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 "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36 #include "nfs4trace.h"
  37
  38 #define NFSDBG_FACILITY         NFSDBG_PNFS
  39 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  40
  41 /* Locking:
  42  *
  43  * pnfs_spinlock:
  44  *      protects pnfs_modules_tbl.
  45  */
  46 static DEFINE_SPINLOCK(pnfs_spinlock);
  47
  48 /*
  49  * pnfs_modules_tbl holds all pnfs modules
  50  */
  51 static LIST_HEAD(pnfs_modules_tbl);
  52
  53 static int
  54 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
  55                        enum pnfs_iomode iomode, bool sync);
  56
  57 /* Return the registered pnfs layout driver module matching given id */
  58 static struct pnfs_layoutdriver_type *
  59 find_pnfs_driver_locked(u32 id)
  60 {
  61         struct pnfs_layoutdriver_type *local;
  62
  63         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  64                 if (local->id == id)
  65                         goto out;
  66         local = NULL;
  67 out:
  68         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  69         return local;
  70 }
  71
  72 static struct pnfs_layoutdriver_type *
  73 find_pnfs_driver(u32 id)
  74 {
  75         struct pnfs_layoutdriver_type *local;
  76
  77         spin_lock(&pnfs_spinlock);
  78         local = find_pnfs_driver_locked(id);
  79         if (local != NULL && !try_module_get(local->owner)) {
  80                 dprintk("%s: Could not grab reference on module\n", __func__);
  81                 local = NULL;
  82         }
  83         spin_unlock(&pnfs_spinlock);
  84         return local;
  85 }
  86
  87 void
  88 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  89 {
  90         if (nfss->pnfs_curr_ld) {
  91                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  92                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  93                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  94                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  95                         nfs4_deviceid_purge_client(nfss->nfs_client);
  96                 module_put(nfss->pnfs_curr_ld->owner);
  97         }
  98         nfss->pnfs_curr_ld = NULL;
  99 }
 100
 101 /*
 102  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 103  * Currently only one pNFS layout driver per filesystem is supported.
 104  *
 105  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 106  */
 107 void
 108 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 109                       u32 id)
 110 {
 111         struct pnfs_layoutdriver_type *ld_type = NULL;
 112
 113         if (id == 0)
 114                 goto out_no_driver;
 115         if (!(server->nfs_client->cl_exchange_flags &
 116                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 117                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 118                         __func__, id, server->nfs_client->cl_exchange_flags);
 119                 goto out_no_driver;
 120         }
 121         ld_type = find_pnfs_driver(id);
 122         if (!ld_type) {
 123                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 124                 ld_type = find_pnfs_driver(id);
 125                 if (!ld_type) {
 126                         dprintk("%s: No pNFS module found for %u.\n",
 127                                 __func__, id);
 128                         goto out_no_driver;
 129                 }
 130         }
 131         server->pnfs_curr_ld = ld_type;
 132         if (ld_type->set_layoutdriver
 133             && ld_type->set_layoutdriver(server, mntfh)) {
 134                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 135                         "driver %u.\n", __func__, id);
 136                 module_put(ld_type->owner);
 137                 goto out_no_driver;
 138         }
 139         /* Bump the MDS count */
 140         atomic_inc(&server->nfs_client->cl_mds_count);
 141
 142         dprintk("%s: pNFS module for %u set\n", __func__, id);
 143         return;
 144
 145 out_no_driver:
 146         dprintk("%s: Using NFSv4 I/O\n", __func__);
 147         server->pnfs_curr_ld = NULL;
 148 }
 149
 150 int
 151 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 152 {
 153         int status = -EINVAL;
 154         struct pnfs_layoutdriver_type *tmp;
 155
 156         if (ld_type->id == 0) {
 157                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 158                 return status;
 159         }
 160         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 161                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 162                        "alloc_lseg and free_lseg.\n", __func__);
 163                 return status;
 164         }
 165
 166         spin_lock(&pnfs_spinlock);
 167         tmp = find_pnfs_driver_locked(ld_type->id);
 168         if (!tmp) {
 169                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 170                 status = 0;
 171                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 172                         ld_type->name);
 173         } else {
 174                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 175                         __func__, ld_type->id);
 176         }
 177         spin_unlock(&pnfs_spinlock);
 178
 179         return status;
 180 }
 181 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 182
 183 void
 184 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 185 {
 186         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 187         spin_lock(&pnfs_spinlock);
 188         list_del(&ld_type->pnfs_tblid);
 189         spin_unlock(&pnfs_spinlock);
 190 }
 191 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 192
 193 /*
 194  * pNFS client layout cache
 195  */
 196
 197 /* Need to hold i_lock if caller does not already hold reference */
 198 void
 199 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 200 {
 201         atomic_inc(&lo->plh_refcount);
 202 }
 203
 204 static struct pnfs_layout_hdr *
 205 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 206 {
 207         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 208         return ld->alloc_layout_hdr(ino, gfp_flags);
 209 }
 210
 211 static void
 212 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 213 {
 214         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 215         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 216
 217         if (!list_empty(&lo->plh_layouts)) {
 218                 struct nfs_client *clp = server->nfs_client;
 219
 220                 spin_lock(&clp->cl_lock);
 221                 list_del_init(&lo->plh_layouts);
 222                 spin_unlock(&clp->cl_lock);
 223         }
 224         put_rpccred(lo->plh_lc_cred);
 225         return ld->free_layout_hdr(lo);
 226 }
 227
 228 static void
 229 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 230 {
 231         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 232         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 233         nfsi->layout = NULL;
 234         /* Reset MDS Threshold I/O counters */
 235         nfsi->write_io = 0;
 236         nfsi->read_io = 0;
 237 }
 238
 239 void
 240 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 241 {
 242         struct inode *inode = lo->plh_inode;
 243
 244         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 245                 if (!list_empty(&lo->plh_segs))
 246                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 247                 pnfs_detach_layout_hdr(lo);
 248                 spin_unlock(&inode->i_lock);
 249                 pnfs_free_layout_hdr(lo);
 250         }
 251 }
 252
 253 static int
 254 pnfs_iomode_to_fail_bit(u32 iomode)
 255 {
 256         return iomode == IOMODE_RW ?
 257                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 258 }
 259
 260 static void
 261 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 262 {
 263         lo->plh_retry_timestamp = jiffies;
 264         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 265                 atomic_inc(&lo->plh_refcount);
 266 }
 267
 268 static void
 269 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 270 {
 271         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 272                 atomic_dec(&lo->plh_refcount);
 273 }
 274
 275 static void
 276 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 277 {
 278         struct inode *inode = lo->plh_inode;
 279         struct pnfs_layout_range range = {
 280                 .iomode = iomode,
 281                 .offset = 0,
 282                 .length = NFS4_MAX_UINT64,
 283         };
 284         LIST_HEAD(head);
 285
 286         spin_lock(&inode->i_lock);
 287         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 288         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 289         spin_unlock(&inode->i_lock);
 290         pnfs_free_lseg_list(&head);
 291         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 292                         iomode == IOMODE_RW ?  "RW" : "READ");
 293 }
 294
 295 static bool
 296 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 297 {
 298         unsigned long start, end;
 299         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 300
 301         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 302                 return false;
 303         end = jiffies;
 304         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 305         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 306                 /* It is time to retry the failed layoutgets */
 307                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 308                 return false;
 309         }
 310         return true;
 311 }
 312
 313 static void
 314 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 315 {
 316         INIT_LIST_HEAD(&lseg->pls_list);
 317         INIT_LIST_HEAD(&lseg->pls_lc_list);
 318         atomic_set(&lseg->pls_refcount, 1);
 319         smp_mb();
 320         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 321         lseg->pls_layout = lo;
 322 }
 323
 324 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 325 {
 326         struct inode *ino = lseg->pls_layout->plh_inode;
 327
 328         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 329 }
 330
 331 static void
 332 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 333                 struct pnfs_layout_segment *lseg)
 334 {
 335         struct inode *inode = lo->plh_inode;
 336
 337         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 338         list_del_init(&lseg->pls_list);
 339         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 340         atomic_dec(&lo->plh_refcount);
 341         if (list_empty(&lo->plh_segs))
 342                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 343         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 344 }
 345
 346 /* Return true if layoutreturn is needed */
 347 static bool
 348 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 349                         struct pnfs_layout_segment *lseg,
 350                         nfs4_stateid *stateid, enum pnfs_iomode *iomode)
 351 {
 352         struct pnfs_layout_segment *s;
 353
 354         if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 355                 return false;
 356
 357         list_for_each_entry(s, &lo->plh_segs, pls_list)
 358                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 359                         return false;
 360
 361         *stateid = lo->plh_stateid;
 362         *iomode = lo->plh_return_iomode;
 363         /* decreased in pnfs_send_layoutreturn() */
 364         lo->plh_block_lgets++;
 365         lo->plh_return_iomode = 0;
 366         return true;
 367 }
 368
 369 void
 370 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 371 {
 372         struct pnfs_layout_hdr *lo;
 373         struct inode *inode;
 374
 375         if (!lseg)
 376                 return;
 377
 378         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 379                 atomic_read(&lseg->pls_refcount),
 380                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 381         lo = lseg->pls_layout;
 382         inode = lo->plh_inode;
 383         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 384                 bool need_return;
 385                 nfs4_stateid stateid;
 386                 enum pnfs_iomode iomode;
 387
 388                 pnfs_get_layout_hdr(lo);
 389                 pnfs_layout_remove_lseg(lo, lseg);
 390                 need_return = pnfs_layout_need_return(lo, lseg,
 391                                                       &stateid, &iomode);
 392                 spin_unlock(&inode->i_lock);
 393                 pnfs_free_lseg(lseg);
 394                 if (need_return)
 395                         pnfs_send_layoutreturn(lo, stateid, iomode,
 396                                                true);
 397                 else
 398                         pnfs_put_layout_hdr(lo);
 399         }
 400 }
 401 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 402
 403 static void pnfs_free_lseg_async_work(struct work_struct *work)
 404 {
 405         struct pnfs_layout_segment *lseg;
 406         struct pnfs_layout_hdr *lo;
 407
 408         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 409         lo = lseg->pls_layout;
 410
 411         pnfs_free_lseg(lseg);
 412         pnfs_put_layout_hdr(lo);
 413 }
 414
 415 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 416 {
 417         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 418         schedule_work(&lseg->pls_work);
 419 }
 420
 421 void
 422 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 423 {
 424         if (!lseg)
 425                 return;
 426
 427         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 428
 429         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 430                 atomic_read(&lseg->pls_refcount),
 431                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 432         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 433                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 434                 pnfs_get_layout_hdr(lo);
 435                 pnfs_layout_remove_lseg(lo, lseg);
 436                 pnfs_free_lseg_async(lseg);
 437         }
 438 }
 439 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 440
 441 static u64
 442 end_offset(u64 start, u64 len)
 443 {
 444         u64 end;
 445
 446         end = start + len;
 447         return end >= start ? end : NFS4_MAX_UINT64;
 448 }
 449
 450 /*
 451  * is l2 fully contained in l1?
 452  *   start1                             end1
 453  *   [----------------------------------)
 454  *           start2           end2
 455  *           [----------------)
 456  */
 457 static bool
 458 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 459                  const struct pnfs_layout_range *l2)
 460 {
 461         u64 start1 = l1->offset;
 462         u64 end1 = end_offset(start1, l1->length);
 463         u64 start2 = l2->offset;
 464         u64 end2 = end_offset(start2, l2->length);
 465
 466         return (start1 <= start2) && (end1 >= end2);
 467 }
 468
 469 /*
 470  * is l1 and l2 intersecting?
 471  *   start1                             end1
 472  *   [----------------------------------)
 473  *                              start2           end2
 474  *                              [----------------)
 475  */
 476 static bool
 477 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 478                     const struct pnfs_layout_range *l2)
 479 {
 480         u64 start1 = l1->offset;
 481         u64 end1 = end_offset(start1, l1->length);
 482         u64 start2 = l2->offset;
 483         u64 end2 = end_offset(start2, l2->length);
 484
 485         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 486                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 487 }
 488
 489 static bool
 490 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 491                  const struct pnfs_layout_range *recall_range)
 492 {
 493         return (recall_range->iomode == IOMODE_ANY ||
 494                 lseg_range->iomode == recall_range->iomode) &&
 495                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 496 }
 497
 498 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 499                 struct list_head *tmp_list)
 500 {
 501         if (!atomic_dec_and_test(&lseg->pls_refcount))
 502                 return false;
 503         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 504         list_add(&lseg->pls_list, tmp_list);
 505         return true;
 506 }
 507
 508 /* Returns 1 if lseg is removed from list, 0 otherwise */
 509 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 510                              struct list_head *tmp_list)
 511 {
 512         int rv = 0;
 513
 514         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 515                 /* Remove the reference keeping the lseg in the
 516                  * list.  It will now be removed when all
 517                  * outstanding io is finished.
 518                  */
 519                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 520                         atomic_read(&lseg->pls_refcount));
 521                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 522                         rv = 1;
 523         }
 524         return rv;
 525 }
 526
 527 /* Returns count of number of matching invalid lsegs remaining in list
 528  * after call.
 529  */
 530 int
 531 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 532                             struct list_head *tmp_list,
 533                             struct pnfs_layout_range *recall_range)
 534 {
 535         struct pnfs_layout_segment *lseg, *next;
 536         int invalid = 0, removed = 0;
 537
 538         dprintk("%s:Begin lo %p\n", __func__, lo);
 539
 540         if (list_empty(&lo->plh_segs))
 541                 return 0;
 542         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 543                 if (!recall_range ||
 544                     should_free_lseg(&lseg->pls_range, recall_range)) {
 545                         dprintk("%s: freeing lseg %p iomode %d "
 546                                 "offset %llu length %llu\n", __func__,
 547                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 548                                 lseg->pls_range.length);
 549                         invalid++;
 550                         removed += mark_lseg_invalid(lseg, tmp_list);
 551                 }
 552         dprintk("%s:Return %i\n", __func__, invalid - removed);
 553         return invalid - removed;
 554 }
 555
 556 /* note free_me must contain lsegs from a single layout_hdr */
 557 void
 558 pnfs_free_lseg_list(struct list_head *free_me)
 559 {
 560         struct pnfs_layout_segment *lseg, *tmp;
 561
 562         if (list_empty(free_me))
 563                 return;
 564
 565         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 566                 list_del(&lseg->pls_list);
 567                 pnfs_free_lseg(lseg);
 568         }
 569 }
 570
 571 void
 572 pnfs_destroy_layout(struct nfs_inode *nfsi)
 573 {
 574         struct pnfs_layout_hdr *lo;
 575         LIST_HEAD(tmp_list);
 576
 577         spin_lock(&nfsi->vfs_inode.i_lock);
 578         lo = nfsi->layout;
 579         if (lo) {
 580                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 581                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 582                 pnfs_get_layout_hdr(lo);
 583                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 584                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 585                 spin_unlock(&nfsi->vfs_inode.i_lock);
 586                 pnfs_free_lseg_list(&tmp_list);
 587                 pnfs_put_layout_hdr(lo);
 588         } else
 589                 spin_unlock(&nfsi->vfs_inode.i_lock);
 590 }
 591 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 592
 593 static bool
 594 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 595                 struct list_head *layout_list)
 596 {
 597         struct pnfs_layout_hdr *lo;
 598         bool ret = false;
 599
 600         spin_lock(&inode->i_lock);
 601         lo = NFS_I(inode)->layout;
 602         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 603                 pnfs_get_layout_hdr(lo);
 604                 list_add(&lo->plh_bulk_destroy, layout_list);
 605                 ret = true;
 606         }
 607         spin_unlock(&inode->i_lock);
 608         return ret;
 609 }
 610
 611 /* Caller must hold rcu_read_lock and clp->cl_lock */
 612 static int
 613 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 614                 struct nfs_server *server,
 615                 struct list_head *layout_list)
 616 {
 617         struct pnfs_layout_hdr *lo, *next;
 618         struct inode *inode;
 619
 620         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 621                 inode = igrab(lo->plh_inode);
 622                 if (inode == NULL)
 623                         continue;
 624                 list_del_init(&lo->plh_layouts);
 625                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 626                         continue;
 627                 rcu_read_unlock();
 628                 spin_unlock(&clp->cl_lock);
 629                 iput(inode);
 630                 spin_lock(&clp->cl_lock);
 631                 rcu_read_lock();
 632                 return -EAGAIN;
 633         }
 634         return 0;
 635 }
 636
 637 static int
 638 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 639                 bool is_bulk_recall)
 640 {
 641         struct pnfs_layout_hdr *lo;
 642         struct inode *inode;
 643         struct pnfs_layout_range range = {
 644                 .iomode = IOMODE_ANY,
 645                 .offset = 0,
 646                 .length = NFS4_MAX_UINT64,
 647         };
 648         LIST_HEAD(lseg_list);
 649         int ret = 0;
 650
 651         while (!list_empty(layout_list)) {
 652                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 653                                 plh_bulk_destroy);
 654                 dprintk("%s freeing layout for inode %lu\n", __func__,
 655                         lo->plh_inode->i_ino);
 656                 inode = lo->plh_inode;
 657
 658                 pnfs_layoutcommit_inode(inode, false);
 659
 660                 spin_lock(&inode->i_lock);
 661                 list_del_init(&lo->plh_bulk_destroy);
 662                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 663                 if (is_bulk_recall)
 664                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 665                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 666                         ret = -EAGAIN;
 667                 spin_unlock(&inode->i_lock);
 668                 pnfs_free_lseg_list(&lseg_list);
 669                 pnfs_put_layout_hdr(lo);
 670                 iput(inode);
 671         }
 672         return ret;
 673 }
 674
 675 int
 676 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 677                 struct nfs_fsid *fsid,
 678                 bool is_recall)
 679 {
 680         struct nfs_server *server;
 681         LIST_HEAD(layout_list);
 682
 683         spin_lock(&clp->cl_lock);
 684         rcu_read_lock();
 685 restart:
 686         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 687                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 688                         continue;
 689                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 690                                 server,
 691                                 &layout_list) != 0)
 692                         goto restart;
 693         }
 694         rcu_read_unlock();
 695         spin_unlock(&clp->cl_lock);
 696
 697         if (list_empty(&layout_list))
 698                 return 0;
 699         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 700 }
 701
 702 int
 703 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 704                 bool is_recall)
 705 {
 706         struct nfs_server *server;
 707         LIST_HEAD(layout_list);
 708
 709         spin_lock(&clp->cl_lock);
 710         rcu_read_lock();
 711 restart:
 712         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 713                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 714                                         server,
 715                                         &layout_list) != 0)
 716                         goto restart;
 717         }
 718         rcu_read_unlock();
 719         spin_unlock(&clp->cl_lock);
 720
 721         if (list_empty(&layout_list))
 722                 return 0;
 723         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 724 }
 725
 726 /*
 727  * Called by the state manger to remove all layouts established under an
 728  * expired lease.
 729  */
 730 void
 731 pnfs_destroy_all_layouts(struct nfs_client *clp)
 732 {
 733         nfs4_deviceid_mark_client_invalid(clp);
 734         nfs4_deviceid_purge_client(clp);
 735
 736         pnfs_destroy_layouts_byclid(clp, false);
 737 }
 738
 739 /*
 740  * Compare 2 layout stateid sequence ids, to see which is newer,
 741  * taking into account wraparound issues.
 742  */
 743 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 744 {
 745         return (s32)(s1 - s2) > 0;
 746 }
 747
 748 /* update lo->plh_stateid with new if is more recent */
 749 void
 750 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 751                         bool update_barrier)
 752 {
 753         u32 oldseq, newseq, new_barrier;
 754         int empty = list_empty(&lo->plh_segs);
 755
 756         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 757         newseq = be32_to_cpu(new->seqid);
 758         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 759                 nfs4_stateid_copy(&lo->plh_stateid, new);
 760                 if (update_barrier) {
 761                         new_barrier = be32_to_cpu(new->seqid);
 762                 } else {
 763                         /* Because of wraparound, we want to keep the barrier
 764                          * "close" to the current seqids.
 765                          */
 766                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 767                 }
 768                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 769                         lo->plh_barrier = new_barrier;
 770         }
 771 }
 772
 773 static bool
 774 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 775                 const nfs4_stateid *stateid)
 776 {
 777         u32 seqid = be32_to_cpu(stateid->seqid);
 778
 779         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 780 }
 781
 782 static bool
 783 pnfs_layout_returning(const struct pnfs_layout_hdr *lo,
 784                       struct pnfs_layout_range *range)
 785 {
 786         return test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 787                 (lo->plh_return_iomode == IOMODE_ANY ||
 788                  lo->plh_return_iomode == range->iomode);
 789 }
 790
 791 /* lget is set to 1 if called from inside send_layoutget call chain */
 792 static bool
 793 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo,
 794                         struct pnfs_layout_range *range, int lget)
 795 {
 796         return lo->plh_block_lgets ||
 797                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 798                 (list_empty(&lo->plh_segs) &&
 799                  (atomic_read(&lo->plh_outstanding) > lget)) ||
 800                 pnfs_layout_returning(lo, range);
 801 }
 802
 803 int
 804 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 805                               struct pnfs_layout_range *range,
 806                               struct nfs4_state *open_state)
 807 {
 808         int status = 0;
 809
 810         dprintk("--> %s\n", __func__);
 811         spin_lock(&lo->plh_inode->i_lock);
 812         if (pnfs_layoutgets_blocked(lo, range, 1)) {
 813                 status = -EAGAIN;
 814         } else if (!nfs4_valid_open_stateid(open_state)) {
 815                 status = -EBADF;
 816         } else if (list_empty(&lo->plh_segs) ||
 817                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 818                 int seq;
 819
 820                 do {
 821                         seq = read_seqbegin(&open_state->seqlock);
 822                         nfs4_stateid_copy(dst, &open_state->stateid);
 823                 } while (read_seqretry(&open_state->seqlock, seq));
 824         } else
 825                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 826         spin_unlock(&lo->plh_inode->i_lock);
 827         dprintk("<-- %s\n", __func__);
 828         return status;
 829 }
 830
 831 /*
 832 * Get layout from server.
 833 *    for now, assume that whole file layouts are requested.
 834 *    arg->offset: 0
 835 *    arg->length: all ones
 836 */
 837 static struct pnfs_layout_segment *
 838 send_layoutget(struct pnfs_layout_hdr *lo,
 839            struct nfs_open_context *ctx,
 840            struct pnfs_layout_range *range,
 841            gfp_t gfp_flags)
 842 {
 843         struct inode *ino = lo->plh_inode;
 844         struct nfs_server *server = NFS_SERVER(ino);
 845         struct nfs4_layoutget *lgp;
 846         struct pnfs_layout_segment *lseg;
 847
 848         dprintk("--> %s\n", __func__);
 849
 850         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 851         if (lgp == NULL)
 852                 return NULL;
 853
 854         lgp->args.minlength = PAGE_CACHE_SIZE;
 855         if (lgp->args.minlength > range->length)
 856                 lgp->args.minlength = range->length;
 857         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 858         lgp->args.range = *range;
 859         lgp->args.type = server->pnfs_curr_ld->id;
 860         lgp->args.inode = ino;
 861         lgp->args.ctx = get_nfs_open_context(ctx);
 862         lgp->gfp_flags = gfp_flags;
 863         lgp->cred = lo->plh_lc_cred;
 864
 865         /* Synchronously retrieve layout information from server and
 866          * store in lseg.
 867          */
 868         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 869         if (IS_ERR(lseg)) {
 870                 switch (PTR_ERR(lseg)) {
 871                 case -ENOMEM:
 872                 case -ERESTARTSYS:
 873                         break;
 874                 default:
 875                         /* remember that LAYOUTGET failed and suspend trying */
 876                         pnfs_layout_io_set_failed(lo, range->iomode);
 877                 }
 878                 return NULL;
 879         }
 880
 881         return lseg;
 882 }
 883
 884 static void pnfs_clear_layoutcommit(struct inode *inode,
 885                 struct list_head *head)
 886 {
 887         struct nfs_inode *nfsi = NFS_I(inode);
 888         struct pnfs_layout_segment *lseg, *tmp;
 889
 890         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 891                 return;
 892         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 893                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 894                         continue;
 895                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 896         }
 897 }
 898
 899 static int
 900 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 901                        enum pnfs_iomode iomode, bool sync)
 902 {
 903         struct inode *ino = lo->plh_inode;
 904         struct nfs4_layoutreturn *lrp;
 905         int status = 0;
 906
 907         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 908         if (unlikely(lrp == NULL)) {
 909                 status = -ENOMEM;
 910                 spin_lock(&ino->i_lock);
 911                 lo->plh_block_lgets--;
 912                 spin_unlock(&ino->i_lock);
 913                 pnfs_put_layout_hdr(lo);
 914                 goto out;
 915         }
 916
 917         lrp->args.stateid = stateid;
 918         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 919         lrp->args.inode = ino;
 920         lrp->args.range.iomode = iomode;
 921         lrp->args.range.offset = 0;
 922         lrp->args.range.length = NFS4_MAX_UINT64;
 923         lrp->args.layout = lo;
 924         lrp->clp = NFS_SERVER(ino)->nfs_client;
 925         lrp->cred = lo->plh_lc_cred;
 926
 927         status = nfs4_proc_layoutreturn(lrp, sync);
 928 out:
 929         if (status) {
 930                 spin_lock(&ino->i_lock);
 931                 clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 932                 spin_unlock(&ino->i_lock);
 933         }
 934         dprintk("<-- %s status: %d\n", __func__, status);
 935         return status;
 936 }
 937
 938 /*
 939  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 940  * when the layout segment list is empty.
 941  *
 942  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 943  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 944  * deviceid is marked invalid.
 945  */
 946 int
 947 _pnfs_return_layout(struct inode *ino)
 948 {
 949         struct pnfs_layout_hdr *lo = NULL;
 950         struct nfs_inode *nfsi = NFS_I(ino);
 951         LIST_HEAD(tmp_list);
 952         nfs4_stateid stateid;
 953         int status = 0, empty;
 954
 955         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 956
 957         spin_lock(&ino->i_lock);
 958         lo = nfsi->layout;
 959         if (!lo) {
 960                 spin_unlock(&ino->i_lock);
 961                 dprintk("NFS: %s no layout to return\n", __func__);
 962                 goto out;
 963         }
 964         stateid = nfsi->layout->plh_stateid;
 965         /* Reference matched in nfs4_layoutreturn_release */
 966         pnfs_get_layout_hdr(lo);
 967         empty = list_empty(&lo->plh_segs);
 968         pnfs_clear_layoutcommit(ino, &tmp_list);
 969         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 970
 971         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 972                 struct pnfs_layout_range range = {
 973                         .iomode         = IOMODE_ANY,
 974                         .offset         = 0,
 975                         .length         = NFS4_MAX_UINT64,
 976                 };
 977                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
 978         }
 979
 980         /* Don't send a LAYOUTRETURN if list was initially empty */
 981         if (empty) {
 982                 spin_unlock(&ino->i_lock);
 983                 pnfs_put_layout_hdr(lo);
 984                 dprintk("NFS: %s no layout segments to return\n", __func__);
 985                 goto out;
 986         }
 987
 988         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 989         lo->plh_block_lgets++;
 990         spin_unlock(&ino->i_lock);
 991         pnfs_free_lseg_list(&tmp_list);
 992
 993         status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
 994 out:
 995         dprintk("<-- %s status: %d\n", __func__, status);
 996         return status;
 997 }
 998 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 999
1000 int
1001 pnfs_commit_and_return_layout(struct inode *inode)
1002 {
1003         struct pnfs_layout_hdr *lo;
1004         int ret;
1005
1006         spin_lock(&inode->i_lock);
1007         lo = NFS_I(inode)->layout;
1008         if (lo == NULL) {
1009                 spin_unlock(&inode->i_lock);
1010                 return 0;
1011         }
1012         pnfs_get_layout_hdr(lo);
1013         /* Block new layoutgets and read/write to ds */
1014         lo->plh_block_lgets++;
1015         spin_unlock(&inode->i_lock);
1016         filemap_fdatawait(inode->i_mapping);
1017         ret = pnfs_layoutcommit_inode(inode, true);
1018         if (ret == 0)
1019                 ret = _pnfs_return_layout(inode);
1020         spin_lock(&inode->i_lock);
1021         lo->plh_block_lgets--;
1022         spin_unlock(&inode->i_lock);
1023         pnfs_put_layout_hdr(lo);
1024         return ret;
1025 }
1026
1027 bool pnfs_roc(struct inode *ino)
1028 {
1029         struct pnfs_layout_hdr *lo;
1030         struct pnfs_layout_segment *lseg, *tmp;
1031         LIST_HEAD(tmp_list);
1032         bool found = false;
1033
1034         spin_lock(&ino->i_lock);
1035         lo = NFS_I(ino)->layout;
1036         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
1037             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1038                 goto out_nolayout;
1039         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1040                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1041                         mark_lseg_invalid(lseg, &tmp_list);
1042                         found = true;
1043                 }
1044         if (!found)
1045                 goto out_nolayout;
1046         lo->plh_block_lgets++;
1047         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
1048         spin_unlock(&ino->i_lock);
1049         pnfs_free_lseg_list(&tmp_list);
1050         return true;
1051
1052 out_nolayout:
1053         spin_unlock(&ino->i_lock);
1054         return false;
1055 }
1056
1057 void pnfs_roc_release(struct inode *ino)
1058 {
1059         struct pnfs_layout_hdr *lo;
1060
1061         spin_lock(&ino->i_lock);
1062         lo = NFS_I(ino)->layout;
1063         lo->plh_block_lgets--;
1064         if (atomic_dec_and_test(&lo->plh_refcount)) {
1065                 pnfs_detach_layout_hdr(lo);
1066                 spin_unlock(&ino->i_lock);
1067                 pnfs_free_layout_hdr(lo);
1068         } else
1069                 spin_unlock(&ino->i_lock);
1070 }
1071
1072 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1073 {
1074         struct pnfs_layout_hdr *lo;
1075
1076         spin_lock(&ino->i_lock);
1077         lo = NFS_I(ino)->layout;
1078         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1079                 lo->plh_barrier = barrier;
1080         spin_unlock(&ino->i_lock);
1081 }
1082
1083 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
1084 {
1085         struct nfs_inode *nfsi = NFS_I(ino);
1086         struct pnfs_layout_hdr *lo;
1087         struct pnfs_layout_segment *lseg;
1088         u32 current_seqid;
1089         bool found = false;
1090
1091         spin_lock(&ino->i_lock);
1092         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1093                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1094                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1095                         found = true;
1096                         goto out;
1097                 }
1098         lo = nfsi->layout;
1099         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1100
1101         /* Since close does not return a layout stateid for use as
1102          * a barrier, we choose the worst-case barrier.
1103          */
1104         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1105 out:
1106         spin_unlock(&ino->i_lock);
1107         return found;
1108 }
1109
1110 /*
1111  * Compare two layout segments for sorting into layout cache.
1112  * We want to preferentially return RW over RO layouts, so ensure those
1113  * are seen first.
1114  */
1115 static s64
1116 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1117            const struct pnfs_layout_range *l2)
1118 {
1119         s64 d;
1120
1121         /* high offset > low offset */
1122         d = l1->offset - l2->offset;
1123         if (d)
1124                 return d;
1125
1126         /* short length > long length */
1127         d = l2->length - l1->length;
1128         if (d)
1129                 return d;
1130
1131         /* read > read/write */
1132         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1133 }
1134
1135 static void
1136 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1137                    struct pnfs_layout_segment *lseg)
1138 {
1139         struct pnfs_layout_segment *lp;
1140
1141         dprintk("%s:Begin\n", __func__);
1142
1143         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1144                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1145                         continue;
1146                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1147                 dprintk("%s: inserted lseg %p "
1148                         "iomode %d offset %llu length %llu before "
1149                         "lp %p iomode %d offset %llu length %llu\n",
1150                         __func__, lseg, lseg->pls_range.iomode,
1151                         lseg->pls_range.offset, lseg->pls_range.length,
1152                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1153                         lp->pls_range.length);
1154                 goto out;
1155         }
1156         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1157         dprintk("%s: inserted lseg %p "
1158                 "iomode %d offset %llu length %llu at tail\n",
1159                 __func__, lseg, lseg->pls_range.iomode,
1160                 lseg->pls_range.offset, lseg->pls_range.length);
1161 out:
1162         pnfs_get_layout_hdr(lo);
1163
1164         dprintk("%s:Return\n", __func__);
1165 }
1166
1167 static struct pnfs_layout_hdr *
1168 alloc_init_layout_hdr(struct inode *ino,
1169                       struct nfs_open_context *ctx,
1170                       gfp_t gfp_flags)
1171 {
1172         struct pnfs_layout_hdr *lo;
1173
1174         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1175         if (!lo)
1176                 return NULL;
1177         atomic_set(&lo->plh_refcount, 1);
1178         INIT_LIST_HEAD(&lo->plh_layouts);
1179         INIT_LIST_HEAD(&lo->plh_segs);
1180         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1181         lo->plh_inode = ino;
1182         lo->plh_lc_cred = get_rpccred(ctx->cred);
1183         return lo;
1184 }
1185
1186 static struct pnfs_layout_hdr *
1187 pnfs_find_alloc_layout(struct inode *ino,
1188                        struct nfs_open_context *ctx,
1189                        gfp_t gfp_flags)
1190 {
1191         struct nfs_inode *nfsi = NFS_I(ino);
1192         struct pnfs_layout_hdr *new = NULL;
1193
1194         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1195
1196         if (nfsi->layout != NULL)
1197                 goto out_existing;
1198         spin_unlock(&ino->i_lock);
1199         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1200         spin_lock(&ino->i_lock);
1201
1202         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1203                 nfsi->layout = new;
1204                 return new;
1205         } else if (new != NULL)
1206                 pnfs_free_layout_hdr(new);
1207 out_existing:
1208         pnfs_get_layout_hdr(nfsi->layout);
1209         return nfsi->layout;
1210 }
1211
1212 /*
1213  * iomode matching rules:
1214  * iomode       lseg    match
1215  * -----        -----   -----
1216  * ANY          READ    true
1217  * ANY          RW      true
1218  * RW           READ    false
1219  * RW           RW      true
1220  * READ         READ    true
1221  * READ         RW      true
1222  */
1223 static bool
1224 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1225                  const struct pnfs_layout_range *range)
1226 {
1227         struct pnfs_layout_range range1;
1228
1229         if ((range->iomode == IOMODE_RW &&
1230              ls_range->iomode != IOMODE_RW) ||
1231             !pnfs_lseg_range_intersecting(ls_range, range))
1232                 return 0;
1233
1234         /* range1 covers only the first byte in the range */
1235         range1 = *range;
1236         range1.length = 1;
1237         return pnfs_lseg_range_contained(ls_range, &range1);
1238 }
1239
1240 /*
1241  * lookup range in layout
1242  */
1243 static struct pnfs_layout_segment *
1244 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1245                 struct pnfs_layout_range *range)
1246 {
1247         struct pnfs_layout_segment *lseg, *ret = NULL;
1248
1249         dprintk("%s:Begin\n", __func__);
1250
1251         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1252                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1253                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1254                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1255                         ret = pnfs_get_lseg(lseg);
1256                         break;
1257                 }
1258                 if (lseg->pls_range.offset > range->offset)
1259                         break;
1260         }
1261
1262         dprintk("%s:Return lseg %p ref %d\n",
1263                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1264         return ret;
1265 }
1266
1267 /*
1268  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1269  * to the MDS or over pNFS
1270  *
1271  * The nfs_inode read_io and write_io fields are cumulative counters reset
1272  * when there are no layout segments. Note that in pnfs_update_layout iomode
1273  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1274  * WRITE request.
1275  *
1276  * A return of true means use MDS I/O.
1277  *
1278  * From rfc 5661:
1279  * If a file's size is smaller than the file size threshold, data accesses
1280  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1281  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1282  * server.  If both file size and I/O size are provided, the client SHOULD
1283  * reach or exceed  both thresholds before sending its read or write
1284  * requests to the data server.
1285  */
1286 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1287                                      struct inode *ino, int iomode)
1288 {
1289         struct nfs4_threshold *t = ctx->mdsthreshold;
1290         struct nfs_inode *nfsi = NFS_I(ino);
1291         loff_t fsize = i_size_read(ino);
1292         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1293
1294         if (t == NULL)
1295                 return ret;
1296
1297         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1298                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1299
1300         switch (iomode) {
1301         case IOMODE_READ:
1302                 if (t->bm & THRESHOLD_RD) {
1303                         dprintk("%s fsize %llu\n", __func__, fsize);
1304                         size_set = true;
1305                         if (fsize < t->rd_sz)
1306                                 size = true;
1307                 }
1308                 if (t->bm & THRESHOLD_RD_IO) {
1309                         dprintk("%s nfsi->read_io %llu\n", __func__,
1310                                 nfsi->read_io);
1311                         io_set = true;
1312                         if (nfsi->read_io < t->rd_io_sz)
1313                                 io = true;
1314                 }
1315                 break;
1316         case IOMODE_RW:
1317                 if (t->bm & THRESHOLD_WR) {
1318                         dprintk("%s fsize %llu\n", __func__, fsize);
1319                         size_set = true;
1320                         if (fsize < t->wr_sz)
1321                                 size = true;
1322                 }
1323                 if (t->bm & THRESHOLD_WR_IO) {
1324                         dprintk("%s nfsi->write_io %llu\n", __func__,
1325                                 nfsi->write_io);
1326                         io_set = true;
1327                         if (nfsi->write_io < t->wr_io_sz)
1328                                 io = true;
1329                 }
1330                 break;
1331         }
1332         if (size_set && io_set) {
1333                 if (size && io)
1334                         ret = true;
1335         } else if (size || io)
1336                 ret = true;
1337
1338         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1339         return ret;
1340 }
1341
1342 /*
1343  * Layout segment is retreived from the server if not cached.
1344  * The appropriate layout segment is referenced and returned to the caller.
1345  */
1346 struct pnfs_layout_segment *
1347 pnfs_update_layout(struct inode *ino,
1348                    struct nfs_open_context *ctx,
1349                    loff_t pos,
1350                    u64 count,
1351                    enum pnfs_iomode iomode,
1352                    gfp_t gfp_flags)
1353 {
1354         struct pnfs_layout_range arg = {
1355                 .iomode = iomode,
1356                 .offset = pos,
1357                 .length = count,
1358         };
1359         unsigned pg_offset;
1360         struct nfs_server *server = NFS_SERVER(ino);
1361         struct nfs_client *clp = server->nfs_client;
1362         struct pnfs_layout_hdr *lo;
1363         struct pnfs_layout_segment *lseg = NULL;
1364         bool first;
1365
1366         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1367                 goto out;
1368
1369         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1370                 goto out;
1371
1372 lookup_again:
1373         first = false;
1374         spin_lock(&ino->i_lock);
1375         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1376         if (lo == NULL) {
1377                 spin_unlock(&ino->i_lock);
1378                 goto out;
1379         }
1380
1381         /* Do we even need to bother with this? */
1382         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1383                 dprintk("%s matches recall, use MDS\n", __func__);
1384                 goto out_unlock;
1385         }
1386
1387         /* if LAYOUTGET already failed once we don't try again */
1388         if (pnfs_layout_io_test_failed(lo, iomode))
1389                 goto out_unlock;
1390
1391         first = list_empty(&lo->plh_segs);
1392         if (first) {
1393                 /* The first layoutget for the file. Need to serialize per
1394                  * RFC 5661 Errata 3208.
1395                  */
1396                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1397                                      &lo->plh_flags)) {
1398                         spin_unlock(&ino->i_lock);
1399                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1400                                     TASK_UNINTERRUPTIBLE);
1401                         pnfs_put_layout_hdr(lo);
1402                         goto lookup_again;
1403                 }
1404         } else {
1405                 /* Check to see if the layout for the given range
1406                  * already exists
1407                  */
1408                 lseg = pnfs_find_lseg(lo, &arg);
1409                 if (lseg)
1410                         goto out_unlock;
1411         }
1412
1413         if (pnfs_layoutgets_blocked(lo, &arg, 0))
1414                 goto out_unlock;
1415         atomic_inc(&lo->plh_outstanding);
1416         spin_unlock(&ino->i_lock);
1417
1418         if (list_empty(&lo->plh_layouts)) {
1419                 /* The lo must be on the clp list if there is any
1420                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1421                  */
1422                 spin_lock(&clp->cl_lock);
1423                 if (list_empty(&lo->plh_layouts))
1424                         list_add_tail(&lo->plh_layouts, &server->layouts);
1425                 spin_unlock(&clp->cl_lock);
1426         }
1427
1428         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1429         if (pg_offset) {
1430                 arg.offset -= pg_offset;
1431                 arg.length += pg_offset;
1432         }
1433         if (arg.length != NFS4_MAX_UINT64)
1434                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1435
1436         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1437         atomic_dec(&lo->plh_outstanding);
1438 out_put_layout_hdr:
1439         if (first) {
1440                 unsigned long *bitlock = &lo->plh_flags;
1441
1442                 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1443                 smp_mb__after_atomic();
1444                 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1445         }
1446         pnfs_put_layout_hdr(lo);
1447 out:
1448         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1449                         "(%s, offset: %llu, length: %llu)\n",
1450                         __func__, ino->i_sb->s_id,
1451                         (unsigned long long)NFS_FILEID(ino),
1452                         lseg == NULL ? "not found" : "found",
1453                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1454                         (unsigned long long)pos,
1455                         (unsigned long long)count);
1456         return lseg;
1457 out_unlock:
1458         spin_unlock(&ino->i_lock);
1459         goto out_put_layout_hdr;
1460 }
1461 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1462
1463 struct pnfs_layout_segment *
1464 pnfs_layout_process(struct nfs4_layoutget *lgp)
1465 {
1466         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1467         struct nfs4_layoutget_res *res = &lgp->res;
1468         struct pnfs_layout_segment *lseg;
1469         struct inode *ino = lo->plh_inode;
1470         LIST_HEAD(free_me);
1471         int status = 0;
1472
1473         /* Inject layout blob into I/O device driver */
1474         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1475         if (!lseg || IS_ERR(lseg)) {
1476                 if (!lseg)
1477                         status = -ENOMEM;
1478                 else
1479                         status = PTR_ERR(lseg);
1480                 dprintk("%s: Could not allocate layout: error %d\n",
1481                        __func__, status);
1482                 goto out;
1483         }
1484
1485         init_lseg(lo, lseg);
1486         lseg->pls_range = res->range;
1487
1488         spin_lock(&ino->i_lock);
1489         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1490                 dprintk("%s forget reply due to recall\n", __func__);
1491                 goto out_forget_reply;
1492         }
1493
1494         if (pnfs_layoutgets_blocked(lo, &lgp->args.range, 1)) {
1495                 dprintk("%s forget reply due to state\n", __func__);
1496                 goto out_forget_reply;
1497         }
1498
1499         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1500                 /* existing state ID, make sure the sequence number matches. */
1501                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1502                         dprintk("%s forget reply due to sequence\n", __func__);
1503                         goto out_forget_reply;
1504                 }
1505                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1506         } else {
1507                 /*
1508                  * We got an entirely new state ID.  Mark all segments for the
1509                  * inode invalid, and don't bother validating the stateid
1510                  * sequence number.
1511                  */
1512                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1513
1514                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1515                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1516         }
1517
1518         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1519
1520         pnfs_get_lseg(lseg);
1521         pnfs_layout_insert_lseg(lo, lseg);
1522
1523         if (res->return_on_close) {
1524                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1525                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1526         }
1527
1528         spin_unlock(&ino->i_lock);
1529         pnfs_free_lseg_list(&free_me);
1530         return lseg;
1531 out:
1532         return ERR_PTR(status);
1533
1534 out_forget_reply:
1535         spin_unlock(&ino->i_lock);
1536         lseg->pls_layout = lo;
1537         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1538         goto out;
1539 }
1540
1541 static void
1542 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1543                                 struct list_head *tmp_list,
1544                                 struct pnfs_layout_range *return_range)
1545 {
1546         struct pnfs_layout_segment *lseg, *next;
1547
1548         dprintk("%s:Begin lo %p\n", __func__, lo);
1549
1550         if (list_empty(&lo->plh_segs))
1551                 return;
1552
1553         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1554                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1555                         dprintk("%s: marking lseg %p iomode %d "
1556                                 "offset %llu length %llu\n", __func__,
1557                                 lseg, lseg->pls_range.iomode,
1558                                 lseg->pls_range.offset,
1559                                 lseg->pls_range.length);
1560                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1561                         mark_lseg_invalid(lseg, tmp_list);
1562                 }
1563 }
1564
1565 void pnfs_error_mark_layout_for_return(struct inode *inode,
1566                                        struct pnfs_layout_segment *lseg)
1567 {
1568         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1569         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1570         struct pnfs_layout_range range = {
1571                 .iomode = lseg->pls_range.iomode,
1572                 .offset = 0,
1573                 .length = NFS4_MAX_UINT64,
1574         };
1575         LIST_HEAD(free_me);
1576
1577         spin_lock(&inode->i_lock);
1578         /* set failure bit so that pnfs path will be retried later */
1579         pnfs_layout_set_fail_bit(lo, iomode);
1580         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1581         if (lo->plh_return_iomode == 0)
1582                 lo->plh_return_iomode = range.iomode;
1583         else if (lo->plh_return_iomode != range.iomode)
1584                 lo->plh_return_iomode = IOMODE_ANY;
1585         /*
1586          * mark all matching lsegs so that we are sure to have no live
1587          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1588          * for how it works.
1589          */
1590         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1591         spin_unlock(&inode->i_lock);
1592         pnfs_free_lseg_list(&free_me);
1593 }
1594 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1595
1596 void
1597 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1598 {
1599         u64 rd_size = req->wb_bytes;
1600
1601         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1602
1603         if (pgio->pg_dreq == NULL)
1604                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1605         else
1606                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1607
1608         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1609                                            req->wb_context,
1610                                            req_offset(req),
1611                                            rd_size,
1612                                            IOMODE_READ,
1613                                            GFP_KERNEL);
1614         /* If no lseg, fall back to read through mds */
1615         if (pgio->pg_lseg == NULL)
1616                 nfs_pageio_reset_read_mds(pgio);
1617
1618 }
1619 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1620
1621 void
1622 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1623                            struct nfs_page *req, u64 wb_size)
1624 {
1625         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1626
1627         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1628                                            req->wb_context,
1629                                            req_offset(req),
1630                                            wb_size,
1631                                            IOMODE_RW,
1632                                            GFP_NOFS);
1633         /* If no lseg, fall back to write through mds */
1634         if (pgio->pg_lseg == NULL)
1635                 nfs_pageio_reset_write_mds(pgio);
1636 }
1637 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1638
1639 void
1640 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1641 {
1642         if (desc->pg_lseg) {
1643                 pnfs_put_lseg(desc->pg_lseg);
1644                 desc->pg_lseg = NULL;
1645         }
1646 }
1647 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1648
1649 /*
1650  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1651  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1652  */
1653 size_t
1654 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1655                      struct nfs_page *prev, struct nfs_page *req)
1656 {
1657         unsigned int size;
1658         u64 seg_end, req_start, seg_left;
1659
1660         size = nfs_generic_pg_test(pgio, prev, req);
1661         if (!size)
1662                 return 0;
1663
1664         /*
1665          * 'size' contains the number of bytes left in the current page (up
1666          * to the original size asked for in @req->wb_bytes).
1667          *
1668          * Calculate how many bytes are left in the layout segment
1669          * and if there are less bytes than 'size', return that instead.
1670          *
1671          * Please also note that 'end_offset' is actually the offset of the
1672          * first byte that lies outside the pnfs_layout_range. FIXME?
1673          *
1674          */
1675         if (pgio->pg_lseg) {
1676                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1677                                      pgio->pg_lseg->pls_range.length);
1678                 req_start = req_offset(req);
1679                 WARN_ON_ONCE(req_start > seg_end);
1680                 /* start of request is past the last byte of this segment */
1681                 if (req_start >= seg_end)
1682                         return 0;
1683
1684                 /* adjust 'size' iff there are fewer bytes left in the
1685                  * segment than what nfs_generic_pg_test returned */
1686                 seg_left = seg_end - req_start;
1687                 if (seg_left < size)
1688                         size = (unsigned int)seg_left;
1689         }
1690
1691         return size;
1692 }
1693 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1694
1695 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1696 {
1697         struct nfs_pageio_descriptor pgio;
1698
1699         /* Resend all requests through the MDS */
1700         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1701                               hdr->completion_ops);
1702         return nfs_pageio_resend(&pgio, hdr);
1703 }
1704 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1705
1706 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1707 {
1708
1709         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1710         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1711             PNFS_LAYOUTRET_ON_ERROR) {
1712                 pnfs_return_layout(hdr->inode);
1713         }
1714         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1715                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1716 }
1717
1718 /*
1719  * Called by non rpc-based layout drivers
1720  */
1721 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1722 {
1723         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1724         if (!hdr->pnfs_error) {
1725                 pnfs_set_layoutcommit(hdr);
1726                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1727         } else
1728                 pnfs_ld_handle_write_error(hdr);
1729         hdr->mds_ops->rpc_release(hdr);
1730 }
1731 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1732
1733 static void
1734 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1735                 struct nfs_pgio_header *hdr)
1736 {
1737         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1738
1739         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1740                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1741                 nfs_pageio_reset_write_mds(desc);
1742                 mirror->pg_recoalesce = 1;
1743         }
1744         nfs_pgio_data_destroy(hdr);
1745 }
1746
1747 static enum pnfs_try_status
1748 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1749                         const struct rpc_call_ops *call_ops,
1750                         struct pnfs_layout_segment *lseg,
1751                         int how)
1752 {
1753         struct inode *inode = hdr->inode;
1754         enum pnfs_try_status trypnfs;
1755         struct nfs_server *nfss = NFS_SERVER(inode);
1756
1757         hdr->mds_ops = call_ops;
1758
1759         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1760                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1761         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1762         if (trypnfs != PNFS_NOT_ATTEMPTED)
1763                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1764         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1765         return trypnfs;
1766 }
1767
1768 static void
1769 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1770               struct nfs_pgio_header *hdr, int how)
1771 {
1772         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1773         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1774         enum pnfs_try_status trypnfs;
1775
1776         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1777         if (trypnfs == PNFS_NOT_ATTEMPTED)
1778                 pnfs_write_through_mds(desc, hdr);
1779 }
1780
1781 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1782 {
1783         pnfs_put_lseg(hdr->lseg);
1784         nfs_pgio_header_free(hdr);
1785 }
1786 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1787
1788 int
1789 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1790 {
1791         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1792
1793         struct nfs_pgio_header *hdr;
1794         int ret;
1795
1796         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1797         if (!hdr) {
1798                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1799                 return -ENOMEM;
1800         }
1801         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1802
1803         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1804         ret = nfs_generic_pgio(desc, hdr);
1805         if (!ret)
1806                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1807
1808         return ret;
1809 }
1810 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1811
1812 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1813 {
1814         struct nfs_pageio_descriptor pgio;
1815
1816         /* Resend all requests through the MDS */
1817         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1818         return nfs_pageio_resend(&pgio, hdr);
1819 }
1820 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1821
1822 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1823 {
1824         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1825         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1826             PNFS_LAYOUTRET_ON_ERROR) {
1827                 pnfs_return_layout(hdr->inode);
1828         }
1829         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1830                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1831 }
1832
1833 /*
1834  * Called by non rpc-based layout drivers
1835  */
1836 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1837 {
1838         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1839         if (likely(!hdr->pnfs_error)) {
1840                 __nfs4_read_done_cb(hdr);
1841                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1842         } else
1843                 pnfs_ld_handle_read_error(hdr);
1844         hdr->mds_ops->rpc_release(hdr);
1845 }
1846 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1847
1848 static void
1849 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1850                 struct nfs_pgio_header *hdr)
1851 {
1852         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1853
1854         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1855                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1856                 nfs_pageio_reset_read_mds(desc);
1857                 mirror->pg_recoalesce = 1;
1858         }
1859         nfs_pgio_data_destroy(hdr);
1860 }
1861
1862 /*
1863  * Call the appropriate parallel I/O subsystem read function.
1864  */
1865 static enum pnfs_try_status
1866 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1867                        const struct rpc_call_ops *call_ops,
1868                        struct pnfs_layout_segment *lseg)
1869 {
1870         struct inode *inode = hdr->inode;
1871         struct nfs_server *nfss = NFS_SERVER(inode);
1872         enum pnfs_try_status trypnfs;
1873
1874         hdr->mds_ops = call_ops;
1875
1876         dprintk("%s: Reading ino:%lu %u@%llu\n",
1877                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1878
1879         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1880         if (trypnfs != PNFS_NOT_ATTEMPTED)
1881                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1882         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1883         return trypnfs;
1884 }
1885
1886 /* Resend all requests through pnfs. */
1887 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
1888 {
1889         struct nfs_pageio_descriptor pgio;
1890
1891         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
1892         return nfs_pageio_resend(&pgio, hdr);
1893 }
1894 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
1895
1896 static void
1897 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1898 {
1899         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1900         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1901         enum pnfs_try_status trypnfs;
1902         int err = 0;
1903
1904         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1905         if (trypnfs == PNFS_TRY_AGAIN)
1906                 err = pnfs_read_resend_pnfs(hdr);
1907         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
1908                 pnfs_read_through_mds(desc, hdr);
1909 }
1910
1911 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1912 {
1913         pnfs_put_lseg(hdr->lseg);
1914         nfs_pgio_header_free(hdr);
1915 }
1916 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1917
1918 int
1919 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1920 {
1921         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1922
1923         struct nfs_pgio_header *hdr;
1924         int ret;
1925
1926         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1927         if (!hdr) {
1928                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1929                 return -ENOMEM;
1930         }
1931         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1932         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1933         ret = nfs_generic_pgio(desc, hdr);
1934         if (!ret)
1935                 pnfs_do_read(desc, hdr);
1936         return ret;
1937 }
1938 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1939
1940 static void pnfs_clear_layoutcommitting(struct inode *inode)
1941 {
1942         unsigned long *bitlock = &NFS_I(inode)->flags;
1943
1944         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1945         smp_mb__after_atomic();
1946         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1947 }
1948
1949 /*
1950  * There can be multiple RW segments.
1951  */
1952 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1953 {
1954         struct pnfs_layout_segment *lseg;
1955
1956         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1957                 if (lseg->pls_range.iomode == IOMODE_RW &&
1958                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1959                         list_add(&lseg->pls_lc_list, listp);
1960         }
1961 }
1962
1963 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1964 {
1965         struct pnfs_layout_segment *lseg, *tmp;
1966
1967         /* Matched by references in pnfs_set_layoutcommit */
1968         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1969                 list_del_init(&lseg->pls_lc_list);
1970                 pnfs_put_lseg(lseg);
1971         }
1972
1973         pnfs_clear_layoutcommitting(inode);
1974 }
1975
1976 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1977 {
1978         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1979 }
1980 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1981
1982 void
1983 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1984 {
1985         struct inode *inode = hdr->inode;
1986         struct nfs_inode *nfsi = NFS_I(inode);
1987         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1988         bool mark_as_dirty = false;
1989
1990         spin_lock(&inode->i_lock);
1991         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1992                 mark_as_dirty = true;
1993                 dprintk("%s: Set layoutcommit for inode %lu ",
1994                         __func__, inode->i_ino);
1995         }
1996         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1997                 /* references matched in nfs4_layoutcommit_release */
1998                 pnfs_get_lseg(hdr->lseg);
1999         }
2000         if (end_pos > nfsi->layout->plh_lwb)
2001                 nfsi->layout->plh_lwb = end_pos;
2002         spin_unlock(&inode->i_lock);
2003         dprintk("%s: lseg %p end_pos %llu\n",
2004                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
2005
2006         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2007          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2008         if (mark_as_dirty)
2009                 mark_inode_dirty_sync(inode);
2010 }
2011 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2012
2013 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
2014 {
2015         struct inode *inode = data->inode;
2016         struct nfs_inode *nfsi = NFS_I(inode);
2017         bool mark_as_dirty = false;
2018
2019         spin_lock(&inode->i_lock);
2020         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2021                 mark_as_dirty = true;
2022                 dprintk("%s: Set layoutcommit for inode %lu ",
2023                         __func__, inode->i_ino);
2024         }
2025         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
2026                 /* references matched in nfs4_layoutcommit_release */
2027                 pnfs_get_lseg(data->lseg);
2028         }
2029         if (data->lwb > nfsi->layout->plh_lwb)
2030                 nfsi->layout->plh_lwb = data->lwb;
2031         spin_unlock(&inode->i_lock);
2032         dprintk("%s: lseg %p end_pos %llu\n",
2033                 __func__, data->lseg, nfsi->layout->plh_lwb);
2034
2035         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2036          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2037         if (mark_as_dirty)
2038                 mark_inode_dirty_sync(inode);
2039 }
2040 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
2041
2042 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2043 {
2044         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2045
2046         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2047                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2048         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2049 }
2050
2051 /*
2052  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2053  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2054  * data to disk to allow the server to recover the data if it crashes.
2055  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2056  * is off, and a COMMIT is sent to a data server, or
2057  * if WRITEs to a data server return NFS_DATA_SYNC.
2058  */
2059 int
2060 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2061 {
2062         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2063         struct nfs4_layoutcommit_data *data;
2064         struct nfs_inode *nfsi = NFS_I(inode);
2065         loff_t end_pos;
2066         int status;
2067
2068         if (!pnfs_layoutcommit_outstanding(inode))
2069                 return 0;
2070
2071         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2072
2073         status = -EAGAIN;
2074         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2075                 if (!sync)
2076                         goto out;
2077                 status = wait_on_bit_lock_action(&nfsi->flags,
2078                                 NFS_INO_LAYOUTCOMMITTING,
2079                                 nfs_wait_bit_killable,
2080                                 TASK_KILLABLE);
2081                 if (status)
2082                         goto out;
2083         }
2084
2085         status = -ENOMEM;
2086         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2087         data = kzalloc(sizeof(*data), GFP_NOFS);
2088         if (!data)
2089                 goto clear_layoutcommitting;
2090
2091         status = 0;
2092         spin_lock(&inode->i_lock);
2093         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2094                 goto out_unlock;
2095
2096         INIT_LIST_HEAD(&data->lseg_list);
2097         pnfs_list_write_lseg(inode, &data->lseg_list);
2098
2099         end_pos = nfsi->layout->plh_lwb;
2100         nfsi->layout->plh_lwb = 0;
2101
2102         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2103         spin_unlock(&inode->i_lock);
2104
2105         data->args.inode = inode;
2106         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2107         nfs_fattr_init(&data->fattr);
2108         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2109         data->res.fattr = &data->fattr;
2110         data->args.lastbytewritten = end_pos - 1;
2111         data->res.server = NFS_SERVER(inode);
2112
2113         if (ld->prepare_layoutcommit) {
2114                 status = ld->prepare_layoutcommit(&data->args);
2115                 if (status) {
2116                         spin_lock(&inode->i_lock);
2117                         if (end_pos < nfsi->layout->plh_lwb)
2118                                 nfsi->layout->plh_lwb = end_pos;
2119                         spin_unlock(&inode->i_lock);
2120                         put_rpccred(data->cred);
2121                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2122                         goto clear_layoutcommitting;
2123                 }
2124         }
2125
2126
2127         status = nfs4_proc_layoutcommit(data, sync);
2128 out:
2129         if (status)
2130                 mark_inode_dirty_sync(inode);
2131         dprintk("<-- %s status %d\n", __func__, status);
2132         return status;
2133 out_unlock:
2134         spin_unlock(&inode->i_lock);
2135         kfree(data);
2136 clear_layoutcommitting:
2137         pnfs_clear_layoutcommitting(inode);
2138         goto out;
2139 }
2140 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2141
2142 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2143 {
2144         struct nfs4_threshold *thp;
2145
2146         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2147         if (!thp) {
2148                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2149                 return NULL;
2150         }
2151         return thp;
2152 }