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 static int pnfs_layout_return_on_reboot(struct pnfs_layout_hdr *lo);
  65
  66 /* Return the registered pnfs layout driver module matching given id */
  67 static struct pnfs_layoutdriver_type *
  68 find_pnfs_driver_locked(u32 id)
  69 {
  70         struct pnfs_layoutdriver_type *local;
  71
  72         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  73                 if (local->id == id)
  74                         goto out;
  75         local = NULL;
  76 out:
  77         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  78         return local;
  79 }
  80
  81 static struct pnfs_layoutdriver_type *
  82 find_pnfs_driver(u32 id)
  83 {
  84         struct pnfs_layoutdriver_type *local;
  85
  86         spin_lock(&pnfs_spinlock);
  87         local = find_pnfs_driver_locked(id);
  88         if (local != NULL && !try_module_get(local->owner)) {
  89                 dprintk("%s: Could not grab reference on module\n", __func__);
  90                 local = NULL;
  91         }
  92         spin_unlock(&pnfs_spinlock);
  93         return local;
  94 }
  95
  96 const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
  97 {
  98         return find_pnfs_driver(id);
  99 }
 100
 101 void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
 102 {
 103         if (ld)
 104                 module_put(ld->owner);
 105 }
 106
 107 void
 108 unset_pnfs_layoutdriver(struct nfs_server *nfss)
 109 {
 110         if (nfss->pnfs_curr_ld) {
 111                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
 112                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 113                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 114                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 115                         nfs4_deviceid_purge_client(nfss->nfs_client);
 116                 module_put(nfss->pnfs_curr_ld->owner);
 117         }
 118         nfss->pnfs_curr_ld = NULL;
 119 }
 120
 121 /*
 122  * When the server sends a list of layout types, we choose one in the order
 123  * given in the list below.
 124  *
 125  * FIXME: should this list be configurable in some fashion? module param?
 126  *        mount option? something else?
 127  */
 128 static const u32 ld_prefs[] = {
 129         LAYOUT_SCSI,
 130         LAYOUT_BLOCK_VOLUME,
 131         LAYOUT_OSD2_OBJECTS,
 132         LAYOUT_FLEX_FILES,
 133         LAYOUT_NFSV4_1_FILES,
 134         0
 135 };
 136
 137 static int
 138 ld_cmp(const void *e1, const void *e2)
 139 {
 140         u32 ld1 = *((u32 *)e1);
 141         u32 ld2 = *((u32 *)e2);
 142         int i;
 143
 144         for (i = 0; ld_prefs[i] != 0; i++) {
 145                 if (ld1 == ld_prefs[i])
 146                         return -1;
 147
 148                 if (ld2 == ld_prefs[i])
 149                         return 1;
 150         }
 151         return 0;
 152 }
 153
 154 /*
 155  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 156  * Currently only one pNFS layout driver per filesystem is supported.
 157  *
 158  * @ids array of layout types supported by MDS.
 159  */
 160 void
 161 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 162                       struct nfs_fsinfo *fsinfo)
 163 {
 164         struct pnfs_layoutdriver_type *ld_type = NULL;
 165         u32 id;
 166         int i;
 167
 168         if (fsinfo->nlayouttypes == 0)
 169                 goto out_no_driver;
 170         if (!(server->nfs_client->cl_exchange_flags &
 171                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 172                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 173                         __func__, server->nfs_client->cl_exchange_flags);
 174                 goto out_no_driver;
 175         }
 176
 177         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 178                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 179
 180         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 181                 id = fsinfo->layouttype[i];
 182                 ld_type = find_pnfs_driver(id);
 183                 if (!ld_type) {
 184                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 185                                         id);
 186                         ld_type = find_pnfs_driver(id);
 187                 }
 188                 if (ld_type)
 189                         break;
 190         }
 191
 192         if (!ld_type) {
 193                 dprintk("%s: No pNFS module found!\n", __func__);
 194                 goto out_no_driver;
 195         }
 196
 197         server->pnfs_curr_ld = ld_type;
 198         if (ld_type->set_layoutdriver
 199             && ld_type->set_layoutdriver(server, mntfh)) {
 200                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 201                         "driver %u.\n", __func__, id);
 202                 module_put(ld_type->owner);
 203                 goto out_no_driver;
 204         }
 205         /* Bump the MDS count */
 206         atomic_inc(&server->nfs_client->cl_mds_count);
 207
 208         dprintk("%s: pNFS module for %u set\n", __func__, id);
 209         return;
 210
 211 out_no_driver:
 212         dprintk("%s: Using NFSv4 I/O\n", __func__);
 213         server->pnfs_curr_ld = NULL;
 214 }
 215
 216 int
 217 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 218 {
 219         int status = -EINVAL;
 220         struct pnfs_layoutdriver_type *tmp;
 221
 222         if (ld_type->id == 0) {
 223                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 224                 return status;
 225         }
 226         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 227                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 228                        "alloc_lseg and free_lseg.\n", __func__);
 229                 return status;
 230         }
 231
 232         spin_lock(&pnfs_spinlock);
 233         tmp = find_pnfs_driver_locked(ld_type->id);
 234         if (!tmp) {
 235                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 236                 status = 0;
 237                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 238                         ld_type->name);
 239         } else {
 240                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 241                         __func__, ld_type->id);
 242         }
 243         spin_unlock(&pnfs_spinlock);
 244
 245         return status;
 246 }
 247 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 248
 249 void
 250 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 251 {
 252         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 253         spin_lock(&pnfs_spinlock);
 254         list_del(&ld_type->pnfs_tblid);
 255         spin_unlock(&pnfs_spinlock);
 256 }
 257 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 258
 259 /*
 260  * pNFS client layout cache
 261  */
 262
 263 /* Need to hold i_lock if caller does not already hold reference */
 264 void
 265 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 266 {
 267         refcount_inc(&lo->plh_refcount);
 268 }
 269
 270 static struct pnfs_layout_hdr *
 271 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 272 {
 273         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 274         return ld->alloc_layout_hdr(ino, gfp_flags);
 275 }
 276
 277 static void
 278 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 279 {
 280         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 281         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 282
 283         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
 284                 struct nfs_client *clp = server->nfs_client;
 285
 286                 spin_lock(&clp->cl_lock);
 287                 list_del_rcu(&lo->plh_layouts);
 288                 spin_unlock(&clp->cl_lock);
 289         }
 290         put_cred(lo->plh_lc_cred);
 291         return ld->free_layout_hdr(lo);
 292 }
 293
 294 static void
 295 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 296 {
 297         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 298         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 299         nfsi->layout = NULL;
 300         /* Reset MDS Threshold I/O counters */
 301         nfsi->write_io = 0;
 302         nfsi->read_io = 0;
 303 }
 304
 305 void
 306 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 307 {
 308         struct inode *inode;
 309         unsigned long i_state;
 310
 311         if (!lo)
 312                 return;
 313         inode = lo->plh_inode;
 314         pnfs_layoutreturn_before_put_layout_hdr(lo);
 315
 316         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 317                 if (!list_empty(&lo->plh_segs))
 318                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 319                 pnfs_detach_layout_hdr(lo);
 320                 i_state = inode->i_state;
 321                 spin_unlock(&inode->i_lock);
 322                 pnfs_free_layout_hdr(lo);
 323                 /* Notify pnfs_destroy_layout_final() that we're done */
 324                 if (i_state & (I_FREEING | I_CLEAR))
 325                         wake_up_var(lo);
 326         }
 327 }
 328
 329 static struct inode *
 330 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
 331 {
 332         struct inode *inode = igrab(lo->plh_inode);
 333         if (inode)
 334                 return inode;
 335         set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 336         return NULL;
 337 }
 338
 339 /*
 340  * Compare 2 layout stateid sequence ids, to see which is newer,
 341  * taking into account wraparound issues.
 342  */
 343 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 344 {
 345         return (s32)(s1 - s2) > 0;
 346 }
 347
 348 static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
 349 {
 350         if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
 351                 lo->plh_barrier = newseq;
 352 }
 353
 354 static void
 355 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 356                          u32 seq)
 357 {
 358         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 359                 iomode = IOMODE_ANY;
 360         lo->plh_return_iomode = iomode;
 361         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 362         /*
 363          * We must set lo->plh_return_seq to avoid livelocks with
 364          * pnfs_layout_need_return()
 365          */
 366         if (seq == 0)
 367                 seq = be32_to_cpu(lo->plh_stateid.seqid);
 368         if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
 369                 lo->plh_return_seq = seq;
 370         pnfs_barrier_update(lo, seq);
 371 }
 372
 373 static void
 374 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 375 {
 376         struct pnfs_layout_segment *lseg;
 377         lo->plh_return_iomode = 0;
 378         lo->plh_return_seq = 0;
 379         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 380         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 381                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 382                         continue;
 383                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 384         }
 385 }
 386
 387 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 388 {
 389         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 390         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 391         smp_mb__after_atomic();
 392         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 393         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 394 }
 395
 396 static void
 397 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 398                 struct list_head *free_me)
 399 {
 400         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 401         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 402         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 403                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 404         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 405                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 406 }
 407
 408 /*
 409  * Update the seqid of a layout stateid after receiving
 410  * NFS4ERR_OLD_STATEID
 411  */
 412 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
 413                 struct pnfs_layout_range *dst_range,
 414                 struct inode *inode)
 415 {
 416         struct pnfs_layout_hdr *lo;
 417         struct pnfs_layout_range range = {
 418                 .iomode = IOMODE_ANY,
 419                 .offset = 0,
 420                 .length = NFS4_MAX_UINT64,
 421         };
 422         bool ret = false;
 423         LIST_HEAD(head);
 424         int err;
 425
 426         spin_lock(&inode->i_lock);
 427         lo = NFS_I(inode)->layout;
 428         if (lo &&  pnfs_layout_is_valid(lo) &&
 429             nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 430                 /* Is our call using the most recent seqid? If so, bump it */
 431                 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
 432                         nfs4_stateid_seqid_inc(dst);
 433                         ret = true;
 434                         goto out;
 435                 }
 436                 /* Try to update the seqid to the most recent */
 437                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 438                 if (err != -EBUSY) {
 439                         dst->seqid = lo->plh_stateid.seqid;
 440                         *dst_range = range;
 441                         ret = true;
 442                 }
 443         }
 444 out:
 445         spin_unlock(&inode->i_lock);
 446         pnfs_free_lseg_list(&head);
 447         return ret;
 448 }
 449
 450 /*
 451  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 452  *
 453  * In order to continue using the pnfs_layout_hdr, a full recovery
 454  * is required.
 455  * Note that caller must hold inode->i_lock.
 456  */
 457 int
 458 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 459                 struct list_head *lseg_list)
 460 {
 461         struct pnfs_layout_range range = {
 462                 .iomode = IOMODE_ANY,
 463                 .offset = 0,
 464                 .length = NFS4_MAX_UINT64,
 465         };
 466         struct pnfs_layout_segment *lseg, *next;
 467
 468         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 469         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 470                 pnfs_clear_lseg_state(lseg, lseg_list);
 471         pnfs_clear_layoutreturn_info(lo);
 472         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 473         set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
 474         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 475             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 476                 pnfs_clear_layoutreturn_waitbit(lo);
 477         return !list_empty(&lo->plh_segs);
 478 }
 479
 480 static int pnfs_mark_layout_stateid_return(struct pnfs_layout_hdr *lo,
 481                                            struct list_head *lseg_list,
 482                                            enum pnfs_iomode iomode, u32 seq)
 483 {
 484         struct pnfs_layout_range range = {
 485                 .iomode = iomode,
 486                 .length = NFS4_MAX_UINT64,
 487         };
 488
 489         return pnfs_mark_matching_lsegs_return(lo, lseg_list, &range, seq);
 490 }
 491
 492 static int
 493 pnfs_iomode_to_fail_bit(u32 iomode)
 494 {
 495         return iomode == IOMODE_RW ?
 496                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 497 }
 498
 499 static void
 500 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 501 {
 502         lo->plh_retry_timestamp = jiffies;
 503         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 504                 refcount_inc(&lo->plh_refcount);
 505 }
 506
 507 static void
 508 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 509 {
 510         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 511                 refcount_dec(&lo->plh_refcount);
 512 }
 513
 514 static void
 515 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 516 {
 517         struct inode *inode = lo->plh_inode;
 518         struct pnfs_layout_range range = {
 519                 .iomode = iomode,
 520                 .offset = 0,
 521                 .length = NFS4_MAX_UINT64,
 522         };
 523         LIST_HEAD(head);
 524
 525         spin_lock(&inode->i_lock);
 526         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 527         pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 528         spin_unlock(&inode->i_lock);
 529         pnfs_free_lseg_list(&head);
 530         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 531                         iomode == IOMODE_RW ?  "RW" : "READ");
 532 }
 533
 534 static bool
 535 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 536 {
 537         unsigned long start, end;
 538         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 539
 540         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 541                 return false;
 542         end = jiffies;
 543         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 544         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 545                 /* It is time to retry the failed layoutgets */
 546                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 547                 return false;
 548         }
 549         return true;
 550 }
 551
 552 static void
 553 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 554                 const struct pnfs_layout_range *range,
 555                 const nfs4_stateid *stateid)
 556 {
 557         INIT_LIST_HEAD(&lseg->pls_list);
 558         INIT_LIST_HEAD(&lseg->pls_lc_list);
 559         INIT_LIST_HEAD(&lseg->pls_commits);
 560         refcount_set(&lseg->pls_refcount, 1);
 561         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 562         lseg->pls_layout = lo;
 563         lseg->pls_range = *range;
 564         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 565 }
 566
 567 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 568 {
 569         if (lseg != NULL) {
 570                 struct inode *inode = lseg->pls_layout->plh_inode;
 571                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 572         }
 573 }
 574
 575 static void
 576 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 577                 struct pnfs_layout_segment *lseg)
 578 {
 579         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 580         list_del_init(&lseg->pls_list);
 581         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 582         refcount_dec(&lo->plh_refcount);
 583         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 584                 return;
 585         if (list_empty(&lo->plh_segs) &&
 586             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 587             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 588                 if (atomic_read(&lo->plh_outstanding) == 0)
 589                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 590                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 591         }
 592 }
 593
 594 static bool
 595 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 596                 struct pnfs_layout_segment *lseg)
 597 {
 598         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 599             pnfs_layout_is_valid(lo)) {
 600                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 601                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 602                 return true;
 603         }
 604         return false;
 605 }
 606
 607 void
 608 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 609 {
 610         struct pnfs_layout_hdr *lo;
 611         struct inode *inode;
 612
 613         if (!lseg)
 614                 return;
 615
 616         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 617                 refcount_read(&lseg->pls_refcount),
 618                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 619
 620         lo = lseg->pls_layout;
 621         inode = lo->plh_inode;
 622
 623         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 624                 pnfs_get_layout_hdr(lo);
 625                 pnfs_layout_remove_lseg(lo, lseg);
 626                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 627                         lseg = NULL;
 628                 spin_unlock(&inode->i_lock);
 629                 pnfs_free_lseg(lseg);
 630                 pnfs_put_layout_hdr(lo);
 631         }
 632 }
 633 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 634
 635 /*
 636  * is l2 fully contained in l1?
 637  *   start1                             end1
 638  *   [----------------------------------)
 639  *           start2           end2
 640  *           [----------------)
 641  */
 642 static bool
 643 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 644                  const struct pnfs_layout_range *l2)
 645 {
 646         u64 start1 = l1->offset;
 647         u64 end1 = pnfs_end_offset(start1, l1->length);
 648         u64 start2 = l2->offset;
 649         u64 end2 = pnfs_end_offset(start2, l2->length);
 650
 651         return (start1 <= start2) && (end1 >= end2);
 652 }
 653
 654 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 655                 struct list_head *tmp_list)
 656 {
 657         if (!refcount_dec_and_test(&lseg->pls_refcount))
 658                 return false;
 659         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 660         list_add(&lseg->pls_list, tmp_list);
 661         return true;
 662 }
 663
 664 /* Returns 1 if lseg is removed from list, 0 otherwise */
 665 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 666                              struct list_head *tmp_list)
 667 {
 668         int rv = 0;
 669
 670         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 671                 /* Remove the reference keeping the lseg in the
 672                  * list.  It will now be removed when all
 673                  * outstanding io is finished.
 674                  */
 675                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 676                         refcount_read(&lseg->pls_refcount));
 677                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 678                         rv = 1;
 679         }
 680         return rv;
 681 }
 682
 683 static bool
 684 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 685                  const struct pnfs_layout_range *recall_range)
 686 {
 687         return (recall_range->iomode == IOMODE_ANY ||
 688                 lseg_range->iomode == recall_range->iomode) &&
 689                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 690 }
 691
 692 static bool
 693 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 694                 const struct pnfs_layout_range *recall_range,
 695                 u32 seq)
 696 {
 697         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 698                 return false;
 699         if (recall_range == NULL)
 700                 return true;
 701         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 702 }
 703
 704 /**
 705  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 706  * @lo: layout header containing the lsegs
 707  * @tmp_list: list head where doomed lsegs should go
 708  * @recall_range: optional recall range argument to match (may be NULL)
 709  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 710  *
 711  * Walk the list of lsegs in the layout header, and tear down any that should
 712  * be destroyed. If "recall_range" is specified then the segment must match
 713  * that range. If "seq" is non-zero, then only match segments that were handed
 714  * out at or before that sequence.
 715  *
 716  * Returns number of matching invalid lsegs remaining in list after scanning
 717  * it and purging them.
 718  */
 719 int
 720 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 721                             struct list_head *tmp_list,
 722                             const struct pnfs_layout_range *recall_range,
 723                             u32 seq)
 724 {
 725         struct pnfs_layout_segment *lseg, *next;
 726         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 727         int remaining = 0;
 728
 729         dprintk("%s:Begin lo %p\n", __func__, lo);
 730
 731         if (list_empty(&lo->plh_segs))
 732                 return 0;
 733         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 734                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 735                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 736                                 "offset %llu length %llu\n", __func__,
 737                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 738                                 lseg->pls_range.offset, lseg->pls_range.length);
 739                         if (mark_lseg_invalid(lseg, tmp_list))
 740                                 continue;
 741                         remaining++;
 742                         pnfs_lseg_cancel_io(server, lseg);
 743                 }
 744         dprintk("%s:Return %i\n", __func__, remaining);
 745         return remaining;
 746 }
 747
 748 static void
 749 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 750                 struct list_head *free_me,
 751                 const struct pnfs_layout_range *range,
 752                 u32 seq)
 753 {
 754         struct pnfs_layout_segment *lseg, *next;
 755
 756         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 757                 if (pnfs_match_lseg_recall(lseg, range, seq))
 758                         list_move_tail(&lseg->pls_list, free_me);
 759         }
 760 }
 761
 762 /* note free_me must contain lsegs from a single layout_hdr */
 763 void
 764 pnfs_free_lseg_list(struct list_head *free_me)
 765 {
 766         struct pnfs_layout_segment *lseg, *tmp;
 767
 768         if (list_empty(free_me))
 769                 return;
 770
 771         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 772                 list_del(&lseg->pls_list);
 773                 pnfs_free_lseg(lseg);
 774         }
 775 }
 776
 777 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 778 {
 779         struct pnfs_layout_hdr *lo;
 780         LIST_HEAD(tmp_list);
 781
 782         spin_lock(&nfsi->vfs_inode.i_lock);
 783         lo = nfsi->layout;
 784         if (lo) {
 785                 pnfs_get_layout_hdr(lo);
 786                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 787                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 788                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 789                 spin_unlock(&nfsi->vfs_inode.i_lock);
 790                 pnfs_free_lseg_list(&tmp_list);
 791                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 792                 pnfs_put_layout_hdr(lo);
 793         } else
 794                 spin_unlock(&nfsi->vfs_inode.i_lock);
 795         return lo;
 796 }
 797
 798 void pnfs_destroy_layout(struct nfs_inode *nfsi)
 799 {
 800         __pnfs_destroy_layout(nfsi);
 801 }
 802 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 803
 804 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
 805                                 struct pnfs_layout_hdr *lo)
 806 {
 807         bool ret;
 808
 809         spin_lock(&nfsi->vfs_inode.i_lock);
 810         ret = nfsi->layout != lo;
 811         spin_unlock(&nfsi->vfs_inode.i_lock);
 812         return ret;
 813 }
 814
 815 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 816 {
 817         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 818
 819         if (lo)
 820                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
 821 }
 822
 823 static bool
 824 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 825                 struct list_head *layout_list)
 826 {
 827         struct pnfs_layout_hdr *lo;
 828         bool ret = false;
 829
 830         spin_lock(&inode->i_lock);
 831         lo = NFS_I(inode)->layout;
 832         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 833                 pnfs_get_layout_hdr(lo);
 834                 list_add(&lo->plh_bulk_destroy, layout_list);
 835                 ret = true;
 836         }
 837         spin_unlock(&inode->i_lock);
 838         return ret;
 839 }
 840
 841 /* Caller must hold rcu_read_lock and clp->cl_lock */
 842 static int
 843 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 844                 struct nfs_server *server,
 845                 struct list_head *layout_list)
 846         __must_hold(&clp->cl_lock)
 847         __must_hold(RCU)
 848 {
 849         struct pnfs_layout_hdr *lo, *next;
 850         struct inode *inode;
 851
 852         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 853                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 854                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 855                     !list_empty(&lo->plh_bulk_destroy))
 856                         continue;
 857                 /* If the sb is being destroyed, just bail */
 858                 if (!nfs_sb_active(server->super))
 859                         break;
 860                 inode = pnfs_grab_inode_layout_hdr(lo);
 861                 if (inode != NULL) {
 862                         if (pnfs_layout_add_bulk_destroy_list(inode,
 863                                                 layout_list))
 864                                 continue;
 865                         rcu_read_unlock();
 866                         spin_unlock(&clp->cl_lock);
 867                         iput(inode);
 868                 } else {
 869                         rcu_read_unlock();
 870                         spin_unlock(&clp->cl_lock);
 871                 }
 872                 nfs_sb_deactive(server->super);
 873                 spin_lock(&clp->cl_lock);
 874                 rcu_read_lock();
 875                 return -EAGAIN;
 876         }
 877         return 0;
 878 }
 879
 880 static int
 881 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 882                                    enum pnfs_layout_destroy_mode mode)
 883 {
 884         struct pnfs_layout_hdr *lo;
 885         struct inode *inode;
 886         LIST_HEAD(lseg_list);
 887         int ret = 0;
 888
 889         while (!list_empty(layout_list)) {
 890                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 891                                 plh_bulk_destroy);
 892                 dprintk("%s freeing layout for inode %lu\n", __func__,
 893                         lo->plh_inode->i_ino);
 894                 inode = lo->plh_inode;
 895
 896                 pnfs_layoutcommit_inode(inode, false);
 897
 898                 spin_lock(&inode->i_lock);
 899                 list_del_init(&lo->plh_bulk_destroy);
 900                 if (mode == PNFS_LAYOUT_FILE_BULK_RETURN) {
 901                         pnfs_mark_layout_stateid_return(lo, &lseg_list,
 902                                                         IOMODE_ANY, 0);
 903                 } else if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 904                         if (mode == PNFS_LAYOUT_BULK_RETURN)
 905                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 906                         ret = -EAGAIN;
 907                 }
 908                 spin_unlock(&inode->i_lock);
 909                 pnfs_free_lseg_list(&lseg_list);
 910                 /* Free all lsegs that are attached to commit buckets */
 911                 nfs_commit_inode(inode, 0);
 912                 pnfs_put_layout_hdr(lo);
 913                 nfs_iput_and_deactive(inode);
 914         }
 915         return ret;
 916 }
 917
 918 int pnfs_layout_destroy_byfsid(struct nfs_client *clp, struct nfs_fsid *fsid,
 919                                enum pnfs_layout_destroy_mode mode)
 920 {
 921         struct nfs_server *server;
 922         LIST_HEAD(layout_list);
 923
 924         spin_lock(&clp->cl_lock);
 925         rcu_read_lock();
 926 restart:
 927         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 928                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 929                         continue;
 930                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 931                                 server,
 932                                 &layout_list) != 0)
 933                         goto restart;
 934         }
 935         rcu_read_unlock();
 936         spin_unlock(&clp->cl_lock);
 937
 938         return pnfs_layout_free_bulk_destroy_list(&layout_list, mode);
 939 }
 940
 941 static void pnfs_layout_build_destroy_list_byclient(struct nfs_client *clp,
 942                                                     struct list_head *list)
 943 {
 944         struct nfs_server *server;
 945
 946         spin_lock(&clp->cl_lock);
 947         rcu_read_lock();
 948 restart:
 949         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 950                 if (pnfs_layout_bulk_destroy_byserver_locked(clp, server,
 951                                                              list) != 0)
 952                         goto restart;
 953         }
 954         rcu_read_unlock();
 955         spin_unlock(&clp->cl_lock);
 956 }
 957
 958 static int pnfs_layout_do_destroy_byclid(struct nfs_client *clp,
 959                                          struct list_head *list,
 960                                          enum pnfs_layout_destroy_mode mode)
 961 {
 962         pnfs_layout_build_destroy_list_byclient(clp, list);
 963         return pnfs_layout_free_bulk_destroy_list(list, mode);
 964 }
 965
 966 int pnfs_layout_destroy_byclid(struct nfs_client *clp,
 967                                enum pnfs_layout_destroy_mode mode)
 968 {
 969         LIST_HEAD(layout_list);
 970
 971         return pnfs_layout_do_destroy_byclid(clp, &layout_list, mode);
 972 }
 973
 974 /*
 975  * Called by the state manager to remove all layouts established under an
 976  * expired lease.
 977  */
 978 void
 979 pnfs_destroy_all_layouts(struct nfs_client *clp)
 980 {
 981         nfs4_deviceid_mark_client_invalid(clp);
 982         nfs4_deviceid_purge_client(clp);
 983
 984         pnfs_layout_destroy_byclid(clp, PNFS_LAYOUT_INVALIDATE);
 985 }
 986
 987 static void pnfs_layout_build_recover_list_byclient(struct nfs_client *clp,
 988                                                     struct list_head *list)
 989 {
 990         struct nfs_server *server;
 991
 992         spin_lock(&clp->cl_lock);
 993         rcu_read_lock();
 994 restart:
 995         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 996                 if (!(server->caps & NFS_CAP_REBOOT_LAYOUTRETURN))
 997                         continue;
 998                 if (pnfs_layout_bulk_destroy_byserver_locked(clp, server,
 999                                                              list) != 0)
1000                         goto restart;
1001         }
1002         rcu_read_unlock();
1003         spin_unlock(&clp->cl_lock);
1004 }
1005
1006 static int pnfs_layout_bulk_list_reboot(struct list_head *list)
1007 {
1008         struct pnfs_layout_hdr *lo;
1009         struct nfs_server *server;
1010         int ret;
1011
1012         list_for_each_entry(lo, list, plh_bulk_destroy) {
1013                 server = NFS_SERVER(lo->plh_inode);
1014                 ret = pnfs_layout_return_on_reboot(lo);
1015                 switch (ret) {
1016                 case 0:
1017                         continue;
1018                 case -NFS4ERR_BAD_STATEID:
1019                         server->caps &= ~NFS_CAP_REBOOT_LAYOUTRETURN;
1020                         break;
1021                 case -NFS4ERR_NO_GRACE:
1022                         break;
1023                 default:
1024                         goto err;
1025                 }
1026                 break;
1027         }
1028         return 0;
1029 err:
1030         return ret;
1031 }
1032
1033 int pnfs_layout_handle_reboot(struct nfs_client *clp)
1034 {
1035         LIST_HEAD(list);
1036         int ret = 0, ret2;
1037
1038         pnfs_layout_build_recover_list_byclient(clp, &list);
1039         if (!list_empty(&list))
1040                 ret = pnfs_layout_bulk_list_reboot(&list);
1041         ret2 = pnfs_layout_do_destroy_byclid(clp, &list,
1042                                              PNFS_LAYOUT_INVALIDATE);
1043         if (!ret)
1044                 ret = ret2;
1045         return (ret == 0) ?  0 : -EAGAIN;
1046 }
1047
1048 static void
1049 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
1050 {
1051         const struct cred *old;
1052
1053         if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
1054                 old = xchg(&lo->plh_lc_cred, get_cred(cred));
1055                 put_cred(old);
1056         }
1057 }
1058
1059 /* update lo->plh_stateid with new if is more recent */
1060 void
1061 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
1062                         const struct cred *cred, bool update_barrier)
1063 {
1064         u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
1065         u32 newseq = be32_to_cpu(new->seqid);
1066
1067         if (!pnfs_layout_is_valid(lo)) {
1068                 pnfs_set_layout_cred(lo, cred);
1069                 nfs4_stateid_copy(&lo->plh_stateid, new);
1070                 lo->plh_barrier = newseq;
1071                 pnfs_clear_layoutreturn_info(lo);
1072                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1073                 return;
1074         }
1075
1076         if (pnfs_seqid_is_newer(newseq, oldseq))
1077                 nfs4_stateid_copy(&lo->plh_stateid, new);
1078
1079         if (update_barrier) {
1080                 pnfs_barrier_update(lo, newseq);
1081                 return;
1082         }
1083         /*
1084          * Because of wraparound, we want to keep the barrier
1085          * "close" to the current seqids. We really only want to
1086          * get here from a layoutget call.
1087          */
1088         if (atomic_read(&lo->plh_outstanding) == 1)
1089                  pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1090 }
1091
1092 static bool
1093 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1094                 const nfs4_stateid *stateid)
1095 {
1096         u32 seqid = be32_to_cpu(stateid->seqid);
1097
1098         return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1099 }
1100
1101 /* lget is set to 1 if called from inside send_layoutget call chain */
1102 static bool
1103 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1104 {
1105         return lo->plh_block_lgets ||
1106                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1107 }
1108
1109 static struct nfs_server *
1110 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1111 {
1112         struct nfs_server *server;
1113
1114         if (inode) {
1115                 server = NFS_SERVER(inode);
1116         } else {
1117                 struct dentry *parent_dir = dget_parent(ctx->dentry);
1118                 server = NFS_SERVER(parent_dir->d_inode);
1119                 dput(parent_dir);
1120         }
1121         return server;
1122 }
1123
1124 static void nfs4_free_pages(struct page **pages, size_t size)
1125 {
1126         int i;
1127
1128         if (!pages)
1129                 return;
1130
1131         for (i = 0; i < size; i++) {
1132                 if (!pages[i])
1133                         break;
1134                 __free_page(pages[i]);
1135         }
1136         kfree(pages);
1137 }
1138
1139 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1140 {
1141         struct page **pages;
1142         int i;
1143
1144         pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1145         if (!pages) {
1146                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1147                 return NULL;
1148         }
1149
1150         for (i = 0; i < size; i++) {
1151                 pages[i] = alloc_page(gfp_flags);
1152                 if (!pages[i]) {
1153                         dprintk("%s: failed to allocate page\n", __func__);
1154                         nfs4_free_pages(pages, i);
1155                         return NULL;
1156                 }
1157         }
1158
1159         return pages;
1160 }
1161
1162 static struct nfs4_layoutget *
1163 pnfs_alloc_init_layoutget_args(struct inode *ino,
1164            struct nfs_open_context *ctx,
1165            const nfs4_stateid *stateid,
1166            const struct pnfs_layout_range *range,
1167            gfp_t gfp_flags)
1168 {
1169         struct nfs_server *server = pnfs_find_server(ino, ctx);
1170         size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1171         size_t max_pages = max_response_pages(server);
1172         struct nfs4_layoutget *lgp;
1173
1174         dprintk("--> %s\n", __func__);
1175
1176         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1177         if (lgp == NULL)
1178                 return NULL;
1179
1180         if (max_reply_sz) {
1181                 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1182                 if (npages < max_pages)
1183                         max_pages = npages;
1184         }
1185
1186         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1187         if (!lgp->args.layout.pages) {
1188                 kfree(lgp);
1189                 return NULL;
1190         }
1191         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1192         lgp->res.layoutp = &lgp->args.layout;
1193
1194         /* Don't confuse uninitialised result and success */
1195         lgp->res.status = -NFS4ERR_DELAY;
1196
1197         lgp->args.minlength = PAGE_SIZE;
1198         if (lgp->args.minlength > range->length)
1199                 lgp->args.minlength = range->length;
1200         if (ino) {
1201                 loff_t i_size = i_size_read(ino);
1202
1203                 if (range->iomode == IOMODE_READ) {
1204                         if (range->offset >= i_size)
1205                                 lgp->args.minlength = 0;
1206                         else if (i_size - range->offset < lgp->args.minlength)
1207                                 lgp->args.minlength = i_size - range->offset;
1208                 }
1209         }
1210         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1211         pnfs_copy_range(&lgp->args.range, range);
1212         lgp->args.type = server->pnfs_curr_ld->id;
1213         lgp->args.inode = ino;
1214         lgp->args.ctx = get_nfs_open_context(ctx);
1215         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1216         lgp->gfp_flags = gfp_flags;
1217         lgp->cred = ctx->cred;
1218         return lgp;
1219 }
1220
1221 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1222 {
1223         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1224
1225         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1226         pnfs_put_layout_hdr(lgp->lo);
1227         put_nfs_open_context(lgp->args.ctx);
1228         kfree(lgp);
1229 }
1230
1231 static void pnfs_clear_layoutcommit(struct inode *inode,
1232                 struct list_head *head)
1233 {
1234         struct nfs_inode *nfsi = NFS_I(inode);
1235         struct pnfs_layout_segment *lseg, *tmp;
1236
1237         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1238                 return;
1239         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1240                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1241                         continue;
1242                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1243         }
1244 }
1245
1246 static void
1247 pnfs_layoutreturn_retry_later_locked(struct pnfs_layout_hdr *lo,
1248                                      const nfs4_stateid *arg_stateid,
1249                                      const struct pnfs_layout_range *range)
1250 {
1251         const struct pnfs_layout_segment *lseg;
1252         u32 seq = be32_to_cpu(arg_stateid->seqid);
1253
1254         if (pnfs_layout_is_valid(lo) &&
1255             nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid)) {
1256                 list_for_each_entry(lseg, &lo->plh_return_segs, pls_list) {
1257                         if (pnfs_seqid_is_newer(lseg->pls_seq, seq) ||
1258                             !pnfs_should_free_range(&lseg->pls_range, range))
1259                                 continue;
1260                         pnfs_set_plh_return_info(lo, range->iomode, seq);
1261                         break;
1262                 }
1263         }
1264 }
1265
1266 void pnfs_layoutreturn_retry_later(struct pnfs_layout_hdr *lo,
1267                                    const nfs4_stateid *arg_stateid,
1268                                    const struct pnfs_layout_range *range)
1269 {
1270         struct inode *inode = lo->plh_inode;
1271
1272         spin_lock(&inode->i_lock);
1273         pnfs_layoutreturn_retry_later_locked(lo, arg_stateid, range);
1274         pnfs_clear_layoutreturn_waitbit(lo);
1275         spin_unlock(&inode->i_lock);
1276 }
1277
1278 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1279                 const nfs4_stateid *arg_stateid,
1280                 const struct pnfs_layout_range *range,
1281                 const nfs4_stateid *stateid)
1282 {
1283         struct inode *inode = lo->plh_inode;
1284         LIST_HEAD(freeme);
1285
1286         spin_lock(&inode->i_lock);
1287         if (!nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1288                 goto out_unlock;
1289         if (stateid && pnfs_layout_is_valid(lo)) {
1290                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1291
1292                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1293                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1294                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1295         } else
1296                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1297 out_unlock:
1298         pnfs_clear_layoutreturn_waitbit(lo);
1299         spin_unlock(&inode->i_lock);
1300         pnfs_free_lseg_list(&freeme);
1301
1302 }
1303
1304 static bool
1305 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1306                 nfs4_stateid *stateid,
1307                 const struct cred **cred,
1308                 enum pnfs_iomode *iomode)
1309 {
1310         /* Serialise LAYOUTGET/LAYOUTRETURN */
1311         if (atomic_read(&lo->plh_outstanding) != 0)
1312                 return false;
1313         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1314                 return false;
1315         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1316         pnfs_get_layout_hdr(lo);
1317         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1318         *cred = get_cred(lo->plh_lc_cred);
1319         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1320                 if (lo->plh_return_seq != 0)
1321                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1322                 if (iomode != NULL)
1323                         *iomode = lo->plh_return_iomode;
1324                 pnfs_clear_layoutreturn_info(lo);
1325         } else if (iomode != NULL)
1326                 *iomode = IOMODE_ANY;
1327         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1328         return true;
1329 }
1330
1331 static void
1332 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1333                 struct pnfs_layout_hdr *lo,
1334                 const nfs4_stateid *stateid,
1335                 enum pnfs_iomode iomode)
1336 {
1337         struct inode *inode = lo->plh_inode;
1338
1339         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1340         args->inode = inode;
1341         args->range.iomode = iomode;
1342         args->range.offset = 0;
1343         args->range.length = NFS4_MAX_UINT64;
1344         args->layout = lo;
1345         nfs4_stateid_copy(&args->stateid, stateid);
1346 }
1347
1348 static int
1349 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1350                        const nfs4_stateid *stateid,
1351                        const struct cred **pcred,
1352                        enum pnfs_iomode iomode,
1353                        unsigned int flags)
1354 {
1355         struct inode *ino = lo->plh_inode;
1356         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1357         struct nfs4_layoutreturn *lrp;
1358         const struct cred *cred = *pcred;
1359         int status = 0;
1360
1361         *pcred = NULL;
1362         lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1363         if (unlikely(lrp == NULL)) {
1364                 status = -ENOMEM;
1365                 spin_lock(&ino->i_lock);
1366                 pnfs_clear_layoutreturn_waitbit(lo);
1367                 spin_unlock(&ino->i_lock);
1368                 put_cred(cred);
1369                 pnfs_put_layout_hdr(lo);
1370                 goto out;
1371         }
1372
1373         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1374         lrp->args.ld_private = &lrp->ld_private;
1375         lrp->clp = NFS_SERVER(ino)->nfs_client;
1376         lrp->cred = cred;
1377         if (ld->prepare_layoutreturn)
1378                 ld->prepare_layoutreturn(&lrp->args);
1379
1380         status = nfs4_proc_layoutreturn(lrp, flags);
1381 out:
1382         dprintk("<-- %s status: %d\n", __func__, status);
1383         return status;
1384 }
1385
1386 /* Return true if layoutreturn is needed */
1387 static bool
1388 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1389 {
1390         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1391                 return false;
1392         return pnfs_mark_layout_stateid_return(lo, &lo->plh_return_segs,
1393                                                lo->plh_return_iomode,
1394                                                lo->plh_return_seq) != EBUSY;
1395 }
1396
1397 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1398 {
1399         struct inode *inode= lo->plh_inode;
1400
1401         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1402                 return;
1403         spin_lock(&inode->i_lock);
1404         if (pnfs_layout_need_return(lo)) {
1405                 const struct cred *cred;
1406                 nfs4_stateid stateid;
1407                 enum pnfs_iomode iomode;
1408                 bool send;
1409
1410                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1411                 spin_unlock(&inode->i_lock);
1412                 if (send) {
1413                         /* Send an async layoutreturn so we dont deadlock */
1414                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
1415                                                PNFS_FL_LAYOUTRETURN_ASYNC);
1416                 }
1417         } else
1418                 spin_unlock(&inode->i_lock);
1419 }
1420
1421 /*
1422  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1423  * when the layout segment list is empty.
1424  *
1425  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1426  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1427  * deviceid is marked invalid.
1428  */
1429 int
1430 _pnfs_return_layout(struct inode *ino)
1431 {
1432         struct pnfs_layout_hdr *lo = NULL;
1433         struct nfs_inode *nfsi = NFS_I(ino);
1434         struct pnfs_layout_range range = {
1435                 .iomode         = IOMODE_ANY,
1436                 .offset         = 0,
1437                 .length         = NFS4_MAX_UINT64,
1438         };
1439         LIST_HEAD(tmp_list);
1440         const struct cred *cred;
1441         nfs4_stateid stateid;
1442         int status = 0;
1443         bool send, valid_layout;
1444
1445         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1446
1447         spin_lock(&ino->i_lock);
1448         lo = nfsi->layout;
1449         if (!lo) {
1450                 spin_unlock(&ino->i_lock);
1451                 dprintk("NFS: %s no layout to return\n", __func__);
1452                 goto out;
1453         }
1454         /* Reference matched in nfs4_layoutreturn_release */
1455         pnfs_get_layout_hdr(lo);
1456         /* Is there an outstanding layoutreturn ? */
1457         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1458                 spin_unlock(&ino->i_lock);
1459                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1460                                         TASK_UNINTERRUPTIBLE))
1461                         goto out_put_layout_hdr;
1462                 spin_lock(&ino->i_lock);
1463         }
1464         valid_layout = pnfs_layout_is_valid(lo);
1465         pnfs_clear_layoutcommit(ino, &tmp_list);
1466         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1467
1468         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1469                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1470
1471         /* Don't send a LAYOUTRETURN if list was initially empty */
1472         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1473                         !valid_layout) {
1474                 spin_unlock(&ino->i_lock);
1475                 dprintk("NFS: %s no layout segments to return\n", __func__);
1476                 goto out_wait_layoutreturn;
1477         }
1478
1479         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1480         spin_unlock(&ino->i_lock);
1481         if (send)
1482                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY,
1483                                                 0);
1484 out_wait_layoutreturn:
1485         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1486 out_put_layout_hdr:
1487         pnfs_free_lseg_list(&tmp_list);
1488         pnfs_put_layout_hdr(lo);
1489 out:
1490         dprintk("<-- %s status: %d\n", __func__, status);
1491         return status;
1492 }
1493
1494 int
1495 pnfs_commit_and_return_layout(struct inode *inode)
1496 {
1497         struct pnfs_layout_hdr *lo;
1498         int ret;
1499
1500         spin_lock(&inode->i_lock);
1501         lo = NFS_I(inode)->layout;
1502         if (lo == NULL) {
1503                 spin_unlock(&inode->i_lock);
1504                 return 0;
1505         }
1506         pnfs_get_layout_hdr(lo);
1507         /* Block new layoutgets and read/write to ds */
1508         lo->plh_block_lgets++;
1509         spin_unlock(&inode->i_lock);
1510         filemap_fdatawait(inode->i_mapping);
1511         ret = pnfs_layoutcommit_inode(inode, true);
1512         if (ret == 0)
1513                 ret = _pnfs_return_layout(inode);
1514         spin_lock(&inode->i_lock);
1515         lo->plh_block_lgets--;
1516         spin_unlock(&inode->i_lock);
1517         pnfs_put_layout_hdr(lo);
1518         return ret;
1519 }
1520
1521 static int pnfs_layout_return_on_reboot(struct pnfs_layout_hdr *lo)
1522 {
1523         struct inode *inode = lo->plh_inode;
1524         const struct cred *cred;
1525
1526         spin_lock(&inode->i_lock);
1527         if (!pnfs_layout_is_valid(lo)) {
1528                 spin_unlock(&inode->i_lock);
1529                 return 0;
1530         }
1531         cred = get_cred(lo->plh_lc_cred);
1532         pnfs_get_layout_hdr(lo);
1533         spin_unlock(&inode->i_lock);
1534
1535         return pnfs_send_layoutreturn(lo, &zero_stateid, &cred, IOMODE_ANY,
1536                                       PNFS_FL_LAYOUTRETURN_PRIVILEGED);
1537 }
1538
1539 bool pnfs_roc(struct inode *ino,
1540                 struct nfs4_layoutreturn_args *args,
1541                 struct nfs4_layoutreturn_res *res,
1542                 const struct cred *cred)
1543 {
1544         struct nfs_inode *nfsi = NFS_I(ino);
1545         struct nfs_open_context *ctx;
1546         struct nfs4_state *state;
1547         struct pnfs_layout_hdr *lo;
1548         struct pnfs_layout_segment *lseg, *next;
1549         const struct cred *lc_cred;
1550         nfs4_stateid stateid;
1551         enum pnfs_iomode iomode = 0;
1552         bool layoutreturn = false, roc = false;
1553         bool skip_read = false;
1554
1555         if (!nfs_have_layout(ino))
1556                 return false;
1557 retry:
1558         rcu_read_lock();
1559         spin_lock(&ino->i_lock);
1560         lo = nfsi->layout;
1561         if (!lo || !pnfs_layout_is_valid(lo) ||
1562             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1563                 lo = NULL;
1564                 goto out_noroc;
1565         }
1566         pnfs_get_layout_hdr(lo);
1567         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1568                 spin_unlock(&ino->i_lock);
1569                 rcu_read_unlock();
1570                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1571                                 TASK_UNINTERRUPTIBLE);
1572                 pnfs_put_layout_hdr(lo);
1573                 goto retry;
1574         }
1575
1576         /* no roc if we hold a delegation */
1577         if (nfs4_check_delegation(ino, FMODE_READ)) {
1578                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1579                         goto out_noroc;
1580                 skip_read = true;
1581         }
1582
1583         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1584                 state = ctx->state;
1585                 if (state == NULL)
1586                         continue;
1587                 /* Don't return layout if there is open file state */
1588                 if (state->state & FMODE_WRITE)
1589                         goto out_noroc;
1590                 if (state->state & FMODE_READ)
1591                         skip_read = true;
1592         }
1593
1594
1595         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1596                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1597                         continue;
1598                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1599                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1600                         continue;
1601                 /*
1602                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1603                  * doesn't invalidate the layout for us.
1604                  */
1605                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1606                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1607                         continue;
1608                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1609         }
1610
1611         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1612                 goto out_noroc;
1613
1614         /* ROC in two conditions:
1615          * 1. there are ROC lsegs
1616          * 2. we don't send layoutreturn
1617          */
1618         /* lo ref dropped in pnfs_roc_release() */
1619         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1620         /* If the creds don't match, we can't compound the layoutreturn */
1621         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1622                 goto out_noroc;
1623
1624         roc = layoutreturn;
1625         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1626         res->lrs_present = 0;
1627         layoutreturn = false;
1628         put_cred(lc_cred);
1629
1630 out_noroc:
1631         spin_unlock(&ino->i_lock);
1632         rcu_read_unlock();
1633         pnfs_layoutcommit_inode(ino, true);
1634         if (roc) {
1635                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1636                 if (ld->prepare_layoutreturn)
1637                         ld->prepare_layoutreturn(args);
1638                 pnfs_put_layout_hdr(lo);
1639                 return true;
1640         }
1641         if (layoutreturn)
1642                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, 0);
1643         pnfs_put_layout_hdr(lo);
1644         return false;
1645 }
1646
1647 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1648                   struct nfs4_layoutreturn_res **respp, int *ret)
1649 {
1650         struct nfs4_layoutreturn_args *arg = *argpp;
1651         int retval = -EAGAIN;
1652
1653         if (!arg)
1654                 return 0;
1655         /* Handle Layoutreturn errors */
1656         switch (*ret) {
1657         case 0:
1658                 retval = 0;
1659                 break;
1660         case -NFS4ERR_NOMATCHING_LAYOUT:
1661                 /* Was there an RPC level error? If not, retry */
1662                 if (task->tk_rpc_status == 0)
1663                         break;
1664                 /* If the call was not sent, let caller handle it */
1665                 if (!RPC_WAS_SENT(task))
1666                         return 0;
1667                 /*
1668                  * Otherwise, assume the call succeeded and
1669                  * that we need to release the layout
1670                  */
1671                 *ret = 0;
1672                 (*respp)->lrs_present = 0;
1673                 retval = 0;
1674                 break;
1675         case -NFS4ERR_DELAY:
1676                 /* Let the caller handle the retry */
1677                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1678                 return 0;
1679         case -NFS4ERR_OLD_STATEID:
1680                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1681                                                      &arg->range, arg->inode))
1682                         break;
1683                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1684                 return -EAGAIN;
1685         }
1686         *argpp = NULL;
1687         *respp = NULL;
1688         return retval;
1689 }
1690
1691 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1692                       struct nfs4_layoutreturn_res *res, int ret)
1693 {
1694         struct pnfs_layout_hdr *lo = args->layout;
1695         struct inode *inode = args->inode;
1696         const nfs4_stateid *res_stateid = NULL;
1697         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1698
1699         switch (ret) {
1700         case -NFS4ERR_BADSESSION:
1701         case -NFS4ERR_DEADSESSION:
1702         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1703         case -NFS4ERR_NOMATCHING_LAYOUT:
1704                 spin_lock(&inode->i_lock);
1705                 pnfs_layoutreturn_retry_later_locked(lo, &args->stateid,
1706                                                      &args->range);
1707                 pnfs_clear_layoutreturn_waitbit(lo);
1708                 spin_unlock(&inode->i_lock);
1709                 break;
1710         case 0:
1711                 if (res->lrs_present)
1712                         res_stateid = &res->stateid;
1713                 fallthrough;
1714         default:
1715                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1716                                              res_stateid);
1717         }
1718         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1719         if (ld_private && ld_private->ops && ld_private->ops->free)
1720                 ld_private->ops->free(ld_private);
1721         pnfs_put_layout_hdr(lo);
1722 }
1723
1724 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1725 {
1726         struct nfs_inode *nfsi = NFS_I(ino);
1727         struct pnfs_layout_hdr *lo;
1728         bool sleep = false;
1729
1730         /* we might not have grabbed lo reference. so need to check under
1731          * i_lock */
1732         spin_lock(&ino->i_lock);
1733         lo = nfsi->layout;
1734         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1735                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1736                 sleep = true;
1737         }
1738         spin_unlock(&ino->i_lock);
1739         return sleep;
1740 }
1741
1742 /*
1743  * Compare two layout segments for sorting into layout cache.
1744  * We want to preferentially return RW over RO layouts, so ensure those
1745  * are seen first.
1746  */
1747 static s64
1748 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1749            const struct pnfs_layout_range *l2)
1750 {
1751         s64 d;
1752
1753         /* high offset > low offset */
1754         d = l1->offset - l2->offset;
1755         if (d)
1756                 return d;
1757
1758         /* short length > long length */
1759         d = l2->length - l1->length;
1760         if (d)
1761                 return d;
1762
1763         /* read > read/write */
1764         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1765 }
1766
1767 static bool
1768 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1769                 const struct pnfs_layout_range *l2)
1770 {
1771         return pnfs_lseg_range_cmp(l1, l2) > 0;
1772 }
1773
1774 static bool
1775 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1776                 struct pnfs_layout_segment *old)
1777 {
1778         return false;
1779 }
1780
1781 void
1782 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1783                    struct pnfs_layout_segment *lseg,
1784                    bool (*is_after)(const struct pnfs_layout_range *,
1785                            const struct pnfs_layout_range *),
1786                    bool (*do_merge)(struct pnfs_layout_segment *,
1787                            struct pnfs_layout_segment *),
1788                    struct list_head *free_me)
1789 {
1790         struct pnfs_layout_segment *lp, *tmp;
1791
1792         dprintk("%s:Begin\n", __func__);
1793
1794         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1795                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1796                         continue;
1797                 if (do_merge(lseg, lp)) {
1798                         mark_lseg_invalid(lp, free_me);
1799                         continue;
1800                 }
1801                 if (is_after(&lseg->pls_range, &lp->pls_range))
1802                         continue;
1803                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1804                 dprintk("%s: inserted lseg %p "
1805                         "iomode %d offset %llu length %llu before "
1806                         "lp %p iomode %d offset %llu length %llu\n",
1807                         __func__, lseg, lseg->pls_range.iomode,
1808                         lseg->pls_range.offset, lseg->pls_range.length,
1809                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1810                         lp->pls_range.length);
1811                 goto out;
1812         }
1813         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1814         dprintk("%s: inserted lseg %p "
1815                 "iomode %d offset %llu length %llu at tail\n",
1816                 __func__, lseg, lseg->pls_range.iomode,
1817                 lseg->pls_range.offset, lseg->pls_range.length);
1818 out:
1819         pnfs_get_layout_hdr(lo);
1820
1821         dprintk("%s:Return\n", __func__);
1822 }
1823 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1824
1825 static void
1826 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1827                    struct pnfs_layout_segment *lseg,
1828                    struct list_head *free_me)
1829 {
1830         struct inode *inode = lo->plh_inode;
1831         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1832
1833         if (ld->add_lseg != NULL)
1834                 ld->add_lseg(lo, lseg, free_me);
1835         else
1836                 pnfs_generic_layout_insert_lseg(lo, lseg,
1837                                 pnfs_lseg_range_is_after,
1838                                 pnfs_lseg_no_merge,
1839                                 free_me);
1840 }
1841
1842 static struct pnfs_layout_hdr *
1843 alloc_init_layout_hdr(struct inode *ino,
1844                       struct nfs_open_context *ctx,
1845                       gfp_t gfp_flags)
1846 {
1847         struct pnfs_layout_hdr *lo;
1848
1849         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1850         if (!lo)
1851                 return NULL;
1852         refcount_set(&lo->plh_refcount, 1);
1853         INIT_LIST_HEAD(&lo->plh_layouts);
1854         INIT_LIST_HEAD(&lo->plh_segs);
1855         INIT_LIST_HEAD(&lo->plh_return_segs);
1856         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1857         lo->plh_inode = ino;
1858         lo->plh_lc_cred = get_cred(ctx->cred);
1859         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1860         return lo;
1861 }
1862
1863 static struct pnfs_layout_hdr *
1864 pnfs_find_alloc_layout(struct inode *ino,
1865                        struct nfs_open_context *ctx,
1866                        gfp_t gfp_flags)
1867         __releases(&ino->i_lock)
1868         __acquires(&ino->i_lock)
1869 {
1870         struct nfs_inode *nfsi = NFS_I(ino);
1871         struct pnfs_layout_hdr *new = NULL;
1872
1873         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1874
1875         if (nfsi->layout != NULL)
1876                 goto out_existing;
1877         spin_unlock(&ino->i_lock);
1878         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1879         spin_lock(&ino->i_lock);
1880
1881         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1882                 nfsi->layout = new;
1883                 return new;
1884         } else if (new != NULL)
1885                 pnfs_free_layout_hdr(new);
1886 out_existing:
1887         pnfs_get_layout_hdr(nfsi->layout);
1888         return nfsi->layout;
1889 }
1890
1891 /*
1892  * iomode matching rules:
1893  * iomode       lseg    strict match
1894  *                      iomode
1895  * -----        -----   ------ -----
1896  * ANY          READ    N/A    true
1897  * ANY          RW      N/A    true
1898  * RW           READ    N/A    false
1899  * RW           RW      N/A    true
1900  * READ         READ    N/A    true
1901  * READ         RW      true   false
1902  * READ         RW      false  true
1903  */
1904 static bool
1905 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1906                  const struct pnfs_layout_range *range,
1907                  bool strict_iomode)
1908 {
1909         struct pnfs_layout_range range1;
1910
1911         if ((range->iomode == IOMODE_RW &&
1912              ls_range->iomode != IOMODE_RW) ||
1913             (range->iomode != ls_range->iomode &&
1914              strict_iomode) ||
1915             !pnfs_lseg_range_intersecting(ls_range, range))
1916                 return false;
1917
1918         /* range1 covers only the first byte in the range */
1919         range1 = *range;
1920         range1.length = 1;
1921         return pnfs_lseg_range_contained(ls_range, &range1);
1922 }
1923
1924 /*
1925  * lookup range in layout
1926  */
1927 static struct pnfs_layout_segment *
1928 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1929                 struct pnfs_layout_range *range,
1930                 bool strict_iomode)
1931 {
1932         struct pnfs_layout_segment *lseg, *ret = NULL;
1933
1934         dprintk("%s:Begin\n", __func__);
1935
1936         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1937                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1938                     pnfs_lseg_range_match(&lseg->pls_range, range,
1939                                           strict_iomode)) {
1940                         ret = pnfs_get_lseg(lseg);
1941                         break;
1942                 }
1943         }
1944
1945         dprintk("%s:Return lseg %p ref %d\n",
1946                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1947         return ret;
1948 }
1949
1950 /*
1951  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1952  * to the MDS or over pNFS
1953  *
1954  * The nfs_inode read_io and write_io fields are cumulative counters reset
1955  * when there are no layout segments. Note that in pnfs_update_layout iomode
1956  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1957  * WRITE request.
1958  *
1959  * A return of true means use MDS I/O.
1960  *
1961  * From rfc 5661:
1962  * If a file's size is smaller than the file size threshold, data accesses
1963  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1964  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1965  * server.  If both file size and I/O size are provided, the client SHOULD
1966  * reach or exceed  both thresholds before sending its read or write
1967  * requests to the data server.
1968  */
1969 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1970                                      struct inode *ino, int iomode)
1971 {
1972         struct nfs4_threshold *t = ctx->mdsthreshold;
1973         struct nfs_inode *nfsi = NFS_I(ino);
1974         loff_t fsize = i_size_read(ino);
1975         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1976
1977         if (t == NULL)
1978                 return ret;
1979
1980         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1981                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1982
1983         switch (iomode) {
1984         case IOMODE_READ:
1985                 if (t->bm & THRESHOLD_RD) {
1986                         dprintk("%s fsize %llu\n", __func__, fsize);
1987                         size_set = true;
1988                         if (fsize < t->rd_sz)
1989                                 size = true;
1990                 }
1991                 if (t->bm & THRESHOLD_RD_IO) {
1992                         dprintk("%s nfsi->read_io %llu\n", __func__,
1993                                 nfsi->read_io);
1994                         io_set = true;
1995                         if (nfsi->read_io < t->rd_io_sz)
1996                                 io = true;
1997                 }
1998                 break;
1999         case IOMODE_RW:
2000                 if (t->bm & THRESHOLD_WR) {
2001                         dprintk("%s fsize %llu\n", __func__, fsize);
2002                         size_set = true;
2003                         if (fsize < t->wr_sz)
2004                                 size = true;
2005                 }
2006                 if (t->bm & THRESHOLD_WR_IO) {
2007                         dprintk("%s nfsi->write_io %llu\n", __func__,
2008                                 nfsi->write_io);
2009                         io_set = true;
2010                         if (nfsi->write_io < t->wr_io_sz)
2011                                 io = true;
2012                 }
2013                 break;
2014         }
2015         if (size_set && io_set) {
2016                 if (size && io)
2017                         ret = true;
2018         } else if (size || io)
2019                 ret = true;
2020
2021         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
2022         return ret;
2023 }
2024
2025 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
2026 {
2027         /*
2028          * send layoutcommit as it can hold up layoutreturn due to lseg
2029          * reference
2030          */
2031         pnfs_layoutcommit_inode(lo->plh_inode, false);
2032         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
2033                                    nfs_wait_bit_killable,
2034                                    TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
2035 }
2036
2037 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
2038 {
2039         atomic_inc(&lo->plh_outstanding);
2040 }
2041
2042 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
2043 {
2044         if (atomic_dec_and_test(&lo->plh_outstanding) &&
2045             test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
2046                 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
2047 }
2048
2049 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
2050 {
2051         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
2052 }
2053
2054 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
2055 {
2056         unsigned long *bitlock = &lo->plh_flags;
2057
2058         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
2059         smp_mb__after_atomic();
2060         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
2061 }
2062
2063 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
2064                                 struct nfs_server *server)
2065 {
2066         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
2067                 struct nfs_client *clp = server->nfs_client;
2068
2069                 /* The lo must be on the clp list if there is any
2070                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
2071                  */
2072                 spin_lock(&clp->cl_lock);
2073                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
2074                 spin_unlock(&clp->cl_lock);
2075         }
2076 }
2077
2078 /*
2079  * Layout segment is retreived from the server if not cached.
2080  * The appropriate layout segment is referenced and returned to the caller.
2081  */
2082 struct pnfs_layout_segment *
2083 pnfs_update_layout(struct inode *ino,
2084                    struct nfs_open_context *ctx,
2085                    loff_t pos,
2086                    u64 count,
2087                    enum pnfs_iomode iomode,
2088                    bool strict_iomode,
2089                    gfp_t gfp_flags)
2090 {
2091         struct pnfs_layout_range arg = {
2092                 .iomode = iomode,
2093                 .offset = pos,
2094                 .length = count,
2095         };
2096         unsigned pg_offset;
2097         struct nfs_server *server = NFS_SERVER(ino);
2098         struct nfs_client *clp = server->nfs_client;
2099         struct pnfs_layout_hdr *lo = NULL;
2100         struct pnfs_layout_segment *lseg = NULL;
2101         struct nfs4_layoutget *lgp;
2102         nfs4_stateid stateid;
2103         struct nfs4_exception exception = {
2104                 .inode = ino,
2105         };
2106         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
2107         bool first;
2108
2109         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
2110                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2111                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
2112                 goto out;
2113         }
2114
2115         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
2116                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2117                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
2118                 goto out;
2119         }
2120
2121 lookup_again:
2122         if (!nfs4_valid_open_stateid(ctx->state)) {
2123                 trace_pnfs_update_layout(ino, pos, count,
2124                                          iomode, lo, lseg,
2125                                          PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2126                 lseg = ERR_PTR(-EIO);
2127                 goto out;
2128         }
2129
2130         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2131         if (IS_ERR(lseg))
2132                 goto out;
2133         first = false;
2134         spin_lock(&ino->i_lock);
2135         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2136         if (lo == NULL) {
2137                 spin_unlock(&ino->i_lock);
2138                 lseg = ERR_PTR(-ENOMEM);
2139                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2140                                  PNFS_UPDATE_LAYOUT_NOMEM);
2141                 goto out;
2142         }
2143
2144         /* Do we even need to bother with this? */
2145         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2146                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2147                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2148                 dprintk("%s matches recall, use MDS\n", __func__);
2149                 goto out_unlock;
2150         }
2151
2152         /* if LAYOUTGET already failed once we don't try again */
2153         if (pnfs_layout_io_test_failed(lo, iomode)) {
2154                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2155                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2156                 goto out_unlock;
2157         }
2158
2159         /*
2160          * If the layout segment list is empty, but there are outstanding
2161          * layoutget calls, then they might be subject to a layoutrecall.
2162          */
2163         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2164             atomic_read(&lo->plh_outstanding) != 0) {
2165                 spin_unlock(&ino->i_lock);
2166                 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2167                                            TASK_KILLABLE));
2168                 if (IS_ERR(lseg))
2169                         goto out_put_layout_hdr;
2170                 pnfs_put_layout_hdr(lo);
2171                 goto lookup_again;
2172         }
2173
2174         /*
2175          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2176          * for LAYOUTRETURN.
2177          */
2178         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2179                 spin_unlock(&ino->i_lock);
2180                 dprintk("%s wait for layoutreturn\n", __func__);
2181                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2182                 if (!IS_ERR(lseg)) {
2183                         pnfs_put_layout_hdr(lo);
2184                         dprintk("%s retrying\n", __func__);
2185                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2186                                                  lseg,
2187                                                  PNFS_UPDATE_LAYOUT_RETRY);
2188                         goto lookup_again;
2189                 }
2190                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2191                                          PNFS_UPDATE_LAYOUT_RETURN);
2192                 goto out_put_layout_hdr;
2193         }
2194
2195         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2196         if (lseg) {
2197                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2198                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2199                 goto out_unlock;
2200         }
2201
2202         /*
2203          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2204          * stateid, or it has been invalidated, then we must use the open
2205          * stateid.
2206          */
2207         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2208                 int status;
2209
2210                 /*
2211                  * The first layoutget for the file. Need to serialize per
2212                  * RFC 5661 Errata 3208.
2213                  */
2214                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2215                                      &lo->plh_flags)) {
2216                         spin_unlock(&ino->i_lock);
2217                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2218                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2219                                                 TASK_KILLABLE));
2220                         if (IS_ERR(lseg))
2221                                 goto out_put_layout_hdr;
2222                         pnfs_put_layout_hdr(lo);
2223                         dprintk("%s retrying\n", __func__);
2224                         goto lookup_again;
2225                 }
2226
2227                 spin_unlock(&ino->i_lock);
2228                 first = true;
2229                 status = nfs4_select_rw_stateid(ctx->state,
2230                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2231                                         NULL, &stateid, NULL);
2232                 if (status != 0) {
2233                         lseg = ERR_PTR(status);
2234                         trace_pnfs_update_layout(ino, pos, count,
2235                                         iomode, lo, lseg,
2236                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2237                         nfs4_schedule_stateid_recovery(server, ctx->state);
2238                         pnfs_clear_first_layoutget(lo);
2239                         pnfs_put_layout_hdr(lo);
2240                         goto lookup_again;
2241                 }
2242                 spin_lock(&ino->i_lock);
2243         } else {
2244                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2245         }
2246
2247         if (pnfs_layoutgets_blocked(lo)) {
2248                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2249                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2250                 goto out_unlock;
2251         }
2252         nfs_layoutget_begin(lo);
2253         spin_unlock(&ino->i_lock);
2254
2255         _add_to_server_list(lo, server);
2256
2257         pg_offset = arg.offset & ~PAGE_MASK;
2258         if (pg_offset) {
2259                 arg.offset -= pg_offset;
2260                 arg.length += pg_offset;
2261         }
2262         if (arg.length != NFS4_MAX_UINT64)
2263                 arg.length = PAGE_ALIGN(arg.length);
2264
2265         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2266         if (!lgp) {
2267                 lseg = ERR_PTR(-ENOMEM);
2268                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2269                                          PNFS_UPDATE_LAYOUT_NOMEM);
2270                 nfs_layoutget_end(lo);
2271                 goto out_put_layout_hdr;
2272         }
2273
2274         lgp->lo = lo;
2275         pnfs_get_layout_hdr(lo);
2276
2277         lseg = nfs4_proc_layoutget(lgp, &exception);
2278         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2279                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2280         nfs_layoutget_end(lo);
2281         if (IS_ERR(lseg)) {
2282                 switch(PTR_ERR(lseg)) {
2283                 case -EBUSY:
2284                         if (time_after(jiffies, giveup))
2285                                 lseg = NULL;
2286                         break;
2287                 case -ERECALLCONFLICT:
2288                 case -EAGAIN:
2289                         break;
2290                 case -ENODATA:
2291                         /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2292                         pnfs_layout_set_fail_bit(
2293                                 lo, pnfs_iomode_to_fail_bit(iomode));
2294                         lseg = NULL;
2295                         goto out_put_layout_hdr;
2296                 default:
2297                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2298                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2299                                 lseg = NULL;
2300                         }
2301                         goto out_put_layout_hdr;
2302                 }
2303                 if (lseg) {
2304                         if (!exception.retry)
2305                                 goto out_put_layout_hdr;
2306                         if (first)
2307                                 pnfs_clear_first_layoutget(lo);
2308                         trace_pnfs_update_layout(ino, pos, count,
2309                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2310                         pnfs_put_layout_hdr(lo);
2311                         goto lookup_again;
2312                 }
2313         } else {
2314                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2315         }
2316
2317 out_put_layout_hdr:
2318         if (first)
2319                 pnfs_clear_first_layoutget(lo);
2320         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2321                                  PNFS_UPDATE_LAYOUT_EXIT);
2322         pnfs_put_layout_hdr(lo);
2323 out:
2324         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2325                         "(%s, offset: %llu, length: %llu)\n",
2326                         __func__, ino->i_sb->s_id,
2327                         (unsigned long long)NFS_FILEID(ino),
2328                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2329                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2330                         (unsigned long long)pos,
2331                         (unsigned long long)count);
2332         return lseg;
2333 out_unlock:
2334         spin_unlock(&ino->i_lock);
2335         goto out_put_layout_hdr;
2336 }
2337 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2338
2339 static bool
2340 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2341 {
2342         switch (range->iomode) {
2343         case IOMODE_READ:
2344         case IOMODE_RW:
2345                 break;
2346         default:
2347                 return false;
2348         }
2349         if (range->offset == NFS4_MAX_UINT64)
2350                 return false;
2351         if (range->length == 0)
2352                 return false;
2353         if (range->length != NFS4_MAX_UINT64 &&
2354             range->length > NFS4_MAX_UINT64 - range->offset)
2355                 return false;
2356         return true;
2357 }
2358
2359 static struct pnfs_layout_hdr *
2360 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2361 {
2362         struct pnfs_layout_hdr *lo;
2363
2364         spin_lock(&ino->i_lock);
2365         lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2366         if (!lo)
2367                 goto out_unlock;
2368         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2369                 goto out_unlock;
2370         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2371                 goto out_unlock;
2372         if (pnfs_layoutgets_blocked(lo))
2373                 goto out_unlock;
2374         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2375                 goto out_unlock;
2376         nfs_layoutget_begin(lo);
2377         spin_unlock(&ino->i_lock);
2378         _add_to_server_list(lo, NFS_SERVER(ino));
2379         return lo;
2380
2381 out_unlock:
2382         spin_unlock(&ino->i_lock);
2383         pnfs_put_layout_hdr(lo);
2384         return NULL;
2385 }
2386
2387 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2388                                      struct nfs_open_context *ctx)
2389 {
2390         struct inode *ino = data->dentry->d_inode;
2391         struct pnfs_layout_range rng = {
2392                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2393                           IOMODE_RW: IOMODE_READ,
2394                 .offset = 0,
2395                 .length = NFS4_MAX_UINT64,
2396         };
2397         struct nfs4_layoutget *lgp;
2398         struct pnfs_layout_hdr *lo;
2399
2400         /* Heuristic: don't send layoutget if we have cached data */
2401         if (rng.iomode == IOMODE_READ &&
2402            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2403                 return;
2404
2405         lo = _pnfs_grab_empty_layout(ino, ctx);
2406         if (!lo)
2407                 return;
2408         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2409                                              nfs_io_gfp_mask());
2410         if (!lgp) {
2411                 pnfs_clear_first_layoutget(lo);
2412                 nfs_layoutget_end(lo);
2413                 pnfs_put_layout_hdr(lo);
2414                 return;
2415         }
2416         lgp->lo = lo;
2417         data->lgp = lgp;
2418         data->o_arg.lg_args = &lgp->args;
2419         data->o_res.lg_res = &lgp->res;
2420 }
2421
2422 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2423                                      struct nfs_open_context *ctx)
2424 {
2425         struct inode *ino = data->dentry->d_inode;
2426         struct pnfs_layout_range rng = {
2427                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2428                           IOMODE_RW: IOMODE_READ,
2429                 .offset = 0,
2430                 .length = NFS4_MAX_UINT64,
2431         };
2432         struct nfs4_layoutget *lgp;
2433
2434         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2435                                              nfs_io_gfp_mask());
2436         if (!lgp)
2437                 return;
2438         data->lgp = lgp;
2439         data->o_arg.lg_args = &lgp->args;
2440         data->o_res.lg_res = &lgp->res;
2441 }
2442
2443 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2444                          struct nfs_open_context *ctx)
2445 {
2446         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2447
2448         if (!(pnfs_enabled_sb(server) &&
2449               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2450                 return;
2451         /* Could check on max_ops, but currently hardcoded high enough */
2452         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2453                 return;
2454         if (data->lgp)
2455                 return;
2456         if (data->state)
2457                 _lgopen_prepare_attached(data, ctx);
2458         else
2459                 _lgopen_prepare_floating(data, ctx);
2460 }
2461
2462 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2463                        struct nfs_open_context *ctx)
2464 {
2465         struct pnfs_layout_hdr *lo;
2466         struct pnfs_layout_segment *lseg;
2467         struct nfs_server *srv = NFS_SERVER(ino);
2468         u32 iomode;
2469
2470         if (!lgp)
2471                 return;
2472         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2473         if (lgp->res.status) {
2474                 switch (lgp->res.status) {
2475                 default:
2476                         break;
2477                 /*
2478                  * Halt lgopen attempts if the server doesn't recognise
2479                  * the "current stateid" value, the layout type, or the
2480                  * layoutget operation as being valid.
2481                  * Also if it complains about too many ops in the compound
2482                  * or of the request/reply being too big.
2483                  */
2484                 case -NFS4ERR_BAD_STATEID:
2485                 case -NFS4ERR_NOTSUPP:
2486                 case -NFS4ERR_REP_TOO_BIG:
2487                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2488                 case -NFS4ERR_REQ_TOO_BIG:
2489                 case -NFS4ERR_TOO_MANY_OPS:
2490                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2491                         srv->caps &= ~NFS_CAP_LGOPEN;
2492                 }
2493                 return;
2494         }
2495         if (!lgp->lo) {
2496                 lo = _pnfs_grab_empty_layout(ino, ctx);
2497                 if (!lo)
2498                         return;
2499                 lgp->lo = lo;
2500         } else
2501                 lo = lgp->lo;
2502
2503         lseg = pnfs_layout_process(lgp);
2504         if (!IS_ERR(lseg)) {
2505                 iomode = lgp->args.range.iomode;
2506                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2507                 pnfs_put_lseg(lseg);
2508         }
2509 }
2510
2511 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2512 {
2513         if (lgp != NULL) {
2514                 if (lgp->lo) {
2515                         pnfs_clear_first_layoutget(lgp->lo);
2516                         nfs_layoutget_end(lgp->lo);
2517                 }
2518                 pnfs_layoutget_free(lgp);
2519         }
2520 }
2521
2522 struct pnfs_layout_segment *
2523 pnfs_layout_process(struct nfs4_layoutget *lgp)
2524 {
2525         struct pnfs_layout_hdr *lo = lgp->lo;
2526         struct nfs4_layoutget_res *res = &lgp->res;
2527         struct pnfs_layout_segment *lseg;
2528         struct inode *ino = lo->plh_inode;
2529         LIST_HEAD(free_me);
2530
2531         if (!pnfs_sanity_check_layout_range(&res->range))
2532                 return ERR_PTR(-EINVAL);
2533
2534         /* Inject layout blob into I/O device driver */
2535         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2536         if (IS_ERR_OR_NULL(lseg)) {
2537                 if (!lseg)
2538                         lseg = ERR_PTR(-ENOMEM);
2539
2540                 dprintk("%s: Could not allocate layout: error %ld\n",
2541                        __func__, PTR_ERR(lseg));
2542                 return lseg;
2543         }
2544
2545         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2546
2547         spin_lock(&ino->i_lock);
2548         if (pnfs_layoutgets_blocked(lo)) {
2549                 dprintk("%s forget reply due to state\n", __func__);
2550                 goto out_forget;
2551         }
2552
2553         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2554             !pnfs_is_first_layoutget(lo))
2555                 goto out_forget;
2556
2557         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2558                 /* existing state ID, make sure the sequence number matches. */
2559                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2560                         if (!pnfs_layout_is_valid(lo))
2561                                 lo->plh_barrier = 0;
2562                         dprintk("%s forget reply due to sequence\n", __func__);
2563                         goto out_forget;
2564                 }
2565                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2566         } else if (pnfs_layout_is_valid(lo)) {
2567                 /*
2568                  * We got an entirely new state ID.  Mark all segments for the
2569                  * inode invalid, and retry the layoutget
2570                  */
2571                 struct pnfs_layout_range range = {
2572                         .iomode = IOMODE_ANY,
2573                         .length = NFS4_MAX_UINT64,
2574                 };
2575                 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2576                 goto out_forget;
2577         } else {
2578                 /* We have a completely new layout */
2579                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2580         }
2581
2582         pnfs_get_lseg(lseg);
2583         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2584
2585
2586         if (res->return_on_close)
2587                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2588
2589         spin_unlock(&ino->i_lock);
2590         pnfs_free_lseg_list(&free_me);
2591         return lseg;
2592
2593 out_forget:
2594         spin_unlock(&ino->i_lock);
2595         lseg->pls_layout = lo;
2596         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2597         return ERR_PTR(-EAGAIN);
2598 }
2599
2600 /**
2601  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2602  * @lo: pointer to layout header
2603  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2604  * @return_range: describe layout segment ranges to be returned
2605  * @seq: stateid seqid to match
2606  *
2607  * This function is mainly intended for use by layoutrecall. It attempts
2608  * to free the layout segment immediately, or else to mark it for return
2609  * as soon as its reference count drops to zero.
2610  *
2611  * Returns
2612  * - 0: a layoutreturn needs to be scheduled.
2613  * - EBUSY: there are layout segment that are still in use.
2614  * - ENOENT: there are no layout segments that need to be returned.
2615  */
2616 int
2617 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2618                                 struct list_head *tmp_list,
2619                                 const struct pnfs_layout_range *return_range,
2620                                 u32 seq)
2621 {
2622         struct pnfs_layout_segment *lseg, *next;
2623         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2624         int remaining = 0;
2625
2626         dprintk("%s:Begin lo %p\n", __func__, lo);
2627
2628         assert_spin_locked(&lo->plh_inode->i_lock);
2629
2630         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2631                 tmp_list = &lo->plh_return_segs;
2632
2633         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2634                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2635                         dprintk("%s: marking lseg %p iomode %d "
2636                                 "offset %llu length %llu\n", __func__,
2637                                 lseg, lseg->pls_range.iomode,
2638                                 lseg->pls_range.offset,
2639                                 lseg->pls_range.length);
2640                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2641                                 tmp_list = &lo->plh_return_segs;
2642                         if (mark_lseg_invalid(lseg, tmp_list))
2643                                 continue;
2644                         remaining++;
2645                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2646                         pnfs_lseg_cancel_io(server, lseg);
2647                 }
2648
2649         if (remaining) {
2650                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2651                 return -EBUSY;
2652         }
2653
2654         if (!list_empty(&lo->plh_return_segs)) {
2655                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2656                 return 0;
2657         }
2658
2659         return -ENOENT;
2660 }
2661
2662 static void
2663 pnfs_mark_layout_for_return(struct inode *inode,
2664                             const struct pnfs_layout_range *range)
2665 {
2666         struct pnfs_layout_hdr *lo;
2667         bool return_now = false;
2668
2669         spin_lock(&inode->i_lock);
2670         lo = NFS_I(inode)->layout;
2671         if (!pnfs_layout_is_valid(lo)) {
2672                 spin_unlock(&inode->i_lock);
2673                 return;
2674         }
2675         pnfs_set_plh_return_info(lo, range->iomode, 0);
2676         /*
2677          * mark all matching lsegs so that we are sure to have no live
2678          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2679          * for how it works.
2680          */
2681         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2682                 const struct cred *cred;
2683                 nfs4_stateid stateid;
2684                 enum pnfs_iomode iomode;
2685
2686                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2687                 spin_unlock(&inode->i_lock);
2688                 if (return_now)
2689                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
2690                                                PNFS_FL_LAYOUTRETURN_ASYNC);
2691         } else {
2692                 spin_unlock(&inode->i_lock);
2693                 nfs_commit_inode(inode, 0);
2694         }
2695 }
2696
2697 void pnfs_error_mark_layout_for_return(struct inode *inode,
2698                                        struct pnfs_layout_segment *lseg)
2699 {
2700         struct pnfs_layout_range range = {
2701                 .iomode = lseg->pls_range.iomode,
2702                 .offset = 0,
2703                 .length = NFS4_MAX_UINT64,
2704         };
2705
2706         pnfs_mark_layout_for_return(inode, &range);
2707 }
2708 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2709
2710 static bool
2711 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2712 {
2713         return pnfs_layout_is_valid(lo) &&
2714                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2715                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2716 }
2717
2718 static struct pnfs_layout_segment *
2719 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2720                      const struct pnfs_layout_range *range,
2721                      enum pnfs_iomode iomode)
2722 {
2723         struct pnfs_layout_segment *lseg;
2724
2725         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2726                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2727                         continue;
2728                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2729                         continue;
2730                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2731                         continue;
2732                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2733                         return lseg;
2734         }
2735         return NULL;
2736 }
2737
2738 /* Find open file states whose mode matches that of the range */
2739 static bool
2740 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2741                                  const struct pnfs_layout_range *range)
2742 {
2743         struct list_head *head;
2744         struct nfs_open_context *ctx;
2745         fmode_t mode = 0;
2746
2747         if (!pnfs_layout_can_be_returned(lo) ||
2748             !pnfs_find_first_lseg(lo, range, range->iomode))
2749                 return false;
2750
2751         head = &NFS_I(lo->plh_inode)->open_files;
2752         list_for_each_entry_rcu(ctx, head, list) {
2753                 if (ctx->state)
2754                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2755         }
2756
2757         switch (range->iomode) {
2758         default:
2759                 break;
2760         case IOMODE_READ:
2761                 mode &= ~FMODE_WRITE;
2762                 break;
2763         case IOMODE_RW:
2764                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2765                         mode &= ~FMODE_READ;
2766         }
2767         return mode == 0;
2768 }
2769
2770 static int pnfs_layout_return_unused_byserver(struct nfs_server *server,
2771                                               void *data)
2772 {
2773         const struct pnfs_layout_range *range = data;
2774         const struct cred *cred;
2775         struct pnfs_layout_hdr *lo;
2776         struct inode *inode;
2777         nfs4_stateid stateid;
2778         enum pnfs_iomode iomode;
2779
2780 restart:
2781         rcu_read_lock();
2782         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2783                 inode = lo->plh_inode;
2784                 if (!inode || !pnfs_layout_can_be_returned(lo) ||
2785                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2786                         continue;
2787                 spin_lock(&inode->i_lock);
2788                 if (!lo->plh_inode ||
2789                     !pnfs_should_return_unused_layout(lo, range)) {
2790                         spin_unlock(&inode->i_lock);
2791                         continue;
2792                 }
2793                 pnfs_get_layout_hdr(lo);
2794                 pnfs_set_plh_return_info(lo, range->iomode, 0);
2795                 if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2796                                                     range, 0) != 0 ||
2797                     !pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode)) {
2798                         spin_unlock(&inode->i_lock);
2799                         rcu_read_unlock();
2800                         pnfs_put_layout_hdr(lo);
2801                         cond_resched();
2802                         goto restart;
2803                 }
2804                 spin_unlock(&inode->i_lock);
2805                 rcu_read_unlock();
2806                 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
2807                                        PNFS_FL_LAYOUTRETURN_ASYNC);
2808                 pnfs_put_layout_hdr(lo);
2809                 cond_resched();
2810                 goto restart;
2811         }
2812         rcu_read_unlock();
2813         return 0;
2814 }
2815
2816 void
2817 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2818                                  enum pnfs_iomode iomode)
2819 {
2820         struct pnfs_layout_range range = {
2821                 .iomode = iomode,
2822                 .offset = 0,
2823                 .length = NFS4_MAX_UINT64,
2824         };
2825
2826         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2827                         &range);
2828 }
2829
2830 /* Check if we have we have a valid layout but if there isn't an intersection
2831  * between the request and the pgio->pg_lseg, put this pgio->pg_lseg away.
2832  */
2833 void
2834 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio,
2835                              struct nfs_page *req)
2836 {
2837         if (pgio->pg_lseg == NULL ||
2838             (test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags) &&
2839             pnfs_lseg_request_intersecting(pgio->pg_lseg, req)))
2840                 return;
2841         pnfs_put_lseg(pgio->pg_lseg);
2842         pgio->pg_lseg = NULL;
2843 }
2844 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2845
2846 void
2847 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2848 {
2849         u64 rd_size;
2850
2851         pnfs_generic_pg_check_layout(pgio, req);
2852         if (pgio->pg_lseg == NULL) {
2853                 if (pgio->pg_dreq == NULL)
2854                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2855                 else
2856                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq,
2857                                                       req_offset(req));
2858
2859                 pgio->pg_lseg =
2860                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2861                                            req_offset(req), rd_size,
2862                                            IOMODE_READ, false,
2863                                            nfs_io_gfp_mask());
2864                 if (IS_ERR(pgio->pg_lseg)) {
2865                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2866                         pgio->pg_lseg = NULL;
2867                         return;
2868                 }
2869         }
2870         /* If no lseg, fall back to read through mds */
2871         if (pgio->pg_lseg == NULL)
2872                 nfs_pageio_reset_read_mds(pgio);
2873
2874 }
2875 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2876
2877 void
2878 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2879                            struct nfs_page *req, u64 wb_size)
2880 {
2881         pnfs_generic_pg_check_layout(pgio, req);
2882         if (pgio->pg_lseg == NULL) {
2883                 pgio->pg_lseg =
2884                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2885                                            req_offset(req), wb_size, IOMODE_RW,
2886                                            false, nfs_io_gfp_mask());
2887                 if (IS_ERR(pgio->pg_lseg)) {
2888                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2889                         pgio->pg_lseg = NULL;
2890                         return;
2891                 }
2892         }
2893         /* If no lseg, fall back to write through mds */
2894         if (pgio->pg_lseg == NULL)
2895                 nfs_pageio_reset_write_mds(pgio);
2896 }
2897 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2898
2899 void
2900 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2901 {
2902         if (desc->pg_lseg) {
2903                 pnfs_put_lseg(desc->pg_lseg);
2904                 desc->pg_lseg = NULL;
2905         }
2906 }
2907 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2908
2909 /*
2910  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2911  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2912  */
2913 size_t
2914 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2915                      struct nfs_page *prev, struct nfs_page *req)
2916 {
2917         unsigned int size;
2918         u64 seg_end, req_start, seg_left;
2919
2920         size = nfs_generic_pg_test(pgio, prev, req);
2921         if (!size)
2922                 return 0;
2923
2924         /*
2925          * 'size' contains the number of bytes left in the current page (up
2926          * to the original size asked for in @req->wb_bytes).
2927          *
2928          * Calculate how many bytes are left in the layout segment
2929          * and if there are less bytes than 'size', return that instead.
2930          *
2931          * Please also note that 'end_offset' is actually the offset of the
2932          * first byte that lies outside the pnfs_layout_range. FIXME?
2933          *
2934          */
2935         if (pgio->pg_lseg) {
2936                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2937                                      pgio->pg_lseg->pls_range.length);
2938                 req_start = req_offset(req);
2939
2940                 /* start of request is past the last byte of this segment */
2941                 if (req_start >= seg_end)
2942                         return 0;
2943
2944                 /* adjust 'size' iff there are fewer bytes left in the
2945                  * segment than what nfs_generic_pg_test returned */
2946                 seg_left = seg_end - req_start;
2947                 if (seg_left < size)
2948                         size = (unsigned int)seg_left;
2949         }
2950
2951         return size;
2952 }
2953 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2954
2955 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2956 {
2957         struct nfs_pageio_descriptor pgio;
2958
2959         /* Resend all requests through the MDS */
2960         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2961                               hdr->completion_ops);
2962         return nfs_pageio_resend(&pgio, hdr);
2963 }
2964 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2965
2966 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2967 {
2968
2969         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2970         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2971             PNFS_LAYOUTRET_ON_ERROR) {
2972                 pnfs_return_layout(hdr->inode);
2973         }
2974         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2975                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2976 }
2977
2978 /*
2979  * Called by non rpc-based layout drivers
2980  */
2981 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2982 {
2983         if (likely(!hdr->pnfs_error)) {
2984                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2985                                 hdr->mds_offset + hdr->res.count);
2986                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2987         }
2988         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2989         if (unlikely(hdr->pnfs_error))
2990                 pnfs_ld_handle_write_error(hdr);
2991         hdr->mds_ops->rpc_release(hdr);
2992 }
2993 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2994
2995 static void
2996 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2997                 struct nfs_pgio_header *hdr)
2998 {
2999         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3000
3001         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3002                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3003                 nfs_pageio_reset_write_mds(desc);
3004                 mirror->pg_recoalesce = 1;
3005         }
3006         hdr->completion_ops->completion(hdr);
3007 }
3008
3009 static enum pnfs_try_status
3010 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
3011                         const struct rpc_call_ops *call_ops,
3012                         struct pnfs_layout_segment *lseg,
3013                         int how)
3014 {
3015         struct inode *inode = hdr->inode;
3016         enum pnfs_try_status trypnfs;
3017         struct nfs_server *nfss = NFS_SERVER(inode);
3018
3019         hdr->mds_ops = call_ops;
3020
3021         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
3022                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
3023         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
3024         if (trypnfs != PNFS_NOT_ATTEMPTED)
3025                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
3026         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3027         return trypnfs;
3028 }
3029
3030 static void
3031 pnfs_do_write(struct nfs_pageio_descriptor *desc,
3032               struct nfs_pgio_header *hdr, int how)
3033 {
3034         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3035         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3036         enum pnfs_try_status trypnfs;
3037
3038         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
3039         switch (trypnfs) {
3040         case PNFS_NOT_ATTEMPTED:
3041                 pnfs_write_through_mds(desc, hdr);
3042                 break;
3043         case PNFS_ATTEMPTED:
3044                 break;
3045         case PNFS_TRY_AGAIN:
3046                 /* cleanup hdr and prepare to redo pnfs */
3047                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3048                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3049                         list_splice_init(&hdr->pages, &mirror->pg_list);
3050                         mirror->pg_recoalesce = 1;
3051                 }
3052                 hdr->mds_ops->rpc_release(hdr);
3053         }
3054 }
3055
3056 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
3057 {
3058         pnfs_put_lseg(hdr->lseg);
3059         nfs_pgio_header_free(hdr);
3060 }
3061
3062 int
3063 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
3064 {
3065         struct nfs_pgio_header *hdr;
3066         int ret;
3067
3068         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3069         if (!hdr) {
3070                 desc->pg_error = -ENOMEM;
3071                 return desc->pg_error;
3072         }
3073         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
3074
3075         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3076         ret = nfs_generic_pgio(desc, hdr);
3077         if (!ret)
3078                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
3079
3080         return ret;
3081 }
3082 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
3083
3084 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
3085 {
3086         struct nfs_pageio_descriptor pgio;
3087
3088         /* Resend all requests through the MDS */
3089         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
3090         return nfs_pageio_resend(&pgio, hdr);
3091 }
3092 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
3093
3094 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
3095 {
3096         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
3097         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
3098             PNFS_LAYOUTRET_ON_ERROR) {
3099                 pnfs_return_layout(hdr->inode);
3100         }
3101         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
3102                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
3103 }
3104
3105 /*
3106  * Called by non rpc-based layout drivers
3107  */
3108 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
3109 {
3110         if (likely(!hdr->pnfs_error))
3111                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
3112         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
3113         if (unlikely(hdr->pnfs_error))
3114                 pnfs_ld_handle_read_error(hdr);
3115         hdr->mds_ops->rpc_release(hdr);
3116 }
3117 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
3118
3119 static void
3120 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
3121                 struct nfs_pgio_header *hdr)
3122 {
3123         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3124
3125         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3126                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3127                 nfs_pageio_reset_read_mds(desc);
3128                 mirror->pg_recoalesce = 1;
3129         }
3130         hdr->completion_ops->completion(hdr);
3131 }
3132
3133 /*
3134  * Call the appropriate parallel I/O subsystem read function.
3135  */
3136 static enum pnfs_try_status
3137 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3138                        const struct rpc_call_ops *call_ops,
3139                        struct pnfs_layout_segment *lseg)
3140 {
3141         struct inode *inode = hdr->inode;
3142         struct nfs_server *nfss = NFS_SERVER(inode);
3143         enum pnfs_try_status trypnfs;
3144
3145         hdr->mds_ops = call_ops;
3146
3147         dprintk("%s: Reading ino:%lu %u@%llu\n",
3148                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3149
3150         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3151         if (trypnfs != PNFS_NOT_ATTEMPTED)
3152                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3153         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3154         return trypnfs;
3155 }
3156
3157 /* Resend all requests through pnfs. */
3158 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3159                            unsigned int mirror_idx)
3160 {
3161         struct nfs_pageio_descriptor pgio;
3162
3163         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3164                 /* Prevent deadlocks with layoutreturn! */
3165                 pnfs_put_lseg(hdr->lseg);
3166                 hdr->lseg = NULL;
3167
3168                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3169                                         hdr->completion_ops);
3170                 pgio.pg_mirror_idx = mirror_idx;
3171                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3172         }
3173 }
3174 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3175
3176 static void
3177 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3178 {
3179         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3180         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3181         enum pnfs_try_status trypnfs;
3182
3183         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3184         switch (trypnfs) {
3185         case PNFS_NOT_ATTEMPTED:
3186                 pnfs_read_through_mds(desc, hdr);
3187                 break;
3188         case PNFS_ATTEMPTED:
3189                 break;
3190         case PNFS_TRY_AGAIN:
3191                 /* cleanup hdr and prepare to redo pnfs */
3192                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3193                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3194                         list_splice_init(&hdr->pages, &mirror->pg_list);
3195                         mirror->pg_recoalesce = 1;
3196                 }
3197                 hdr->mds_ops->rpc_release(hdr);
3198         }
3199 }
3200
3201 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3202 {
3203         pnfs_put_lseg(hdr->lseg);
3204         nfs_pgio_header_free(hdr);
3205 }
3206
3207 int
3208 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3209 {
3210         struct nfs_pgio_header *hdr;
3211         int ret;
3212
3213         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3214         if (!hdr) {
3215                 desc->pg_error = -ENOMEM;
3216                 return desc->pg_error;
3217         }
3218         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3219         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3220         ret = nfs_generic_pgio(desc, hdr);
3221         if (!ret)
3222                 pnfs_do_read(desc, hdr);
3223         return ret;
3224 }
3225 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3226
3227 static void pnfs_clear_layoutcommitting(struct inode *inode)
3228 {
3229         unsigned long *bitlock = &NFS_I(inode)->flags;
3230
3231         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3232         smp_mb__after_atomic();
3233         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3234 }
3235
3236 /*
3237  * There can be multiple RW segments.
3238  */
3239 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3240 {
3241         struct pnfs_layout_segment *lseg;
3242
3243         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3244                 if (lseg->pls_range.iomode == IOMODE_RW &&
3245                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3246                         list_add(&lseg->pls_lc_list, listp);
3247         }
3248 }
3249
3250 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3251 {
3252         struct pnfs_layout_segment *lseg, *tmp;
3253
3254         /* Matched by references in pnfs_set_layoutcommit */
3255         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3256                 list_del_init(&lseg->pls_lc_list);
3257                 pnfs_put_lseg(lseg);
3258         }
3259
3260         pnfs_clear_layoutcommitting(inode);
3261 }
3262
3263 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3264 {
3265         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3266 }
3267 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3268
3269 void
3270 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3271                 loff_t end_pos)
3272 {
3273         struct nfs_inode *nfsi = NFS_I(inode);
3274         bool mark_as_dirty = false;
3275
3276         spin_lock(&inode->i_lock);
3277         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3278                 nfsi->layout->plh_lwb = end_pos;
3279                 mark_as_dirty = true;
3280                 dprintk("%s: Set layoutcommit for inode %lu ",
3281                         __func__, inode->i_ino);
3282         } else if (end_pos > nfsi->layout->plh_lwb)
3283                 nfsi->layout->plh_lwb = end_pos;
3284         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3285                 /* references matched in nfs4_layoutcommit_release */
3286                 pnfs_get_lseg(lseg);
3287         }
3288         spin_unlock(&inode->i_lock);
3289         dprintk("%s: lseg %p end_pos %llu\n",
3290                 __func__, lseg, nfsi->layout->plh_lwb);
3291
3292         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3293          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3294         if (mark_as_dirty)
3295                 mark_inode_dirty_sync(inode);
3296 }
3297 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3298
3299 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3300 {
3301         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3302
3303         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3304                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3305         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3306 }
3307
3308 /*
3309  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3310  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3311  * data to disk to allow the server to recover the data if it crashes.
3312  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3313  * is off, and a COMMIT is sent to a data server, or
3314  * if WRITEs to a data server return NFS_DATA_SYNC.
3315  */
3316 int
3317 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3318 {
3319         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3320         struct nfs4_layoutcommit_data *data;
3321         struct nfs_inode *nfsi = NFS_I(inode);
3322         loff_t end_pos;
3323         int status;
3324
3325         if (!pnfs_layoutcommit_outstanding(inode))
3326                 return 0;
3327
3328         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3329
3330         status = -EAGAIN;
3331         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3332                 if (!sync)
3333                         goto out;
3334                 status = wait_on_bit_lock_action(&nfsi->flags,
3335                                 NFS_INO_LAYOUTCOMMITTING,
3336                                 nfs_wait_bit_killable,
3337                                 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3338                 if (status)
3339                         goto out;
3340         }
3341
3342         status = -ENOMEM;
3343         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3344         data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3345         if (!data)
3346                 goto clear_layoutcommitting;
3347
3348         status = 0;
3349         spin_lock(&inode->i_lock);
3350         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3351                 goto out_unlock;
3352
3353         INIT_LIST_HEAD(&data->lseg_list);
3354         pnfs_list_write_lseg(inode, &data->lseg_list);
3355
3356         end_pos = nfsi->layout->plh_lwb;
3357
3358         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3359         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3360         spin_unlock(&inode->i_lock);
3361
3362         data->args.inode = inode;
3363         nfs_fattr_init(&data->fattr);
3364         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3365         data->res.fattr = &data->fattr;
3366         if (end_pos != 0)
3367                 data->args.lastbytewritten = end_pos - 1;
3368         else
3369                 data->args.lastbytewritten = U64_MAX;
3370         data->res.server = NFS_SERVER(inode);
3371
3372         if (ld->prepare_layoutcommit) {
3373                 status = ld->prepare_layoutcommit(&data->args);
3374                 if (status) {
3375                         put_cred(data->cred);
3376                         spin_lock(&inode->i_lock);
3377                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3378                         if (end_pos > nfsi->layout->plh_lwb)
3379                                 nfsi->layout->plh_lwb = end_pos;
3380                         goto out_unlock;
3381                 }
3382         }
3383
3384
3385         status = nfs4_proc_layoutcommit(data, sync);
3386 out:
3387         if (status)
3388                 mark_inode_dirty_sync(inode);
3389         dprintk("<-- %s status %d\n", __func__, status);
3390         return status;
3391 out_unlock:
3392         spin_unlock(&inode->i_lock);
3393         kfree(data);
3394 clear_layoutcommitting:
3395         pnfs_clear_layoutcommitting(inode);
3396         goto out;
3397 }
3398 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3399
3400 int
3401 pnfs_generic_sync(struct inode *inode, bool datasync)
3402 {
3403         return pnfs_layoutcommit_inode(inode, true);
3404 }
3405 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3406
3407 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3408 {
3409         struct nfs4_threshold *thp;
3410
3411         thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3412         if (!thp) {
3413                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3414                 return NULL;
3415         }
3416         return thp;
3417 }
3418
3419 #if IS_ENABLED(CONFIG_NFS_V4_2)
3420 int
3421 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3422 {
3423         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3424         struct nfs_server *server = NFS_SERVER(inode);
3425         struct nfs_inode *nfsi = NFS_I(inode);
3426         struct nfs42_layoutstat_data *data;
3427         struct pnfs_layout_hdr *hdr;
3428         int status = 0;
3429
3430         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3431                 goto out;
3432
3433         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3434                 goto out;
3435
3436         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3437                 goto out;
3438
3439         spin_lock(&inode->i_lock);
3440         if (!NFS_I(inode)->layout) {
3441                 spin_unlock(&inode->i_lock);
3442                 goto out_clear_layoutstats;
3443         }
3444         hdr = NFS_I(inode)->layout;
3445         pnfs_get_layout_hdr(hdr);
3446         spin_unlock(&inode->i_lock);
3447
3448         data = kzalloc(sizeof(*data), gfp_flags);
3449         if (!data) {
3450                 status = -ENOMEM;
3451                 goto out_put;
3452         }
3453
3454         data->args.fh = NFS_FH(inode);
3455         data->args.inode = inode;
3456         status = ld->prepare_layoutstats(&data->args);
3457         if (status)
3458                 goto out_free;
3459
3460         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3461
3462 out:
3463         dprintk("%s returns %d\n", __func__, status);
3464         return status;
3465
3466 out_free:
3467         kfree(data);
3468 out_put:
3469         pnfs_put_layout_hdr(hdr);
3470 out_clear_layoutstats:
3471         smp_mb__before_atomic();
3472         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3473         smp_mb__after_atomic();
3474         goto out;
3475 }
3476 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3477 #endif
3478
3479 unsigned int layoutstats_timer;
3480 module_param(layoutstats_timer, uint, 0644);
3481 EXPORT_SYMBOL_GPL(layoutstats_timer);