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3047         smp_mb__after_atomic();
3048         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3049 }
3050
3051 /*
3052  * There can be multiple RW segments.
3053  */
3054 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3055 {
3056         struct pnfs_layout_segment *lseg;
3057
3058         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3059                 if (lseg->pls_range.iomode == IOMODE_RW &&
3060                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3061                         list_add(&lseg->pls_lc_list, listp);
3062         }
3063 }
3064
3065 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3066 {
3067         struct pnfs_layout_segment *lseg, *tmp;
3068
3069         /* Matched by references in pnfs_set_layoutcommit */
3070         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3071                 list_del_init(&lseg->pls_lc_list);
3072                 pnfs_put_lseg(lseg);
3073         }
3074
3075         pnfs_clear_layoutcommitting(inode);
3076 }
3077
3078 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3079 {
3080         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3081 }
3082 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3083
3084 void
3085 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3086                 loff_t end_pos)
3087 {
3088         struct nfs_inode *nfsi = NFS_I(inode);
3089         bool mark_as_dirty = false;
3090
3091         spin_lock(&inode->i_lock);
3092         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3093                 nfsi->layout->plh_lwb = end_pos;
3094                 mark_as_dirty = true;
3095                 dprintk("%s: Set layoutcommit for inode %lu ",
3096                         __func__, inode->i_ino);
3097         } else if (end_pos > nfsi->layout->plh_lwb)
3098                 nfsi->layout->plh_lwb = end_pos;
3099         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3100                 /* references matched in nfs4_layoutcommit_release */
3101                 pnfs_get_lseg(lseg);
3102         }
3103         spin_unlock(&inode->i_lock);
3104         dprintk("%s: lseg %p end_pos %llu\n",
3105                 __func__, lseg, nfsi->layout->plh_lwb);
3106
3107         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3108          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3109         if (mark_as_dirty)
3110                 mark_inode_dirty_sync(inode);
3111 }
3112 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3113
3114 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3115 {
3116         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3117
3118         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3119                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3120         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3121 }
3122
3123 /*
3124  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3125  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3126  * data to disk to allow the server to recover the data if it crashes.
3127  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3128  * is off, and a COMMIT is sent to a data server, or
3129  * if WRITEs to a data server return NFS_DATA_SYNC.
3130  */
3131 int
3132 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3133 {
3134         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3135         struct nfs4_layoutcommit_data *data;
3136         struct nfs_inode *nfsi = NFS_I(inode);
3137         loff_t end_pos;
3138         int status;
3139
3140         if (!pnfs_layoutcommit_outstanding(inode))
3141                 return 0;
3142
3143         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3144
3145         status = -EAGAIN;
3146         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3147                 if (!sync)
3148                         goto out;
3149                 status = wait_on_bit_lock_action(&nfsi->flags,
3150                                 NFS_INO_LAYOUTCOMMITTING,
3151                                 nfs_wait_bit_killable,
3152                                 TASK_KILLABLE);
3153                 if (status)
3154                         goto out;
3155         }
3156
3157         status = -ENOMEM;
3158         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3159         data = kzalloc(sizeof(*data), GFP_NOFS);
3160         if (!data)
3161                 goto clear_layoutcommitting;
3162
3163         status = 0;
3164         spin_lock(&inode->i_lock);
3165         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3166                 goto out_unlock;
3167
3168         INIT_LIST_HEAD(&data->lseg_list);
3169         pnfs_list_write_lseg(inode, &data->lseg_list);
3170
3171         end_pos = nfsi->layout->plh_lwb;
3172
3173         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3174         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3175         spin_unlock(&inode->i_lock);
3176
3177         data->args.inode = inode;
3178         nfs_fattr_init(&data->fattr);
3179         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3180         data->res.fattr = &data->fattr;
3181         if (end_pos != 0)
3182                 data->args.lastbytewritten = end_pos - 1;
3183         else
3184                 data->args.lastbytewritten = U64_MAX;
3185         data->res.server = NFS_SERVER(inode);
3186
3187         if (ld->prepare_layoutcommit) {
3188                 status = ld->prepare_layoutcommit(&data->args);
3189                 if (status) {
3190                         put_cred(data->cred);
3191                         spin_lock(&inode->i_lock);
3192                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3193                         if (end_pos > nfsi->layout->plh_lwb)
3194                                 nfsi->layout->plh_lwb = end_pos;
3195                         goto out_unlock;
3196                 }
3197         }
3198
3199
3200         status = nfs4_proc_layoutcommit(data, sync);
3201 out:
3202         if (status)
3203                 mark_inode_dirty_sync(inode);
3204         dprintk("<-- %s status %d\n", __func__, status);
3205         return status;
3206 out_unlock:
3207         spin_unlock(&inode->i_lock);
3208         kfree(data);
3209 clear_layoutcommitting:
3210         pnfs_clear_layoutcommitting(inode);
3211         goto out;
3212 }
3213 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3214
3215 int
3216 pnfs_generic_sync(struct inode *inode, bool datasync)
3217 {
3218         return pnfs_layoutcommit_inode(inode, true);
3219 }
3220 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3221
3222 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3223 {
3224         struct nfs4_threshold *thp;
3225
3226         thp = kzalloc(sizeof(*thp), GFP_NOFS);
3227         if (!thp) {
3228                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3229                 return NULL;
3230         }
3231         return thp;
3232 }
3233
3234 #if IS_ENABLED(CONFIG_NFS_V4_2)
3235 int
3236 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3237 {
3238         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3239         struct nfs_server *server = NFS_SERVER(inode);
3240         struct nfs_inode *nfsi = NFS_I(inode);
3241         struct nfs42_layoutstat_data *data;
3242         struct pnfs_layout_hdr *hdr;
3243         int status = 0;
3244
3245         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3246                 goto out;
3247
3248         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3249                 goto out;
3250
3251         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3252                 goto out;
3253
3254         spin_lock(&inode->i_lock);
3255         if (!NFS_I(inode)->layout) {
3256                 spin_unlock(&inode->i_lock);
3257                 goto out_clear_layoutstats;
3258         }
3259         hdr = NFS_I(inode)->layout;
3260         pnfs_get_layout_hdr(hdr);
3261         spin_unlock(&inode->i_lock);
3262
3263         data = kzalloc(sizeof(*data), gfp_flags);
3264         if (!data) {
3265                 status = -ENOMEM;
3266                 goto out_put;
3267         }
3268
3269         data->args.fh = NFS_FH(inode);
3270         data->args.inode = inode;
3271         status = ld->prepare_layoutstats(&data->args);
3272         if (status)
3273                 goto out_free;
3274
3275         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3276
3277 out:
3278         dprintk("%s returns %d\n", __func__, status);
3279         return status;
3280
3281 out_free:
3282         kfree(data);
3283 out_put:
3284         pnfs_put_layout_hdr(hdr);
3285 out_clear_layoutstats:
3286         smp_mb__before_atomic();
3287         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3288         smp_mb__after_atomic();
3289         goto out;
3290 }
3291 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3292 #endif
3293
3294 unsigned int layoutstats_timer;
3295 module_param(layoutstats_timer, uint, 0644);
3296 EXPORT_SYMBOL_GPL(layoutstats_timer);