fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include <linux/sort.h>
  34 #include "internal.h"
  35 #include "pnfs.h"
  36 #include "iostat.h"
  37 #include "nfs4trace.h"
  38 #include "delegation.h"
  39 #include "nfs42.h"
  40 #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 (!pnfs_layout_is_valid(lo) ||
1288             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1289                 goto out_unlock;
1290         if (stateid) {
1291                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1292
1293                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1294                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1295                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1296         } else
1297                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1298 out_unlock:
1299         pnfs_clear_layoutreturn_waitbit(lo);
1300         spin_unlock(&inode->i_lock);
1301         pnfs_free_lseg_list(&freeme);
1302
1303 }
1304
1305 static bool
1306 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1307                 nfs4_stateid *stateid,
1308                 const struct cred **cred,
1309                 enum pnfs_iomode *iomode)
1310 {
1311         /* Serialise LAYOUTGET/LAYOUTRETURN */
1312         if (atomic_read(&lo->plh_outstanding) != 0)
1313                 return false;
1314         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1315                 return false;
1316         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1317         pnfs_get_layout_hdr(lo);
1318         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1319         *cred = get_cred(lo->plh_lc_cred);
1320         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1321                 if (lo->plh_return_seq != 0)
1322                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1323                 if (iomode != NULL)
1324                         *iomode = lo->plh_return_iomode;
1325                 pnfs_clear_layoutreturn_info(lo);
1326         } else if (iomode != NULL)
1327                 *iomode = IOMODE_ANY;
1328         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1329         return true;
1330 }
1331
1332 static void
1333 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1334                 struct pnfs_layout_hdr *lo,
1335                 const nfs4_stateid *stateid,
1336                 enum pnfs_iomode iomode)
1337 {
1338         struct inode *inode = lo->plh_inode;
1339
1340         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1341         args->inode = inode;
1342         args->range.iomode = iomode;
1343         args->range.offset = 0;
1344         args->range.length = NFS4_MAX_UINT64;
1345         args->layout = lo;
1346         nfs4_stateid_copy(&args->stateid, stateid);
1347 }
1348
1349 static int
1350 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1351                        const nfs4_stateid *stateid,
1352                        const struct cred **pcred,
1353                        enum pnfs_iomode iomode,
1354                        unsigned int flags)
1355 {
1356         struct inode *ino = lo->plh_inode;
1357         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1358         struct nfs4_layoutreturn *lrp;
1359         const struct cred *cred = *pcred;
1360         int status = 0;
1361
1362         *pcred = NULL;
1363         lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1364         if (unlikely(lrp == NULL)) {
1365                 status = -ENOMEM;
1366                 spin_lock(&ino->i_lock);
1367                 pnfs_clear_layoutreturn_waitbit(lo);
1368                 spin_unlock(&ino->i_lock);
1369                 put_cred(cred);
1370                 pnfs_put_layout_hdr(lo);
1371                 goto out;
1372         }
1373
1374         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1375         lrp->args.ld_private = &lrp->ld_private;
1376         lrp->clp = NFS_SERVER(ino)->nfs_client;
1377         lrp->cred = cred;
1378         if (ld->prepare_layoutreturn)
1379                 ld->prepare_layoutreturn(&lrp->args);
1380
1381         status = nfs4_proc_layoutreturn(lrp, flags);
1382 out:
1383         dprintk("<-- %s status: %d\n", __func__, status);
1384         return status;
1385 }
1386
1387 /* Return true if layoutreturn is needed */
1388 static bool
1389 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1390 {
1391         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1392                 return false;
1393         return pnfs_mark_layout_stateid_return(lo, &lo->plh_return_segs,
1394                                                lo->plh_return_iomode,
1395                                                lo->plh_return_seq) != EBUSY;
1396 }
1397
1398 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1399 {
1400         struct inode *inode= lo->plh_inode;
1401
1402         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1403                 return;
1404         spin_lock(&inode->i_lock);
1405         if (pnfs_layout_need_return(lo)) {
1406                 const struct cred *cred;
1407                 nfs4_stateid stateid;
1408                 enum pnfs_iomode iomode;
1409                 bool send;
1410
1411                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1412                 spin_unlock(&inode->i_lock);
1413                 if (send) {
1414                         /* Send an async layoutreturn so we dont deadlock */
1415                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
1416                                                PNFS_FL_LAYOUTRETURN_ASYNC);
1417                 }
1418         } else
1419                 spin_unlock(&inode->i_lock);
1420 }
1421
1422 /*
1423  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1424  * when the layout segment list is empty.
1425  *
1426  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1427  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1428  * deviceid is marked invalid.
1429  */
1430 int
1431 _pnfs_return_layout(struct inode *ino)
1432 {
1433         struct pnfs_layout_hdr *lo = NULL;
1434         struct nfs_inode *nfsi = NFS_I(ino);
1435         struct pnfs_layout_range range = {
1436                 .iomode         = IOMODE_ANY,
1437                 .offset         = 0,
1438                 .length         = NFS4_MAX_UINT64,
1439         };
1440         LIST_HEAD(tmp_list);
1441         const struct cred *cred;
1442         nfs4_stateid stateid;
1443         int status = 0;
1444         bool send, valid_layout;
1445
1446         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1447
1448         spin_lock(&ino->i_lock);
1449         lo = nfsi->layout;
1450         if (!lo) {
1451                 spin_unlock(&ino->i_lock);
1452                 dprintk("NFS: %s no layout to return\n", __func__);
1453                 goto out;
1454         }
1455         /* Reference matched in nfs4_layoutreturn_release */
1456         pnfs_get_layout_hdr(lo);
1457         /* Is there an outstanding layoutreturn ? */
1458         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1459                 spin_unlock(&ino->i_lock);
1460                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1461                                         TASK_UNINTERRUPTIBLE))
1462                         goto out_put_layout_hdr;
1463                 spin_lock(&ino->i_lock);
1464         }
1465         valid_layout = pnfs_layout_is_valid(lo);
1466         pnfs_clear_layoutcommit(ino, &tmp_list);
1467         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1468
1469         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1470                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1471
1472         /* Don't send a LAYOUTRETURN if list was initially empty */
1473         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1474                         !valid_layout) {
1475                 spin_unlock(&ino->i_lock);
1476                 dprintk("NFS: %s no layout segments to return\n", __func__);
1477                 goto out_wait_layoutreturn;
1478         }
1479
1480         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1481         spin_unlock(&ino->i_lock);
1482         if (send)
1483                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY,
1484                                                 0);
1485 out_wait_layoutreturn:
1486         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1487 out_put_layout_hdr:
1488         pnfs_free_lseg_list(&tmp_list);
1489         pnfs_put_layout_hdr(lo);
1490 out:
1491         dprintk("<-- %s status: %d\n", __func__, status);
1492         return status;
1493 }
1494
1495 int
1496 pnfs_commit_and_return_layout(struct inode *inode)
1497 {
1498         struct pnfs_layout_hdr *lo;
1499         int ret;
1500
1501         spin_lock(&inode->i_lock);
1502         lo = NFS_I(inode)->layout;
1503         if (lo == NULL) {
1504                 spin_unlock(&inode->i_lock);
1505                 return 0;
1506         }
1507         pnfs_get_layout_hdr(lo);
1508         /* Block new layoutgets and read/write to ds */
1509         lo->plh_block_lgets++;
1510         spin_unlock(&inode->i_lock);
1511         filemap_fdatawait(inode->i_mapping);
1512         ret = pnfs_layoutcommit_inode(inode, true);
1513         if (ret == 0)
1514                 ret = _pnfs_return_layout(inode);
1515         spin_lock(&inode->i_lock);
1516         lo->plh_block_lgets--;
1517         spin_unlock(&inode->i_lock);
1518         pnfs_put_layout_hdr(lo);
1519         return ret;
1520 }
1521
1522 static int pnfs_layout_return_on_reboot(struct pnfs_layout_hdr *lo)
1523 {
1524         struct inode *inode = lo->plh_inode;
1525         const struct cred *cred;
1526
1527         spin_lock(&inode->i_lock);
1528         if (!pnfs_layout_is_valid(lo)) {
1529                 spin_unlock(&inode->i_lock);
1530                 return 0;
1531         }
1532         cred = get_cred(lo->plh_lc_cred);
1533         pnfs_get_layout_hdr(lo);
1534         spin_unlock(&inode->i_lock);
1535
1536         return pnfs_send_layoutreturn(lo, &zero_stateid, &cred, IOMODE_ANY,
1537                                       PNFS_FL_LAYOUTRETURN_PRIVILEGED);
1538 }
1539
1540 bool pnfs_roc(struct inode *ino,
1541                 struct nfs4_layoutreturn_args *args,
1542                 struct nfs4_layoutreturn_res *res,
1543                 const struct cred *cred)
1544 {
1545         struct nfs_inode *nfsi = NFS_I(ino);
1546         struct nfs_open_context *ctx;
1547         struct nfs4_state *state;
1548         struct pnfs_layout_hdr *lo;
1549         struct pnfs_layout_segment *lseg, *next;
1550         const struct cred *lc_cred;
1551         nfs4_stateid stateid;
1552         enum pnfs_iomode iomode = 0;
1553         bool layoutreturn = false, roc = false;
1554         bool skip_read = false;
1555
1556         if (!nfs_have_layout(ino))
1557                 return false;
1558 retry:
1559         rcu_read_lock();
1560         spin_lock(&ino->i_lock);
1561         lo = nfsi->layout;
1562         if (!lo || !pnfs_layout_is_valid(lo) ||
1563             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1564                 lo = NULL;
1565                 goto out_noroc;
1566         }
1567         pnfs_get_layout_hdr(lo);
1568         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1569                 spin_unlock(&ino->i_lock);
1570                 rcu_read_unlock();
1571                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1572                                 TASK_UNINTERRUPTIBLE);
1573                 pnfs_put_layout_hdr(lo);
1574                 goto retry;
1575         }
1576
1577         /* no roc if we hold a delegation */
1578         if (nfs4_check_delegation(ino, FMODE_READ)) {
1579                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1580                         goto out_noroc;
1581                 skip_read = true;
1582         }
1583
1584         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1585                 state = ctx->state;
1586                 if (state == NULL)
1587                         continue;
1588                 /* Don't return layout if there is open file state */
1589                 if (state->state & FMODE_WRITE)
1590                         goto out_noroc;
1591                 if (state->state & FMODE_READ)
1592                         skip_read = true;
1593         }
1594
1595
1596         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1597                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1598                         continue;
1599                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1600                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1601                         continue;
1602                 /*
1603                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1604                  * doesn't invalidate the layout for us.
1605                  */
1606                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1607                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1608                         continue;
1609                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1610         }
1611
1612         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1613                 goto out_noroc;
1614
1615         /* ROC in two conditions:
1616          * 1. there are ROC lsegs
1617          * 2. we don't send layoutreturn
1618          */
1619         /* lo ref dropped in pnfs_roc_release() */
1620         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1621         /* If the creds don't match, we can't compound the layoutreturn */
1622         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1623                 goto out_noroc;
1624
1625         roc = layoutreturn;
1626         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1627         res->lrs_present = 0;
1628         layoutreturn = false;
1629         put_cred(lc_cred);
1630
1631 out_noroc:
1632         spin_unlock(&ino->i_lock);
1633         rcu_read_unlock();
1634         pnfs_layoutcommit_inode(ino, true);
1635         if (roc) {
1636                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1637                 if (ld->prepare_layoutreturn)
1638                         ld->prepare_layoutreturn(args);
1639                 pnfs_put_layout_hdr(lo);
1640                 return true;
1641         }
1642         if (layoutreturn)
1643                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, 0);
1644         pnfs_put_layout_hdr(lo);
1645         return false;
1646 }
1647
1648 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1649                   struct nfs4_layoutreturn_res **respp, int *ret)
1650 {
1651         struct nfs4_layoutreturn_args *arg = *argpp;
1652         int retval = -EAGAIN;
1653
1654         if (!arg)
1655                 return 0;
1656         /* Handle Layoutreturn errors */
1657         switch (*ret) {
1658         case 0:
1659                 retval = 0;
1660                 break;
1661         case -NFS4ERR_NOMATCHING_LAYOUT:
1662                 /* Was there an RPC level error? If not, retry */
1663                 if (task->tk_rpc_status == 0)
1664                         break;
1665                 /* If the call was not sent, let caller handle it */
1666                 if (!RPC_WAS_SENT(task))
1667                         return 0;
1668                 /*
1669                  * Otherwise, assume the call succeeded and
1670                  * that we need to release the layout
1671                  */
1672                 *ret = 0;
1673                 (*respp)->lrs_present = 0;
1674                 retval = 0;
1675                 break;
1676         case -NFS4ERR_DELAY:
1677                 /* Let the caller handle the retry */
1678                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1679                 return 0;
1680         case -NFS4ERR_OLD_STATEID:
1681                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1682                                                      &arg->range, arg->inode))
1683                         break;
1684                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1685                 return -EAGAIN;
1686         }
1687         *argpp = NULL;
1688         *respp = NULL;
1689         return retval;
1690 }
1691
1692 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1693                       struct nfs4_layoutreturn_res *res, int ret)
1694 {
1695         struct pnfs_layout_hdr *lo = args->layout;
1696         struct inode *inode = args->inode;
1697         const nfs4_stateid *res_stateid = NULL;
1698         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1699
1700         switch (ret) {
1701         case -NFS4ERR_BADSESSION:
1702         case -NFS4ERR_DEADSESSION:
1703         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1704         case -NFS4ERR_NOMATCHING_LAYOUT:
1705                 spin_lock(&inode->i_lock);
1706                 pnfs_layoutreturn_retry_later_locked(lo, &args->stateid,
1707                                                      &args->range);
1708                 pnfs_clear_layoutreturn_waitbit(lo);
1709                 spin_unlock(&inode->i_lock);
1710                 break;
1711         case 0:
1712                 if (res->lrs_present)
1713                         res_stateid = &res->stateid;
1714                 fallthrough;
1715         default:
1716                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1717                                              res_stateid);
1718         }
1719         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1720         if (ld_private && ld_private->ops && ld_private->ops->free)
1721                 ld_private->ops->free(ld_private);
1722         pnfs_put_layout_hdr(lo);
1723 }
1724
1725 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1726 {
1727         struct nfs_inode *nfsi = NFS_I(ino);
1728         struct pnfs_layout_hdr *lo;
1729         bool sleep = false;
1730
1731         /* we might not have grabbed lo reference. so need to check under
1732          * i_lock */
1733         spin_lock(&ino->i_lock);
1734         lo = nfsi->layout;
1735         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1736                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1737                 sleep = true;
1738         }
1739         spin_unlock(&ino->i_lock);
1740         return sleep;
1741 }
1742
1743 /*
1744  * Compare two layout segments for sorting into layout cache.
1745  * We want to preferentially return RW over RO layouts, so ensure those
1746  * are seen first.
1747  */
1748 static s64
1749 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1750            const struct pnfs_layout_range *l2)
1751 {
1752         s64 d;
1753
1754         /* high offset > low offset */
1755         d = l1->offset - l2->offset;
1756         if (d)
1757                 return d;
1758
1759         /* short length > long length */
1760         d = l2->length - l1->length;
1761         if (d)
1762                 return d;
1763
1764         /* read > read/write */
1765         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1766 }
1767
1768 static bool
1769 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1770                 const struct pnfs_layout_range *l2)
1771 {
1772         return pnfs_lseg_range_cmp(l1, l2) > 0;
1773 }
1774
1775 static bool
1776 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1777                 struct pnfs_layout_segment *old)
1778 {
1779         return false;
1780 }
1781
1782 void
1783 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1784                    struct pnfs_layout_segment *lseg,
1785                    bool (*is_after)(const struct pnfs_layout_range *,
1786                            const struct pnfs_layout_range *),
1787                    bool (*do_merge)(struct pnfs_layout_segment *,
1788                            struct pnfs_layout_segment *),
1789                    struct list_head *free_me)
1790 {
1791         struct pnfs_layout_segment *lp, *tmp;
1792
1793         dprintk("%s:Begin\n", __func__);
1794
1795         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1796                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1797                         continue;
1798                 if (do_merge(lseg, lp)) {
1799                         mark_lseg_invalid(lp, free_me);
1800                         continue;
1801                 }
1802                 if (is_after(&lseg->pls_range, &lp->pls_range))
1803                         continue;
1804                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1805                 dprintk("%s: inserted lseg %p "
1806                         "iomode %d offset %llu length %llu before "
1807                         "lp %p iomode %d offset %llu length %llu\n",
1808                         __func__, lseg, lseg->pls_range.iomode,
1809                         lseg->pls_range.offset, lseg->pls_range.length,
1810                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1811                         lp->pls_range.length);
1812                 goto out;
1813         }
1814         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1815         dprintk("%s: inserted lseg %p "
1816                 "iomode %d offset %llu length %llu at tail\n",
1817                 __func__, lseg, lseg->pls_range.iomode,
1818                 lseg->pls_range.offset, lseg->pls_range.length);
1819 out:
1820         pnfs_get_layout_hdr(lo);
1821
1822         dprintk("%s:Return\n", __func__);
1823 }
1824 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1825
1826 static void
1827 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1828                    struct pnfs_layout_segment *lseg,
1829                    struct list_head *free_me)
1830 {
1831         struct inode *inode = lo->plh_inode;
1832         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1833
1834         if (ld->add_lseg != NULL)
1835                 ld->add_lseg(lo, lseg, free_me);
1836         else
1837                 pnfs_generic_layout_insert_lseg(lo, lseg,
1838                                 pnfs_lseg_range_is_after,
1839                                 pnfs_lseg_no_merge,
1840                                 free_me);
1841 }
1842
1843 static struct pnfs_layout_hdr *
1844 alloc_init_layout_hdr(struct inode *ino,
1845                       struct nfs_open_context *ctx,
1846                       gfp_t gfp_flags)
1847 {
1848         struct pnfs_layout_hdr *lo;
1849
1850         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1851         if (!lo)
1852                 return NULL;
1853         refcount_set(&lo->plh_refcount, 1);
1854         INIT_LIST_HEAD(&lo->plh_layouts);
1855         INIT_LIST_HEAD(&lo->plh_segs);
1856         INIT_LIST_HEAD(&lo->plh_return_segs);
1857         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1858         lo->plh_inode = ino;
1859         lo->plh_lc_cred = get_cred(ctx->cred);
1860         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1861         return lo;
1862 }
1863
1864 static struct pnfs_layout_hdr *
1865 pnfs_find_alloc_layout(struct inode *ino,
1866                        struct nfs_open_context *ctx,
1867                        gfp_t gfp_flags)
1868         __releases(&ino->i_lock)
1869         __acquires(&ino->i_lock)
1870 {
1871         struct nfs_inode *nfsi = NFS_I(ino);
1872         struct pnfs_layout_hdr *new = NULL;
1873
1874         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1875
1876         if (nfsi->layout != NULL)
1877                 goto out_existing;
1878         spin_unlock(&ino->i_lock);
1879         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1880         spin_lock(&ino->i_lock);
1881
1882         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1883                 nfsi->layout = new;
1884                 return new;
1885         } else if (new != NULL)
1886                 pnfs_free_layout_hdr(new);
1887 out_existing:
1888         pnfs_get_layout_hdr(nfsi->layout);
1889         return nfsi->layout;
1890 }
1891
1892 /*
1893  * iomode matching rules:
1894  * iomode       lseg    strict match
1895  *                      iomode
1896  * -----        -----   ------ -----
1897  * ANY          READ    N/A    true
1898  * ANY          RW      N/A    true
1899  * RW           READ    N/A    false
1900  * RW           RW      N/A    true
1901  * READ         READ    N/A    true
1902  * READ         RW      true   false
1903  * READ         RW      false  true
1904  */
1905 static bool
1906 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1907                  const struct pnfs_layout_range *range,
1908                  bool strict_iomode)
1909 {
1910         struct pnfs_layout_range range1;
1911
1912         if ((range->iomode == IOMODE_RW &&
1913              ls_range->iomode != IOMODE_RW) ||
1914             (range->iomode != ls_range->iomode &&
1915              strict_iomode) ||
1916             !pnfs_lseg_range_intersecting(ls_range, range))
1917                 return false;
1918
1919         /* range1 covers only the first byte in the range */
1920         range1 = *range;
1921         range1.length = 1;
1922         return pnfs_lseg_range_contained(ls_range, &range1);
1923 }
1924
1925 /*
1926  * lookup range in layout
1927  */
1928 static struct pnfs_layout_segment *
1929 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1930                 struct pnfs_layout_range *range,
1931                 bool strict_iomode)
1932 {
1933         struct pnfs_layout_segment *lseg, *ret = NULL;
1934
1935         dprintk("%s:Begin\n", __func__);
1936
1937         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1938                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1939                     pnfs_lseg_range_match(&lseg->pls_range, range,
1940                                           strict_iomode)) {
1941                         ret = pnfs_get_lseg(lseg);
1942                         break;
1943                 }
1944         }
1945
1946         dprintk("%s:Return lseg %p ref %d\n",
1947                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1948         return ret;
1949 }
1950
1951 /*
1952  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1953  * to the MDS or over pNFS
1954  *
1955  * The nfs_inode read_io and write_io fields are cumulative counters reset
1956  * when there are no layout segments. Note that in pnfs_update_layout iomode
1957  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1958  * WRITE request.
1959  *
1960  * A return of true means use MDS I/O.
1961  *
1962  * From rfc 5661:
1963  * If a file's size is smaller than the file size threshold, data accesses
1964  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1965  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1966  * server.  If both file size and I/O size are provided, the client SHOULD
1967  * reach or exceed  both thresholds before sending its read or write
1968  * requests to the data server.
1969  */
1970 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1971                                      struct inode *ino, int iomode)
1972 {
1973         struct nfs4_threshold *t = ctx->mdsthreshold;
1974         struct nfs_inode *nfsi = NFS_I(ino);
1975         loff_t fsize = i_size_read(ino);
1976         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1977
1978         if (t == NULL)
1979                 return ret;
1980
1981         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1982                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1983
1984         switch (iomode) {
1985         case IOMODE_READ:
1986                 if (t->bm & THRESHOLD_RD) {
1987                         dprintk("%s fsize %llu\n", __func__, fsize);
1988                         size_set = true;
1989                         if (fsize < t->rd_sz)
1990                                 size = true;
1991                 }
1992                 if (t->bm & THRESHOLD_RD_IO) {
1993                         dprintk("%s nfsi->read_io %llu\n", __func__,
1994                                 nfsi->read_io);
1995                         io_set = true;
1996                         if (nfsi->read_io < t->rd_io_sz)
1997                                 io = true;
1998                 }
1999                 break;
2000         case IOMODE_RW:
2001                 if (t->bm & THRESHOLD_WR) {
2002                         dprintk("%s fsize %llu\n", __func__, fsize);
2003                         size_set = true;
2004                         if (fsize < t->wr_sz)
2005                                 size = true;
2006                 }
2007                 if (t->bm & THRESHOLD_WR_IO) {
2008                         dprintk("%s nfsi->write_io %llu\n", __func__,
2009                                 nfsi->write_io);
2010                         io_set = true;
2011                         if (nfsi->write_io < t->wr_io_sz)
2012                                 io = true;
2013                 }
2014                 break;
2015         }
2016         if (size_set && io_set) {
2017                 if (size && io)
2018                         ret = true;
2019         } else if (size || io)
2020                 ret = true;
2021
2022         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
2023         return ret;
2024 }
2025
2026 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
2027 {
2028         /*
2029          * send layoutcommit as it can hold up layoutreturn due to lseg
2030          * reference
2031          */
2032         pnfs_layoutcommit_inode(lo->plh_inode, false);
2033         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
2034                                    nfs_wait_bit_killable,
2035                                    TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
2036 }
2037
2038 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
2039 {
2040         atomic_inc(&lo->plh_outstanding);
2041 }
2042
2043 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
2044 {
2045         if (atomic_dec_and_test(&lo->plh_outstanding) &&
2046             test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
2047                 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
2048 }
2049
2050 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
2051 {
2052         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
2053 }
2054
2055 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
2056 {
2057         unsigned long *bitlock = &lo->plh_flags;
2058
2059         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
2060         smp_mb__after_atomic();
2061         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
2062 }
2063
2064 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
2065                                 struct nfs_server *server)
2066 {
2067         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
2068                 struct nfs_client *clp = server->nfs_client;
2069
2070                 /* The lo must be on the clp list if there is any
2071                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
2072                  */
2073                 spin_lock(&clp->cl_lock);
2074                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
2075                 spin_unlock(&clp->cl_lock);
2076         }
2077 }
2078
2079 /*
2080  * Layout segment is retreived from the server if not cached.
2081  * The appropriate layout segment is referenced and returned to the caller.
2082  */
2083 struct pnfs_layout_segment *
2084 pnfs_update_layout(struct inode *ino,
2085                    struct nfs_open_context *ctx,
2086                    loff_t pos,
2087                    u64 count,
2088                    enum pnfs_iomode iomode,
2089                    bool strict_iomode,
2090                    gfp_t gfp_flags)
2091 {
2092         struct pnfs_layout_range arg = {
2093                 .iomode = iomode,
2094                 .offset = pos,
2095                 .length = count,
2096         };
2097         unsigned pg_offset;
2098         struct nfs_server *server = NFS_SERVER(ino);
2099         struct nfs_client *clp = server->nfs_client;
2100         struct pnfs_layout_hdr *lo = NULL;
2101         struct pnfs_layout_segment *lseg = NULL;
2102         struct nfs4_layoutget *lgp;
2103         nfs4_stateid stateid;
2104         struct nfs4_exception exception = {
2105                 .inode = ino,
2106         };
2107         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
2108         bool first;
2109
2110         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
2111                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2112                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
2113                 goto out;
2114         }
2115
2116         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
2117                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2118                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
2119                 goto out;
2120         }
2121
2122 lookup_again:
2123         if (!nfs4_valid_open_stateid(ctx->state)) {
2124                 trace_pnfs_update_layout(ino, pos, count,
2125                                          iomode, lo, lseg,
2126                                          PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2127                 lseg = ERR_PTR(-EIO);
2128                 goto out;
2129         }
2130
2131         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2132         if (IS_ERR(lseg))
2133                 goto out;
2134         first = false;
2135         spin_lock(&ino->i_lock);
2136         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2137         if (lo == NULL) {
2138                 spin_unlock(&ino->i_lock);
2139                 lseg = ERR_PTR(-ENOMEM);
2140                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2141                                  PNFS_UPDATE_LAYOUT_NOMEM);
2142                 goto out;
2143         }
2144
2145         /* Do we even need to bother with this? */
2146         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2147                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2148                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2149                 dprintk("%s matches recall, use MDS\n", __func__);
2150                 goto out_unlock;
2151         }
2152
2153         /* if LAYOUTGET already failed once we don't try again */
2154         if (pnfs_layout_io_test_failed(lo, iomode)) {
2155                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2156                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2157                 goto out_unlock;
2158         }
2159
2160         /*
2161          * If the layout segment list is empty, but there are outstanding
2162          * layoutget calls, then they might be subject to a layoutrecall.
2163          */
2164         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2165             atomic_read(&lo->plh_outstanding) != 0) {
2166                 spin_unlock(&ino->i_lock);
2167                 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2168                                            TASK_KILLABLE));
2169                 if (IS_ERR(lseg))
2170                         goto out_put_layout_hdr;
2171                 pnfs_put_layout_hdr(lo);
2172                 goto lookup_again;
2173         }
2174
2175         /*
2176          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2177          * for LAYOUTRETURN.
2178          */
2179         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2180                 spin_unlock(&ino->i_lock);
2181                 dprintk("%s wait for layoutreturn\n", __func__);
2182                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2183                 if (!IS_ERR(lseg)) {
2184                         pnfs_put_layout_hdr(lo);
2185                         dprintk("%s retrying\n", __func__);
2186                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2187                                                  lseg,
2188                                                  PNFS_UPDATE_LAYOUT_RETRY);
2189                         goto lookup_again;
2190                 }
2191                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2192                                          PNFS_UPDATE_LAYOUT_RETURN);
2193                 goto out_put_layout_hdr;
2194         }
2195
2196         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2197         if (lseg) {
2198                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2199                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2200                 goto out_unlock;
2201         }
2202
2203         /*
2204          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2205          * stateid, or it has been invalidated, then we must use the open
2206          * stateid.
2207          */
2208         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2209                 int status;
2210
2211                 /*
2212                  * The first layoutget for the file. Need to serialize per
2213                  * RFC 5661 Errata 3208.
2214                  */
2215                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2216                                      &lo->plh_flags)) {
2217                         spin_unlock(&ino->i_lock);
2218                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2219                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2220                                                 TASK_KILLABLE));
2221                         if (IS_ERR(lseg))
2222                                 goto out_put_layout_hdr;
2223                         pnfs_put_layout_hdr(lo);
2224                         dprintk("%s retrying\n", __func__);
2225                         goto lookup_again;
2226                 }
2227
2228                 spin_unlock(&ino->i_lock);
2229                 first = true;
2230                 status = nfs4_select_rw_stateid(ctx->state,
2231                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2232                                         NULL, &stateid, NULL);
2233                 if (status != 0) {
2234                         lseg = ERR_PTR(status);
2235                         trace_pnfs_update_layout(ino, pos, count,
2236                                         iomode, lo, lseg,
2237                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2238                         nfs4_schedule_stateid_recovery(server, ctx->state);
2239                         pnfs_clear_first_layoutget(lo);
2240                         pnfs_put_layout_hdr(lo);
2241                         goto lookup_again;
2242                 }
2243                 spin_lock(&ino->i_lock);
2244         } else {
2245                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2246         }
2247
2248         if (pnfs_layoutgets_blocked(lo)) {
2249                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2250                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2251                 goto out_unlock;
2252         }
2253         nfs_layoutget_begin(lo);
2254         spin_unlock(&ino->i_lock);
2255
2256         _add_to_server_list(lo, server);
2257
2258         pg_offset = arg.offset & ~PAGE_MASK;
2259         if (pg_offset) {
2260                 arg.offset -= pg_offset;
2261                 arg.length += pg_offset;
2262         }
2263         if (arg.length != NFS4_MAX_UINT64)
2264                 arg.length = PAGE_ALIGN(arg.length);
2265
2266         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2267         if (!lgp) {
2268                 lseg = ERR_PTR(-ENOMEM);
2269                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2270                                          PNFS_UPDATE_LAYOUT_NOMEM);
2271                 nfs_layoutget_end(lo);
2272                 goto out_put_layout_hdr;
2273         }
2274
2275         lgp->lo = lo;
2276         pnfs_get_layout_hdr(lo);
2277
2278         lseg = nfs4_proc_layoutget(lgp, &exception);
2279         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2280                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2281         nfs_layoutget_end(lo);
2282         if (IS_ERR(lseg)) {
2283                 switch(PTR_ERR(lseg)) {
2284                 case -EBUSY:
2285                         if (time_after(jiffies, giveup))
2286                                 lseg = NULL;
2287                         break;
2288                 case -ERECALLCONFLICT:
2289                 case -EAGAIN:
2290                         break;
2291                 case -ENODATA:
2292                         /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2293                         pnfs_layout_set_fail_bit(
2294                                 lo, pnfs_iomode_to_fail_bit(iomode));
2295                         lseg = NULL;
2296                         goto out_put_layout_hdr;
2297                 default:
2298                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2299                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2300                                 lseg = NULL;
2301                         }
2302                         goto out_put_layout_hdr;
2303                 }
2304                 if (lseg) {
2305                         if (!exception.retry)
2306                                 goto out_put_layout_hdr;
2307                         if (first)
2308                                 pnfs_clear_first_layoutget(lo);
2309                         trace_pnfs_update_layout(ino, pos, count,
2310                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2311                         pnfs_put_layout_hdr(lo);
2312                         goto lookup_again;
2313                 }
2314         } else {
2315                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2316         }
2317
2318 out_put_layout_hdr:
2319         if (first)
2320                 pnfs_clear_first_layoutget(lo);
2321         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2322                                  PNFS_UPDATE_LAYOUT_EXIT);
2323         pnfs_put_layout_hdr(lo);
2324 out:
2325         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2326                         "(%s, offset: %llu, length: %llu)\n",
2327                         __func__, ino->i_sb->s_id,
2328                         (unsigned long long)NFS_FILEID(ino),
2329                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2330                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2331                         (unsigned long long)pos,
2332                         (unsigned long long)count);
2333         return lseg;
2334 out_unlock:
2335         spin_unlock(&ino->i_lock);
2336         goto out_put_layout_hdr;
2337 }
2338 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2339
2340 static bool
2341 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2342 {
2343         switch (range->iomode) {
2344         case IOMODE_READ:
2345         case IOMODE_RW:
2346                 break;
2347         default:
2348                 return false;
2349         }
2350         if (range->offset == NFS4_MAX_UINT64)
2351                 return false;
2352         if (range->length == 0)
2353                 return false;
2354         if (range->length != NFS4_MAX_UINT64 &&
2355             range->length > NFS4_MAX_UINT64 - range->offset)
2356                 return false;
2357         return true;
2358 }
2359
2360 static struct pnfs_layout_hdr *
2361 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2362 {
2363         struct pnfs_layout_hdr *lo;
2364
2365         spin_lock(&ino->i_lock);
2366         lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2367         if (!lo)
2368                 goto out_unlock;
2369         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2370                 goto out_unlock;
2371         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2372                 goto out_unlock;
2373         if (pnfs_layoutgets_blocked(lo))
2374                 goto out_unlock;
2375         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2376                 goto out_unlock;
2377         nfs_layoutget_begin(lo);
2378         spin_unlock(&ino->i_lock);
2379         _add_to_server_list(lo, NFS_SERVER(ino));
2380         return lo;
2381
2382 out_unlock:
2383         spin_unlock(&ino->i_lock);
2384         pnfs_put_layout_hdr(lo);
2385         return NULL;
2386 }
2387
2388 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2389                                      struct nfs_open_context *ctx)
2390 {
2391         struct inode *ino = data->dentry->d_inode;
2392         struct pnfs_layout_range rng = {
2393                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2394                           IOMODE_RW: IOMODE_READ,
2395                 .offset = 0,
2396                 .length = NFS4_MAX_UINT64,
2397         };
2398         struct nfs4_layoutget *lgp;
2399         struct pnfs_layout_hdr *lo;
2400
2401         /* Heuristic: don't send layoutget if we have cached data */
2402         if (rng.iomode == IOMODE_READ &&
2403            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2404                 return;
2405
2406         lo = _pnfs_grab_empty_layout(ino, ctx);
2407         if (!lo)
2408                 return;
2409         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2410                                              nfs_io_gfp_mask());
2411         if (!lgp) {
2412                 pnfs_clear_first_layoutget(lo);
2413                 nfs_layoutget_end(lo);
2414                 pnfs_put_layout_hdr(lo);
2415                 return;
2416         }
2417         lgp->lo = lo;
2418         data->lgp = lgp;
2419         data->o_arg.lg_args = &lgp->args;
2420         data->o_res.lg_res = &lgp->res;
2421 }
2422
2423 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2424                                      struct nfs_open_context *ctx)
2425 {
2426         struct inode *ino = data->dentry->d_inode;
2427         struct pnfs_layout_range rng = {
2428                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2429                           IOMODE_RW: IOMODE_READ,
2430                 .offset = 0,
2431                 .length = NFS4_MAX_UINT64,
2432         };
2433         struct nfs4_layoutget *lgp;
2434
2435         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2436                                              nfs_io_gfp_mask());
2437         if (!lgp)
2438                 return;
2439         data->lgp = lgp;
2440         data->o_arg.lg_args = &lgp->args;
2441         data->o_res.lg_res = &lgp->res;
2442 }
2443
2444 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2445                          struct nfs_open_context *ctx)
2446 {
2447         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2448
2449         if (!(pnfs_enabled_sb(server) &&
2450               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2451                 return;
2452         /* Could check on max_ops, but currently hardcoded high enough */
2453         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2454                 return;
2455         if (data->lgp)
2456                 return;
2457         if (data->state)
2458                 _lgopen_prepare_attached(data, ctx);
2459         else
2460                 _lgopen_prepare_floating(data, ctx);
2461 }
2462
2463 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2464                        struct nfs_open_context *ctx)
2465 {
2466         struct pnfs_layout_hdr *lo;
2467         struct pnfs_layout_segment *lseg;
2468         struct nfs_server *srv = NFS_SERVER(ino);
2469         u32 iomode;
2470
2471         if (!lgp)
2472                 return;
2473         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2474         if (lgp->res.status) {
2475                 switch (lgp->res.status) {
2476                 default:
2477                         break;
2478                 /*
2479                  * Halt lgopen attempts if the server doesn't recognise
2480                  * the "current stateid" value, the layout type, or the
2481                  * layoutget operation as being valid.
2482                  * Also if it complains about too many ops in the compound
2483                  * or of the request/reply being too big.
2484                  */
2485                 case -NFS4ERR_BAD_STATEID:
2486                 case -NFS4ERR_NOTSUPP:
2487                 case -NFS4ERR_REP_TOO_BIG:
2488                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2489                 case -NFS4ERR_REQ_TOO_BIG:
2490                 case -NFS4ERR_TOO_MANY_OPS:
2491                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2492                         srv->caps &= ~NFS_CAP_LGOPEN;
2493                 }
2494                 return;
2495         }
2496         if (!lgp->lo) {
2497                 lo = _pnfs_grab_empty_layout(ino, ctx);
2498                 if (!lo)
2499                         return;
2500                 lgp->lo = lo;
2501         } else
2502                 lo = lgp->lo;
2503
2504         lseg = pnfs_layout_process(lgp);
2505         if (!IS_ERR(lseg)) {
2506                 iomode = lgp->args.range.iomode;
2507                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2508                 pnfs_put_lseg(lseg);
2509         }
2510 }
2511
2512 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2513 {
2514         if (lgp != NULL) {
2515                 if (lgp->lo) {
2516                         pnfs_clear_first_layoutget(lgp->lo);
2517                         nfs_layoutget_end(lgp->lo);
2518                 }
2519                 pnfs_layoutget_free(lgp);
2520         }
2521 }
2522
2523 struct pnfs_layout_segment *
2524 pnfs_layout_process(struct nfs4_layoutget *lgp)
2525 {
2526         struct pnfs_layout_hdr *lo = lgp->lo;
2527         struct nfs4_layoutget_res *res = &lgp->res;
2528         struct pnfs_layout_segment *lseg;
2529         struct inode *ino = lo->plh_inode;
2530         LIST_HEAD(free_me);
2531
2532         if (!pnfs_sanity_check_layout_range(&res->range))
2533                 return ERR_PTR(-EINVAL);
2534
2535         /* Inject layout blob into I/O device driver */
2536         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2537         if (IS_ERR_OR_NULL(lseg)) {
2538                 if (!lseg)
2539                         lseg = ERR_PTR(-ENOMEM);
2540
2541                 dprintk("%s: Could not allocate layout: error %ld\n",
2542                        __func__, PTR_ERR(lseg));
2543                 return lseg;
2544         }
2545
2546         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2547
2548         spin_lock(&ino->i_lock);
2549         if (pnfs_layoutgets_blocked(lo)) {
2550                 dprintk("%s forget reply due to state\n", __func__);
2551                 goto out_forget;
2552         }
2553
2554         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2555             !pnfs_is_first_layoutget(lo))
2556                 goto out_forget;
2557
2558         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2559                 /* existing state ID, make sure the sequence number matches. */
2560                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2561                         if (!pnfs_layout_is_valid(lo))
2562                                 lo->plh_barrier = 0;
2563                         dprintk("%s forget reply due to sequence\n", __func__);
2564                         goto out_forget;
2565                 }
2566                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2567         } else if (pnfs_layout_is_valid(lo)) {
2568                 /*
2569                  * We got an entirely new state ID.  Mark all segments for the
2570                  * inode invalid, and retry the layoutget
2571                  */
2572                 struct pnfs_layout_range range = {
2573                         .iomode = IOMODE_ANY,
2574                         .length = NFS4_MAX_UINT64,
2575                 };
2576                 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2577                 goto out_forget;
2578         } else {
2579                 /* We have a completely new layout */
2580                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2581         }
2582
2583         pnfs_get_lseg(lseg);
2584         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2585
2586
2587         if (res->return_on_close)
2588                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2589
2590         spin_unlock(&ino->i_lock);
2591         pnfs_free_lseg_list(&free_me);
2592         return lseg;
2593
2594 out_forget:
2595         spin_unlock(&ino->i_lock);
2596         lseg->pls_layout = lo;
2597         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2598         return ERR_PTR(-EAGAIN);
2599 }
2600
2601 /**
2602  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2603  * @lo: pointer to layout header
2604  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2605  * @return_range: describe layout segment ranges to be returned
2606  * @seq: stateid seqid to match
2607  *
2608  * This function is mainly intended for use by layoutrecall. It attempts
2609  * to free the layout segment immediately, or else to mark it for return
2610  * as soon as its reference count drops to zero.
2611  *
2612  * Returns
2613  * - 0: a layoutreturn needs to be scheduled.
2614  * - EBUSY: there are layout segment that are still in use.
2615  * - ENOENT: there are no layout segments that need to be returned.
2616  */
2617 int
2618 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2619                                 struct list_head *tmp_list,
2620                                 const struct pnfs_layout_range *return_range,
2621                                 u32 seq)
2622 {
2623         struct pnfs_layout_segment *lseg, *next;
2624         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2625         int remaining = 0;
2626
2627         dprintk("%s:Begin lo %p\n", __func__, lo);
2628
2629         assert_spin_locked(&lo->plh_inode->i_lock);
2630
2631         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2632                 tmp_list = &lo->plh_return_segs;
2633
2634         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2635                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2636                         dprintk("%s: marking lseg %p iomode %d "
2637                                 "offset %llu length %llu\n", __func__,
2638                                 lseg, lseg->pls_range.iomode,
2639                                 lseg->pls_range.offset,
2640                                 lseg->pls_range.length);
2641                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2642                                 tmp_list = &lo->plh_return_segs;
2643                         if (mark_lseg_invalid(lseg, tmp_list))
2644                                 continue;
2645                         remaining++;
2646                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2647                         pnfs_lseg_cancel_io(server, lseg);
2648                 }
2649
2650         if (remaining) {
2651                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2652                 return -EBUSY;
2653         }
2654
2655         if (!list_empty(&lo->plh_return_segs)) {
2656                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2657                 return 0;
2658         }
2659
2660         return -ENOENT;
2661 }
2662
2663 static void
2664 pnfs_mark_layout_for_return(struct inode *inode,
2665                             const struct pnfs_layout_range *range)
2666 {
2667         struct pnfs_layout_hdr *lo;
2668         bool return_now = false;
2669
2670         spin_lock(&inode->i_lock);
2671         lo = NFS_I(inode)->layout;
2672         if (!pnfs_layout_is_valid(lo)) {
2673                 spin_unlock(&inode->i_lock);
2674                 return;
2675         }
2676         pnfs_set_plh_return_info(lo, range->iomode, 0);
2677         /*
2678          * mark all matching lsegs so that we are sure to have no live
2679          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2680          * for how it works.
2681          */
2682         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2683                 const struct cred *cred;
2684                 nfs4_stateid stateid;
2685                 enum pnfs_iomode iomode;
2686
2687                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2688                 spin_unlock(&inode->i_lock);
2689                 if (return_now)
2690                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
2691                                                PNFS_FL_LAYOUTRETURN_ASYNC);
2692         } else {
2693                 spin_unlock(&inode->i_lock);
2694                 nfs_commit_inode(inode, 0);
2695         }
2696 }
2697
2698 void pnfs_error_mark_layout_for_return(struct inode *inode,
2699                                        struct pnfs_layout_segment *lseg)
2700 {
2701         struct pnfs_layout_range range = {
2702                 .iomode = lseg->pls_range.iomode,
2703                 .offset = 0,
2704                 .length = NFS4_MAX_UINT64,
2705         };
2706
2707         pnfs_mark_layout_for_return(inode, &range);
2708 }
2709 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2710
2711 static bool
2712 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2713 {
2714         return pnfs_layout_is_valid(lo) &&
2715                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2716                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2717 }
2718
2719 static struct pnfs_layout_segment *
2720 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2721                      const struct pnfs_layout_range *range,
2722                      enum pnfs_iomode iomode)
2723 {
2724         struct pnfs_layout_segment *lseg;
2725
2726         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2727                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2728                         continue;
2729                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2730                         continue;
2731                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2732                         continue;
2733                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2734                         return lseg;
2735         }
2736         return NULL;
2737 }
2738
2739 /* Find open file states whose mode matches that of the range */
2740 static bool
2741 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2742                                  const struct pnfs_layout_range *range)
2743 {
2744         struct list_head *head;
2745         struct nfs_open_context *ctx;
2746         fmode_t mode = 0;
2747
2748         if (!pnfs_layout_can_be_returned(lo) ||
2749             !pnfs_find_first_lseg(lo, range, range->iomode))
2750                 return false;
2751
2752         head = &NFS_I(lo->plh_inode)->open_files;
2753         list_for_each_entry_rcu(ctx, head, list) {
2754                 if (ctx->state)
2755                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2756         }
2757
2758         switch (range->iomode) {
2759         default:
2760                 break;
2761         case IOMODE_READ:
2762                 mode &= ~FMODE_WRITE;
2763                 break;
2764         case IOMODE_RW:
2765                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2766                         mode &= ~FMODE_READ;
2767         }
2768         return mode == 0;
2769 }
2770
2771 static int pnfs_layout_return_unused_byserver(struct nfs_server *server,
2772                                               void *data)
2773 {
2774         const struct pnfs_layout_range *range = data;
2775         const struct cred *cred;
2776         struct pnfs_layout_hdr *lo;
2777         struct inode *inode;
2778         nfs4_stateid stateid;
2779         enum pnfs_iomode iomode;
2780
2781 restart:
2782         rcu_read_lock();
2783         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2784                 inode = lo->plh_inode;
2785                 if (!inode || !pnfs_layout_can_be_returned(lo) ||
2786                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2787                         continue;
2788                 spin_lock(&inode->i_lock);
2789                 if (!lo->plh_inode ||
2790                     !pnfs_should_return_unused_layout(lo, range)) {
2791                         spin_unlock(&inode->i_lock);
2792                         continue;
2793                 }
2794                 pnfs_get_layout_hdr(lo);
2795                 pnfs_set_plh_return_info(lo, range->iomode, 0);
2796                 if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2797                                                     range, 0) != 0 ||
2798                     !pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode)) {
2799                         spin_unlock(&inode->i_lock);
2800                         rcu_read_unlock();
2801                         pnfs_put_layout_hdr(lo);
2802                         cond_resched();
2803                         goto restart;
2804                 }
2805                 spin_unlock(&inode->i_lock);
2806                 rcu_read_unlock();
2807                 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
2808                                        PNFS_FL_LAYOUTRETURN_ASYNC);
2809                 pnfs_put_layout_hdr(lo);
2810                 cond_resched();
2811                 goto restart;
2812         }
2813         rcu_read_unlock();
2814         return 0;
2815 }
2816
2817 void
2818 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2819                                  enum pnfs_iomode iomode)
2820 {
2821         struct pnfs_layout_range range = {
2822                 .iomode = iomode,
2823                 .offset = 0,
2824                 .length = NFS4_MAX_UINT64,
2825         };
2826
2827         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2828                         &range);
2829 }
2830
2831 /* Check if we have we have a valid layout but if there isn't an intersection
2832  * between the request and the pgio->pg_lseg, put this pgio->pg_lseg away.
2833  */
2834 void
2835 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio,
2836                              struct nfs_page *req)
2837 {
2838         if (pgio->pg_lseg == NULL ||
2839             (test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags) &&
2840             pnfs_lseg_request_intersecting(pgio->pg_lseg, req)))
2841                 return;
2842         pnfs_put_lseg(pgio->pg_lseg);
2843         pgio->pg_lseg = NULL;
2844 }
2845 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2846
2847 void
2848 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2849 {
2850         u64 rd_size;
2851
2852         pnfs_generic_pg_check_layout(pgio, req);
2853         if (pgio->pg_lseg == NULL) {
2854                 if (pgio->pg_dreq == NULL)
2855                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2856                 else
2857                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq,
2858                                                       req_offset(req));
2859
2860                 pgio->pg_lseg =
2861                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2862                                            req_offset(req), rd_size,
2863                                            IOMODE_READ, false,
2864                                            nfs_io_gfp_mask());
2865                 if (IS_ERR(pgio->pg_lseg)) {
2866                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2867                         pgio->pg_lseg = NULL;
2868                         return;
2869                 }
2870         }
2871         /* If no lseg, fall back to read through mds */
2872         if (pgio->pg_lseg == NULL)
2873                 nfs_pageio_reset_read_mds(pgio);
2874
2875 }
2876 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2877
2878 void
2879 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2880                            struct nfs_page *req, u64 wb_size)
2881 {
2882         pnfs_generic_pg_check_layout(pgio, req);
2883         if (pgio->pg_lseg == NULL) {
2884                 pgio->pg_lseg =
2885                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2886                                            req_offset(req), wb_size, IOMODE_RW,
2887                                            false, nfs_io_gfp_mask());
2888                 if (IS_ERR(pgio->pg_lseg)) {
2889                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2890                         pgio->pg_lseg = NULL;
2891                         return;
2892                 }
2893         }
2894         /* If no lseg, fall back to write through mds */
2895         if (pgio->pg_lseg == NULL)
2896                 nfs_pageio_reset_write_mds(pgio);
2897 }
2898 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2899
2900 void
2901 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2902 {
2903         if (desc->pg_lseg) {
2904                 pnfs_put_lseg(desc->pg_lseg);
2905                 desc->pg_lseg = NULL;
2906         }
2907 }
2908 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2909
2910 /*
2911  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2912  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2913  */
2914 size_t
2915 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2916                      struct nfs_page *prev, struct nfs_page *req)
2917 {
2918         unsigned int size;
2919         u64 seg_end, req_start, seg_left;
2920
2921         size = nfs_generic_pg_test(pgio, prev, req);
2922         if (!size)
2923                 return 0;
2924
2925         /*
2926          * 'size' contains the number of bytes left in the current page (up
2927          * to the original size asked for in @req->wb_bytes).
2928          *
2929          * Calculate how many bytes are left in the layout segment
2930          * and if there are less bytes than 'size', return that instead.
2931          *
2932          * Please also note that 'end_offset' is actually the offset of the
2933          * first byte that lies outside the pnfs_layout_range. FIXME?
2934          *
2935          */
2936         if (pgio->pg_lseg) {
2937                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2938                                      pgio->pg_lseg->pls_range.length);
2939                 req_start = req_offset(req);
2940
2941                 /* start of request is past the last byte of this segment */
2942                 if (req_start >= seg_end)
2943                         return 0;
2944
2945                 /* adjust 'size' iff there are fewer bytes left in the
2946                  * segment than what nfs_generic_pg_test returned */
2947                 seg_left = seg_end - req_start;
2948                 if (seg_left < size)
2949                         size = (unsigned int)seg_left;
2950         }
2951
2952         return size;
2953 }
2954 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2955
2956 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2957 {
2958         struct nfs_pageio_descriptor pgio;
2959
2960         /* Resend all requests through the MDS */
2961         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2962                               hdr->completion_ops);
2963         return nfs_pageio_resend(&pgio, hdr);
2964 }
2965 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2966
2967 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2968 {
2969
2970         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2971         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2972             PNFS_LAYOUTRET_ON_ERROR) {
2973                 pnfs_return_layout(hdr->inode);
2974         }
2975         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2976                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2977 }
2978
2979 /*
2980  * Called by non rpc-based layout drivers
2981  */
2982 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2983 {
2984         if (likely(!hdr->pnfs_error)) {
2985                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2986                                 hdr->mds_offset + hdr->res.count);
2987                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2988         }
2989         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2990         if (unlikely(hdr->pnfs_error))
2991                 pnfs_ld_handle_write_error(hdr);
2992         hdr->mds_ops->rpc_release(hdr);
2993 }
2994 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2995
2996 static void
2997 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2998                 struct nfs_pgio_header *hdr)
2999 {
3000         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3001
3002         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3003                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3004                 nfs_pageio_reset_write_mds(desc);
3005                 mirror->pg_recoalesce = 1;
3006         }
3007         hdr->completion_ops->completion(hdr);
3008 }
3009
3010 static enum pnfs_try_status
3011 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
3012                         const struct rpc_call_ops *call_ops,
3013                         struct pnfs_layout_segment *lseg,
3014                         int how)
3015 {
3016         struct inode *inode = hdr->inode;
3017         enum pnfs_try_status trypnfs;
3018         struct nfs_server *nfss = NFS_SERVER(inode);
3019
3020         hdr->mds_ops = call_ops;
3021
3022         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
3023                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
3024         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
3025         if (trypnfs != PNFS_NOT_ATTEMPTED)
3026                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
3027         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3028         return trypnfs;
3029 }
3030
3031 static void
3032 pnfs_do_write(struct nfs_pageio_descriptor *desc,
3033               struct nfs_pgio_header *hdr, int how)
3034 {
3035         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3036         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3037         enum pnfs_try_status trypnfs;
3038
3039         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
3040         switch (trypnfs) {
3041         case PNFS_NOT_ATTEMPTED:
3042                 pnfs_write_through_mds(desc, hdr);
3043                 break;
3044         case PNFS_ATTEMPTED:
3045                 break;
3046         case PNFS_TRY_AGAIN:
3047                 /* cleanup hdr and prepare to redo pnfs */
3048                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3049                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3050                         list_splice_init(&hdr->pages, &mirror->pg_list);
3051                         mirror->pg_recoalesce = 1;
3052                 }
3053                 hdr->mds_ops->rpc_release(hdr);
3054         }
3055 }
3056
3057 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
3058 {
3059         pnfs_put_lseg(hdr->lseg);
3060         nfs_pgio_header_free(hdr);
3061 }
3062
3063 int
3064 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
3065 {
3066         struct nfs_pgio_header *hdr;
3067         int ret;
3068
3069         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3070         if (!hdr) {
3071                 desc->pg_error = -ENOMEM;
3072                 return desc->pg_error;
3073         }
3074         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
3075
3076         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3077         ret = nfs_generic_pgio(desc, hdr);
3078         if (!ret)
3079                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
3080
3081         return ret;
3082 }
3083 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
3084
3085 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
3086 {
3087         struct nfs_pageio_descriptor pgio;
3088
3089         /* Resend all requests through the MDS */
3090         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
3091         return nfs_pageio_resend(&pgio, hdr);
3092 }
3093 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
3094
3095 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
3096 {
3097         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
3098         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
3099             PNFS_LAYOUTRET_ON_ERROR) {
3100                 pnfs_return_layout(hdr->inode);
3101         }
3102         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
3103                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
3104 }
3105
3106 /*
3107  * Called by non rpc-based layout drivers
3108  */
3109 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
3110 {
3111         if (likely(!hdr->pnfs_error))
3112                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
3113         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
3114         if (unlikely(hdr->pnfs_error))
3115                 pnfs_ld_handle_read_error(hdr);
3116         hdr->mds_ops->rpc_release(hdr);
3117 }
3118 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
3119
3120 static void
3121 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
3122                 struct nfs_pgio_header *hdr)
3123 {
3124         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3125
3126         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3127                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3128                 nfs_pageio_reset_read_mds(desc);
3129                 mirror->pg_recoalesce = 1;
3130         }
3131         hdr->completion_ops->completion(hdr);
3132 }
3133
3134 /*
3135  * Call the appropriate parallel I/O subsystem read function.
3136  */
3137 static enum pnfs_try_status
3138 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3139                        const struct rpc_call_ops *call_ops,
3140                        struct pnfs_layout_segment *lseg)
3141 {
3142         struct inode *inode = hdr->inode;
3143         struct nfs_server *nfss = NFS_SERVER(inode);
3144         enum pnfs_try_status trypnfs;
3145
3146         hdr->mds_ops = call_ops;
3147
3148         dprintk("%s: Reading ino:%lu %u@%llu\n",
3149                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3150
3151         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3152         if (trypnfs != PNFS_NOT_ATTEMPTED)
3153                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3154         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3155         return trypnfs;
3156 }
3157
3158 /* Resend all requests through pnfs. */
3159 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3160                            unsigned int mirror_idx)
3161 {
3162         struct nfs_pageio_descriptor pgio;
3163
3164         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3165                 /* Prevent deadlocks with layoutreturn! */
3166                 pnfs_put_lseg(hdr->lseg);
3167                 hdr->lseg = NULL;
3168
3169                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3170                                         hdr->completion_ops);
3171                 pgio.pg_mirror_idx = mirror_idx;
3172                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3173         }
3174 }
3175 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3176
3177 static void
3178 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3179 {
3180         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3181         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3182         enum pnfs_try_status trypnfs;
3183
3184         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3185         switch (trypnfs) {
3186         case PNFS_NOT_ATTEMPTED:
3187                 pnfs_read_through_mds(desc, hdr);
3188                 break;
3189         case PNFS_ATTEMPTED:
3190                 break;
3191         case PNFS_TRY_AGAIN:
3192                 /* cleanup hdr and prepare to redo pnfs */
3193                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3194                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3195                         list_splice_init(&hdr->pages, &mirror->pg_list);
3196                         mirror->pg_recoalesce = 1;
3197                 }
3198                 hdr->mds_ops->rpc_release(hdr);
3199         }
3200 }
3201
3202 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3203 {
3204         pnfs_put_lseg(hdr->lseg);
3205         nfs_pgio_header_free(hdr);
3206 }
3207
3208 int
3209 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3210 {
3211         struct nfs_pgio_header *hdr;
3212         int ret;
3213
3214         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3215         if (!hdr) {
3216                 desc->pg_error = -ENOMEM;
3217                 return desc->pg_error;
3218         }
3219         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3220         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3221         ret = nfs_generic_pgio(desc, hdr);
3222         if (!ret)
3223                 pnfs_do_read(desc, hdr);
3224         return ret;
3225 }
3226 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3227
3228 static void pnfs_clear_layoutcommitting(struct inode *inode)
3229 {
3230         unsigned long *bitlock = &NFS_I(inode)->flags;
3231
3232         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3233         smp_mb__after_atomic();
3234         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3235 }
3236
3237 /*
3238  * There can be multiple RW segments.
3239  */
3240 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3241 {
3242         struct pnfs_layout_segment *lseg;
3243
3244         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3245                 if (lseg->pls_range.iomode == IOMODE_RW &&
3246                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3247                         list_add(&lseg->pls_lc_list, listp);
3248         }
3249 }
3250
3251 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3252 {
3253         struct pnfs_layout_segment *lseg, *tmp;
3254
3255         /* Matched by references in pnfs_set_layoutcommit */
3256         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3257                 list_del_init(&lseg->pls_lc_list);
3258                 pnfs_put_lseg(lseg);
3259         }
3260
3261         pnfs_clear_layoutcommitting(inode);
3262 }
3263
3264 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3265 {
3266         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3267 }
3268 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3269
3270 void
3271 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3272                 loff_t end_pos)
3273 {
3274         struct nfs_inode *nfsi = NFS_I(inode);
3275         bool mark_as_dirty = false;
3276
3277         spin_lock(&inode->i_lock);
3278         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3279                 nfsi->layout->plh_lwb = end_pos;
3280                 mark_as_dirty = true;
3281                 dprintk("%s: Set layoutcommit for inode %lu ",
3282                         __func__, inode->i_ino);
3283         } else if (end_pos > nfsi->layout->plh_lwb)
3284                 nfsi->layout->plh_lwb = end_pos;
3285         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3286                 /* references matched in nfs4_layoutcommit_release */
3287                 pnfs_get_lseg(lseg);
3288         }
3289         spin_unlock(&inode->i_lock);
3290         dprintk("%s: lseg %p end_pos %llu\n",
3291                 __func__, lseg, nfsi->layout->plh_lwb);
3292
3293         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3294          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3295         if (mark_as_dirty)
3296                 mark_inode_dirty_sync(inode);
3297 }
3298 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3299
3300 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3301 {
3302         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3303
3304         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3305                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3306         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3307 }
3308
3309 /*
3310  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3311  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3312  * data to disk to allow the server to recover the data if it crashes.
3313  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3314  * is off, and a COMMIT is sent to a data server, or
3315  * if WRITEs to a data server return NFS_DATA_SYNC.
3316  */
3317 int
3318 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3319 {
3320         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3321         struct nfs4_layoutcommit_data *data;
3322         struct nfs_inode *nfsi = NFS_I(inode);
3323         loff_t end_pos;
3324         int status;
3325
3326         if (!pnfs_layoutcommit_outstanding(inode))
3327                 return 0;
3328
3329         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3330
3331         status = -EAGAIN;
3332         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3333                 if (!sync)
3334                         goto out;
3335                 status = wait_on_bit_lock_action(&nfsi->flags,
3336                                 NFS_INO_LAYOUTCOMMITTING,
3337                                 nfs_wait_bit_killable,
3338                                 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3339                 if (status)
3340                         goto out;
3341         }
3342
3343         status = -ENOMEM;
3344         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3345         data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3346         if (!data)
3347                 goto clear_layoutcommitting;
3348
3349         status = 0;
3350         spin_lock(&inode->i_lock);
3351         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3352                 goto out_unlock;
3353
3354         INIT_LIST_HEAD(&data->lseg_list);
3355         pnfs_list_write_lseg(inode, &data->lseg_list);
3356
3357         end_pos = nfsi->layout->plh_lwb;
3358
3359         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3360         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3361         spin_unlock(&inode->i_lock);
3362
3363         data->args.inode = inode;
3364         nfs_fattr_init(&data->fattr);
3365         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3366         data->res.fattr = &data->fattr;
3367         if (end_pos != 0)
3368                 data->args.lastbytewritten = end_pos - 1;
3369         else
3370                 data->args.lastbytewritten = U64_MAX;
3371         data->res.server = NFS_SERVER(inode);
3372
3373         if (ld->prepare_layoutcommit) {
3374                 status = ld->prepare_layoutcommit(&data->args);
3375                 if (status) {
3376                         put_cred(data->cred);
3377                         spin_lock(&inode->i_lock);
3378                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3379                         if (end_pos > nfsi->layout->plh_lwb)
3380                                 nfsi->layout->plh_lwb = end_pos;
3381                         goto out_unlock;
3382                 }
3383         }
3384
3385
3386         status = nfs4_proc_layoutcommit(data, sync);
3387 out:
3388         if (status)
3389                 mark_inode_dirty_sync(inode);
3390         dprintk("<-- %s status %d\n", __func__, status);
3391         return status;
3392 out_unlock:
3393         spin_unlock(&inode->i_lock);
3394         kfree(data);
3395 clear_layoutcommitting:
3396         pnfs_clear_layoutcommitting(inode);
3397         goto out;
3398 }
3399 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3400
3401 int
3402 pnfs_generic_sync(struct inode *inode, bool datasync)
3403 {
3404         return pnfs_layoutcommit_inode(inode, true);
3405 }
3406 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3407
3408 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3409 {
3410         struct nfs4_threshold *thp;
3411
3412         thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3413         if (!thp) {
3414                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3415                 return NULL;
3416         }
3417         return thp;
3418 }
3419
3420 #if IS_ENABLED(CONFIG_NFS_V4_2)
3421 int
3422 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3423 {
3424         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3425         struct nfs_server *server = NFS_SERVER(inode);
3426         struct nfs_inode *nfsi = NFS_I(inode);
3427         struct nfs42_layoutstat_data *data;
3428         struct pnfs_layout_hdr *hdr;
3429         int status = 0;
3430
3431         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3432                 goto out;
3433
3434         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3435                 goto out;
3436
3437         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3438                 goto out;
3439
3440         spin_lock(&inode->i_lock);
3441         if (!NFS_I(inode)->layout) {
3442                 spin_unlock(&inode->i_lock);
3443                 goto out_clear_layoutstats;
3444         }
3445         hdr = NFS_I(inode)->layout;
3446         pnfs_get_layout_hdr(hdr);
3447         spin_unlock(&inode->i_lock);
3448
3449         data = kzalloc(sizeof(*data), gfp_flags);
3450         if (!data) {
3451                 status = -ENOMEM;
3452                 goto out_put;
3453         }
3454
3455         data->args.fh = NFS_FH(inode);
3456         data->args.inode = inode;
3457         status = ld->prepare_layoutstats(&data->args);
3458         if (status)
3459                 goto out_free;
3460
3461         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3462
3463 out:
3464         dprintk("%s returns %d\n", __func__, status);
3465         return status;
3466
3467 out_free:
3468         kfree(data);
3469 out_put:
3470         pnfs_put_layout_hdr(hdr);
3471 out_clear_layoutstats:
3472         smp_mb__before_atomic();
3473         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3474         smp_mb__after_atomic();
3475         goto out;
3476 }
3477 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3478 #endif
3479
3480 unsigned int layoutstats_timer;
3481 module_param(layoutstats_timer, uint, 0644);
3482 EXPORT_SYMBOL_GPL(layoutstats_timer);