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
  65 /* Return the registered pnfs layout driver module matching given id */
  66 static struct pnfs_layoutdriver_type *
  67 find_pnfs_driver_locked(u32 id)
  68 {
  69         struct pnfs_layoutdriver_type *local;
  70
  71         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  72                 if (local->id == id)
  73                         goto out;
  74         local = NULL;
  75 out:
  76         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  77         return local;
  78 }
  79
  80 static struct pnfs_layoutdriver_type *
  81 find_pnfs_driver(u32 id)
  82 {
  83         struct pnfs_layoutdriver_type *local;
  84
  85         spin_lock(&pnfs_spinlock);
  86         local = find_pnfs_driver_locked(id);
  87         if (local != NULL && !try_module_get(local->owner)) {
  88                 dprintk("%s: Could not grab reference on module\n", __func__);
  89                 local = NULL;
  90         }
  91         spin_unlock(&pnfs_spinlock);
  92         return local;
  93 }
  94
  95 const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
  96 {
  97         return find_pnfs_driver(id);
  98 }
  99
 100 void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
 101 {
 102         if (ld)
 103                 module_put(ld->owner);
 104 }
 105
 106 void
 107 unset_pnfs_layoutdriver(struct nfs_server *nfss)
 108 {
 109         if (nfss->pnfs_curr_ld) {
 110                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
 111                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 112                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 113                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 114                         nfs4_deviceid_purge_client(nfss->nfs_client);
 115                 module_put(nfss->pnfs_curr_ld->owner);
 116         }
 117         nfss->pnfs_curr_ld = NULL;
 118 }
 119
 120 /*
 121  * When the server sends a list of layout types, we choose one in the order
 122  * given in the list below.
 123  *
 124  * FIXME: should this list be configurable in some fashion? module param?
 125  *        mount option? something else?
 126  */
 127 static const u32 ld_prefs[] = {
 128         LAYOUT_SCSI,
 129         LAYOUT_BLOCK_VOLUME,
 130         LAYOUT_OSD2_OBJECTS,
 131         LAYOUT_FLEX_FILES,
 132         LAYOUT_NFSV4_1_FILES,
 133         0
 134 };
 135
 136 static int
 137 ld_cmp(const void *e1, const void *e2)
 138 {
 139         u32 ld1 = *((u32 *)e1);
 140         u32 ld2 = *((u32 *)e2);
 141         int i;
 142
 143         for (i = 0; ld_prefs[i] != 0; i++) {
 144                 if (ld1 == ld_prefs[i])
 145                         return -1;
 146
 147                 if (ld2 == ld_prefs[i])
 148                         return 1;
 149         }
 150         return 0;
 151 }
 152
 153 /*
 154  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 155  * Currently only one pNFS layout driver per filesystem is supported.
 156  *
 157  * @ids array of layout types supported by MDS.
 158  */
 159 void
 160 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 161                       struct nfs_fsinfo *fsinfo)
 162 {
 163         struct pnfs_layoutdriver_type *ld_type = NULL;
 164         u32 id;
 165         int i;
 166
 167         if (fsinfo->nlayouttypes == 0)
 168                 goto out_no_driver;
 169         if (!(server->nfs_client->cl_exchange_flags &
 170                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 171                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 172                         __func__, server->nfs_client->cl_exchange_flags);
 173                 goto out_no_driver;
 174         }
 175
 176         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 177                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 178
 179         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 180                 id = fsinfo->layouttype[i];
 181                 ld_type = find_pnfs_driver(id);
 182                 if (!ld_type) {
 183                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 184                                         id);
 185                         ld_type = find_pnfs_driver(id);
 186                 }
 187                 if (ld_type)
 188                         break;
 189         }
 190
 191         if (!ld_type) {
 192                 dprintk("%s: No pNFS module found!\n", __func__);
 193                 goto out_no_driver;
 194         }
 195
 196         server->pnfs_curr_ld = ld_type;
 197         if (ld_type->set_layoutdriver
 198             && ld_type->set_layoutdriver(server, mntfh)) {
 199                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 200                         "driver %u.\n", __func__, id);
 201                 module_put(ld_type->owner);
 202                 goto out_no_driver;
 203         }
 204         /* Bump the MDS count */
 205         atomic_inc(&server->nfs_client->cl_mds_count);
 206
 207         dprintk("%s: pNFS module for %u set\n", __func__, id);
 208         return;
 209
 210 out_no_driver:
 211         dprintk("%s: Using NFSv4 I/O\n", __func__);
 212         server->pnfs_curr_ld = NULL;
 213 }
 214
 215 int
 216 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 217 {
 218         int status = -EINVAL;
 219         struct pnfs_layoutdriver_type *tmp;
 220
 221         if (ld_type->id == 0) {
 222                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 223                 return status;
 224         }
 225         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 226                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 227                        "alloc_lseg and free_lseg.\n", __func__);
 228                 return status;
 229         }
 230
 231         spin_lock(&pnfs_spinlock);
 232         tmp = find_pnfs_driver_locked(ld_type->id);
 233         if (!tmp) {
 234                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 235                 status = 0;
 236                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 237                         ld_type->name);
 238         } else {
 239                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 240                         __func__, ld_type->id);
 241         }
 242         spin_unlock(&pnfs_spinlock);
 243
 244         return status;
 245 }
 246 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 247
 248 void
 249 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 250 {
 251         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 252         spin_lock(&pnfs_spinlock);
 253         list_del(&ld_type->pnfs_tblid);
 254         spin_unlock(&pnfs_spinlock);
 255 }
 256 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 257
 258 /*
 259  * pNFS client layout cache
 260  */
 261
 262 /* Need to hold i_lock if caller does not already hold reference */
 263 void
 264 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 265 {
 266         refcount_inc(&lo->plh_refcount);
 267 }
 268
 269 static struct pnfs_layout_hdr *
 270 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 271 {
 272         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 273         return ld->alloc_layout_hdr(ino, gfp_flags);
 274 }
 275
 276 static void
 277 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 278 {
 279         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 280         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 281
 282         if (!list_empty(&lo->plh_layouts)) {
 283                 struct nfs_client *clp = server->nfs_client;
 284
 285                 spin_lock(&clp->cl_lock);
 286                 list_del_init(&lo->plh_layouts);
 287                 spin_unlock(&clp->cl_lock);
 288         }
 289         put_rpccred(lo->plh_lc_cred);
 290         return ld->free_layout_hdr(lo);
 291 }
 292
 293 static void
 294 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 295 {
 296         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 297         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 298         nfsi->layout = NULL;
 299         /* Reset MDS Threshold I/O counters */
 300         nfsi->write_io = 0;
 301         nfsi->read_io = 0;
 302 }
 303
 304 void
 305 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 306 {
 307         struct inode *inode;
 308         unsigned long i_state;
 309
 310         if (!lo)
 311                 return;
 312         inode = lo->plh_inode;
 313         pnfs_layoutreturn_before_put_layout_hdr(lo);
 314
 315         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 316                 if (!list_empty(&lo->plh_segs))
 317                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 318                 pnfs_detach_layout_hdr(lo);
 319                 i_state = inode->i_state;
 320                 spin_unlock(&inode->i_lock);
 321                 pnfs_free_layout_hdr(lo);
 322                 /* Notify pnfs_destroy_layout_final() that we're done */
 323                 if (i_state & (I_FREEING | I_CLEAR))
 324                         wake_up_var(lo);
 325         }
 326 }
 327
 328 static void
 329 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 330                          u32 seq)
 331 {
 332         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 333                 iomode = IOMODE_ANY;
 334         lo->plh_return_iomode = iomode;
 335         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 336         if (seq != 0) {
 337                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 338                 lo->plh_return_seq = seq;
 339         }
 340 }
 341
 342 static void
 343 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 344 {
 345         struct pnfs_layout_segment *lseg;
 346         lo->plh_return_iomode = 0;
 347         lo->plh_return_seq = 0;
 348         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 349         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 350                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 351                         continue;
 352                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 353         }
 354 }
 355
 356 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 357 {
 358         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 359         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 360         smp_mb__after_atomic();
 361         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 362         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 363 }
 364
 365 static void
 366 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 367                 struct list_head *free_me)
 368 {
 369         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 370         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 371         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 372                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 373         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 374                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 375 }
 376
 377 /*
 378  * Update the seqid of a layout stateid
 379  */
 380 bool nfs4_layoutreturn_refresh_stateid(nfs4_stateid *dst,
 381                 struct pnfs_layout_range *dst_range,
 382                 struct inode *inode)
 383 {
 384         struct pnfs_layout_hdr *lo;
 385         struct pnfs_layout_range range = {
 386                 .iomode = IOMODE_ANY,
 387                 .offset = 0,
 388                 .length = NFS4_MAX_UINT64,
 389         };
 390         bool ret = false;
 391         LIST_HEAD(head);
 392         int err;
 393
 394         spin_lock(&inode->i_lock);
 395         lo = NFS_I(inode)->layout;
 396         if (lo && nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 397                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 398                 if (err != -EBUSY) {
 399                         dst->seqid = lo->plh_stateid.seqid;
 400                         *dst_range = range;
 401                         ret = true;
 402                 }
 403         }
 404         spin_unlock(&inode->i_lock);
 405         pnfs_free_lseg_list(&head);
 406         return ret;
 407 }
 408
 409 /*
 410  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 411  *
 412  * In order to continue using the pnfs_layout_hdr, a full recovery
 413  * is required.
 414  * Note that caller must hold inode->i_lock.
 415  */
 416 int
 417 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 418                 struct list_head *lseg_list)
 419 {
 420         struct pnfs_layout_range range = {
 421                 .iomode = IOMODE_ANY,
 422                 .offset = 0,
 423                 .length = NFS4_MAX_UINT64,
 424         };
 425         struct pnfs_layout_segment *lseg, *next;
 426
 427         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 428         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 429                 pnfs_clear_lseg_state(lseg, lseg_list);
 430         pnfs_clear_layoutreturn_info(lo);
 431         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 432         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 433             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 434                 pnfs_clear_layoutreturn_waitbit(lo);
 435         return !list_empty(&lo->plh_segs);
 436 }
 437
 438 static int
 439 pnfs_iomode_to_fail_bit(u32 iomode)
 440 {
 441         return iomode == IOMODE_RW ?
 442                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 443 }
 444
 445 static void
 446 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 447 {
 448         lo->plh_retry_timestamp = jiffies;
 449         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 450                 refcount_inc(&lo->plh_refcount);
 451 }
 452
 453 static void
 454 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 455 {
 456         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 457                 refcount_dec(&lo->plh_refcount);
 458 }
 459
 460 static void
 461 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 462 {
 463         struct inode *inode = lo->plh_inode;
 464         struct pnfs_layout_range range = {
 465                 .iomode = iomode,
 466                 .offset = 0,
 467                 .length = NFS4_MAX_UINT64,
 468         };
 469         LIST_HEAD(head);
 470
 471         spin_lock(&inode->i_lock);
 472         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 473         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 474         spin_unlock(&inode->i_lock);
 475         pnfs_free_lseg_list(&head);
 476         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 477                         iomode == IOMODE_RW ?  "RW" : "READ");
 478 }
 479
 480 static bool
 481 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 482 {
 483         unsigned long start, end;
 484         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 485
 486         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 487                 return false;
 488         end = jiffies;
 489         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 490         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 491                 /* It is time to retry the failed layoutgets */
 492                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 493                 return false;
 494         }
 495         return true;
 496 }
 497
 498 static void
 499 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 500                 const struct pnfs_layout_range *range,
 501                 const nfs4_stateid *stateid)
 502 {
 503         INIT_LIST_HEAD(&lseg->pls_list);
 504         INIT_LIST_HEAD(&lseg->pls_lc_list);
 505         refcount_set(&lseg->pls_refcount, 1);
 506         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 507         lseg->pls_layout = lo;
 508         lseg->pls_range = *range;
 509         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 510 }
 511
 512 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 513 {
 514         if (lseg != NULL) {
 515                 struct inode *inode = lseg->pls_layout->plh_inode;
 516                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 517         }
 518 }
 519
 520 static void
 521 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 522                 struct pnfs_layout_segment *lseg)
 523 {
 524         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 525         list_del_init(&lseg->pls_list);
 526         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 527         refcount_dec(&lo->plh_refcount);
 528         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 529                 return;
 530         if (list_empty(&lo->plh_segs) &&
 531             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 532             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 533                 if (atomic_read(&lo->plh_outstanding) == 0)
 534                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 535                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 536         }
 537 }
 538
 539 static bool
 540 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 541                 struct pnfs_layout_segment *lseg)
 542 {
 543         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 544             pnfs_layout_is_valid(lo)) {
 545                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 546                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 547                 return true;
 548         }
 549         return false;
 550 }
 551
 552 void
 553 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 554 {
 555         struct pnfs_layout_hdr *lo;
 556         struct inode *inode;
 557
 558         if (!lseg)
 559                 return;
 560
 561         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 562                 refcount_read(&lseg->pls_refcount),
 563                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 564
 565         lo = lseg->pls_layout;
 566         inode = lo->plh_inode;
 567
 568         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 569                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 570                         spin_unlock(&inode->i_lock);
 571                         return;
 572                 }
 573                 pnfs_get_layout_hdr(lo);
 574                 pnfs_layout_remove_lseg(lo, lseg);
 575                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 576                         lseg = NULL;
 577                 spin_unlock(&inode->i_lock);
 578                 pnfs_free_lseg(lseg);
 579                 pnfs_put_layout_hdr(lo);
 580         }
 581 }
 582 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 583
 584 /*
 585  * is l2 fully contained in l1?
 586  *   start1                             end1
 587  *   [----------------------------------)
 588  *           start2           end2
 589  *           [----------------)
 590  */
 591 static bool
 592 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 593                  const struct pnfs_layout_range *l2)
 594 {
 595         u64 start1 = l1->offset;
 596         u64 end1 = pnfs_end_offset(start1, l1->length);
 597         u64 start2 = l2->offset;
 598         u64 end2 = pnfs_end_offset(start2, l2->length);
 599
 600         return (start1 <= start2) && (end1 >= end2);
 601 }
 602
 603 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 604                 struct list_head *tmp_list)
 605 {
 606         if (!refcount_dec_and_test(&lseg->pls_refcount))
 607                 return false;
 608         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 609         list_add(&lseg->pls_list, tmp_list);
 610         return true;
 611 }
 612
 613 /* Returns 1 if lseg is removed from list, 0 otherwise */
 614 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 615                              struct list_head *tmp_list)
 616 {
 617         int rv = 0;
 618
 619         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 620                 /* Remove the reference keeping the lseg in the
 621                  * list.  It will now be removed when all
 622                  * outstanding io is finished.
 623                  */
 624                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 625                         refcount_read(&lseg->pls_refcount));
 626                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 627                         rv = 1;
 628         }
 629         return rv;
 630 }
 631
 632 /*
 633  * Compare 2 layout stateid sequence ids, to see which is newer,
 634  * taking into account wraparound issues.
 635  */
 636 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 637 {
 638         return (s32)(s1 - s2) > 0;
 639 }
 640
 641 static bool
 642 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 643                  const struct pnfs_layout_range *recall_range)
 644 {
 645         return (recall_range->iomode == IOMODE_ANY ||
 646                 lseg_range->iomode == recall_range->iomode) &&
 647                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 648 }
 649
 650 static bool
 651 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 652                 const struct pnfs_layout_range *recall_range,
 653                 u32 seq)
 654 {
 655         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 656                 return false;
 657         if (recall_range == NULL)
 658                 return true;
 659         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 660 }
 661
 662 /**
 663  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 664  * @lo: layout header containing the lsegs
 665  * @tmp_list: list head where doomed lsegs should go
 666  * @recall_range: optional recall range argument to match (may be NULL)
 667  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 668  *
 669  * Walk the list of lsegs in the layout header, and tear down any that should
 670  * be destroyed. If "recall_range" is specified then the segment must match
 671  * that range. If "seq" is non-zero, then only match segments that were handed
 672  * out at or before that sequence.
 673  *
 674  * Returns number of matching invalid lsegs remaining in list after scanning
 675  * it and purging them.
 676  */
 677 int
 678 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 679                             struct list_head *tmp_list,
 680                             const struct pnfs_layout_range *recall_range,
 681                             u32 seq)
 682 {
 683         struct pnfs_layout_segment *lseg, *next;
 684         int remaining = 0;
 685
 686         dprintk("%s:Begin lo %p\n", __func__, lo);
 687
 688         if (list_empty(&lo->plh_segs))
 689                 return 0;
 690         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 691                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 692                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 693                                 "offset %llu length %llu\n", __func__,
 694                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 695                                 lseg->pls_range.offset, lseg->pls_range.length);
 696                         if (!mark_lseg_invalid(lseg, tmp_list))
 697                                 remaining++;
 698                 }
 699         dprintk("%s:Return %i\n", __func__, remaining);
 700         return remaining;
 701 }
 702
 703 static void
 704 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 705                 struct list_head *free_me,
 706                 const struct pnfs_layout_range *range,
 707                 u32 seq)
 708 {
 709         struct pnfs_layout_segment *lseg, *next;
 710
 711         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 712                 if (pnfs_match_lseg_recall(lseg, range, seq))
 713                         list_move_tail(&lseg->pls_list, free_me);
 714         }
 715 }
 716
 717 /* note free_me must contain lsegs from a single layout_hdr */
 718 void
 719 pnfs_free_lseg_list(struct list_head *free_me)
 720 {
 721         struct pnfs_layout_segment *lseg, *tmp;
 722
 723         if (list_empty(free_me))
 724                 return;
 725
 726         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 727                 list_del(&lseg->pls_list);
 728                 pnfs_free_lseg(lseg);
 729         }
 730 }
 731
 732 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 733 {
 734         struct pnfs_layout_hdr *lo;
 735         LIST_HEAD(tmp_list);
 736
 737         spin_lock(&nfsi->vfs_inode.i_lock);
 738         lo = nfsi->layout;
 739         if (lo) {
 740                 pnfs_get_layout_hdr(lo);
 741                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 742                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 743                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 744                 spin_unlock(&nfsi->vfs_inode.i_lock);
 745                 pnfs_free_lseg_list(&tmp_list);
 746                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 747                 pnfs_put_layout_hdr(lo);
 748         } else
 749                 spin_unlock(&nfsi->vfs_inode.i_lock);
 750         return lo;
 751 }
 752
 753 void pnfs_destroy_layout(struct nfs_inode *nfsi)
 754 {
 755         __pnfs_destroy_layout(nfsi);
 756 }
 757 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 758
 759 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
 760                                 struct pnfs_layout_hdr *lo)
 761 {
 762         bool ret;
 763
 764         spin_lock(&nfsi->vfs_inode.i_lock);
 765         ret = nfsi->layout != lo;
 766         spin_unlock(&nfsi->vfs_inode.i_lock);
 767         return ret;
 768 }
 769
 770 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 771 {
 772         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 773
 774         if (lo)
 775                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
 776 }
 777
 778 static bool
 779 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 780                 struct list_head *layout_list)
 781 {
 782         struct pnfs_layout_hdr *lo;
 783         bool ret = false;
 784
 785         spin_lock(&inode->i_lock);
 786         lo = NFS_I(inode)->layout;
 787         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 788                 pnfs_get_layout_hdr(lo);
 789                 list_add(&lo->plh_bulk_destroy, layout_list);
 790                 ret = true;
 791         }
 792         spin_unlock(&inode->i_lock);
 793         return ret;
 794 }
 795
 796 /* Caller must hold rcu_read_lock and clp->cl_lock */
 797 static int
 798 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 799                 struct nfs_server *server,
 800                 struct list_head *layout_list)
 801         __must_hold(&clp->cl_lock)
 802         __must_hold(RCU)
 803 {
 804         struct pnfs_layout_hdr *lo, *next;
 805         struct inode *inode;
 806
 807         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 808                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 809                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 810                     !list_empty(&lo->plh_bulk_destroy))
 811                         continue;
 812                 /* If the sb is being destroyed, just bail */
 813                 if (!nfs_sb_active(server->super))
 814                         break;
 815                 inode = igrab(lo->plh_inode);
 816                 if (inode != NULL) {
 817                         list_del_init(&lo->plh_layouts);
 818                         if (pnfs_layout_add_bulk_destroy_list(inode,
 819                                                 layout_list))
 820                                 continue;
 821                         rcu_read_unlock();
 822                         spin_unlock(&clp->cl_lock);
 823                         iput(inode);
 824                 } else {
 825                         rcu_read_unlock();
 826                         spin_unlock(&clp->cl_lock);
 827                         set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 828                 }
 829                 nfs_sb_deactive(server->super);
 830                 spin_lock(&clp->cl_lock);
 831                 rcu_read_lock();
 832                 return -EAGAIN;
 833         }
 834         return 0;
 835 }
 836
 837 static int
 838 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 839                 bool is_bulk_recall)
 840 {
 841         struct pnfs_layout_hdr *lo;
 842         struct inode *inode;
 843         LIST_HEAD(lseg_list);
 844         int ret = 0;
 845
 846         while (!list_empty(layout_list)) {
 847                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 848                                 plh_bulk_destroy);
 849                 dprintk("%s freeing layout for inode %lu\n", __func__,
 850                         lo->plh_inode->i_ino);
 851                 inode = lo->plh_inode;
 852
 853                 pnfs_layoutcommit_inode(inode, false);
 854
 855                 spin_lock(&inode->i_lock);
 856                 list_del_init(&lo->plh_bulk_destroy);
 857                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 858                         if (is_bulk_recall)
 859                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 860                         ret = -EAGAIN;
 861                 }
 862                 spin_unlock(&inode->i_lock);
 863                 pnfs_free_lseg_list(&lseg_list);
 864                 /* Free all lsegs that are attached to commit buckets */
 865                 nfs_commit_inode(inode, 0);
 866                 pnfs_put_layout_hdr(lo);
 867                 nfs_iput_and_deactive(inode);
 868         }
 869         return ret;
 870 }
 871
 872 int
 873 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 874                 struct nfs_fsid *fsid,
 875                 bool is_recall)
 876 {
 877         struct nfs_server *server;
 878         LIST_HEAD(layout_list);
 879
 880         spin_lock(&clp->cl_lock);
 881         rcu_read_lock();
 882 restart:
 883         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 884                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 885                         continue;
 886                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 887                                 server,
 888                                 &layout_list) != 0)
 889                         goto restart;
 890         }
 891         rcu_read_unlock();
 892         spin_unlock(&clp->cl_lock);
 893
 894         if (list_empty(&layout_list))
 895                 return 0;
 896         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 897 }
 898
 899 int
 900 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 901                 bool is_recall)
 902 {
 903         struct nfs_server *server;
 904         LIST_HEAD(layout_list);
 905
 906         spin_lock(&clp->cl_lock);
 907         rcu_read_lock();
 908 restart:
 909         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 910                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 911                                         server,
 912                                         &layout_list) != 0)
 913                         goto restart;
 914         }
 915         rcu_read_unlock();
 916         spin_unlock(&clp->cl_lock);
 917
 918         if (list_empty(&layout_list))
 919                 return 0;
 920         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 921 }
 922
 923 /*
 924  * Called by the state manger to remove all layouts established under an
 925  * expired lease.
 926  */
 927 void
 928 pnfs_destroy_all_layouts(struct nfs_client *clp)
 929 {
 930         nfs4_deviceid_mark_client_invalid(clp);
 931         nfs4_deviceid_purge_client(clp);
 932
 933         pnfs_destroy_layouts_byclid(clp, false);
 934 }
 935
 936 /* update lo->plh_stateid with new if is more recent */
 937 void
 938 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 939                         bool update_barrier)
 940 {
 941         u32 oldseq, newseq, new_barrier = 0;
 942
 943         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 944         newseq = be32_to_cpu(new->seqid);
 945
 946         if (!pnfs_layout_is_valid(lo)) {
 947                 nfs4_stateid_copy(&lo->plh_stateid, new);
 948                 lo->plh_barrier = newseq;
 949                 pnfs_clear_layoutreturn_info(lo);
 950                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 951                 return;
 952         }
 953         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 954                 nfs4_stateid_copy(&lo->plh_stateid, new);
 955                 /*
 956                  * Because of wraparound, we want to keep the barrier
 957                  * "close" to the current seqids.
 958                  */
 959                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 960         }
 961         if (update_barrier)
 962                 new_barrier = be32_to_cpu(new->seqid);
 963         else if (new_barrier == 0)
 964                 return;
 965         if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 966                 lo->plh_barrier = new_barrier;
 967 }
 968
 969 static bool
 970 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 971                 const nfs4_stateid *stateid)
 972 {
 973         u32 seqid = be32_to_cpu(stateid->seqid);
 974
 975         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 976 }
 977
 978 /* lget is set to 1 if called from inside send_layoutget call chain */
 979 static bool
 980 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 981 {
 982         return lo->plh_block_lgets ||
 983                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 984 }
 985
 986 static struct nfs_server *
 987 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
 988 {
 989         struct nfs_server *server;
 990
 991         if (inode) {
 992                 server = NFS_SERVER(inode);
 993         } else {
 994                 struct dentry *parent_dir = dget_parent(ctx->dentry);
 995                 server = NFS_SERVER(parent_dir->d_inode);
 996                 dput(parent_dir);
 997         }
 998         return server;
 999 }
1000
1001 static void nfs4_free_pages(struct page **pages, size_t size)
1002 {
1003         int i;
1004
1005         if (!pages)
1006                 return;
1007
1008         for (i = 0; i < size; i++) {
1009                 if (!pages[i])
1010                         break;
1011                 __free_page(pages[i]);
1012         }
1013         kfree(pages);
1014 }
1015
1016 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1017 {
1018         struct page **pages;
1019         int i;
1020
1021         pages = kcalloc(size, sizeof(struct page *), gfp_flags);
1022         if (!pages) {
1023                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1024                 return NULL;
1025         }
1026
1027         for (i = 0; i < size; i++) {
1028                 pages[i] = alloc_page(gfp_flags);
1029                 if (!pages[i]) {
1030                         dprintk("%s: failed to allocate page\n", __func__);
1031                         nfs4_free_pages(pages, size);
1032                         return NULL;
1033                 }
1034         }
1035
1036         return pages;
1037 }
1038
1039 static struct nfs4_layoutget *
1040 pnfs_alloc_init_layoutget_args(struct inode *ino,
1041            struct nfs_open_context *ctx,
1042            const nfs4_stateid *stateid,
1043            const struct pnfs_layout_range *range,
1044            gfp_t gfp_flags)
1045 {
1046         struct nfs_server *server = pnfs_find_server(ino, ctx);
1047         size_t max_pages = max_response_pages(server);
1048         struct nfs4_layoutget *lgp;
1049
1050         dprintk("--> %s\n", __func__);
1051
1052         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1053         if (lgp == NULL)
1054                 return NULL;
1055
1056         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1057         if (!lgp->args.layout.pages) {
1058                 kfree(lgp);
1059                 return NULL;
1060         }
1061         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1062         lgp->res.layoutp = &lgp->args.layout;
1063
1064         /* Don't confuse uninitialised result and success */
1065         lgp->res.status = -NFS4ERR_DELAY;
1066
1067         lgp->args.minlength = PAGE_SIZE;
1068         if (lgp->args.minlength > range->length)
1069                 lgp->args.minlength = range->length;
1070         if (ino) {
1071                 loff_t i_size = i_size_read(ino);
1072
1073                 if (range->iomode == IOMODE_READ) {
1074                         if (range->offset >= i_size)
1075                                 lgp->args.minlength = 0;
1076                         else if (i_size - range->offset < lgp->args.minlength)
1077                                 lgp->args.minlength = i_size - range->offset;
1078                 }
1079         }
1080         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1081         pnfs_copy_range(&lgp->args.range, range);
1082         lgp->args.type = server->pnfs_curr_ld->id;
1083         lgp->args.inode = ino;
1084         lgp->args.ctx = get_nfs_open_context(ctx);
1085         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1086         lgp->gfp_flags = gfp_flags;
1087         lgp->cred = get_rpccred(ctx->cred);
1088         return lgp;
1089 }
1090
1091 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1092 {
1093         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1094
1095         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1096         if (lgp->args.inode)
1097                 pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1098         put_rpccred(lgp->cred);
1099         put_nfs_open_context(lgp->args.ctx);
1100         kfree(lgp);
1101 }
1102
1103 static void pnfs_clear_layoutcommit(struct inode *inode,
1104                 struct list_head *head)
1105 {
1106         struct nfs_inode *nfsi = NFS_I(inode);
1107         struct pnfs_layout_segment *lseg, *tmp;
1108
1109         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1110                 return;
1111         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1112                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1113                         continue;
1114                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1115         }
1116 }
1117
1118 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1119                 const nfs4_stateid *arg_stateid,
1120                 const struct pnfs_layout_range *range,
1121                 const nfs4_stateid *stateid)
1122 {
1123         struct inode *inode = lo->plh_inode;
1124         LIST_HEAD(freeme);
1125
1126         spin_lock(&inode->i_lock);
1127         if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1128             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1129                 goto out_unlock;
1130         if (stateid) {
1131                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1132
1133                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1134                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1135                 pnfs_set_layout_stateid(lo, stateid, true);
1136         } else
1137                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1138 out_unlock:
1139         pnfs_clear_layoutreturn_waitbit(lo);
1140         spin_unlock(&inode->i_lock);
1141         pnfs_free_lseg_list(&freeme);
1142
1143 }
1144
1145 static bool
1146 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1147                 nfs4_stateid *stateid,
1148                 enum pnfs_iomode *iomode)
1149 {
1150         /* Serialise LAYOUTGET/LAYOUTRETURN */
1151         if (atomic_read(&lo->plh_outstanding) != 0)
1152                 return false;
1153         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1154                 return false;
1155         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1156         pnfs_get_layout_hdr(lo);
1157         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1158                 if (stateid != NULL) {
1159                         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1160                         if (lo->plh_return_seq != 0)
1161                                 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1162                 }
1163                 if (iomode != NULL)
1164                         *iomode = lo->plh_return_iomode;
1165                 pnfs_clear_layoutreturn_info(lo);
1166                 return true;
1167         }
1168         if (stateid != NULL)
1169                 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1170         if (iomode != NULL)
1171                 *iomode = IOMODE_ANY;
1172         return true;
1173 }
1174
1175 static void
1176 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1177                 struct pnfs_layout_hdr *lo,
1178                 const nfs4_stateid *stateid,
1179                 enum pnfs_iomode iomode)
1180 {
1181         struct inode *inode = lo->plh_inode;
1182
1183         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1184         args->inode = inode;
1185         args->range.iomode = iomode;
1186         args->range.offset = 0;
1187         args->range.length = NFS4_MAX_UINT64;
1188         args->layout = lo;
1189         nfs4_stateid_copy(&args->stateid, stateid);
1190 }
1191
1192 static int
1193 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1194                        enum pnfs_iomode iomode, bool sync)
1195 {
1196         struct inode *ino = lo->plh_inode;
1197         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1198         struct nfs4_layoutreturn *lrp;
1199         int status = 0;
1200
1201         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1202         if (unlikely(lrp == NULL)) {
1203                 status = -ENOMEM;
1204                 spin_lock(&ino->i_lock);
1205                 pnfs_clear_layoutreturn_waitbit(lo);
1206                 spin_unlock(&ino->i_lock);
1207                 pnfs_put_layout_hdr(lo);
1208                 goto out;
1209         }
1210
1211         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1212         lrp->args.ld_private = &lrp->ld_private;
1213         lrp->clp = NFS_SERVER(ino)->nfs_client;
1214         lrp->cred = lo->plh_lc_cred;
1215         if (ld->prepare_layoutreturn)
1216                 ld->prepare_layoutreturn(&lrp->args);
1217
1218         status = nfs4_proc_layoutreturn(lrp, sync);
1219 out:
1220         dprintk("<-- %s status: %d\n", __func__, status);
1221         return status;
1222 }
1223
1224 static bool
1225 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1226                                 enum pnfs_iomode iomode,
1227                                 u32 seq)
1228 {
1229         struct pnfs_layout_range recall_range = {
1230                 .length = NFS4_MAX_UINT64,
1231                 .iomode = iomode,
1232         };
1233         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1234                                                &recall_range, seq) != -EBUSY;
1235 }
1236
1237 /* Return true if layoutreturn is needed */
1238 static bool
1239 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1240 {
1241         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1242                 return false;
1243         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1244                                                lo->plh_return_seq);
1245 }
1246
1247 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1248 {
1249         struct inode *inode= lo->plh_inode;
1250
1251         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1252                 return;
1253         spin_lock(&inode->i_lock);
1254         if (pnfs_layout_need_return(lo)) {
1255                 nfs4_stateid stateid;
1256                 enum pnfs_iomode iomode;
1257                 bool send;
1258
1259                 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1260                 spin_unlock(&inode->i_lock);
1261                 if (send) {
1262                         /* Send an async layoutreturn so we dont deadlock */
1263                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1264                 }
1265         } else
1266                 spin_unlock(&inode->i_lock);
1267 }
1268
1269 /*
1270  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1271  * when the layout segment list is empty.
1272  *
1273  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1274  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1275  * deviceid is marked invalid.
1276  */
1277 int
1278 _pnfs_return_layout(struct inode *ino)
1279 {
1280         struct pnfs_layout_hdr *lo = NULL;
1281         struct nfs_inode *nfsi = NFS_I(ino);
1282         struct pnfs_layout_range range = {
1283                 .iomode         = IOMODE_ANY,
1284                 .offset         = 0,
1285                 .length         = NFS4_MAX_UINT64,
1286         };
1287         LIST_HEAD(tmp_list);
1288         nfs4_stateid stateid;
1289         int status = 0;
1290         bool send, valid_layout;
1291
1292         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1293
1294         spin_lock(&ino->i_lock);
1295         lo = nfsi->layout;
1296         if (!lo) {
1297                 spin_unlock(&ino->i_lock);
1298                 dprintk("NFS: %s no layout to return\n", __func__);
1299                 goto out;
1300         }
1301         /* Reference matched in nfs4_layoutreturn_release */
1302         pnfs_get_layout_hdr(lo);
1303         /* Is there an outstanding layoutreturn ? */
1304         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1305                 spin_unlock(&ino->i_lock);
1306                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1307                                         TASK_UNINTERRUPTIBLE))
1308                         goto out_put_layout_hdr;
1309                 spin_lock(&ino->i_lock);
1310         }
1311         valid_layout = pnfs_layout_is_valid(lo);
1312         pnfs_clear_layoutcommit(ino, &tmp_list);
1313         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1314
1315         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1316                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1317
1318         /* Don't send a LAYOUTRETURN if list was initially empty */
1319         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1320                         !valid_layout) {
1321                 spin_unlock(&ino->i_lock);
1322                 dprintk("NFS: %s no layout segments to return\n", __func__);
1323                 goto out_put_layout_hdr;
1324         }
1325
1326         send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1327         spin_unlock(&ino->i_lock);
1328         if (send)
1329                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1330 out_put_layout_hdr:
1331         pnfs_free_lseg_list(&tmp_list);
1332         pnfs_put_layout_hdr(lo);
1333 out:
1334         dprintk("<-- %s status: %d\n", __func__, status);
1335         return status;
1336 }
1337
1338 int
1339 pnfs_commit_and_return_layout(struct inode *inode)
1340 {
1341         struct pnfs_layout_hdr *lo;
1342         int ret;
1343
1344         spin_lock(&inode->i_lock);
1345         lo = NFS_I(inode)->layout;
1346         if (lo == NULL) {
1347                 spin_unlock(&inode->i_lock);
1348                 return 0;
1349         }
1350         pnfs_get_layout_hdr(lo);
1351         /* Block new layoutgets and read/write to ds */
1352         lo->plh_block_lgets++;
1353         spin_unlock(&inode->i_lock);
1354         filemap_fdatawait(inode->i_mapping);
1355         ret = pnfs_layoutcommit_inode(inode, true);
1356         if (ret == 0)
1357                 ret = _pnfs_return_layout(inode);
1358         spin_lock(&inode->i_lock);
1359         lo->plh_block_lgets--;
1360         spin_unlock(&inode->i_lock);
1361         pnfs_put_layout_hdr(lo);
1362         return ret;
1363 }
1364
1365 bool pnfs_roc(struct inode *ino,
1366                 struct nfs4_layoutreturn_args *args,
1367                 struct nfs4_layoutreturn_res *res,
1368                 const struct rpc_cred *cred)
1369 {
1370         struct nfs_inode *nfsi = NFS_I(ino);
1371         struct nfs_open_context *ctx;
1372         struct nfs4_state *state;
1373         struct pnfs_layout_hdr *lo;
1374         struct pnfs_layout_segment *lseg, *next;
1375         nfs4_stateid stateid;
1376         enum pnfs_iomode iomode = 0;
1377         bool layoutreturn = false, roc = false;
1378         bool skip_read = false;
1379
1380         if (!nfs_have_layout(ino))
1381                 return false;
1382 retry:
1383         spin_lock(&ino->i_lock);
1384         lo = nfsi->layout;
1385         if (!lo || !pnfs_layout_is_valid(lo) ||
1386             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1387                 lo = NULL;
1388                 goto out_noroc;
1389         }
1390         pnfs_get_layout_hdr(lo);
1391         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1392                 spin_unlock(&ino->i_lock);
1393                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1394                                 TASK_UNINTERRUPTIBLE);
1395                 pnfs_put_layout_hdr(lo);
1396                 goto retry;
1397         }
1398
1399         /* no roc if we hold a delegation */
1400         if (nfs4_check_delegation(ino, FMODE_READ)) {
1401                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1402                         goto out_noroc;
1403                 skip_read = true;
1404         }
1405
1406         list_for_each_entry(ctx, &nfsi->open_files, list) {
1407                 state = ctx->state;
1408                 if (state == NULL)
1409                         continue;
1410                 /* Don't return layout if there is open file state */
1411                 if (state->state & FMODE_WRITE)
1412                         goto out_noroc;
1413                 if (state->state & FMODE_READ)
1414                         skip_read = true;
1415         }
1416
1417
1418         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1419                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1420                         continue;
1421                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1422                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1423                         continue;
1424                 /*
1425                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1426                  * doesn't invalidate the layout for us.
1427                  */
1428                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1429                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1430                         continue;
1431                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1432         }
1433
1434         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1435                 goto out_noroc;
1436
1437         /* ROC in two conditions:
1438          * 1. there are ROC lsegs
1439          * 2. we don't send layoutreturn
1440          */
1441         /* lo ref dropped in pnfs_roc_release() */
1442         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1443         /* If the creds don't match, we can't compound the layoutreturn */
1444         if (!layoutreturn || cred != lo->plh_lc_cred)
1445                 goto out_noroc;
1446
1447         roc = layoutreturn;
1448         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1449         res->lrs_present = 0;
1450         layoutreturn = false;
1451
1452 out_noroc:
1453         spin_unlock(&ino->i_lock);
1454         pnfs_layoutcommit_inode(ino, true);
1455         if (roc) {
1456                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1457                 if (ld->prepare_layoutreturn)
1458                         ld->prepare_layoutreturn(args);
1459                 pnfs_put_layout_hdr(lo);
1460                 return true;
1461         }
1462         if (layoutreturn)
1463                 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1464         pnfs_put_layout_hdr(lo);
1465         return false;
1466 }
1467
1468 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1469                 struct nfs4_layoutreturn_res *res,
1470                 int ret)
1471 {
1472         struct pnfs_layout_hdr *lo = args->layout;
1473         struct inode *inode = args->inode;
1474         const nfs4_stateid *arg_stateid = NULL;
1475         const nfs4_stateid *res_stateid = NULL;
1476         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1477
1478         switch (ret) {
1479         case -NFS4ERR_NOMATCHING_LAYOUT:
1480                 spin_lock(&inode->i_lock);
1481                 if (pnfs_layout_is_valid(lo) &&
1482                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1483                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1484                 spin_unlock(&inode->i_lock);
1485                 break;
1486         case 0:
1487                 if (res->lrs_present)
1488                         res_stateid = &res->stateid;
1489                 /* Fallthrough */
1490         default:
1491                 arg_stateid = &args->stateid;
1492         }
1493         pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1494                         res_stateid);
1495         if (ld_private && ld_private->ops && ld_private->ops->free)
1496                 ld_private->ops->free(ld_private);
1497         pnfs_put_layout_hdr(lo);
1498         trace_nfs4_layoutreturn_on_close(args->inode, 0);
1499 }
1500
1501 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1502 {
1503         struct nfs_inode *nfsi = NFS_I(ino);
1504         struct pnfs_layout_hdr *lo;
1505         bool sleep = false;
1506
1507         /* we might not have grabbed lo reference. so need to check under
1508          * i_lock */
1509         spin_lock(&ino->i_lock);
1510         lo = nfsi->layout;
1511         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1512                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1513                 sleep = true;
1514         }
1515         spin_unlock(&ino->i_lock);
1516         return sleep;
1517 }
1518
1519 /*
1520  * Compare two layout segments for sorting into layout cache.
1521  * We want to preferentially return RW over RO layouts, so ensure those
1522  * are seen first.
1523  */
1524 static s64
1525 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1526            const struct pnfs_layout_range *l2)
1527 {
1528         s64 d;
1529
1530         /* high offset > low offset */
1531         d = l1->offset - l2->offset;
1532         if (d)
1533                 return d;
1534
1535         /* short length > long length */
1536         d = l2->length - l1->length;
1537         if (d)
1538                 return d;
1539
1540         /* read > read/write */
1541         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1542 }
1543
1544 static bool
1545 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1546                 const struct pnfs_layout_range *l2)
1547 {
1548         return pnfs_lseg_range_cmp(l1, l2) > 0;
1549 }
1550
1551 static bool
1552 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1553                 struct pnfs_layout_segment *old)
1554 {
1555         return false;
1556 }
1557
1558 void
1559 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1560                    struct pnfs_layout_segment *lseg,
1561                    bool (*is_after)(const struct pnfs_layout_range *,
1562                            const struct pnfs_layout_range *),
1563                    bool (*do_merge)(struct pnfs_layout_segment *,
1564                            struct pnfs_layout_segment *),
1565                    struct list_head *free_me)
1566 {
1567         struct pnfs_layout_segment *lp, *tmp;
1568
1569         dprintk("%s:Begin\n", __func__);
1570
1571         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1572                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1573                         continue;
1574                 if (do_merge(lseg, lp)) {
1575                         mark_lseg_invalid(lp, free_me);
1576                         continue;
1577                 }
1578                 if (is_after(&lseg->pls_range, &lp->pls_range))
1579                         continue;
1580                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1581                 dprintk("%s: inserted lseg %p "
1582                         "iomode %d offset %llu length %llu before "
1583                         "lp %p iomode %d offset %llu length %llu\n",
1584                         __func__, lseg, lseg->pls_range.iomode,
1585                         lseg->pls_range.offset, lseg->pls_range.length,
1586                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1587                         lp->pls_range.length);
1588                 goto out;
1589         }
1590         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1591         dprintk("%s: inserted lseg %p "
1592                 "iomode %d offset %llu length %llu at tail\n",
1593                 __func__, lseg, lseg->pls_range.iomode,
1594                 lseg->pls_range.offset, lseg->pls_range.length);
1595 out:
1596         pnfs_get_layout_hdr(lo);
1597
1598         dprintk("%s:Return\n", __func__);
1599 }
1600 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1601
1602 static void
1603 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1604                    struct pnfs_layout_segment *lseg,
1605                    struct list_head *free_me)
1606 {
1607         struct inode *inode = lo->plh_inode;
1608         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1609
1610         if (ld->add_lseg != NULL)
1611                 ld->add_lseg(lo, lseg, free_me);
1612         else
1613                 pnfs_generic_layout_insert_lseg(lo, lseg,
1614                                 pnfs_lseg_range_is_after,
1615                                 pnfs_lseg_no_merge,
1616                                 free_me);
1617 }
1618
1619 static struct pnfs_layout_hdr *
1620 alloc_init_layout_hdr(struct inode *ino,
1621                       struct nfs_open_context *ctx,
1622                       gfp_t gfp_flags)
1623 {
1624         struct pnfs_layout_hdr *lo;
1625
1626         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1627         if (!lo)
1628                 return NULL;
1629         refcount_set(&lo->plh_refcount, 1);
1630         INIT_LIST_HEAD(&lo->plh_layouts);
1631         INIT_LIST_HEAD(&lo->plh_segs);
1632         INIT_LIST_HEAD(&lo->plh_return_segs);
1633         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1634         lo->plh_inode = ino;
1635         lo->plh_lc_cred = get_rpccred(ctx->cred);
1636         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1637         return lo;
1638 }
1639
1640 static struct pnfs_layout_hdr *
1641 pnfs_find_alloc_layout(struct inode *ino,
1642                        struct nfs_open_context *ctx,
1643                        gfp_t gfp_flags)
1644         __releases(&ino->i_lock)
1645         __acquires(&ino->i_lock)
1646 {
1647         struct nfs_inode *nfsi = NFS_I(ino);
1648         struct pnfs_layout_hdr *new = NULL;
1649
1650         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1651
1652         if (nfsi->layout != NULL)
1653                 goto out_existing;
1654         spin_unlock(&ino->i_lock);
1655         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1656         spin_lock(&ino->i_lock);
1657
1658         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1659                 nfsi->layout = new;
1660                 return new;
1661         } else if (new != NULL)
1662                 pnfs_free_layout_hdr(new);
1663 out_existing:
1664         pnfs_get_layout_hdr(nfsi->layout);
1665         return nfsi->layout;
1666 }
1667
1668 /*
1669  * iomode matching rules:
1670  * iomode       lseg    strict match
1671  *                      iomode
1672  * -----        -----   ------ -----
1673  * ANY          READ    N/A    true
1674  * ANY          RW      N/A    true
1675  * RW           READ    N/A    false
1676  * RW           RW      N/A    true
1677  * READ         READ    N/A    true
1678  * READ         RW      true   false
1679  * READ         RW      false  true
1680  */
1681 static bool
1682 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1683                  const struct pnfs_layout_range *range,
1684                  bool strict_iomode)
1685 {
1686         struct pnfs_layout_range range1;
1687
1688         if ((range->iomode == IOMODE_RW &&
1689              ls_range->iomode != IOMODE_RW) ||
1690             (range->iomode != ls_range->iomode &&
1691              strict_iomode) ||
1692             !pnfs_lseg_range_intersecting(ls_range, range))
1693                 return false;
1694
1695         /* range1 covers only the first byte in the range */
1696         range1 = *range;
1697         range1.length = 1;
1698         return pnfs_lseg_range_contained(ls_range, &range1);
1699 }
1700
1701 /*
1702  * lookup range in layout
1703  */
1704 static struct pnfs_layout_segment *
1705 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1706                 struct pnfs_layout_range *range,
1707                 bool strict_iomode)
1708 {
1709         struct pnfs_layout_segment *lseg, *ret = NULL;
1710
1711         dprintk("%s:Begin\n", __func__);
1712
1713         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1714                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1715                     pnfs_lseg_range_match(&lseg->pls_range, range,
1716                                           strict_iomode)) {
1717                         ret = pnfs_get_lseg(lseg);
1718                         break;
1719                 }
1720         }
1721
1722         dprintk("%s:Return lseg %p ref %d\n",
1723                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1724         return ret;
1725 }
1726
1727 /*
1728  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1729  * to the MDS or over pNFS
1730  *
1731  * The nfs_inode read_io and write_io fields are cumulative counters reset
1732  * when there are no layout segments. Note that in pnfs_update_layout iomode
1733  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1734  * WRITE request.
1735  *
1736  * A return of true means use MDS I/O.
1737  *
1738  * From rfc 5661:
1739  * If a file's size is smaller than the file size threshold, data accesses
1740  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1741  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1742  * server.  If both file size and I/O size are provided, the client SHOULD
1743  * reach or exceed  both thresholds before sending its read or write
1744  * requests to the data server.
1745  */
1746 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1747                                      struct inode *ino, int iomode)
1748 {
1749         struct nfs4_threshold *t = ctx->mdsthreshold;
1750         struct nfs_inode *nfsi = NFS_I(ino);
1751         loff_t fsize = i_size_read(ino);
1752         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1753
1754         if (t == NULL)
1755                 return ret;
1756
1757         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1758                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1759
1760         switch (iomode) {
1761         case IOMODE_READ:
1762                 if (t->bm & THRESHOLD_RD) {
1763                         dprintk("%s fsize %llu\n", __func__, fsize);
1764                         size_set = true;
1765                         if (fsize < t->rd_sz)
1766                                 size = true;
1767                 }
1768                 if (t->bm & THRESHOLD_RD_IO) {
1769                         dprintk("%s nfsi->read_io %llu\n", __func__,
1770                                 nfsi->read_io);
1771                         io_set = true;
1772                         if (nfsi->read_io < t->rd_io_sz)
1773                                 io = true;
1774                 }
1775                 break;
1776         case IOMODE_RW:
1777                 if (t->bm & THRESHOLD_WR) {
1778                         dprintk("%s fsize %llu\n", __func__, fsize);
1779                         size_set = true;
1780                         if (fsize < t->wr_sz)
1781                                 size = true;
1782                 }
1783                 if (t->bm & THRESHOLD_WR_IO) {
1784                         dprintk("%s nfsi->write_io %llu\n", __func__,
1785                                 nfsi->write_io);
1786                         io_set = true;
1787                         if (nfsi->write_io < t->wr_io_sz)
1788                                 io = true;
1789                 }
1790                 break;
1791         }
1792         if (size_set && io_set) {
1793                 if (size && io)
1794                         ret = true;
1795         } else if (size || io)
1796                 ret = true;
1797
1798         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1799         return ret;
1800 }
1801
1802 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1803 {
1804         /*
1805          * send layoutcommit as it can hold up layoutreturn due to lseg
1806          * reference
1807          */
1808         pnfs_layoutcommit_inode(lo->plh_inode, false);
1809         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1810                                    nfs_wait_bit_killable,
1811                                    TASK_KILLABLE);
1812 }
1813
1814 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1815 {
1816         atomic_inc(&lo->plh_outstanding);
1817 }
1818
1819 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1820 {
1821         if (atomic_dec_and_test(&lo->plh_outstanding))
1822                 wake_up_var(&lo->plh_outstanding);
1823 }
1824
1825 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1826 {
1827         unsigned long *bitlock = &lo->plh_flags;
1828
1829         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1830         smp_mb__after_atomic();
1831         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1832 }
1833
1834 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1835                                 struct nfs_server *server)
1836 {
1837         if (list_empty(&lo->plh_layouts)) {
1838                 struct nfs_client *clp = server->nfs_client;
1839
1840                 /* The lo must be on the clp list if there is any
1841                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1842                  */
1843                 spin_lock(&clp->cl_lock);
1844                 if (list_empty(&lo->plh_layouts))
1845                         list_add_tail(&lo->plh_layouts, &server->layouts);
1846                 spin_unlock(&clp->cl_lock);
1847         }
1848 }
1849
1850 /*
1851  * Layout segment is retreived from the server if not cached.
1852  * The appropriate layout segment is referenced and returned to the caller.
1853  */
1854 struct pnfs_layout_segment *
1855 pnfs_update_layout(struct inode *ino,
1856                    struct nfs_open_context *ctx,
1857                    loff_t pos,
1858                    u64 count,
1859                    enum pnfs_iomode iomode,
1860                    bool strict_iomode,
1861                    gfp_t gfp_flags)
1862 {
1863         struct pnfs_layout_range arg = {
1864                 .iomode = iomode,
1865                 .offset = pos,
1866                 .length = count,
1867         };
1868         unsigned pg_offset;
1869         struct nfs_server *server = NFS_SERVER(ino);
1870         struct nfs_client *clp = server->nfs_client;
1871         struct pnfs_layout_hdr *lo = NULL;
1872         struct pnfs_layout_segment *lseg = NULL;
1873         struct nfs4_layoutget *lgp;
1874         nfs4_stateid stateid;
1875         long timeout = 0;
1876         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1877         bool first;
1878
1879         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1880                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1881                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1882                 goto out;
1883         }
1884
1885         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1886                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1887                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1888                 goto out;
1889         }
1890
1891 lookup_again:
1892         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1893         if (IS_ERR(lseg))
1894                 goto out;
1895         first = false;
1896         spin_lock(&ino->i_lock);
1897         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1898         if (lo == NULL) {
1899                 spin_unlock(&ino->i_lock);
1900                 lseg = ERR_PTR(-ENOMEM);
1901                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1902                                  PNFS_UPDATE_LAYOUT_NOMEM);
1903                 goto out;
1904         }
1905
1906         /* Do we even need to bother with this? */
1907         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1908                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1909                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1910                 dprintk("%s matches recall, use MDS\n", __func__);
1911                 goto out_unlock;
1912         }
1913
1914         /* if LAYOUTGET already failed once we don't try again */
1915         if (pnfs_layout_io_test_failed(lo, iomode)) {
1916                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1917                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1918                 goto out_unlock;
1919         }
1920
1921         /*
1922          * If the layout segment list is empty, but there are outstanding
1923          * layoutget calls, then they might be subject to a layoutrecall.
1924          */
1925         if (list_empty(&lo->plh_segs) &&
1926             atomic_read(&lo->plh_outstanding) != 0) {
1927                 spin_unlock(&ino->i_lock);
1928                 lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
1929                                         !atomic_read(&lo->plh_outstanding)));
1930                 if (IS_ERR(lseg))
1931                         goto out_put_layout_hdr;
1932                 pnfs_put_layout_hdr(lo);
1933                 goto lookup_again;
1934         }
1935
1936         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1937         if (lseg) {
1938                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1939                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1940                 goto out_unlock;
1941         }
1942
1943         if (!nfs4_valid_open_stateid(ctx->state)) {
1944                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1945                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1946                 goto out_unlock;
1947         }
1948
1949         /*
1950          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1951          * stateid, or it has been invalidated, then we must use the open
1952          * stateid.
1953          */
1954         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1955
1956                 /*
1957                  * The first layoutget for the file. Need to serialize per
1958                  * RFC 5661 Errata 3208.
1959                  */
1960                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1961                                      &lo->plh_flags)) {
1962                         spin_unlock(&ino->i_lock);
1963                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
1964                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
1965                                                 TASK_KILLABLE));
1966                         if (IS_ERR(lseg))
1967                                 goto out_put_layout_hdr;
1968                         pnfs_put_layout_hdr(lo);
1969                         dprintk("%s retrying\n", __func__);
1970                         goto lookup_again;
1971                 }
1972
1973                 first = true;
1974                 if (nfs4_select_rw_stateid(ctx->state,
1975                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1976                                         NULL, &stateid, NULL) != 0) {
1977                         trace_pnfs_update_layout(ino, pos, count,
1978                                         iomode, lo, lseg,
1979                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1980                         goto out_unlock;
1981                 }
1982         } else {
1983                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1984         }
1985
1986         /*
1987          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1988          * for LAYOUTRETURN even if first is true.
1989          */
1990         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1991                 spin_unlock(&ino->i_lock);
1992                 dprintk("%s wait for layoutreturn\n", __func__);
1993                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
1994                 if (!IS_ERR(lseg)) {
1995                         if (first)
1996                                 pnfs_clear_first_layoutget(lo);
1997                         pnfs_put_layout_hdr(lo);
1998                         dprintk("%s retrying\n", __func__);
1999                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2000                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
2001                         goto lookup_again;
2002                 }
2003                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2004                                 PNFS_UPDATE_LAYOUT_RETURN);
2005                 goto out_put_layout_hdr;
2006         }
2007
2008         if (pnfs_layoutgets_blocked(lo)) {
2009                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2010                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2011                 goto out_unlock;
2012         }
2013         nfs_layoutget_begin(lo);
2014         spin_unlock(&ino->i_lock);
2015
2016         _add_to_server_list(lo, server);
2017
2018         pg_offset = arg.offset & ~PAGE_MASK;
2019         if (pg_offset) {
2020                 arg.offset -= pg_offset;
2021                 arg.length += pg_offset;
2022         }
2023         if (arg.length != NFS4_MAX_UINT64)
2024                 arg.length = PAGE_ALIGN(arg.length);
2025
2026         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2027         if (!lgp) {
2028                 lseg = ERR_PTR(-ENOMEM);
2029                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2030                                          PNFS_UPDATE_LAYOUT_NOMEM);
2031                 nfs_layoutget_end(lo);
2032                 goto out_put_layout_hdr;
2033         }
2034
2035         lseg = nfs4_proc_layoutget(lgp, &timeout);
2036         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2037                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2038         nfs_layoutget_end(lo);
2039         if (IS_ERR(lseg)) {
2040                 switch(PTR_ERR(lseg)) {
2041                 case -EBUSY:
2042                         if (time_after(jiffies, giveup))
2043                                 lseg = NULL;
2044                         break;
2045                 case -ERECALLCONFLICT:
2046                 case -EAGAIN:
2047                         break;
2048                 case -ENODATA:
2049                         /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2050                         pnfs_layout_set_fail_bit(
2051                                 lo, pnfs_iomode_to_fail_bit(iomode));
2052                         lseg = NULL;
2053                         goto out_put_layout_hdr;
2054                 default:
2055                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2056                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2057                                 lseg = NULL;
2058                         }
2059                         goto out_put_layout_hdr;
2060                 }
2061                 if (lseg) {
2062                         if (first)
2063                                 pnfs_clear_first_layoutget(lo);
2064                         trace_pnfs_update_layout(ino, pos, count,
2065                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2066                         pnfs_put_layout_hdr(lo);
2067                         goto lookup_again;
2068                 }
2069         } else {
2070                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2071         }
2072
2073 out_put_layout_hdr:
2074         if (first)
2075                 pnfs_clear_first_layoutget(lo);
2076         pnfs_put_layout_hdr(lo);
2077 out:
2078         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2079                         "(%s, offset: %llu, length: %llu)\n",
2080                         __func__, ino->i_sb->s_id,
2081                         (unsigned long long)NFS_FILEID(ino),
2082                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2083                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2084                         (unsigned long long)pos,
2085                         (unsigned long long)count);
2086         return lseg;
2087 out_unlock:
2088         spin_unlock(&ino->i_lock);
2089         goto out_put_layout_hdr;
2090 }
2091 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2092
2093 static bool
2094 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2095 {
2096         switch (range->iomode) {
2097         case IOMODE_READ:
2098         case IOMODE_RW:
2099                 break;
2100         default:
2101                 return false;
2102         }
2103         if (range->offset == NFS4_MAX_UINT64)
2104                 return false;
2105         if (range->length == 0)
2106                 return false;
2107         if (range->length != NFS4_MAX_UINT64 &&
2108             range->length > NFS4_MAX_UINT64 - range->offset)
2109                 return false;
2110         return true;
2111 }
2112
2113 static struct pnfs_layout_hdr *
2114 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2115 {
2116         struct pnfs_layout_hdr *lo;
2117
2118         spin_lock(&ino->i_lock);
2119         lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2120         if (!lo)
2121                 goto out_unlock;
2122         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2123                 goto out_unlock;
2124         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2125                 goto out_unlock;
2126         if (pnfs_layoutgets_blocked(lo))
2127                 goto out_unlock;
2128         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2129                 goto out_unlock;
2130         nfs_layoutget_begin(lo);
2131         spin_unlock(&ino->i_lock);
2132         _add_to_server_list(lo, NFS_SERVER(ino));
2133         return lo;
2134
2135 out_unlock:
2136         spin_unlock(&ino->i_lock);
2137         pnfs_put_layout_hdr(lo);
2138         return NULL;
2139 }
2140
2141 extern const nfs4_stateid current_stateid;
2142
2143 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2144                                      struct nfs_open_context *ctx)
2145 {
2146         struct inode *ino = data->dentry->d_inode;
2147         struct pnfs_layout_range rng = {
2148                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2149                           IOMODE_RW: IOMODE_READ,
2150                 .offset = 0,
2151                 .length = NFS4_MAX_UINT64,
2152         };
2153         struct nfs4_layoutget *lgp;
2154         struct pnfs_layout_hdr *lo;
2155
2156         /* Heuristic: don't send layoutget if we have cached data */
2157         if (rng.iomode == IOMODE_READ &&
2158            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2159                 return;
2160
2161         lo = _pnfs_grab_empty_layout(ino, ctx);
2162         if (!lo)
2163                 return;
2164         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2165                                              &rng, GFP_KERNEL);
2166         if (!lgp) {
2167                 pnfs_clear_first_layoutget(lo);
2168                 nfs_layoutget_end(lo);
2169                 pnfs_put_layout_hdr(lo);
2170                 return;
2171         }
2172         data->lgp = lgp;
2173         data->o_arg.lg_args = &lgp->args;
2174         data->o_res.lg_res = &lgp->res;
2175 }
2176
2177 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2178                                      struct nfs_open_context *ctx)
2179 {
2180         struct pnfs_layout_range rng = {
2181                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2182                           IOMODE_RW: IOMODE_READ,
2183                 .offset = 0,
2184                 .length = NFS4_MAX_UINT64,
2185         };
2186         struct nfs4_layoutget *lgp;
2187
2188         lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2189                                              &rng, GFP_KERNEL);
2190         if (!lgp)
2191                 return;
2192         data->lgp = lgp;
2193         data->o_arg.lg_args = &lgp->args;
2194         data->o_res.lg_res = &lgp->res;
2195 }
2196
2197 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2198                          struct nfs_open_context *ctx)
2199 {
2200         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2201
2202         if (!(pnfs_enabled_sb(server) &&
2203               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2204                 return;
2205         /* Could check on max_ops, but currently hardcoded high enough */
2206         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2207                 return;
2208         if (data->state)
2209                 _lgopen_prepare_attached(data, ctx);
2210         else
2211                 _lgopen_prepare_floating(data, ctx);
2212 }
2213
2214 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2215                        struct nfs_open_context *ctx)
2216 {
2217         struct pnfs_layout_hdr *lo;
2218         struct pnfs_layout_segment *lseg;
2219         struct nfs_server *srv = NFS_SERVER(ino);
2220         u32 iomode;
2221
2222         if (!lgp)
2223                 return;
2224         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2225         if (lgp->res.status) {
2226                 switch (lgp->res.status) {
2227                 default:
2228                         break;
2229                 /*
2230                  * Halt lgopen attempts if the server doesn't recognise
2231                  * the "current stateid" value, the layout type, or the
2232                  * layoutget operation as being valid.
2233                  * Also if it complains about too many ops in the compound
2234                  * or of the request/reply being too big.
2235                  */
2236                 case -NFS4ERR_BAD_STATEID:
2237                 case -NFS4ERR_NOTSUPP:
2238                 case -NFS4ERR_REP_TOO_BIG:
2239                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2240                 case -NFS4ERR_REQ_TOO_BIG:
2241                 case -NFS4ERR_TOO_MANY_OPS:
2242                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2243                         srv->caps &= ~NFS_CAP_LGOPEN;
2244                 }
2245                 return;
2246         }
2247         if (!lgp->args.inode) {
2248                 lo = _pnfs_grab_empty_layout(ino, ctx);
2249                 if (!lo)
2250                         return;
2251                 lgp->args.inode = ino;
2252         } else
2253                 lo = NFS_I(lgp->args.inode)->layout;
2254
2255         lseg = pnfs_layout_process(lgp);
2256         if (!IS_ERR(lseg)) {
2257                 iomode = lgp->args.range.iomode;
2258                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2259                 pnfs_put_lseg(lseg);
2260         }
2261 }
2262
2263 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2264 {
2265         if (lgp != NULL) {
2266                 struct inode *inode = lgp->args.inode;
2267                 if (inode) {
2268                         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2269                         pnfs_clear_first_layoutget(lo);
2270                         nfs_layoutget_end(lo);
2271                 }
2272                 pnfs_layoutget_free(lgp);
2273         }
2274 }
2275
2276 struct pnfs_layout_segment *
2277 pnfs_layout_process(struct nfs4_layoutget *lgp)
2278 {
2279         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2280         struct nfs4_layoutget_res *res = &lgp->res;
2281         struct pnfs_layout_segment *lseg;
2282         struct inode *ino = lo->plh_inode;
2283         LIST_HEAD(free_me);
2284
2285         if (!pnfs_sanity_check_layout_range(&res->range))
2286                 return ERR_PTR(-EINVAL);
2287
2288         /* Inject layout blob into I/O device driver */
2289         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2290         if (IS_ERR_OR_NULL(lseg)) {
2291                 if (!lseg)
2292                         lseg = ERR_PTR(-ENOMEM);
2293
2294                 dprintk("%s: Could not allocate layout: error %ld\n",
2295                        __func__, PTR_ERR(lseg));
2296                 return lseg;
2297         }
2298
2299         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2300
2301         spin_lock(&ino->i_lock);
2302         if (pnfs_layoutgets_blocked(lo)) {
2303                 dprintk("%s forget reply due to state\n", __func__);
2304                 goto out_forget;
2305         }
2306
2307         if (!pnfs_layout_is_valid(lo)) {
2308                 /* We have a completely new layout */
2309                 pnfs_set_layout_stateid(lo, &res->stateid, true);
2310         } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2311                 /* existing state ID, make sure the sequence number matches. */
2312                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2313                         dprintk("%s forget reply due to sequence\n", __func__);
2314                         goto out_forget;
2315                 }
2316                 pnfs_set_layout_stateid(lo, &res->stateid, false);
2317         } else {
2318                 /*
2319                  * We got an entirely new state ID.  Mark all segments for the
2320                  * inode invalid, and retry the layoutget
2321                  */
2322                 struct pnfs_layout_range range = {
2323                         .iomode = IOMODE_ANY,
2324                         .length = NFS4_MAX_UINT64,
2325                 };
2326                 pnfs_set_plh_return_info(lo, IOMODE_ANY, 0);
2327                 pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2328                                                 &range, 0);
2329                 goto out_forget;
2330         }
2331
2332         pnfs_get_lseg(lseg);
2333         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2334
2335
2336         if (res->return_on_close)
2337                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2338
2339         spin_unlock(&ino->i_lock);
2340         pnfs_free_lseg_list(&free_me);
2341         return lseg;
2342
2343 out_forget:
2344         spin_unlock(&ino->i_lock);
2345         lseg->pls_layout = lo;
2346         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2347         pnfs_free_lseg_list(&free_me);
2348         return ERR_PTR(-EAGAIN);
2349 }
2350
2351 /**
2352  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2353  * @lo: pointer to layout header
2354  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2355  * @return_range: describe layout segment ranges to be returned
2356  * @seq: stateid seqid to match
2357  *
2358  * This function is mainly intended for use by layoutrecall. It attempts
2359  * to free the layout segment immediately, or else to mark it for return
2360  * as soon as its reference count drops to zero.
2361  *
2362  * Returns
2363  * - 0: a layoutreturn needs to be scheduled.
2364  * - EBUSY: there are layout segment that are still in use.
2365  * - ENOENT: there are no layout segments that need to be returned.
2366  */
2367 int
2368 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2369                                 struct list_head *tmp_list,
2370                                 const struct pnfs_layout_range *return_range,
2371                                 u32 seq)
2372 {
2373         struct pnfs_layout_segment *lseg, *next;
2374         int remaining = 0;
2375
2376         dprintk("%s:Begin lo %p\n", __func__, lo);
2377
2378         assert_spin_locked(&lo->plh_inode->i_lock);
2379
2380         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2381                 tmp_list = &lo->plh_return_segs;
2382
2383         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2384                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2385                         dprintk("%s: marking lseg %p iomode %d "
2386                                 "offset %llu length %llu\n", __func__,
2387                                 lseg, lseg->pls_range.iomode,
2388                                 lseg->pls_range.offset,
2389                                 lseg->pls_range.length);
2390                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2391                                 tmp_list = &lo->plh_return_segs;
2392                         if (mark_lseg_invalid(lseg, tmp_list))
2393                                 continue;
2394                         remaining++;
2395                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2396                 }
2397
2398         if (remaining) {
2399                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2400                 return -EBUSY;
2401         }
2402
2403         if (!list_empty(&lo->plh_return_segs)) {
2404                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2405                 return 0;
2406         }
2407
2408         return -ENOENT;
2409 }
2410
2411 void pnfs_error_mark_layout_for_return(struct inode *inode,
2412                                        struct pnfs_layout_segment *lseg)
2413 {
2414         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2415         struct pnfs_layout_range range = {
2416                 .iomode = lseg->pls_range.iomode,
2417                 .offset = 0,
2418                 .length = NFS4_MAX_UINT64,
2419         };
2420         bool return_now = false;
2421
2422         spin_lock(&inode->i_lock);
2423         if (!pnfs_layout_is_valid(lo)) {
2424                 spin_unlock(&inode->i_lock);
2425                 return;
2426         }
2427         pnfs_set_plh_return_info(lo, range.iomode, 0);
2428         /*
2429          * mark all matching lsegs so that we are sure to have no live
2430          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2431          * for how it works.
2432          */
2433         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0) != -EBUSY) {
2434                 nfs4_stateid stateid;
2435                 enum pnfs_iomode iomode;
2436
2437                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2438                 spin_unlock(&inode->i_lock);
2439                 if (return_now)
2440                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2441         } else {
2442                 spin_unlock(&inode->i_lock);
2443                 nfs_commit_inode(inode, 0);
2444         }
2445 }
2446 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2447
2448 void
2449 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2450 {
2451         if (pgio->pg_lseg == NULL ||
2452             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2453                 return;
2454         pnfs_put_lseg(pgio->pg_lseg);
2455         pgio->pg_lseg = NULL;
2456 }
2457 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2458
2459 /*
2460  * Check for any intersection between the request and the pgio->pg_lseg,
2461  * and if none, put this pgio->pg_lseg away.
2462  */
2463 static void
2464 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2465 {
2466         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2467                 pnfs_put_lseg(pgio->pg_lseg);
2468                 pgio->pg_lseg = NULL;
2469         }
2470 }
2471
2472 void
2473 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2474 {
2475         u64 rd_size = req->wb_bytes;
2476
2477         pnfs_generic_pg_check_layout(pgio);
2478         pnfs_generic_pg_check_range(pgio, req);
2479         if (pgio->pg_lseg == NULL) {
2480                 if (pgio->pg_dreq == NULL)
2481                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2482                 else
2483                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2484
2485                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2486                                                    req->wb_context,
2487                                                    req_offset(req),
2488                                                    rd_size,
2489                                                    IOMODE_READ,
2490                                                    false,
2491                                                    GFP_KERNEL);
2492                 if (IS_ERR(pgio->pg_lseg)) {
2493                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2494                         pgio->pg_lseg = NULL;
2495                         return;
2496                 }
2497         }
2498         /* If no lseg, fall back to read through mds */
2499         if (pgio->pg_lseg == NULL)
2500                 nfs_pageio_reset_read_mds(pgio);
2501
2502 }
2503 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2504
2505 void
2506 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2507                            struct nfs_page *req, u64 wb_size)
2508 {
2509         pnfs_generic_pg_check_layout(pgio);
2510         pnfs_generic_pg_check_range(pgio, req);
2511         if (pgio->pg_lseg == NULL) {
2512                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2513                                                    req->wb_context,
2514                                                    req_offset(req),
2515                                                    wb_size,
2516                                                    IOMODE_RW,
2517                                                    false,
2518                                                    GFP_NOFS);
2519                 if (IS_ERR(pgio->pg_lseg)) {
2520                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2521                         pgio->pg_lseg = NULL;
2522                         return;
2523                 }
2524         }
2525         /* If no lseg, fall back to write through mds */
2526         if (pgio->pg_lseg == NULL)
2527                 nfs_pageio_reset_write_mds(pgio);
2528 }
2529 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2530
2531 void
2532 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2533 {
2534         if (desc->pg_lseg) {
2535                 pnfs_put_lseg(desc->pg_lseg);
2536                 desc->pg_lseg = NULL;
2537         }
2538 }
2539 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2540
2541 /*
2542  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2543  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2544  */
2545 size_t
2546 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2547                      struct nfs_page *prev, struct nfs_page *req)
2548 {
2549         unsigned int size;
2550         u64 seg_end, req_start, seg_left;
2551
2552         size = nfs_generic_pg_test(pgio, prev, req);
2553         if (!size)
2554                 return 0;
2555
2556         /*
2557          * 'size' contains the number of bytes left in the current page (up
2558          * to the original size asked for in @req->wb_bytes).
2559          *
2560          * Calculate how many bytes are left in the layout segment
2561          * and if there are less bytes than 'size', return that instead.
2562          *
2563          * Please also note that 'end_offset' is actually the offset of the
2564          * first byte that lies outside the pnfs_layout_range. FIXME?
2565          *
2566          */
2567         if (pgio->pg_lseg) {
2568                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2569                                      pgio->pg_lseg->pls_range.length);
2570                 req_start = req_offset(req);
2571
2572                 /* start of request is past the last byte of this segment */
2573                 if (req_start >= seg_end)
2574                         return 0;
2575
2576                 /* adjust 'size' iff there are fewer bytes left in the
2577                  * segment than what nfs_generic_pg_test returned */
2578                 seg_left = seg_end - req_start;
2579                 if (seg_left < size)
2580                         size = (unsigned int)seg_left;
2581         }
2582
2583         return size;
2584 }
2585 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2586
2587 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2588 {
2589         struct nfs_pageio_descriptor pgio;
2590
2591         /* Resend all requests through the MDS */
2592         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2593                               hdr->completion_ops);
2594         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2595         return nfs_pageio_resend(&pgio, hdr);
2596 }
2597 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2598
2599 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2600 {
2601
2602         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2603         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2604             PNFS_LAYOUTRET_ON_ERROR) {
2605                 pnfs_return_layout(hdr->inode);
2606         }
2607         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2608                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2609 }
2610
2611 /*
2612  * Called by non rpc-based layout drivers
2613  */
2614 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2615 {
2616         if (likely(!hdr->pnfs_error)) {
2617                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2618                                 hdr->mds_offset + hdr->res.count);
2619                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2620         }
2621         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2622         if (unlikely(hdr->pnfs_error))
2623                 pnfs_ld_handle_write_error(hdr);
2624         hdr->mds_ops->rpc_release(hdr);
2625 }
2626 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2627
2628 static void
2629 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2630                 struct nfs_pgio_header *hdr)
2631 {
2632         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2633
2634         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2635                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2636                 nfs_pageio_reset_write_mds(desc);
2637                 mirror->pg_recoalesce = 1;
2638         }
2639         hdr->completion_ops->completion(hdr);
2640 }
2641
2642 static enum pnfs_try_status
2643 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2644                         const struct rpc_call_ops *call_ops,
2645                         struct pnfs_layout_segment *lseg,
2646                         int how)
2647 {
2648         struct inode *inode = hdr->inode;
2649         enum pnfs_try_status trypnfs;
2650         struct nfs_server *nfss = NFS_SERVER(inode);
2651
2652         hdr->mds_ops = call_ops;
2653
2654         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2655                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2656         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2657         if (trypnfs != PNFS_NOT_ATTEMPTED)
2658                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2659         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2660         return trypnfs;
2661 }
2662
2663 static void
2664 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2665               struct nfs_pgio_header *hdr, int how)
2666 {
2667         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2668         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2669         enum pnfs_try_status trypnfs;
2670
2671         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2672         switch (trypnfs) {
2673         case PNFS_NOT_ATTEMPTED:
2674                 pnfs_write_through_mds(desc, hdr);
2675         case PNFS_ATTEMPTED:
2676                 break;
2677         case PNFS_TRY_AGAIN:
2678                 /* cleanup hdr and prepare to redo pnfs */
2679                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2680                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2681                         list_splice_init(&hdr->pages, &mirror->pg_list);
2682                         mirror->pg_recoalesce = 1;
2683                 }
2684                 hdr->mds_ops->rpc_release(hdr);
2685         }
2686 }
2687
2688 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2689 {
2690         pnfs_put_lseg(hdr->lseg);
2691         nfs_pgio_header_free(hdr);
2692 }
2693
2694 int
2695 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2696 {
2697         struct nfs_pgio_header *hdr;
2698         int ret;
2699
2700         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2701         if (!hdr) {
2702                 desc->pg_error = -ENOMEM;
2703                 return desc->pg_error;
2704         }
2705         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2706
2707         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2708         ret = nfs_generic_pgio(desc, hdr);
2709         if (!ret)
2710                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2711
2712         return ret;
2713 }
2714 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2715
2716 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2717 {
2718         struct nfs_pageio_descriptor pgio;
2719
2720         /* Resend all requests through the MDS */
2721         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2722         return nfs_pageio_resend(&pgio, hdr);
2723 }
2724 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2725
2726 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2727 {
2728         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2729         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2730             PNFS_LAYOUTRET_ON_ERROR) {
2731                 pnfs_return_layout(hdr->inode);
2732         }
2733         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2734                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2735 }
2736
2737 /*
2738  * Called by non rpc-based layout drivers
2739  */
2740 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2741 {
2742         if (likely(!hdr->pnfs_error))
2743                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2744         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2745         if (unlikely(hdr->pnfs_error))
2746                 pnfs_ld_handle_read_error(hdr);
2747         hdr->mds_ops->rpc_release(hdr);
2748 }
2749 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2750
2751 static void
2752 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2753                 struct nfs_pgio_header *hdr)
2754 {
2755         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2756
2757         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2758                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2759                 nfs_pageio_reset_read_mds(desc);
2760                 mirror->pg_recoalesce = 1;
2761         }
2762         hdr->completion_ops->completion(hdr);
2763 }
2764
2765 /*
2766  * Call the appropriate parallel I/O subsystem read function.
2767  */
2768 static enum pnfs_try_status
2769 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2770                        const struct rpc_call_ops *call_ops,
2771                        struct pnfs_layout_segment *lseg)
2772 {
2773         struct inode *inode = hdr->inode;
2774         struct nfs_server *nfss = NFS_SERVER(inode);
2775         enum pnfs_try_status trypnfs;
2776
2777         hdr->mds_ops = call_ops;
2778
2779         dprintk("%s: Reading ino:%lu %u@%llu\n",
2780                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2781
2782         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2783         if (trypnfs != PNFS_NOT_ATTEMPTED)
2784                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2785         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2786         return trypnfs;
2787 }
2788
2789 /* Resend all requests through pnfs. */
2790 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2791 {
2792         struct nfs_pageio_descriptor pgio;
2793
2794         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2795                 /* Prevent deadlocks with layoutreturn! */
2796                 pnfs_put_lseg(hdr->lseg);
2797                 hdr->lseg = NULL;
2798
2799                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2800                                         hdr->completion_ops);
2801                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2802         }
2803 }
2804 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2805
2806 static void
2807 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2808 {
2809         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2810         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2811         enum pnfs_try_status trypnfs;
2812
2813         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2814         switch (trypnfs) {
2815         case PNFS_NOT_ATTEMPTED:
2816                 pnfs_read_through_mds(desc, hdr);
2817         case PNFS_ATTEMPTED:
2818                 break;
2819         case PNFS_TRY_AGAIN:
2820                 /* cleanup hdr and prepare to redo pnfs */
2821                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2822                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2823                         list_splice_init(&hdr->pages, &mirror->pg_list);
2824                         mirror->pg_recoalesce = 1;
2825                 }
2826                 hdr->mds_ops->rpc_release(hdr);
2827         }
2828 }
2829
2830 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2831 {
2832         pnfs_put_lseg(hdr->lseg);
2833         nfs_pgio_header_free(hdr);
2834 }
2835
2836 int
2837 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2838 {
2839         struct nfs_pgio_header *hdr;
2840         int ret;
2841
2842         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2843         if (!hdr) {
2844                 desc->pg_error = -ENOMEM;
2845                 return desc->pg_error;
2846         }
2847         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2848         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2849         ret = nfs_generic_pgio(desc, hdr);
2850         if (!ret)
2851                 pnfs_do_read(desc, hdr);
2852         return ret;
2853 }
2854 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2855
2856 static void pnfs_clear_layoutcommitting(struct inode *inode)
2857 {
2858         unsigned long *bitlock = &NFS_I(inode)->flags;
2859
2860         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2861         smp_mb__after_atomic();
2862         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2863 }
2864
2865 /*
2866  * There can be multiple RW segments.
2867  */
2868 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2869 {
2870         struct pnfs_layout_segment *lseg;
2871
2872         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2873                 if (lseg->pls_range.iomode == IOMODE_RW &&
2874                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2875                         list_add(&lseg->pls_lc_list, listp);
2876         }
2877 }
2878
2879 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2880 {
2881         struct pnfs_layout_segment *lseg, *tmp;
2882
2883         /* Matched by references in pnfs_set_layoutcommit */
2884         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2885                 list_del_init(&lseg->pls_lc_list);
2886                 pnfs_put_lseg(lseg);
2887         }
2888
2889         pnfs_clear_layoutcommitting(inode);
2890 }
2891
2892 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2893 {
2894         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2895 }
2896 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2897
2898 void
2899 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2900                 loff_t end_pos)
2901 {
2902         struct nfs_inode *nfsi = NFS_I(inode);
2903         bool mark_as_dirty = false;
2904
2905         spin_lock(&inode->i_lock);
2906         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2907                 nfsi->layout->plh_lwb = end_pos;
2908                 mark_as_dirty = true;
2909                 dprintk("%s: Set layoutcommit for inode %lu ",
2910                         __func__, inode->i_ino);
2911         } else if (end_pos > nfsi->layout->plh_lwb)
2912                 nfsi->layout->plh_lwb = end_pos;
2913         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2914                 /* references matched in nfs4_layoutcommit_release */
2915                 pnfs_get_lseg(lseg);
2916         }
2917         spin_unlock(&inode->i_lock);
2918         dprintk("%s: lseg %p end_pos %llu\n",
2919                 __func__, lseg, nfsi->layout->plh_lwb);
2920
2921         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2922          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2923         if (mark_as_dirty)
2924                 mark_inode_dirty_sync(inode);
2925 }
2926 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2927
2928 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2929 {
2930         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2931
2932         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2933                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2934         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2935 }
2936
2937 /*
2938  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2939  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2940  * data to disk to allow the server to recover the data if it crashes.
2941  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2942  * is off, and a COMMIT is sent to a data server, or
2943  * if WRITEs to a data server return NFS_DATA_SYNC.
2944  */
2945 int
2946 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2947 {
2948         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2949         struct nfs4_layoutcommit_data *data;
2950         struct nfs_inode *nfsi = NFS_I(inode);
2951         loff_t end_pos;
2952         int status;
2953
2954         if (!pnfs_layoutcommit_outstanding(inode))
2955                 return 0;
2956
2957         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2958
2959         status = -EAGAIN;
2960         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2961                 if (!sync)
2962                         goto out;
2963                 status = wait_on_bit_lock_action(&nfsi->flags,
2964                                 NFS_INO_LAYOUTCOMMITTING,
2965                                 nfs_wait_bit_killable,
2966                                 TASK_KILLABLE);
2967                 if (status)
2968                         goto out;
2969         }
2970
2971         status = -ENOMEM;
2972         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2973         data = kzalloc(sizeof(*data), GFP_NOFS);
2974         if (!data)
2975                 goto clear_layoutcommitting;
2976
2977         status = 0;
2978         spin_lock(&inode->i_lock);
2979         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2980                 goto out_unlock;
2981
2982         INIT_LIST_HEAD(&data->lseg_list);
2983         pnfs_list_write_lseg(inode, &data->lseg_list);
2984
2985         end_pos = nfsi->layout->plh_lwb;
2986
2987         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2988         spin_unlock(&inode->i_lock);
2989
2990         data->args.inode = inode;
2991         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2992         nfs_fattr_init(&data->fattr);
2993         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2994         data->res.fattr = &data->fattr;
2995         if (end_pos != 0)
2996                 data->args.lastbytewritten = end_pos - 1;
2997         else
2998                 data->args.lastbytewritten = U64_MAX;
2999         data->res.server = NFS_SERVER(inode);
3000
3001         if (ld->prepare_layoutcommit) {
3002                 status = ld->prepare_layoutcommit(&data->args);
3003                 if (status) {
3004                         put_rpccred(data->cred);
3005                         spin_lock(&inode->i_lock);
3006                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3007                         if (end_pos > nfsi->layout->plh_lwb)
3008                                 nfsi->layout->plh_lwb = end_pos;
3009                         goto out_unlock;
3010                 }
3011         }
3012
3013
3014         status = nfs4_proc_layoutcommit(data, sync);
3015 out:
3016         if (status)
3017                 mark_inode_dirty_sync(inode);
3018         dprintk("<-- %s status %d\n", __func__, status);
3019         return status;
3020 out_unlock:
3021         spin_unlock(&inode->i_lock);
3022         kfree(data);
3023 clear_layoutcommitting:
3024         pnfs_clear_layoutcommitting(inode);
3025         goto out;
3026 }
3027 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3028
3029 int
3030 pnfs_generic_sync(struct inode *inode, bool datasync)
3031 {
3032         return pnfs_layoutcommit_inode(inode, true);
3033 }
3034 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3035
3036 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3037 {
3038         struct nfs4_threshold *thp;
3039
3040         thp = kzalloc(sizeof(*thp), GFP_NOFS);
3041         if (!thp) {
3042                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3043                 return NULL;
3044         }
3045         return thp;
3046 }
3047
3048 #if IS_ENABLED(CONFIG_NFS_V4_2)
3049 int
3050 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3051 {
3052         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3053         struct nfs_server *server = NFS_SERVER(inode);
3054         struct nfs_inode *nfsi = NFS_I(inode);
3055         struct nfs42_layoutstat_data *data;
3056         struct pnfs_layout_hdr *hdr;
3057         int status = 0;
3058
3059         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3060                 goto out;
3061
3062         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3063                 goto out;
3064
3065         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3066                 goto out;
3067
3068         spin_lock(&inode->i_lock);
3069         if (!NFS_I(inode)->layout) {
3070                 spin_unlock(&inode->i_lock);
3071                 goto out_clear_layoutstats;
3072         }
3073         hdr = NFS_I(inode)->layout;
3074         pnfs_get_layout_hdr(hdr);
3075         spin_unlock(&inode->i_lock);
3076
3077         data = kzalloc(sizeof(*data), gfp_flags);
3078         if (!data) {
3079                 status = -ENOMEM;
3080                 goto out_put;
3081         }
3082
3083         data->args.fh = NFS_FH(inode);
3084         data->args.inode = inode;
3085         status = ld->prepare_layoutstats(&data->args);
3086         if (status)
3087                 goto out_free;
3088
3089         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3090
3091 out:
3092         dprintk("%s returns %d\n", __func__, status);
3093         return status;
3094
3095 out_free:
3096         kfree(data);
3097 out_put:
3098         pnfs_put_layout_hdr(hdr);
3099 out_clear_layoutstats:
3100         smp_mb__before_atomic();
3101         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3102         smp_mb__after_atomic();
3103         goto out;
3104 }
3105 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3106 #endif
3107
3108 unsigned int layoutstats_timer;
3109 module_param(layoutstats_timer, uint, 0644);
3110 EXPORT_SYMBOL_GPL(layoutstats_timer);