fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include <linux/sort.h>
  34 #include "internal.h"
  35 #include "pnfs.h"
  36 #include "iostat.h"
  37 #include "nfs4trace.h"
  38 #include "delegation.h"
  39 #include "nfs42.h"
  40
  41 #define NFSDBG_FACILITY         NFSDBG_PNFS
  42 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  43
  44 /* Locking:
  45  *
  46  * pnfs_spinlock:
  47  *      protects pnfs_modules_tbl.
  48  */
  49 static DEFINE_SPINLOCK(pnfs_spinlock);
  50
  51 /*
  52  * pnfs_modules_tbl holds all pnfs modules
  53  */
  54 static LIST_HEAD(pnfs_modules_tbl);
  55
  56 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  57 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  58                 struct list_head *free_me,
  59                 const struct pnfs_layout_range *range,
  60                 u32 seq);
  61 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  62                                 struct list_head *tmp_list);
  63
  64 /* Return the registered pnfs layout driver module matching given id */
  65 static struct pnfs_layoutdriver_type *
  66 find_pnfs_driver_locked(u32 id)
  67 {
  68         struct pnfs_layoutdriver_type *local;
  69
  70         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  71                 if (local->id == id)
  72                         goto out;
  73         local = NULL;
  74 out:
  75         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  76         return local;
  77 }
  78
  79 static struct pnfs_layoutdriver_type *
  80 find_pnfs_driver(u32 id)
  81 {
  82         struct pnfs_layoutdriver_type *local;
  83
  84         spin_lock(&pnfs_spinlock);
  85         local = find_pnfs_driver_locked(id);
  86         if (local != NULL && !try_module_get(local->owner)) {
  87                 dprintk("%s: Could not grab reference on module\n", __func__);
  88                 local = NULL;
  89         }
  90         spin_unlock(&pnfs_spinlock);
  91         return local;
  92 }
  93
  94 void
  95 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  96 {
  97         if (nfss->pnfs_curr_ld) {
  98                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  99                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 100                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 101                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 102                         nfs4_deviceid_purge_client(nfss->nfs_client);
 103                 module_put(nfss->pnfs_curr_ld->owner);
 104         }
 105         nfss->pnfs_curr_ld = NULL;
 106 }
 107
 108 /*
 109  * When the server sends a list of layout types, we choose one in the order
 110  * given in the list below.
 111  *
 112  * FIXME: should this list be configurable in some fashion? module param?
 113  *        mount option? something else?
 114  */
 115 static const u32 ld_prefs[] = {
 116         LAYOUT_SCSI,
 117         LAYOUT_BLOCK_VOLUME,
 118         LAYOUT_OSD2_OBJECTS,
 119         LAYOUT_FLEX_FILES,
 120         LAYOUT_NFSV4_1_FILES,
 121         0
 122 };
 123
 124 static int
 125 ld_cmp(const void *e1, const void *e2)
 126 {
 127         u32 ld1 = *((u32 *)e1);
 128         u32 ld2 = *((u32 *)e2);
 129         int i;
 130
 131         for (i = 0; ld_prefs[i] != 0; i++) {
 132                 if (ld1 == ld_prefs[i])
 133                         return -1;
 134
 135                 if (ld2 == ld_prefs[i])
 136                         return 1;
 137         }
 138         return 0;
 139 }
 140
 141 /*
 142  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 143  * Currently only one pNFS layout driver per filesystem is supported.
 144  *
 145  * @ids array of layout types supported by MDS.
 146  */
 147 void
 148 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 149                       struct nfs_fsinfo *fsinfo)
 150 {
 151         struct pnfs_layoutdriver_type *ld_type = NULL;
 152         u32 id;
 153         int i;
 154
 155         if (fsinfo->nlayouttypes == 0)
 156                 goto out_no_driver;
 157         if (!(server->nfs_client->cl_exchange_flags &
 158                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 159                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 160                         __func__, server->nfs_client->cl_exchange_flags);
 161                 goto out_no_driver;
 162         }
 163
 164         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 165                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 166
 167         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 168                 id = fsinfo->layouttype[i];
 169                 ld_type = find_pnfs_driver(id);
 170                 if (!ld_type) {
 171                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 172                                         id);
 173                         ld_type = find_pnfs_driver(id);
 174                 }
 175                 if (ld_type)
 176                         break;
 177         }
 178
 179         if (!ld_type) {
 180                 dprintk("%s: No pNFS module found!\n", __func__);
 181                 goto out_no_driver;
 182         }
 183
 184         server->pnfs_curr_ld = ld_type;
 185         if (ld_type->set_layoutdriver
 186             && ld_type->set_layoutdriver(server, mntfh)) {
 187                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 188                         "driver %u.\n", __func__, id);
 189                 module_put(ld_type->owner);
 190                 goto out_no_driver;
 191         }
 192         /* Bump the MDS count */
 193         atomic_inc(&server->nfs_client->cl_mds_count);
 194
 195         dprintk("%s: pNFS module for %u set\n", __func__, id);
 196         return;
 197
 198 out_no_driver:
 199         dprintk("%s: Using NFSv4 I/O\n", __func__);
 200         server->pnfs_curr_ld = NULL;
 201 }
 202
 203 int
 204 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 205 {
 206         int status = -EINVAL;
 207         struct pnfs_layoutdriver_type *tmp;
 208
 209         if (ld_type->id == 0) {
 210                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 211                 return status;
 212         }
 213         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 214                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 215                        "alloc_lseg and free_lseg.\n", __func__);
 216                 return status;
 217         }
 218
 219         spin_lock(&pnfs_spinlock);
 220         tmp = find_pnfs_driver_locked(ld_type->id);
 221         if (!tmp) {
 222                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 223                 status = 0;
 224                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 225                         ld_type->name);
 226         } else {
 227                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 228                         __func__, ld_type->id);
 229         }
 230         spin_unlock(&pnfs_spinlock);
 231
 232         return status;
 233 }
 234 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 235
 236 void
 237 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 238 {
 239         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 240         spin_lock(&pnfs_spinlock);
 241         list_del(&ld_type->pnfs_tblid);
 242         spin_unlock(&pnfs_spinlock);
 243 }
 244 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 245
 246 /*
 247  * pNFS client layout cache
 248  */
 249
 250 /* Need to hold i_lock if caller does not already hold reference */
 251 void
 252 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 253 {
 254         refcount_inc(&lo->plh_refcount);
 255 }
 256
 257 static struct pnfs_layout_hdr *
 258 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 259 {
 260         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 261         return ld->alloc_layout_hdr(ino, gfp_flags);
 262 }
 263
 264 static void
 265 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 266 {
 267         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 268         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 269
 270         if (!list_empty(&lo->plh_layouts)) {
 271                 struct nfs_client *clp = server->nfs_client;
 272
 273                 spin_lock(&clp->cl_lock);
 274                 list_del_init(&lo->plh_layouts);
 275                 spin_unlock(&clp->cl_lock);
 276         }
 277         put_rpccred(lo->plh_lc_cred);
 278         return ld->free_layout_hdr(lo);
 279 }
 280
 281 static void
 282 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 283 {
 284         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 285         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 286         nfsi->layout = NULL;
 287         /* Reset MDS Threshold I/O counters */
 288         nfsi->write_io = 0;
 289         nfsi->read_io = 0;
 290 }
 291
 292 void
 293 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 294 {
 295         struct inode *inode;
 296
 297         if (!lo)
 298                 return;
 299         inode = lo->plh_inode;
 300         pnfs_layoutreturn_before_put_layout_hdr(lo);
 301
 302         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 303                 if (!list_empty(&lo->plh_segs))
 304                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 305                 pnfs_detach_layout_hdr(lo);
 306                 spin_unlock(&inode->i_lock);
 307                 pnfs_free_layout_hdr(lo);
 308         }
 309 }
 310
 311 static void
 312 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 313                          u32 seq)
 314 {
 315         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 316                 iomode = IOMODE_ANY;
 317         lo->plh_return_iomode = iomode;
 318         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 319         if (seq != 0) {
 320                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 321                 lo->plh_return_seq = seq;
 322         }
 323 }
 324
 325 static void
 326 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 327 {
 328         struct pnfs_layout_segment *lseg;
 329         lo->plh_return_iomode = 0;
 330         lo->plh_return_seq = 0;
 331         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 332         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 333                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 334                         continue;
 335                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 336         }
 337 }
 338
 339 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 340 {
 341         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 342         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 343         smp_mb__after_atomic();
 344         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 345         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 346 }
 347
 348 static void
 349 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 350                 struct list_head *free_me)
 351 {
 352         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 353         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 354         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 355                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 356         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 357                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 358 }
 359
 360 /*
 361  * Update the seqid of a layout stateid
 362  */
 363 bool nfs4_refresh_layout_stateid(nfs4_stateid *dst, struct inode *inode)
 364 {
 365         struct pnfs_layout_hdr *lo;
 366         bool ret = false;
 367
 368         spin_lock(&inode->i_lock);
 369         lo = NFS_I(inode)->layout;
 370         if (lo && nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 371                 dst->seqid = lo->plh_stateid.seqid;
 372                 ret = true;
 373         }
 374         spin_unlock(&inode->i_lock);
 375         return ret;
 376 }
 377
 378 /*
 379  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 380  *
 381  * In order to continue using the pnfs_layout_hdr, a full recovery
 382  * is required.
 383  * Note that caller must hold inode->i_lock.
 384  */
 385 int
 386 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 387                 struct list_head *lseg_list)
 388 {
 389         struct pnfs_layout_range range = {
 390                 .iomode = IOMODE_ANY,
 391                 .offset = 0,
 392                 .length = NFS4_MAX_UINT64,
 393         };
 394         struct pnfs_layout_segment *lseg, *next;
 395
 396         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 397         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 398                 pnfs_clear_lseg_state(lseg, lseg_list);
 399         pnfs_clear_layoutreturn_info(lo);
 400         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 401         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 402             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 403                 pnfs_clear_layoutreturn_waitbit(lo);
 404         return !list_empty(&lo->plh_segs);
 405 }
 406
 407 static int
 408 pnfs_iomode_to_fail_bit(u32 iomode)
 409 {
 410         return iomode == IOMODE_RW ?
 411                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 412 }
 413
 414 static void
 415 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 416 {
 417         lo->plh_retry_timestamp = jiffies;
 418         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 419                 refcount_inc(&lo->plh_refcount);
 420 }
 421
 422 static void
 423 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 424 {
 425         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 426                 refcount_dec(&lo->plh_refcount);
 427 }
 428
 429 static void
 430 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 431 {
 432         struct inode *inode = lo->plh_inode;
 433         struct pnfs_layout_range range = {
 434                 .iomode = iomode,
 435                 .offset = 0,
 436                 .length = NFS4_MAX_UINT64,
 437         };
 438         LIST_HEAD(head);
 439
 440         spin_lock(&inode->i_lock);
 441         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 442         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 443         spin_unlock(&inode->i_lock);
 444         pnfs_free_lseg_list(&head);
 445         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 446                         iomode == IOMODE_RW ?  "RW" : "READ");
 447 }
 448
 449 static bool
 450 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 451 {
 452         unsigned long start, end;
 453         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 454
 455         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 456                 return false;
 457         end = jiffies;
 458         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 459         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 460                 /* It is time to retry the failed layoutgets */
 461                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 462                 return false;
 463         }
 464         return true;
 465 }
 466
 467 static void
 468 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 469                 const struct pnfs_layout_range *range,
 470                 const nfs4_stateid *stateid)
 471 {
 472         INIT_LIST_HEAD(&lseg->pls_list);
 473         INIT_LIST_HEAD(&lseg->pls_lc_list);
 474         refcount_set(&lseg->pls_refcount, 1);
 475         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 476         lseg->pls_layout = lo;
 477         lseg->pls_range = *range;
 478         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 479 }
 480
 481 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 482 {
 483         if (lseg != NULL) {
 484                 struct inode *inode = lseg->pls_layout->plh_inode;
 485                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 486         }
 487 }
 488
 489 static void
 490 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 491                 struct pnfs_layout_segment *lseg)
 492 {
 493         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 494         list_del_init(&lseg->pls_list);
 495         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 496         refcount_dec(&lo->plh_refcount);
 497         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 498                 return;
 499         if (list_empty(&lo->plh_segs) &&
 500             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 501             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 502                 if (atomic_read(&lo->plh_outstanding) == 0)
 503                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 504                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 505         }
 506 }
 507
 508 static bool
 509 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 510                 struct pnfs_layout_segment *lseg)
 511 {
 512         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 513             pnfs_layout_is_valid(lo)) {
 514                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 515                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 516                 return true;
 517         }
 518         return false;
 519 }
 520
 521 void
 522 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 523 {
 524         struct pnfs_layout_hdr *lo;
 525         struct inode *inode;
 526
 527         if (!lseg)
 528                 return;
 529
 530         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 531                 refcount_read(&lseg->pls_refcount),
 532                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 533
 534         lo = lseg->pls_layout;
 535         inode = lo->plh_inode;
 536
 537         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 538                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 539                         spin_unlock(&inode->i_lock);
 540                         return;
 541                 }
 542                 pnfs_get_layout_hdr(lo);
 543                 pnfs_layout_remove_lseg(lo, lseg);
 544                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 545                         lseg = NULL;
 546                 spin_unlock(&inode->i_lock);
 547                 pnfs_free_lseg(lseg);
 548                 pnfs_put_layout_hdr(lo);
 549         }
 550 }
 551 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 552
 553 /*
 554  * is l2 fully contained in l1?
 555  *   start1                             end1
 556  *   [----------------------------------)
 557  *           start2           end2
 558  *           [----------------)
 559  */
 560 static bool
 561 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 562                  const struct pnfs_layout_range *l2)
 563 {
 564         u64 start1 = l1->offset;
 565         u64 end1 = pnfs_end_offset(start1, l1->length);
 566         u64 start2 = l2->offset;
 567         u64 end2 = pnfs_end_offset(start2, l2->length);
 568
 569         return (start1 <= start2) && (end1 >= end2);
 570 }
 571
 572 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 573                 struct list_head *tmp_list)
 574 {
 575         if (!refcount_dec_and_test(&lseg->pls_refcount))
 576                 return false;
 577         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 578         list_add(&lseg->pls_list, tmp_list);
 579         return true;
 580 }
 581
 582 /* Returns 1 if lseg is removed from list, 0 otherwise */
 583 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 584                              struct list_head *tmp_list)
 585 {
 586         int rv = 0;
 587
 588         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 589                 /* Remove the reference keeping the lseg in the
 590                  * list.  It will now be removed when all
 591                  * outstanding io is finished.
 592                  */
 593                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 594                         refcount_read(&lseg->pls_refcount));
 595                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 596                         rv = 1;
 597         }
 598         return rv;
 599 }
 600
 601 /*
 602  * Compare 2 layout stateid sequence ids, to see which is newer,
 603  * taking into account wraparound issues.
 604  */
 605 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 606 {
 607         return (s32)(s1 - s2) > 0;
 608 }
 609
 610 static bool
 611 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 612                  const struct pnfs_layout_range *recall_range)
 613 {
 614         return (recall_range->iomode == IOMODE_ANY ||
 615                 lseg_range->iomode == recall_range->iomode) &&
 616                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 617 }
 618
 619 static bool
 620 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 621                 const struct pnfs_layout_range *recall_range,
 622                 u32 seq)
 623 {
 624         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 625                 return false;
 626         if (recall_range == NULL)
 627                 return true;
 628         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 629 }
 630
 631 /**
 632  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 633  * @lo: layout header containing the lsegs
 634  * @tmp_list: list head where doomed lsegs should go
 635  * @recall_range: optional recall range argument to match (may be NULL)
 636  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 637  *
 638  * Walk the list of lsegs in the layout header, and tear down any that should
 639  * be destroyed. If "recall_range" is specified then the segment must match
 640  * that range. If "seq" is non-zero, then only match segments that were handed
 641  * out at or before that sequence.
 642  *
 643  * Returns number of matching invalid lsegs remaining in list after scanning
 644  * it and purging them.
 645  */
 646 int
 647 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 648                             struct list_head *tmp_list,
 649                             const struct pnfs_layout_range *recall_range,
 650                             u32 seq)
 651 {
 652         struct pnfs_layout_segment *lseg, *next;
 653         int remaining = 0;
 654
 655         dprintk("%s:Begin lo %p\n", __func__, lo);
 656
 657         if (list_empty(&lo->plh_segs))
 658                 return 0;
 659         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 660                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 661                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 662                                 "offset %llu length %llu\n", __func__,
 663                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 664                                 lseg->pls_range.offset, lseg->pls_range.length);
 665                         if (!mark_lseg_invalid(lseg, tmp_list))
 666                                 remaining++;
 667                 }
 668         dprintk("%s:Return %i\n", __func__, remaining);
 669         return remaining;
 670 }
 671
 672 static void
 673 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 674                 struct list_head *free_me,
 675                 const struct pnfs_layout_range *range,
 676                 u32 seq)
 677 {
 678         struct pnfs_layout_segment *lseg, *next;
 679
 680         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 681                 if (pnfs_match_lseg_recall(lseg, range, seq))
 682                         list_move_tail(&lseg->pls_list, free_me);
 683         }
 684 }
 685
 686 /* note free_me must contain lsegs from a single layout_hdr */
 687 void
 688 pnfs_free_lseg_list(struct list_head *free_me)
 689 {
 690         struct pnfs_layout_segment *lseg, *tmp;
 691
 692         if (list_empty(free_me))
 693                 return;
 694
 695         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 696                 list_del(&lseg->pls_list);
 697                 pnfs_free_lseg(lseg);
 698         }
 699 }
 700
 701 void
 702 pnfs_destroy_layout(struct nfs_inode *nfsi)
 703 {
 704         struct pnfs_layout_hdr *lo;
 705         LIST_HEAD(tmp_list);
 706
 707         spin_lock(&nfsi->vfs_inode.i_lock);
 708         lo = nfsi->layout;
 709         if (lo) {
 710                 pnfs_get_layout_hdr(lo);
 711                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 712                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 713                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 714                 spin_unlock(&nfsi->vfs_inode.i_lock);
 715                 pnfs_free_lseg_list(&tmp_list);
 716                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 717                 pnfs_put_layout_hdr(lo);
 718         } else
 719                 spin_unlock(&nfsi->vfs_inode.i_lock);
 720 }
 721 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 722
 723 static bool
 724 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 725                 struct list_head *layout_list)
 726 {
 727         struct pnfs_layout_hdr *lo;
 728         bool ret = false;
 729
 730         spin_lock(&inode->i_lock);
 731         lo = NFS_I(inode)->layout;
 732         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 733                 pnfs_get_layout_hdr(lo);
 734                 list_add(&lo->plh_bulk_destroy, layout_list);
 735                 ret = true;
 736         }
 737         spin_unlock(&inode->i_lock);
 738         return ret;
 739 }
 740
 741 /* Caller must hold rcu_read_lock and clp->cl_lock */
 742 static int
 743 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 744                 struct nfs_server *server,
 745                 struct list_head *layout_list)
 746 {
 747         struct pnfs_layout_hdr *lo, *next;
 748         struct inode *inode;
 749
 750         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 751                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
 752                         continue;
 753                 inode = igrab(lo->plh_inode);
 754                 if (inode == NULL)
 755                         continue;
 756                 list_del_init(&lo->plh_layouts);
 757                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 758                         continue;
 759                 rcu_read_unlock();
 760                 spin_unlock(&clp->cl_lock);
 761                 iput(inode);
 762                 spin_lock(&clp->cl_lock);
 763                 rcu_read_lock();
 764                 return -EAGAIN;
 765         }
 766         return 0;
 767 }
 768
 769 static int
 770 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 771                 bool is_bulk_recall)
 772 {
 773         struct pnfs_layout_hdr *lo;
 774         struct inode *inode;
 775         LIST_HEAD(lseg_list);
 776         int ret = 0;
 777
 778         while (!list_empty(layout_list)) {
 779                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 780                                 plh_bulk_destroy);
 781                 dprintk("%s freeing layout for inode %lu\n", __func__,
 782                         lo->plh_inode->i_ino);
 783                 inode = lo->plh_inode;
 784
 785                 pnfs_layoutcommit_inode(inode, false);
 786
 787                 spin_lock(&inode->i_lock);
 788                 list_del_init(&lo->plh_bulk_destroy);
 789                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 790                         if (is_bulk_recall)
 791                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 792                         ret = -EAGAIN;
 793                 }
 794                 spin_unlock(&inode->i_lock);
 795                 pnfs_free_lseg_list(&lseg_list);
 796                 /* Free all lsegs that are attached to commit buckets */
 797                 nfs_commit_inode(inode, 0);
 798                 pnfs_put_layout_hdr(lo);
 799                 iput(inode);
 800         }
 801         return ret;
 802 }
 803
 804 int
 805 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 806                 struct nfs_fsid *fsid,
 807                 bool is_recall)
 808 {
 809         struct nfs_server *server;
 810         LIST_HEAD(layout_list);
 811
 812         spin_lock(&clp->cl_lock);
 813         rcu_read_lock();
 814 restart:
 815         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 816                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 817                         continue;
 818                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 819                                 server,
 820                                 &layout_list) != 0)
 821                         goto restart;
 822         }
 823         rcu_read_unlock();
 824         spin_unlock(&clp->cl_lock);
 825
 826         if (list_empty(&layout_list))
 827                 return 0;
 828         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 829 }
 830
 831 int
 832 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 833                 bool is_recall)
 834 {
 835         struct nfs_server *server;
 836         LIST_HEAD(layout_list);
 837
 838         spin_lock(&clp->cl_lock);
 839         rcu_read_lock();
 840 restart:
 841         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 842                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 843                                         server,
 844                                         &layout_list) != 0)
 845                         goto restart;
 846         }
 847         rcu_read_unlock();
 848         spin_unlock(&clp->cl_lock);
 849
 850         if (list_empty(&layout_list))
 851                 return 0;
 852         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 853 }
 854
 855 /*
 856  * Called by the state manger to remove all layouts established under an
 857  * expired lease.
 858  */
 859 void
 860 pnfs_destroy_all_layouts(struct nfs_client *clp)
 861 {
 862         nfs4_deviceid_mark_client_invalid(clp);
 863         nfs4_deviceid_purge_client(clp);
 864
 865         pnfs_destroy_layouts_byclid(clp, false);
 866 }
 867
 868 /* update lo->plh_stateid with new if is more recent */
 869 void
 870 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 871                         bool update_barrier)
 872 {
 873         u32 oldseq, newseq, new_barrier = 0;
 874
 875         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 876         newseq = be32_to_cpu(new->seqid);
 877
 878         if (!pnfs_layout_is_valid(lo)) {
 879                 nfs4_stateid_copy(&lo->plh_stateid, new);
 880                 lo->plh_barrier = newseq;
 881                 pnfs_clear_layoutreturn_info(lo);
 882                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 883                 return;
 884         }
 885         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 886                 nfs4_stateid_copy(&lo->plh_stateid, new);
 887                 /*
 888                  * Because of wraparound, we want to keep the barrier
 889                  * "close" to the current seqids.
 890                  */
 891                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 892         }
 893         if (update_barrier)
 894                 new_barrier = be32_to_cpu(new->seqid);
 895         else if (new_barrier == 0)
 896                 return;
 897         if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 898                 lo->plh_barrier = new_barrier;
 899 }
 900
 901 static bool
 902 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 903                 const nfs4_stateid *stateid)
 904 {
 905         u32 seqid = be32_to_cpu(stateid->seqid);
 906
 907         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 908 }
 909
 910 /* lget is set to 1 if called from inside send_layoutget call chain */
 911 static bool
 912 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 913 {
 914         return lo->plh_block_lgets ||
 915                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 916 }
 917
 918 /*
 919  * Get layout from server.
 920  *    for now, assume that whole file layouts are requested.
 921  *    arg->offset: 0
 922  *    arg->length: all ones
 923  */
 924 static struct pnfs_layout_segment *
 925 send_layoutget(struct pnfs_layout_hdr *lo,
 926            struct nfs_open_context *ctx,
 927            nfs4_stateid *stateid,
 928            const struct pnfs_layout_range *range,
 929            long *timeout, gfp_t gfp_flags)
 930 {
 931         struct inode *ino = lo->plh_inode;
 932         struct nfs_server *server = NFS_SERVER(ino);
 933         struct nfs4_layoutget *lgp;
 934         loff_t i_size;
 935
 936         dprintk("--> %s\n", __func__);
 937
 938         /*
 939          * Synchronously retrieve layout information from server and
 940          * store in lseg. If we race with a concurrent seqid morphing
 941          * op, then re-send the LAYOUTGET.
 942          */
 943         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 944         if (lgp == NULL)
 945                 return ERR_PTR(-ENOMEM);
 946
 947         i_size = i_size_read(ino);
 948
 949         lgp->args.minlength = PAGE_SIZE;
 950         if (lgp->args.minlength > range->length)
 951                 lgp->args.minlength = range->length;
 952         if (range->iomode == IOMODE_READ) {
 953                 if (range->offset >= i_size)
 954                         lgp->args.minlength = 0;
 955                 else if (i_size - range->offset < lgp->args.minlength)
 956                         lgp->args.minlength = i_size - range->offset;
 957         }
 958         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 959         pnfs_copy_range(&lgp->args.range, range);
 960         lgp->args.type = server->pnfs_curr_ld->id;
 961         lgp->args.inode = ino;
 962         lgp->args.ctx = get_nfs_open_context(ctx);
 963         nfs4_stateid_copy(&lgp->args.stateid, stateid);
 964         lgp->gfp_flags = gfp_flags;
 965         lgp->cred = lo->plh_lc_cred;
 966
 967         return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
 968 }
 969
 970 static void pnfs_clear_layoutcommit(struct inode *inode,
 971                 struct list_head *head)
 972 {
 973         struct nfs_inode *nfsi = NFS_I(inode);
 974         struct pnfs_layout_segment *lseg, *tmp;
 975
 976         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 977                 return;
 978         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 979                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 980                         continue;
 981                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 982         }
 983 }
 984
 985 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
 986                 const nfs4_stateid *arg_stateid,
 987                 const struct pnfs_layout_range *range,
 988                 const nfs4_stateid *stateid)
 989 {
 990         struct inode *inode = lo->plh_inode;
 991         LIST_HEAD(freeme);
 992
 993         spin_lock(&inode->i_lock);
 994         if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
 995             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
 996                 goto out_unlock;
 997         if (stateid) {
 998                 u32 seq = be32_to_cpu(arg_stateid->seqid);
 999
1000                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1001                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1002                 pnfs_set_layout_stateid(lo, stateid, true);
1003         } else
1004                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1005 out_unlock:
1006         pnfs_clear_layoutreturn_waitbit(lo);
1007         spin_unlock(&inode->i_lock);
1008         pnfs_free_lseg_list(&freeme);
1009
1010 }
1011
1012 static bool
1013 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1014                 nfs4_stateid *stateid,
1015                 enum pnfs_iomode *iomode)
1016 {
1017         /* Serialise LAYOUTGET/LAYOUTRETURN */
1018         if (atomic_read(&lo->plh_outstanding) != 0)
1019                 return false;
1020         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1021                 return false;
1022         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1023         pnfs_get_layout_hdr(lo);
1024         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1025                 if (stateid != NULL) {
1026                         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1027                         if (lo->plh_return_seq != 0)
1028                                 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1029                 }
1030                 if (iomode != NULL)
1031                         *iomode = lo->plh_return_iomode;
1032                 pnfs_clear_layoutreturn_info(lo);
1033                 return true;
1034         }
1035         if (stateid != NULL)
1036                 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1037         if (iomode != NULL)
1038                 *iomode = IOMODE_ANY;
1039         return true;
1040 }
1041
1042 static void
1043 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1044                 struct pnfs_layout_hdr *lo,
1045                 const nfs4_stateid *stateid,
1046                 enum pnfs_iomode iomode)
1047 {
1048         struct inode *inode = lo->plh_inode;
1049
1050         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1051         args->inode = inode;
1052         args->range.iomode = iomode;
1053         args->range.offset = 0;
1054         args->range.length = NFS4_MAX_UINT64;
1055         args->layout = lo;
1056         nfs4_stateid_copy(&args->stateid, stateid);
1057 }
1058
1059 static int
1060 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1061                        enum pnfs_iomode iomode, bool sync)
1062 {
1063         struct inode *ino = lo->plh_inode;
1064         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1065         struct nfs4_layoutreturn *lrp;
1066         int status = 0;
1067
1068         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1069         if (unlikely(lrp == NULL)) {
1070                 status = -ENOMEM;
1071                 spin_lock(&ino->i_lock);
1072                 pnfs_clear_layoutreturn_waitbit(lo);
1073                 spin_unlock(&ino->i_lock);
1074                 pnfs_put_layout_hdr(lo);
1075                 goto out;
1076         }
1077
1078         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1079         lrp->args.ld_private = &lrp->ld_private;
1080         lrp->clp = NFS_SERVER(ino)->nfs_client;
1081         lrp->cred = lo->plh_lc_cred;
1082         if (ld->prepare_layoutreturn)
1083                 ld->prepare_layoutreturn(&lrp->args);
1084
1085         status = nfs4_proc_layoutreturn(lrp, sync);
1086 out:
1087         dprintk("<-- %s status: %d\n", __func__, status);
1088         return status;
1089 }
1090
1091 /* Return true if layoutreturn is needed */
1092 static bool
1093 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1094 {
1095         struct pnfs_layout_segment *s;
1096
1097         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1098                 return false;
1099
1100         /* Defer layoutreturn until all lsegs are done */
1101         list_for_each_entry(s, &lo->plh_segs, pls_list) {
1102                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1103                         return false;
1104         }
1105
1106         return true;
1107 }
1108
1109 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1110 {
1111         struct inode *inode= lo->plh_inode;
1112
1113         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1114                 return;
1115         spin_lock(&inode->i_lock);
1116         if (pnfs_layout_need_return(lo)) {
1117                 nfs4_stateid stateid;
1118                 enum pnfs_iomode iomode;
1119                 bool send;
1120
1121                 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1122                 spin_unlock(&inode->i_lock);
1123                 if (send) {
1124                         /* Send an async layoutreturn so we dont deadlock */
1125                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1126                 }
1127         } else
1128                 spin_unlock(&inode->i_lock);
1129 }
1130
1131 /*
1132  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1133  * when the layout segment list is empty.
1134  *
1135  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1136  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1137  * deviceid is marked invalid.
1138  */
1139 int
1140 _pnfs_return_layout(struct inode *ino)
1141 {
1142         struct pnfs_layout_hdr *lo = NULL;
1143         struct nfs_inode *nfsi = NFS_I(ino);
1144         LIST_HEAD(tmp_list);
1145         nfs4_stateid stateid;
1146         int status = 0;
1147         bool send;
1148
1149         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1150
1151         spin_lock(&ino->i_lock);
1152         lo = nfsi->layout;
1153         if (!lo) {
1154                 spin_unlock(&ino->i_lock);
1155                 dprintk("NFS: %s no layout to return\n", __func__);
1156                 goto out;
1157         }
1158         /* Reference matched in nfs4_layoutreturn_release */
1159         pnfs_get_layout_hdr(lo);
1160         /* Is there an outstanding layoutreturn ? */
1161         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1162                 spin_unlock(&ino->i_lock);
1163                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1164                                         TASK_UNINTERRUPTIBLE))
1165                         goto out_put_layout_hdr;
1166                 spin_lock(&ino->i_lock);
1167         }
1168         pnfs_clear_layoutcommit(ino, &tmp_list);
1169         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1170
1171         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1172                 struct pnfs_layout_range range = {
1173                         .iomode         = IOMODE_ANY,
1174                         .offset         = 0,
1175                         .length         = NFS4_MAX_UINT64,
1176                 };
1177                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1178         }
1179
1180         /* Don't send a LAYOUTRETURN if list was initially empty */
1181         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1182                 spin_unlock(&ino->i_lock);
1183                 dprintk("NFS: %s no layout segments to return\n", __func__);
1184                 goto out_put_layout_hdr;
1185         }
1186
1187         send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1188         spin_unlock(&ino->i_lock);
1189         if (send)
1190                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1191 out_put_layout_hdr:
1192         pnfs_free_lseg_list(&tmp_list);
1193         pnfs_put_layout_hdr(lo);
1194 out:
1195         dprintk("<-- %s status: %d\n", __func__, status);
1196         return status;
1197 }
1198
1199 int
1200 pnfs_commit_and_return_layout(struct inode *inode)
1201 {
1202         struct pnfs_layout_hdr *lo;
1203         int ret;
1204
1205         spin_lock(&inode->i_lock);
1206         lo = NFS_I(inode)->layout;
1207         if (lo == NULL) {
1208                 spin_unlock(&inode->i_lock);
1209                 return 0;
1210         }
1211         pnfs_get_layout_hdr(lo);
1212         /* Block new layoutgets and read/write to ds */
1213         lo->plh_block_lgets++;
1214         spin_unlock(&inode->i_lock);
1215         filemap_fdatawait(inode->i_mapping);
1216         ret = pnfs_layoutcommit_inode(inode, true);
1217         if (ret == 0)
1218                 ret = _pnfs_return_layout(inode);
1219         spin_lock(&inode->i_lock);
1220         lo->plh_block_lgets--;
1221         spin_unlock(&inode->i_lock);
1222         pnfs_put_layout_hdr(lo);
1223         return ret;
1224 }
1225
1226 bool pnfs_roc(struct inode *ino,
1227                 struct nfs4_layoutreturn_args *args,
1228                 struct nfs4_layoutreturn_res *res,
1229                 const struct rpc_cred *cred)
1230 {
1231         struct nfs_inode *nfsi = NFS_I(ino);
1232         struct nfs_open_context *ctx;
1233         struct nfs4_state *state;
1234         struct pnfs_layout_hdr *lo;
1235         struct pnfs_layout_segment *lseg, *next;
1236         nfs4_stateid stateid;
1237         enum pnfs_iomode iomode = 0;
1238         bool layoutreturn = false, roc = false;
1239         bool skip_read = false;
1240
1241         if (!nfs_have_layout(ino))
1242                 return false;
1243 retry:
1244         spin_lock(&ino->i_lock);
1245         lo = nfsi->layout;
1246         if (!lo || !pnfs_layout_is_valid(lo) ||
1247             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1248                 lo = NULL;
1249                 goto out_noroc;
1250         }
1251         pnfs_get_layout_hdr(lo);
1252         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1253                 spin_unlock(&ino->i_lock);
1254                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1255                                 TASK_UNINTERRUPTIBLE);
1256                 pnfs_put_layout_hdr(lo);
1257                 goto retry;
1258         }
1259
1260         /* no roc if we hold a delegation */
1261         if (nfs4_check_delegation(ino, FMODE_READ)) {
1262                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1263                         goto out_noroc;
1264                 skip_read = true;
1265         }
1266
1267         list_for_each_entry(ctx, &nfsi->open_files, list) {
1268                 state = ctx->state;
1269                 if (state == NULL)
1270                         continue;
1271                 /* Don't return layout if there is open file state */
1272                 if (state->state & FMODE_WRITE)
1273                         goto out_noroc;
1274                 if (state->state & FMODE_READ)
1275                         skip_read = true;
1276         }
1277
1278
1279         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1280                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1281                         continue;
1282                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1283                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1284                         continue;
1285                 /*
1286                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1287                  * doesn't invalidate the layout for us.
1288                  */
1289                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1290                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1291                         continue;
1292                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1293         }
1294
1295         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1296                 goto out_noroc;
1297
1298         /* ROC in two conditions:
1299          * 1. there are ROC lsegs
1300          * 2. we don't send layoutreturn
1301          */
1302         /* lo ref dropped in pnfs_roc_release() */
1303         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1304         /* If the creds don't match, we can't compound the layoutreturn */
1305         if (!layoutreturn || cred != lo->plh_lc_cred)
1306                 goto out_noroc;
1307
1308         roc = layoutreturn;
1309         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1310         res->lrs_present = 0;
1311         layoutreturn = false;
1312
1313 out_noroc:
1314         spin_unlock(&ino->i_lock);
1315         pnfs_layoutcommit_inode(ino, true);
1316         if (roc) {
1317                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1318                 if (ld->prepare_layoutreturn)
1319                         ld->prepare_layoutreturn(args);
1320                 pnfs_put_layout_hdr(lo);
1321                 return true;
1322         }
1323         if (layoutreturn)
1324                 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1325         pnfs_put_layout_hdr(lo);
1326         return false;
1327 }
1328
1329 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1330                 struct nfs4_layoutreturn_res *res,
1331                 int ret)
1332 {
1333         struct pnfs_layout_hdr *lo = args->layout;
1334         const nfs4_stateid *arg_stateid = NULL;
1335         const nfs4_stateid *res_stateid = NULL;
1336         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1337
1338         if (ret == 0) {
1339                 arg_stateid = &args->stateid;
1340                 if (res->lrs_present)
1341                         res_stateid = &res->stateid;
1342         }
1343         pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1344                         res_stateid);
1345         if (ld_private && ld_private->ops && ld_private->ops->free)
1346                 ld_private->ops->free(ld_private);
1347         pnfs_put_layout_hdr(lo);
1348         trace_nfs4_layoutreturn_on_close(args->inode, 0);
1349 }
1350
1351 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1352 {
1353         struct nfs_inode *nfsi = NFS_I(ino);
1354         struct pnfs_layout_hdr *lo;
1355         bool sleep = false;
1356
1357         /* we might not have grabbed lo reference. so need to check under
1358          * i_lock */
1359         spin_lock(&ino->i_lock);
1360         lo = nfsi->layout;
1361         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1362                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1363                 sleep = true;
1364         }
1365         spin_unlock(&ino->i_lock);
1366         return sleep;
1367 }
1368
1369 /*
1370  * Compare two layout segments for sorting into layout cache.
1371  * We want to preferentially return RW over RO layouts, so ensure those
1372  * are seen first.
1373  */
1374 static s64
1375 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1376            const struct pnfs_layout_range *l2)
1377 {
1378         s64 d;
1379
1380         /* high offset > low offset */
1381         d = l1->offset - l2->offset;
1382         if (d)
1383                 return d;
1384
1385         /* short length > long length */
1386         d = l2->length - l1->length;
1387         if (d)
1388                 return d;
1389
1390         /* read > read/write */
1391         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1392 }
1393
1394 static bool
1395 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1396                 const struct pnfs_layout_range *l2)
1397 {
1398         return pnfs_lseg_range_cmp(l1, l2) > 0;
1399 }
1400
1401 static bool
1402 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1403                 struct pnfs_layout_segment *old)
1404 {
1405         return false;
1406 }
1407
1408 void
1409 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1410                    struct pnfs_layout_segment *lseg,
1411                    bool (*is_after)(const struct pnfs_layout_range *,
1412                            const struct pnfs_layout_range *),
1413                    bool (*do_merge)(struct pnfs_layout_segment *,
1414                            struct pnfs_layout_segment *),
1415                    struct list_head *free_me)
1416 {
1417         struct pnfs_layout_segment *lp, *tmp;
1418
1419         dprintk("%s:Begin\n", __func__);
1420
1421         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1422                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1423                         continue;
1424                 if (do_merge(lseg, lp)) {
1425                         mark_lseg_invalid(lp, free_me);
1426                         continue;
1427                 }
1428                 if (is_after(&lseg->pls_range, &lp->pls_range))
1429                         continue;
1430                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1431                 dprintk("%s: inserted lseg %p "
1432                         "iomode %d offset %llu length %llu before "
1433                         "lp %p iomode %d offset %llu length %llu\n",
1434                         __func__, lseg, lseg->pls_range.iomode,
1435                         lseg->pls_range.offset, lseg->pls_range.length,
1436                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1437                         lp->pls_range.length);
1438                 goto out;
1439         }
1440         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1441         dprintk("%s: inserted lseg %p "
1442                 "iomode %d offset %llu length %llu at tail\n",
1443                 __func__, lseg, lseg->pls_range.iomode,
1444                 lseg->pls_range.offset, lseg->pls_range.length);
1445 out:
1446         pnfs_get_layout_hdr(lo);
1447
1448         dprintk("%s:Return\n", __func__);
1449 }
1450 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1451
1452 static void
1453 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1454                    struct pnfs_layout_segment *lseg,
1455                    struct list_head *free_me)
1456 {
1457         struct inode *inode = lo->plh_inode;
1458         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1459
1460         if (ld->add_lseg != NULL)
1461                 ld->add_lseg(lo, lseg, free_me);
1462         else
1463                 pnfs_generic_layout_insert_lseg(lo, lseg,
1464                                 pnfs_lseg_range_is_after,
1465                                 pnfs_lseg_no_merge,
1466                                 free_me);
1467 }
1468
1469 static struct pnfs_layout_hdr *
1470 alloc_init_layout_hdr(struct inode *ino,
1471                       struct nfs_open_context *ctx,
1472                       gfp_t gfp_flags)
1473 {
1474         struct pnfs_layout_hdr *lo;
1475
1476         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1477         if (!lo)
1478                 return NULL;
1479         refcount_set(&lo->plh_refcount, 1);
1480         INIT_LIST_HEAD(&lo->plh_layouts);
1481         INIT_LIST_HEAD(&lo->plh_segs);
1482         INIT_LIST_HEAD(&lo->plh_return_segs);
1483         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1484         lo->plh_inode = ino;
1485         lo->plh_lc_cred = get_rpccred(ctx->cred);
1486         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1487         return lo;
1488 }
1489
1490 static struct pnfs_layout_hdr *
1491 pnfs_find_alloc_layout(struct inode *ino,
1492                        struct nfs_open_context *ctx,
1493                        gfp_t gfp_flags)
1494         __releases(&ino->i_lock)
1495         __acquires(&ino->i_lock)
1496 {
1497         struct nfs_inode *nfsi = NFS_I(ino);
1498         struct pnfs_layout_hdr *new = NULL;
1499
1500         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1501
1502         if (nfsi->layout != NULL)
1503                 goto out_existing;
1504         spin_unlock(&ino->i_lock);
1505         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1506         spin_lock(&ino->i_lock);
1507
1508         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1509                 nfsi->layout = new;
1510                 return new;
1511         } else if (new != NULL)
1512                 pnfs_free_layout_hdr(new);
1513 out_existing:
1514         pnfs_get_layout_hdr(nfsi->layout);
1515         return nfsi->layout;
1516 }
1517
1518 /*
1519  * iomode matching rules:
1520  * iomode       lseg    strict match
1521  *                      iomode
1522  * -----        -----   ------ -----
1523  * ANY          READ    N/A    true
1524  * ANY          RW      N/A    true
1525  * RW           READ    N/A    false
1526  * RW           RW      N/A    true
1527  * READ         READ    N/A    true
1528  * READ         RW      true   false
1529  * READ         RW      false  true
1530  */
1531 static bool
1532 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1533                  const struct pnfs_layout_range *range,
1534                  bool strict_iomode)
1535 {
1536         struct pnfs_layout_range range1;
1537
1538         if ((range->iomode == IOMODE_RW &&
1539              ls_range->iomode != IOMODE_RW) ||
1540             (range->iomode != ls_range->iomode &&
1541              strict_iomode) ||
1542             !pnfs_lseg_range_intersecting(ls_range, range))
1543                 return 0;
1544
1545         /* range1 covers only the first byte in the range */
1546         range1 = *range;
1547         range1.length = 1;
1548         return pnfs_lseg_range_contained(ls_range, &range1);
1549 }
1550
1551 /*
1552  * lookup range in layout
1553  */
1554 static struct pnfs_layout_segment *
1555 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1556                 struct pnfs_layout_range *range,
1557                 bool strict_iomode)
1558 {
1559         struct pnfs_layout_segment *lseg, *ret = NULL;
1560
1561         dprintk("%s:Begin\n", __func__);
1562
1563         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1564                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1565                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1566                     pnfs_lseg_range_match(&lseg->pls_range, range,
1567                                           strict_iomode)) {
1568                         ret = pnfs_get_lseg(lseg);
1569                         break;
1570                 }
1571         }
1572
1573         dprintk("%s:Return lseg %p ref %d\n",
1574                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1575         return ret;
1576 }
1577
1578 /*
1579  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1580  * to the MDS or over pNFS
1581  *
1582  * The nfs_inode read_io and write_io fields are cumulative counters reset
1583  * when there are no layout segments. Note that in pnfs_update_layout iomode
1584  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1585  * WRITE request.
1586  *
1587  * A return of true means use MDS I/O.
1588  *
1589  * From rfc 5661:
1590  * If a file's size is smaller than the file size threshold, data accesses
1591  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1592  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1593  * server.  If both file size and I/O size are provided, the client SHOULD
1594  * reach or exceed  both thresholds before sending its read or write
1595  * requests to the data server.
1596  */
1597 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1598                                      struct inode *ino, int iomode)
1599 {
1600         struct nfs4_threshold *t = ctx->mdsthreshold;
1601         struct nfs_inode *nfsi = NFS_I(ino);
1602         loff_t fsize = i_size_read(ino);
1603         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1604
1605         if (t == NULL)
1606                 return ret;
1607
1608         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1609                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1610
1611         switch (iomode) {
1612         case IOMODE_READ:
1613                 if (t->bm & THRESHOLD_RD) {
1614                         dprintk("%s fsize %llu\n", __func__, fsize);
1615                         size_set = true;
1616                         if (fsize < t->rd_sz)
1617                                 size = true;
1618                 }
1619                 if (t->bm & THRESHOLD_RD_IO) {
1620                         dprintk("%s nfsi->read_io %llu\n", __func__,
1621                                 nfsi->read_io);
1622                         io_set = true;
1623                         if (nfsi->read_io < t->rd_io_sz)
1624                                 io = true;
1625                 }
1626                 break;
1627         case IOMODE_RW:
1628                 if (t->bm & THRESHOLD_WR) {
1629                         dprintk("%s fsize %llu\n", __func__, fsize);
1630                         size_set = true;
1631                         if (fsize < t->wr_sz)
1632                                 size = true;
1633                 }
1634                 if (t->bm & THRESHOLD_WR_IO) {
1635                         dprintk("%s nfsi->write_io %llu\n", __func__,
1636                                 nfsi->write_io);
1637                         io_set = true;
1638                         if (nfsi->write_io < t->wr_io_sz)
1639                                 io = true;
1640                 }
1641                 break;
1642         }
1643         if (size_set && io_set) {
1644                 if (size && io)
1645                         ret = true;
1646         } else if (size || io)
1647                 ret = true;
1648
1649         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1650         return ret;
1651 }
1652
1653 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1654 {
1655         /*
1656          * send layoutcommit as it can hold up layoutreturn due to lseg
1657          * reference
1658          */
1659         pnfs_layoutcommit_inode(lo->plh_inode, false);
1660         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1661                                    nfs_wait_bit_killable,
1662                                    TASK_UNINTERRUPTIBLE);
1663 }
1664
1665 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1666 {
1667         unsigned long *bitlock = &lo->plh_flags;
1668
1669         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1670         smp_mb__after_atomic();
1671         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1672 }
1673
1674 /*
1675  * Layout segment is retreived from the server if not cached.
1676  * The appropriate layout segment is referenced and returned to the caller.
1677  */
1678 struct pnfs_layout_segment *
1679 pnfs_update_layout(struct inode *ino,
1680                    struct nfs_open_context *ctx,
1681                    loff_t pos,
1682                    u64 count,
1683                    enum pnfs_iomode iomode,
1684                    bool strict_iomode,
1685                    gfp_t gfp_flags)
1686 {
1687         struct pnfs_layout_range arg = {
1688                 .iomode = iomode,
1689                 .offset = pos,
1690                 .length = count,
1691         };
1692         unsigned pg_offset;
1693         struct nfs_server *server = NFS_SERVER(ino);
1694         struct nfs_client *clp = server->nfs_client;
1695         struct pnfs_layout_hdr *lo = NULL;
1696         struct pnfs_layout_segment *lseg = NULL;
1697         nfs4_stateid stateid;
1698         long timeout = 0;
1699         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1700         bool first;
1701
1702         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1703                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1704                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1705                 goto out;
1706         }
1707
1708         if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1709                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1710                                  PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1711                 goto out;
1712         }
1713
1714         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1715                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1716                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1717                 goto out;
1718         }
1719
1720 lookup_again:
1721         nfs4_client_recover_expired_lease(clp);
1722         first = false;
1723         spin_lock(&ino->i_lock);
1724         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1725         if (lo == NULL) {
1726                 spin_unlock(&ino->i_lock);
1727                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1728                                  PNFS_UPDATE_LAYOUT_NOMEM);
1729                 goto out;
1730         }
1731
1732         /* Do we even need to bother with this? */
1733         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1734                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1735                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1736                 dprintk("%s matches recall, use MDS\n", __func__);
1737                 goto out_unlock;
1738         }
1739
1740         /* if LAYOUTGET already failed once we don't try again */
1741         if (pnfs_layout_io_test_failed(lo, iomode)) {
1742                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1743                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1744                 goto out_unlock;
1745         }
1746
1747         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1748         if (lseg) {
1749                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1750                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1751                 goto out_unlock;
1752         }
1753
1754         if (!nfs4_valid_open_stateid(ctx->state)) {
1755                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1756                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1757                 goto out_unlock;
1758         }
1759
1760         /*
1761          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1762          * stateid, or it has been invalidated, then we must use the open
1763          * stateid.
1764          */
1765         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1766
1767                 /*
1768                  * The first layoutget for the file. Need to serialize per
1769                  * RFC 5661 Errata 3208.
1770                  */
1771                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1772                                      &lo->plh_flags)) {
1773                         spin_unlock(&ino->i_lock);
1774                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1775                                     TASK_UNINTERRUPTIBLE);
1776                         pnfs_put_layout_hdr(lo);
1777                         dprintk("%s retrying\n", __func__);
1778                         goto lookup_again;
1779                 }
1780
1781                 first = true;
1782                 if (nfs4_select_rw_stateid(ctx->state,
1783                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1784                                         NULL, &stateid, NULL) != 0) {
1785                         trace_pnfs_update_layout(ino, pos, count,
1786                                         iomode, lo, lseg,
1787                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1788                         goto out_unlock;
1789                 }
1790         } else {
1791                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1792         }
1793
1794         /*
1795          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1796          * for LAYOUTRETURN even if first is true.
1797          */
1798         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1799                 spin_unlock(&ino->i_lock);
1800                 dprintk("%s wait for layoutreturn\n", __func__);
1801                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1802                         if (first)
1803                                 pnfs_clear_first_layoutget(lo);
1804                         pnfs_put_layout_hdr(lo);
1805                         dprintk("%s retrying\n", __func__);
1806                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1807                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
1808                         goto lookup_again;
1809                 }
1810                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1811                                 PNFS_UPDATE_LAYOUT_RETURN);
1812                 goto out_put_layout_hdr;
1813         }
1814
1815         if (pnfs_layoutgets_blocked(lo)) {
1816                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1817                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1818                 goto out_unlock;
1819         }
1820         atomic_inc(&lo->plh_outstanding);
1821         spin_unlock(&ino->i_lock);
1822
1823         if (list_empty(&lo->plh_layouts)) {
1824                 /* The lo must be on the clp list if there is any
1825                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1826                  */
1827                 spin_lock(&clp->cl_lock);
1828                 if (list_empty(&lo->plh_layouts))
1829                         list_add_tail(&lo->plh_layouts, &server->layouts);
1830                 spin_unlock(&clp->cl_lock);
1831         }
1832
1833         pg_offset = arg.offset & ~PAGE_MASK;
1834         if (pg_offset) {
1835                 arg.offset -= pg_offset;
1836                 arg.length += pg_offset;
1837         }
1838         if (arg.length != NFS4_MAX_UINT64)
1839                 arg.length = PAGE_ALIGN(arg.length);
1840
1841         lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1842         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1843                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1844         atomic_dec(&lo->plh_outstanding);
1845         if (IS_ERR(lseg)) {
1846                 switch(PTR_ERR(lseg)) {
1847                 case -EBUSY:
1848                         if (time_after(jiffies, giveup))
1849                                 lseg = NULL;
1850                         break;
1851                 case -ERECALLCONFLICT:
1852                         /* Huh? We hold no layouts, how is there a recall? */
1853                         if (first) {
1854                                 lseg = NULL;
1855                                 break;
1856                         }
1857                         /* Destroy the existing layout and start over */
1858                         if (time_after(jiffies, giveup))
1859                                 pnfs_destroy_layout(NFS_I(ino));
1860                         /* Fallthrough */
1861                 case -EAGAIN:
1862                         break;
1863                 default:
1864                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1865                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1866                                 lseg = NULL;
1867                         }
1868                         goto out_put_layout_hdr;
1869                 }
1870                 if (lseg) {
1871                         if (first)
1872                                 pnfs_clear_first_layoutget(lo);
1873                         trace_pnfs_update_layout(ino, pos, count,
1874                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1875                         pnfs_put_layout_hdr(lo);
1876                         goto lookup_again;
1877                 }
1878         } else {
1879                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1880         }
1881
1882 out_put_layout_hdr:
1883         if (first)
1884                 pnfs_clear_first_layoutget(lo);
1885         pnfs_put_layout_hdr(lo);
1886 out:
1887         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1888                         "(%s, offset: %llu, length: %llu)\n",
1889                         __func__, ino->i_sb->s_id,
1890                         (unsigned long long)NFS_FILEID(ino),
1891                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1892                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1893                         (unsigned long long)pos,
1894                         (unsigned long long)count);
1895         return lseg;
1896 out_unlock:
1897         spin_unlock(&ino->i_lock);
1898         goto out_put_layout_hdr;
1899 }
1900 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1901
1902 static bool
1903 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1904 {
1905         switch (range->iomode) {
1906         case IOMODE_READ:
1907         case IOMODE_RW:
1908                 break;
1909         default:
1910                 return false;
1911         }
1912         if (range->offset == NFS4_MAX_UINT64)
1913                 return false;
1914         if (range->length == 0)
1915                 return false;
1916         if (range->length != NFS4_MAX_UINT64 &&
1917             range->length > NFS4_MAX_UINT64 - range->offset)
1918                 return false;
1919         return true;
1920 }
1921
1922 struct pnfs_layout_segment *
1923 pnfs_layout_process(struct nfs4_layoutget *lgp)
1924 {
1925         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1926         struct nfs4_layoutget_res *res = &lgp->res;
1927         struct pnfs_layout_segment *lseg;
1928         struct inode *ino = lo->plh_inode;
1929         LIST_HEAD(free_me);
1930
1931         if (!pnfs_sanity_check_layout_range(&res->range))
1932                 return ERR_PTR(-EINVAL);
1933
1934         /* Inject layout blob into I/O device driver */
1935         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1936         if (IS_ERR_OR_NULL(lseg)) {
1937                 if (!lseg)
1938                         lseg = ERR_PTR(-ENOMEM);
1939
1940                 dprintk("%s: Could not allocate layout: error %ld\n",
1941                        __func__, PTR_ERR(lseg));
1942                 return lseg;
1943         }
1944
1945         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1946
1947         spin_lock(&ino->i_lock);
1948         if (pnfs_layoutgets_blocked(lo)) {
1949                 dprintk("%s forget reply due to state\n", __func__);
1950                 goto out_forget;
1951         }
1952
1953         if (!pnfs_layout_is_valid(lo)) {
1954                 /* We have a completely new layout */
1955                 pnfs_set_layout_stateid(lo, &res->stateid, true);
1956         } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1957                 /* existing state ID, make sure the sequence number matches. */
1958                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1959                         dprintk("%s forget reply due to sequence\n", __func__);
1960                         goto out_forget;
1961                 }
1962                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1963         } else {
1964                 /*
1965                  * We got an entirely new state ID.  Mark all segments for the
1966                  * inode invalid, and retry the layoutget
1967                  */
1968                 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1969                 goto out_forget;
1970         }
1971
1972         pnfs_get_lseg(lseg);
1973         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1974
1975
1976         if (res->return_on_close)
1977                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1978
1979         spin_unlock(&ino->i_lock);
1980         pnfs_free_lseg_list(&free_me);
1981         return lseg;
1982
1983 out_forget:
1984         spin_unlock(&ino->i_lock);
1985         lseg->pls_layout = lo;
1986         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1987         if (!pnfs_layout_is_valid(lo))
1988                 nfs_commit_inode(ino, 0);
1989         return ERR_PTR(-EAGAIN);
1990 }
1991
1992 /**
1993  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1994  * @lo: pointer to layout header
1995  * @tmp_list: list header to be used with pnfs_free_lseg_list()
1996  * @return_range: describe layout segment ranges to be returned
1997  *
1998  * This function is mainly intended for use by layoutrecall. It attempts
1999  * to free the layout segment immediately, or else to mark it for return
2000  * as soon as its reference count drops to zero.
2001  */
2002 int
2003 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2004                                 struct list_head *tmp_list,
2005                                 const struct pnfs_layout_range *return_range,
2006                                 u32 seq)
2007 {
2008         struct pnfs_layout_segment *lseg, *next;
2009         int remaining = 0;
2010
2011         dprintk("%s:Begin lo %p\n", __func__, lo);
2012
2013         if (list_empty(&lo->plh_segs))
2014                 return 0;
2015
2016         assert_spin_locked(&lo->plh_inode->i_lock);
2017
2018         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2019                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2020                         dprintk("%s: marking lseg %p iomode %d "
2021                                 "offset %llu length %llu\n", __func__,
2022                                 lseg, lseg->pls_range.iomode,
2023                                 lseg->pls_range.offset,
2024                                 lseg->pls_range.length);
2025                         if (mark_lseg_invalid(lseg, tmp_list))
2026                                 continue;
2027                         remaining++;
2028                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2029                 }
2030
2031         if (remaining)
2032                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2033
2034         return remaining;
2035 }
2036
2037 void pnfs_error_mark_layout_for_return(struct inode *inode,
2038                                        struct pnfs_layout_segment *lseg)
2039 {
2040         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2041         struct pnfs_layout_range range = {
2042                 .iomode = lseg->pls_range.iomode,
2043                 .offset = 0,
2044                 .length = NFS4_MAX_UINT64,
2045         };
2046         bool return_now = false;
2047
2048         spin_lock(&inode->i_lock);
2049         if (!pnfs_layout_is_valid(lo)) {
2050                 spin_unlock(&inode->i_lock);
2051                 return;
2052         }
2053         pnfs_set_plh_return_info(lo, range.iomode, 0);
2054         /*
2055          * mark all matching lsegs so that we are sure to have no live
2056          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2057          * for how it works.
2058          */
2059         if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2060                 nfs4_stateid stateid;
2061                 enum pnfs_iomode iomode;
2062
2063                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2064                 spin_unlock(&inode->i_lock);
2065                 if (return_now)
2066                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2067         } else {
2068                 spin_unlock(&inode->i_lock);
2069                 nfs_commit_inode(inode, 0);
2070         }
2071 }
2072 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2073
2074 void
2075 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2076 {
2077         if (pgio->pg_lseg == NULL ||
2078             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2079                 return;
2080         pnfs_put_lseg(pgio->pg_lseg);
2081         pgio->pg_lseg = NULL;
2082 }
2083 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2084
2085 /*
2086  * Check for any intersection between the request and the pgio->pg_lseg,
2087  * and if none, put this pgio->pg_lseg away.
2088  */
2089 static void
2090 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2091 {
2092         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2093                 pnfs_put_lseg(pgio->pg_lseg);
2094                 pgio->pg_lseg = NULL;
2095         }
2096 }
2097
2098 void
2099 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2100 {
2101         u64 rd_size = req->wb_bytes;
2102
2103         pnfs_generic_pg_check_layout(pgio);
2104         pnfs_generic_pg_check_range(pgio, req);
2105         if (pgio->pg_lseg == NULL) {
2106                 if (pgio->pg_dreq == NULL)
2107                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2108                 else
2109                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2110
2111                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2112                                                    req->wb_context,
2113                                                    req_offset(req),
2114                                                    rd_size,
2115                                                    IOMODE_READ,
2116                                                    false,
2117                                                    GFP_KERNEL);
2118                 if (IS_ERR(pgio->pg_lseg)) {
2119                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2120                         pgio->pg_lseg = NULL;
2121                         return;
2122                 }
2123         }
2124         /* If no lseg, fall back to read through mds */
2125         if (pgio->pg_lseg == NULL)
2126                 nfs_pageio_reset_read_mds(pgio);
2127
2128 }
2129 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2130
2131 void
2132 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2133                            struct nfs_page *req, u64 wb_size)
2134 {
2135         pnfs_generic_pg_check_layout(pgio);
2136         pnfs_generic_pg_check_range(pgio, req);
2137         if (pgio->pg_lseg == NULL) {
2138                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2139                                                    req->wb_context,
2140                                                    req_offset(req),
2141                                                    wb_size,
2142                                                    IOMODE_RW,
2143                                                    false,
2144                                                    GFP_NOFS);
2145                 if (IS_ERR(pgio->pg_lseg)) {
2146                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2147                         pgio->pg_lseg = NULL;
2148                         return;
2149                 }
2150         }
2151         /* If no lseg, fall back to write through mds */
2152         if (pgio->pg_lseg == NULL)
2153                 nfs_pageio_reset_write_mds(pgio);
2154 }
2155 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2156
2157 void
2158 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2159 {
2160         if (desc->pg_lseg) {
2161                 pnfs_put_lseg(desc->pg_lseg);
2162                 desc->pg_lseg = NULL;
2163         }
2164 }
2165 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2166
2167 /*
2168  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2169  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2170  */
2171 size_t
2172 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2173                      struct nfs_page *prev, struct nfs_page *req)
2174 {
2175         unsigned int size;
2176         u64 seg_end, req_start, seg_left;
2177
2178         size = nfs_generic_pg_test(pgio, prev, req);
2179         if (!size)
2180                 return 0;
2181
2182         /*
2183          * 'size' contains the number of bytes left in the current page (up
2184          * to the original size asked for in @req->wb_bytes).
2185          *
2186          * Calculate how many bytes are left in the layout segment
2187          * and if there are less bytes than 'size', return that instead.
2188          *
2189          * Please also note that 'end_offset' is actually the offset of the
2190          * first byte that lies outside the pnfs_layout_range. FIXME?
2191          *
2192          */
2193         if (pgio->pg_lseg) {
2194                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2195                                      pgio->pg_lseg->pls_range.length);
2196                 req_start = req_offset(req);
2197
2198                 /* start of request is past the last byte of this segment */
2199                 if (req_start >= seg_end)
2200                         return 0;
2201
2202                 /* adjust 'size' iff there are fewer bytes left in the
2203                  * segment than what nfs_generic_pg_test returned */
2204                 seg_left = seg_end - req_start;
2205                 if (seg_left < size)
2206                         size = (unsigned int)seg_left;
2207         }
2208
2209         return size;
2210 }
2211 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2212
2213 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2214 {
2215         struct nfs_pageio_descriptor pgio;
2216
2217         /* Resend all requests through the MDS */
2218         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2219                               hdr->completion_ops);
2220         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2221         return nfs_pageio_resend(&pgio, hdr);
2222 }
2223 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2224
2225 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2226 {
2227
2228         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2229         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2230             PNFS_LAYOUTRET_ON_ERROR) {
2231                 pnfs_return_layout(hdr->inode);
2232         }
2233         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2234                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2235 }
2236
2237 /*
2238  * Called by non rpc-based layout drivers
2239  */
2240 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2241 {
2242         if (likely(!hdr->pnfs_error)) {
2243                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2244                                 hdr->mds_offset + hdr->res.count);
2245                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2246         }
2247         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2248         if (unlikely(hdr->pnfs_error))
2249                 pnfs_ld_handle_write_error(hdr);
2250         hdr->mds_ops->rpc_release(hdr);
2251 }
2252 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2253
2254 static void
2255 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2256                 struct nfs_pgio_header *hdr)
2257 {
2258         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2259
2260         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2261                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2262                 nfs_pageio_reset_write_mds(desc);
2263                 mirror->pg_recoalesce = 1;
2264         }
2265         hdr->completion_ops->completion(hdr);
2266 }
2267
2268 static enum pnfs_try_status
2269 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2270                         const struct rpc_call_ops *call_ops,
2271                         struct pnfs_layout_segment *lseg,
2272                         int how)
2273 {
2274         struct inode *inode = hdr->inode;
2275         enum pnfs_try_status trypnfs;
2276         struct nfs_server *nfss = NFS_SERVER(inode);
2277
2278         hdr->mds_ops = call_ops;
2279
2280         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2281                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2282         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2283         if (trypnfs != PNFS_NOT_ATTEMPTED)
2284                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2285         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2286         return trypnfs;
2287 }
2288
2289 static void
2290 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2291               struct nfs_pgio_header *hdr, int how)
2292 {
2293         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2294         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2295         enum pnfs_try_status trypnfs;
2296
2297         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2298         switch (trypnfs) {
2299         case PNFS_NOT_ATTEMPTED:
2300                 pnfs_write_through_mds(desc, hdr);
2301         case PNFS_ATTEMPTED:
2302                 break;
2303         case PNFS_TRY_AGAIN:
2304                 /* cleanup hdr and prepare to redo pnfs */
2305                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2306                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2307                         list_splice_init(&hdr->pages, &mirror->pg_list);
2308                         mirror->pg_recoalesce = 1;
2309                 }
2310                 hdr->mds_ops->rpc_release(hdr);
2311         }
2312 }
2313
2314 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2315 {
2316         pnfs_put_lseg(hdr->lseg);
2317         nfs_pgio_header_free(hdr);
2318 }
2319
2320 int
2321 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2322 {
2323         struct nfs_pgio_header *hdr;
2324         int ret;
2325
2326         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2327         if (!hdr) {
2328                 desc->pg_error = -ENOMEM;
2329                 return desc->pg_error;
2330         }
2331         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2332
2333         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2334         ret = nfs_generic_pgio(desc, hdr);
2335         if (!ret)
2336                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2337
2338         return ret;
2339 }
2340 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2341
2342 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2343 {
2344         struct nfs_pageio_descriptor pgio;
2345
2346         /* Resend all requests through the MDS */
2347         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2348         return nfs_pageio_resend(&pgio, hdr);
2349 }
2350 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2351
2352 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2353 {
2354         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2355         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2356             PNFS_LAYOUTRET_ON_ERROR) {
2357                 pnfs_return_layout(hdr->inode);
2358         }
2359         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2360                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2361 }
2362
2363 /*
2364  * Called by non rpc-based layout drivers
2365  */
2366 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2367 {
2368         if (likely(!hdr->pnfs_error))
2369                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2370         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2371         if (unlikely(hdr->pnfs_error))
2372                 pnfs_ld_handle_read_error(hdr);
2373         hdr->mds_ops->rpc_release(hdr);
2374 }
2375 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2376
2377 static void
2378 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2379                 struct nfs_pgio_header *hdr)
2380 {
2381         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2382
2383         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2384                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2385                 nfs_pageio_reset_read_mds(desc);
2386                 mirror->pg_recoalesce = 1;
2387         }
2388         hdr->completion_ops->completion(hdr);
2389 }
2390
2391 /*
2392  * Call the appropriate parallel I/O subsystem read function.
2393  */
2394 static enum pnfs_try_status
2395 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2396                        const struct rpc_call_ops *call_ops,
2397                        struct pnfs_layout_segment *lseg)
2398 {
2399         struct inode *inode = hdr->inode;
2400         struct nfs_server *nfss = NFS_SERVER(inode);
2401         enum pnfs_try_status trypnfs;
2402
2403         hdr->mds_ops = call_ops;
2404
2405         dprintk("%s: Reading ino:%lu %u@%llu\n",
2406                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2407
2408         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2409         if (trypnfs != PNFS_NOT_ATTEMPTED)
2410                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2411         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2412         return trypnfs;
2413 }
2414
2415 /* Resend all requests through pnfs. */
2416 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2417 {
2418         struct nfs_pageio_descriptor pgio;
2419
2420         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2421                 /* Prevent deadlocks with layoutreturn! */
2422                 pnfs_put_lseg(hdr->lseg);
2423                 hdr->lseg = NULL;
2424
2425                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2426                                         hdr->completion_ops);
2427                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2428         }
2429 }
2430 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2431
2432 static void
2433 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2434 {
2435         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2436         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2437         enum pnfs_try_status trypnfs;
2438
2439         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2440         switch (trypnfs) {
2441         case PNFS_NOT_ATTEMPTED:
2442                 pnfs_read_through_mds(desc, hdr);
2443         case PNFS_ATTEMPTED:
2444                 break;
2445         case PNFS_TRY_AGAIN:
2446                 /* cleanup hdr and prepare to redo pnfs */
2447                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2448                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2449                         list_splice_init(&hdr->pages, &mirror->pg_list);
2450                         mirror->pg_recoalesce = 1;
2451                 }
2452                 hdr->mds_ops->rpc_release(hdr);
2453         }
2454 }
2455
2456 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2457 {
2458         pnfs_put_lseg(hdr->lseg);
2459         nfs_pgio_header_free(hdr);
2460 }
2461
2462 int
2463 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2464 {
2465         struct nfs_pgio_header *hdr;
2466         int ret;
2467
2468         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2469         if (!hdr) {
2470                 desc->pg_error = -ENOMEM;
2471                 return desc->pg_error;
2472         }
2473         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2474         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2475         ret = nfs_generic_pgio(desc, hdr);
2476         if (!ret)
2477                 pnfs_do_read(desc, hdr);
2478         return ret;
2479 }
2480 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2481
2482 static void pnfs_clear_layoutcommitting(struct inode *inode)
2483 {
2484         unsigned long *bitlock = &NFS_I(inode)->flags;
2485
2486         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2487         smp_mb__after_atomic();
2488         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2489 }
2490
2491 /*
2492  * There can be multiple RW segments.
2493  */
2494 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2495 {
2496         struct pnfs_layout_segment *lseg;
2497
2498         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2499                 if (lseg->pls_range.iomode == IOMODE_RW &&
2500                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2501                         list_add(&lseg->pls_lc_list, listp);
2502         }
2503 }
2504
2505 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2506 {
2507         struct pnfs_layout_segment *lseg, *tmp;
2508
2509         /* Matched by references in pnfs_set_layoutcommit */
2510         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2511                 list_del_init(&lseg->pls_lc_list);
2512                 pnfs_put_lseg(lseg);
2513         }
2514
2515         pnfs_clear_layoutcommitting(inode);
2516 }
2517
2518 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2519 {
2520         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2521 }
2522 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2523
2524 void
2525 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2526                 loff_t end_pos)
2527 {
2528         struct nfs_inode *nfsi = NFS_I(inode);
2529         bool mark_as_dirty = false;
2530
2531         spin_lock(&inode->i_lock);
2532         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2533                 nfsi->layout->plh_lwb = end_pos;
2534                 mark_as_dirty = true;
2535                 dprintk("%s: Set layoutcommit for inode %lu ",
2536                         __func__, inode->i_ino);
2537         } else if (end_pos > nfsi->layout->plh_lwb)
2538                 nfsi->layout->plh_lwb = end_pos;
2539         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2540                 /* references matched in nfs4_layoutcommit_release */
2541                 pnfs_get_lseg(lseg);
2542         }
2543         spin_unlock(&inode->i_lock);
2544         dprintk("%s: lseg %p end_pos %llu\n",
2545                 __func__, lseg, nfsi->layout->plh_lwb);
2546
2547         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2548          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2549         if (mark_as_dirty)
2550                 mark_inode_dirty_sync(inode);
2551 }
2552 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2553
2554 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2555 {
2556         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2557
2558         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2559                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2560         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2561 }
2562
2563 /*
2564  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2565  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2566  * data to disk to allow the server to recover the data if it crashes.
2567  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2568  * is off, and a COMMIT is sent to a data server, or
2569  * if WRITEs to a data server return NFS_DATA_SYNC.
2570  */
2571 int
2572 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2573 {
2574         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2575         struct nfs4_layoutcommit_data *data;
2576         struct nfs_inode *nfsi = NFS_I(inode);
2577         loff_t end_pos;
2578         int status;
2579
2580         if (!pnfs_layoutcommit_outstanding(inode))
2581                 return 0;
2582
2583         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2584
2585         status = -EAGAIN;
2586         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2587                 if (!sync)
2588                         goto out;
2589                 status = wait_on_bit_lock_action(&nfsi->flags,
2590                                 NFS_INO_LAYOUTCOMMITTING,
2591                                 nfs_wait_bit_killable,
2592                                 TASK_KILLABLE);
2593                 if (status)
2594                         goto out;
2595         }
2596
2597         status = -ENOMEM;
2598         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2599         data = kzalloc(sizeof(*data), GFP_NOFS);
2600         if (!data)
2601                 goto clear_layoutcommitting;
2602
2603         status = 0;
2604         spin_lock(&inode->i_lock);
2605         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2606                 goto out_unlock;
2607
2608         INIT_LIST_HEAD(&data->lseg_list);
2609         pnfs_list_write_lseg(inode, &data->lseg_list);
2610
2611         end_pos = nfsi->layout->plh_lwb;
2612
2613         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2614         spin_unlock(&inode->i_lock);
2615
2616         data->args.inode = inode;
2617         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2618         nfs_fattr_init(&data->fattr);
2619         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2620         data->res.fattr = &data->fattr;
2621         if (end_pos != 0)
2622                 data->args.lastbytewritten = end_pos - 1;
2623         else
2624                 data->args.lastbytewritten = U64_MAX;
2625         data->res.server = NFS_SERVER(inode);
2626
2627         if (ld->prepare_layoutcommit) {
2628                 status = ld->prepare_layoutcommit(&data->args);
2629                 if (status) {
2630                         put_rpccred(data->cred);
2631                         spin_lock(&inode->i_lock);
2632                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2633                         if (end_pos > nfsi->layout->plh_lwb)
2634                                 nfsi->layout->plh_lwb = end_pos;
2635                         goto out_unlock;
2636                 }
2637         }
2638
2639
2640         status = nfs4_proc_layoutcommit(data, sync);
2641 out:
2642         if (status)
2643                 mark_inode_dirty_sync(inode);
2644         dprintk("<-- %s status %d\n", __func__, status);
2645         return status;
2646 out_unlock:
2647         spin_unlock(&inode->i_lock);
2648         kfree(data);
2649 clear_layoutcommitting:
2650         pnfs_clear_layoutcommitting(inode);
2651         goto out;
2652 }
2653 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2654
2655 int
2656 pnfs_generic_sync(struct inode *inode, bool datasync)
2657 {
2658         return pnfs_layoutcommit_inode(inode, true);
2659 }
2660 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2661
2662 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2663 {
2664         struct nfs4_threshold *thp;
2665
2666         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2667         if (!thp) {
2668                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2669                 return NULL;
2670         }
2671         return thp;
2672 }
2673
2674 #if IS_ENABLED(CONFIG_NFS_V4_2)
2675 int
2676 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2677 {
2678         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2679         struct nfs_server *server = NFS_SERVER(inode);
2680         struct nfs_inode *nfsi = NFS_I(inode);
2681         struct nfs42_layoutstat_data *data;
2682         struct pnfs_layout_hdr *hdr;
2683         int status = 0;
2684
2685         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2686                 goto out;
2687
2688         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2689                 goto out;
2690
2691         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2692                 goto out;
2693
2694         spin_lock(&inode->i_lock);
2695         if (!NFS_I(inode)->layout) {
2696                 spin_unlock(&inode->i_lock);
2697                 goto out_clear_layoutstats;
2698         }
2699         hdr = NFS_I(inode)->layout;
2700         pnfs_get_layout_hdr(hdr);
2701         spin_unlock(&inode->i_lock);
2702
2703         data = kzalloc(sizeof(*data), gfp_flags);
2704         if (!data) {
2705                 status = -ENOMEM;
2706                 goto out_put;
2707         }
2708
2709         data->args.fh = NFS_FH(inode);
2710         data->args.inode = inode;
2711         status = ld->prepare_layoutstats(&data->args);
2712         if (status)
2713                 goto out_free;
2714
2715         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2716
2717 out:
2718         dprintk("%s returns %d\n", __func__, status);
2719         return status;
2720
2721 out_free:
2722         kfree(data);
2723 out_put:
2724         pnfs_put_layout_hdr(hdr);
2725 out_clear_layoutstats:
2726         smp_mb__before_atomic();
2727         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2728         smp_mb__after_atomic();
2729         goto out;
2730 }
2731 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2732 #endif
2733
2734 unsigned int layoutstats_timer;
2735 module_param(layoutstats_timer, uint, 0644);
2736 EXPORT_SYMBOL_GPL(layoutstats_timer);