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 #include "delegation.h"
  38
  39 #define NFSDBG_FACILITY         NFSDBG_PNFS
  40 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  41
  42 /* Locking:
  43  *
  44  * pnfs_spinlock:
  45  *      protects pnfs_modules_tbl.
  46  */
  47 static DEFINE_SPINLOCK(pnfs_spinlock);
  48
  49 /*
  50  * pnfs_modules_tbl holds all pnfs modules
  51  */
  52 static LIST_HEAD(pnfs_modules_tbl);
  53
  54 static int
  55 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
  56                        enum pnfs_iomode iomode, bool sync);
  57
  58 /* Return the registered pnfs layout driver module matching given id */
  59 static struct pnfs_layoutdriver_type *
  60 find_pnfs_driver_locked(u32 id)
  61 {
  62         struct pnfs_layoutdriver_type *local;
  63
  64         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  65                 if (local->id == id)
  66                         goto out;
  67         local = NULL;
  68 out:
  69         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  70         return local;
  71 }
  72
  73 static struct pnfs_layoutdriver_type *
  74 find_pnfs_driver(u32 id)
  75 {
  76         struct pnfs_layoutdriver_type *local;
  77
  78         spin_lock(&pnfs_spinlock);
  79         local = find_pnfs_driver_locked(id);
  80         if (local != NULL && !try_module_get(local->owner)) {
  81                 dprintk("%s: Could not grab reference on module\n", __func__);
  82                 local = NULL;
  83         }
  84         spin_unlock(&pnfs_spinlock);
  85         return local;
  86 }
  87
  88 void
  89 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  90 {
  91         if (nfss->pnfs_curr_ld) {
  92                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  93                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  94                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  95                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  96                         nfs4_deviceid_purge_client(nfss->nfs_client);
  97                 module_put(nfss->pnfs_curr_ld->owner);
  98         }
  99         nfss->pnfs_curr_ld = NULL;
 100 }
 101
 102 /*
 103  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 104  * Currently only one pNFS layout driver per filesystem is supported.
 105  *
 106  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 107  */
 108 void
 109 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 110                       u32 id)
 111 {
 112         struct pnfs_layoutdriver_type *ld_type = NULL;
 113
 114         if (id == 0)
 115                 goto out_no_driver;
 116         if (!(server->nfs_client->cl_exchange_flags &
 117                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 118                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 119                         __func__, id, server->nfs_client->cl_exchange_flags);
 120                 goto out_no_driver;
 121         }
 122         ld_type = find_pnfs_driver(id);
 123         if (!ld_type) {
 124                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 125                 ld_type = find_pnfs_driver(id);
 126                 if (!ld_type) {
 127                         dprintk("%s: No pNFS module found for %u.\n",
 128                                 __func__, id);
 129                         goto out_no_driver;
 130                 }
 131         }
 132         server->pnfs_curr_ld = ld_type;
 133         if (ld_type->set_layoutdriver
 134             && ld_type->set_layoutdriver(server, mntfh)) {
 135                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 136                         "driver %u.\n", __func__, id);
 137                 module_put(ld_type->owner);
 138                 goto out_no_driver;
 139         }
 140         /* Bump the MDS count */
 141         atomic_inc(&server->nfs_client->cl_mds_count);
 142
 143         dprintk("%s: pNFS module for %u set\n", __func__, id);
 144         return;
 145
 146 out_no_driver:
 147         dprintk("%s: Using NFSv4 I/O\n", __func__);
 148         server->pnfs_curr_ld = NULL;
 149 }
 150
 151 int
 152 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 153 {
 154         int status = -EINVAL;
 155         struct pnfs_layoutdriver_type *tmp;
 156
 157         if (ld_type->id == 0) {
 158                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 159                 return status;
 160         }
 161         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 162                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 163                        "alloc_lseg and free_lseg.\n", __func__);
 164                 return status;
 165         }
 166
 167         spin_lock(&pnfs_spinlock);
 168         tmp = find_pnfs_driver_locked(ld_type->id);
 169         if (!tmp) {
 170                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 171                 status = 0;
 172                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 173                         ld_type->name);
 174         } else {
 175                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 176                         __func__, ld_type->id);
 177         }
 178         spin_unlock(&pnfs_spinlock);
 179
 180         return status;
 181 }
 182 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 183
 184 void
 185 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 186 {
 187         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 188         spin_lock(&pnfs_spinlock);
 189         list_del(&ld_type->pnfs_tblid);
 190         spin_unlock(&pnfs_spinlock);
 191 }
 192 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 193
 194 /*
 195  * pNFS client layout cache
 196  */
 197
 198 /* Need to hold i_lock if caller does not already hold reference */
 199 void
 200 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 201 {
 202         atomic_inc(&lo->plh_refcount);
 203 }
 204
 205 static struct pnfs_layout_hdr *
 206 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 207 {
 208         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 209         return ld->alloc_layout_hdr(ino, gfp_flags);
 210 }
 211
 212 static void
 213 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 214 {
 215         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 216         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 217
 218         if (!list_empty(&lo->plh_layouts)) {
 219                 struct nfs_client *clp = server->nfs_client;
 220
 221                 spin_lock(&clp->cl_lock);
 222                 list_del_init(&lo->plh_layouts);
 223                 spin_unlock(&clp->cl_lock);
 224         }
 225         put_rpccred(lo->plh_lc_cred);
 226         return ld->free_layout_hdr(lo);
 227 }
 228
 229 static void
 230 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 231 {
 232         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 233         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 234         nfsi->layout = NULL;
 235         /* Reset MDS Threshold I/O counters */
 236         nfsi->write_io = 0;
 237         nfsi->read_io = 0;
 238 }
 239
 240 void
 241 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 242 {
 243         struct inode *inode = lo->plh_inode;
 244
 245         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 246                 if (!list_empty(&lo->plh_segs))
 247                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 248                 pnfs_detach_layout_hdr(lo);
 249                 spin_unlock(&inode->i_lock);
 250                 pnfs_free_layout_hdr(lo);
 251         }
 252 }
 253
 254 static int
 255 pnfs_iomode_to_fail_bit(u32 iomode)
 256 {
 257         return iomode == IOMODE_RW ?
 258                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 259 }
 260
 261 static void
 262 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 263 {
 264         lo->plh_retry_timestamp = jiffies;
 265         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 266                 atomic_inc(&lo->plh_refcount);
 267 }
 268
 269 static void
 270 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 271 {
 272         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 273                 atomic_dec(&lo->plh_refcount);
 274 }
 275
 276 static void
 277 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 278 {
 279         struct inode *inode = lo->plh_inode;
 280         struct pnfs_layout_range range = {
 281                 .iomode = iomode,
 282                 .offset = 0,
 283                 .length = NFS4_MAX_UINT64,
 284         };
 285         LIST_HEAD(head);
 286
 287         spin_lock(&inode->i_lock);
 288         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 289         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 290         spin_unlock(&inode->i_lock);
 291         pnfs_free_lseg_list(&head);
 292         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 293                         iomode == IOMODE_RW ?  "RW" : "READ");
 294 }
 295
 296 static bool
 297 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 298 {
 299         unsigned long start, end;
 300         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 301
 302         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 303                 return false;
 304         end = jiffies;
 305         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 306         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 307                 /* It is time to retry the failed layoutgets */
 308                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 309                 return false;
 310         }
 311         return true;
 312 }
 313
 314 static void
 315 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 316 {
 317         INIT_LIST_HEAD(&lseg->pls_list);
 318         INIT_LIST_HEAD(&lseg->pls_lc_list);
 319         atomic_set(&lseg->pls_refcount, 1);
 320         smp_mb();
 321         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 322         lseg->pls_layout = lo;
 323 }
 324
 325 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 326 {
 327         struct inode *ino = lseg->pls_layout->plh_inode;
 328
 329         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 330 }
 331
 332 static void
 333 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 334                 struct pnfs_layout_segment *lseg)
 335 {
 336         struct inode *inode = lo->plh_inode;
 337
 338         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 339         list_del_init(&lseg->pls_list);
 340         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 341         atomic_dec(&lo->plh_refcount);
 342         if (list_empty(&lo->plh_segs))
 343                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 344         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 345 }
 346
 347 /* Return true if layoutreturn is needed */
 348 static bool
 349 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 350                         struct pnfs_layout_segment *lseg)
 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 (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 359                         return false;
 360
 361         return true;
 362 }
 363
 364 static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
 365                 struct pnfs_layout_hdr *lo, struct inode *inode)
 366 {
 367         lo = lseg->pls_layout;
 368         inode = lo->plh_inode;
 369
 370         spin_lock(&inode->i_lock);
 371         if (pnfs_layout_need_return(lo, lseg)) {
 372                 nfs4_stateid stateid;
 373                 enum pnfs_iomode iomode;
 374
 375                 stateid = lo->plh_stateid;
 376                 iomode = lo->plh_return_iomode;
 377                 /* decreased in pnfs_send_layoutreturn() */
 378                 lo->plh_block_lgets++;
 379                 lo->plh_return_iomode = 0;
 380                 spin_unlock(&inode->i_lock);
 381                 pnfs_get_layout_hdr(lo);
 382
 383                 /* Send an async layoutreturn so we dont deadlock */
 384                 pnfs_send_layoutreturn(lo, stateid, iomode, false);
 385         } else
 386                 spin_unlock(&inode->i_lock);
 387 }
 388
 389 void
 390 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 391 {
 392         struct pnfs_layout_hdr *lo;
 393         struct inode *inode;
 394
 395         if (!lseg)
 396                 return;
 397
 398         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 399                 atomic_read(&lseg->pls_refcount),
 400                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 401
 402         /* Handle the case where refcount != 1 */
 403         if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
 404                 return;
 405
 406         lo = lseg->pls_layout;
 407         inode = lo->plh_inode;
 408         /* Do we need a layoutreturn? */
 409         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 410                 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
 411
 412         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 413                 pnfs_get_layout_hdr(lo);
 414                 pnfs_layout_remove_lseg(lo, lseg);
 415                 spin_unlock(&inode->i_lock);
 416                 pnfs_free_lseg(lseg);
 417                 pnfs_put_layout_hdr(lo);
 418         }
 419 }
 420 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 421
 422 static void pnfs_free_lseg_async_work(struct work_struct *work)
 423 {
 424         struct pnfs_layout_segment *lseg;
 425         struct pnfs_layout_hdr *lo;
 426
 427         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 428         lo = lseg->pls_layout;
 429
 430         pnfs_free_lseg(lseg);
 431         pnfs_put_layout_hdr(lo);
 432 }
 433
 434 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 435 {
 436         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 437         schedule_work(&lseg->pls_work);
 438 }
 439
 440 void
 441 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 442 {
 443         if (!lseg)
 444                 return;
 445
 446         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 447
 448         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 449                 atomic_read(&lseg->pls_refcount),
 450                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 451         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 452                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 453                 pnfs_get_layout_hdr(lo);
 454                 pnfs_layout_remove_lseg(lo, lseg);
 455                 pnfs_free_lseg_async(lseg);
 456         }
 457 }
 458 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 459
 460 static u64
 461 end_offset(u64 start, u64 len)
 462 {
 463         u64 end;
 464
 465         end = start + len;
 466         return end >= start ? end : NFS4_MAX_UINT64;
 467 }
 468
 469 /*
 470  * is l2 fully contained in l1?
 471  *   start1                             end1
 472  *   [----------------------------------)
 473  *           start2           end2
 474  *           [----------------)
 475  */
 476 static bool
 477 pnfs_lseg_range_contained(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 (start1 <= start2) && (end1 >= end2);
 486 }
 487
 488 /*
 489  * is l1 and l2 intersecting?
 490  *   start1                             end1
 491  *   [----------------------------------)
 492  *                              start2           end2
 493  *                              [----------------)
 494  */
 495 static bool
 496 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 497                     const struct pnfs_layout_range *l2)
 498 {
 499         u64 start1 = l1->offset;
 500         u64 end1 = end_offset(start1, l1->length);
 501         u64 start2 = l2->offset;
 502         u64 end2 = end_offset(start2, l2->length);
 503
 504         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 505                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 506 }
 507
 508 static bool
 509 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 510                  const struct pnfs_layout_range *recall_range)
 511 {
 512         return (recall_range->iomode == IOMODE_ANY ||
 513                 lseg_range->iomode == recall_range->iomode) &&
 514                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 515 }
 516
 517 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 518                 struct list_head *tmp_list)
 519 {
 520         if (!atomic_dec_and_test(&lseg->pls_refcount))
 521                 return false;
 522         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 523         list_add(&lseg->pls_list, tmp_list);
 524         return true;
 525 }
 526
 527 /* Returns 1 if lseg is removed from list, 0 otherwise */
 528 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 529                              struct list_head *tmp_list)
 530 {
 531         int rv = 0;
 532
 533         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 534                 /* Remove the reference keeping the lseg in the
 535                  * list.  It will now be removed when all
 536                  * outstanding io is finished.
 537                  */
 538                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 539                         atomic_read(&lseg->pls_refcount));
 540                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 541                         rv = 1;
 542         }
 543         return rv;
 544 }
 545
 546 /* Returns count of number of matching invalid lsegs remaining in list
 547  * after call.
 548  */
 549 int
 550 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 551                             struct list_head *tmp_list,
 552                             struct pnfs_layout_range *recall_range)
 553 {
 554         struct pnfs_layout_segment *lseg, *next;
 555         int invalid = 0, removed = 0;
 556
 557         dprintk("%s:Begin lo %p\n", __func__, lo);
 558
 559         if (list_empty(&lo->plh_segs))
 560                 return 0;
 561         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 562                 if (!recall_range ||
 563                     should_free_lseg(&lseg->pls_range, recall_range)) {
 564                         dprintk("%s: freeing lseg %p iomode %d "
 565                                 "offset %llu length %llu\n", __func__,
 566                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 567                                 lseg->pls_range.length);
 568                         invalid++;
 569                         removed += mark_lseg_invalid(lseg, tmp_list);
 570                 }
 571         dprintk("%s:Return %i\n", __func__, invalid - removed);
 572         return invalid - removed;
 573 }
 574
 575 /* note free_me must contain lsegs from a single layout_hdr */
 576 void
 577 pnfs_free_lseg_list(struct list_head *free_me)
 578 {
 579         struct pnfs_layout_segment *lseg, *tmp;
 580
 581         if (list_empty(free_me))
 582                 return;
 583
 584         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 585                 list_del(&lseg->pls_list);
 586                 pnfs_free_lseg(lseg);
 587         }
 588 }
 589
 590 void
 591 pnfs_destroy_layout(struct nfs_inode *nfsi)
 592 {
 593         struct pnfs_layout_hdr *lo;
 594         LIST_HEAD(tmp_list);
 595
 596         spin_lock(&nfsi->vfs_inode.i_lock);
 597         lo = nfsi->layout;
 598         if (lo) {
 599                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 600                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 601                 pnfs_get_layout_hdr(lo);
 602                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 603                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 604                 pnfs_clear_retry_layoutget(lo);
 605                 spin_unlock(&nfsi->vfs_inode.i_lock);
 606                 pnfs_free_lseg_list(&tmp_list);
 607                 pnfs_put_layout_hdr(lo);
 608         } else
 609                 spin_unlock(&nfsi->vfs_inode.i_lock);
 610 }
 611 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 612
 613 static bool
 614 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 615                 struct list_head *layout_list)
 616 {
 617         struct pnfs_layout_hdr *lo;
 618         bool ret = false;
 619
 620         spin_lock(&inode->i_lock);
 621         lo = NFS_I(inode)->layout;
 622         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 623                 pnfs_get_layout_hdr(lo);
 624                 list_add(&lo->plh_bulk_destroy, layout_list);
 625                 ret = true;
 626         }
 627         spin_unlock(&inode->i_lock);
 628         return ret;
 629 }
 630
 631 /* Caller must hold rcu_read_lock and clp->cl_lock */
 632 static int
 633 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 634                 struct nfs_server *server,
 635                 struct list_head *layout_list)
 636 {
 637         struct pnfs_layout_hdr *lo, *next;
 638         struct inode *inode;
 639
 640         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 641                 inode = igrab(lo->plh_inode);
 642                 if (inode == NULL)
 643                         continue;
 644                 list_del_init(&lo->plh_layouts);
 645                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 646                         continue;
 647                 rcu_read_unlock();
 648                 spin_unlock(&clp->cl_lock);
 649                 iput(inode);
 650                 spin_lock(&clp->cl_lock);
 651                 rcu_read_lock();
 652                 return -EAGAIN;
 653         }
 654         return 0;
 655 }
 656
 657 static int
 658 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 659                 bool is_bulk_recall)
 660 {
 661         struct pnfs_layout_hdr *lo;
 662         struct inode *inode;
 663         struct pnfs_layout_range range = {
 664                 .iomode = IOMODE_ANY,
 665                 .offset = 0,
 666                 .length = NFS4_MAX_UINT64,
 667         };
 668         LIST_HEAD(lseg_list);
 669         int ret = 0;
 670
 671         while (!list_empty(layout_list)) {
 672                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 673                                 plh_bulk_destroy);
 674                 dprintk("%s freeing layout for inode %lu\n", __func__,
 675                         lo->plh_inode->i_ino);
 676                 inode = lo->plh_inode;
 677
 678                 pnfs_layoutcommit_inode(inode, false);
 679
 680                 spin_lock(&inode->i_lock);
 681                 list_del_init(&lo->plh_bulk_destroy);
 682                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 683                 if (is_bulk_recall)
 684                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 685                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 686                         ret = -EAGAIN;
 687                 spin_unlock(&inode->i_lock);
 688                 pnfs_free_lseg_list(&lseg_list);
 689                 pnfs_put_layout_hdr(lo);
 690                 iput(inode);
 691         }
 692         return ret;
 693 }
 694
 695 int
 696 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 697                 struct nfs_fsid *fsid,
 698                 bool is_recall)
 699 {
 700         struct nfs_server *server;
 701         LIST_HEAD(layout_list);
 702
 703         spin_lock(&clp->cl_lock);
 704         rcu_read_lock();
 705 restart:
 706         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 707                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 708                         continue;
 709                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 710                                 server,
 711                                 &layout_list) != 0)
 712                         goto restart;
 713         }
 714         rcu_read_unlock();
 715         spin_unlock(&clp->cl_lock);
 716
 717         if (list_empty(&layout_list))
 718                 return 0;
 719         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 720 }
 721
 722 int
 723 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 724                 bool is_recall)
 725 {
 726         struct nfs_server *server;
 727         LIST_HEAD(layout_list);
 728
 729         spin_lock(&clp->cl_lock);
 730         rcu_read_lock();
 731 restart:
 732         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 733                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 734                                         server,
 735                                         &layout_list) != 0)
 736                         goto restart;
 737         }
 738         rcu_read_unlock();
 739         spin_unlock(&clp->cl_lock);
 740
 741         if (list_empty(&layout_list))
 742                 return 0;
 743         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 744 }
 745
 746 /*
 747  * Called by the state manger to remove all layouts established under an
 748  * expired lease.
 749  */
 750 void
 751 pnfs_destroy_all_layouts(struct nfs_client *clp)
 752 {
 753         nfs4_deviceid_mark_client_invalid(clp);
 754         nfs4_deviceid_purge_client(clp);
 755
 756         pnfs_destroy_layouts_byclid(clp, false);
 757 }
 758
 759 /*
 760  * Compare 2 layout stateid sequence ids, to see which is newer,
 761  * taking into account wraparound issues.
 762  */
 763 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 764 {
 765         return (s32)(s1 - s2) > 0;
 766 }
 767
 768 /* update lo->plh_stateid with new if is more recent */
 769 void
 770 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 771                         bool update_barrier)
 772 {
 773         u32 oldseq, newseq, new_barrier;
 774         int empty = list_empty(&lo->plh_segs);
 775
 776         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 777         newseq = be32_to_cpu(new->seqid);
 778         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 779                 nfs4_stateid_copy(&lo->plh_stateid, new);
 780                 if (update_barrier) {
 781                         new_barrier = be32_to_cpu(new->seqid);
 782                 } else {
 783                         /* Because of wraparound, we want to keep the barrier
 784                          * "close" to the current seqids.
 785                          */
 786                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 787                 }
 788                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 789                         lo->plh_barrier = new_barrier;
 790         }
 791 }
 792
 793 static bool
 794 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 795                 const nfs4_stateid *stateid)
 796 {
 797         u32 seqid = be32_to_cpu(stateid->seqid);
 798
 799         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 800 }
 801
 802 static bool
 803 pnfs_layout_returning(const struct pnfs_layout_hdr *lo,
 804                       struct pnfs_layout_range *range)
 805 {
 806         return test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 807                 (lo->plh_return_iomode == IOMODE_ANY ||
 808                  lo->plh_return_iomode == range->iomode);
 809 }
 810
 811 /* lget is set to 1 if called from inside send_layoutget call chain */
 812 static bool
 813 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo,
 814                         struct pnfs_layout_range *range, int lget)
 815 {
 816         return lo->plh_block_lgets ||
 817                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 818                 (list_empty(&lo->plh_segs) &&
 819                  (atomic_read(&lo->plh_outstanding) > lget)) ||
 820                 pnfs_layout_returning(lo, range);
 821 }
 822
 823 int
 824 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 825                               struct pnfs_layout_range *range,
 826                               struct nfs4_state *open_state)
 827 {
 828         int status = 0;
 829
 830         dprintk("--> %s\n", __func__);
 831         spin_lock(&lo->plh_inode->i_lock);
 832         if (pnfs_layoutgets_blocked(lo, range, 1)) {
 833                 status = -EAGAIN;
 834         } else if (!nfs4_valid_open_stateid(open_state)) {
 835                 status = -EBADF;
 836         } else if (list_empty(&lo->plh_segs) ||
 837                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 838                 int seq;
 839
 840                 do {
 841                         seq = read_seqbegin(&open_state->seqlock);
 842                         nfs4_stateid_copy(dst, &open_state->stateid);
 843                 } while (read_seqretry(&open_state->seqlock, seq));
 844         } else
 845                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 846         spin_unlock(&lo->plh_inode->i_lock);
 847         dprintk("<-- %s\n", __func__);
 848         return status;
 849 }
 850
 851 /*
 852 * Get layout from server.
 853 *    for now, assume that whole file layouts are requested.
 854 *    arg->offset: 0
 855 *    arg->length: all ones
 856 */
 857 static struct pnfs_layout_segment *
 858 send_layoutget(struct pnfs_layout_hdr *lo,
 859            struct nfs_open_context *ctx,
 860            struct pnfs_layout_range *range,
 861            gfp_t gfp_flags)
 862 {
 863         struct inode *ino = lo->plh_inode;
 864         struct nfs_server *server = NFS_SERVER(ino);
 865         struct nfs4_layoutget *lgp;
 866         struct pnfs_layout_segment *lseg;
 867
 868         dprintk("--> %s\n", __func__);
 869
 870         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 871         if (lgp == NULL)
 872                 return NULL;
 873
 874         lgp->args.minlength = PAGE_CACHE_SIZE;
 875         if (lgp->args.minlength > range->length)
 876                 lgp->args.minlength = range->length;
 877         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 878         lgp->args.range = *range;
 879         lgp->args.type = server->pnfs_curr_ld->id;
 880         lgp->args.inode = ino;
 881         lgp->args.ctx = get_nfs_open_context(ctx);
 882         lgp->gfp_flags = gfp_flags;
 883         lgp->cred = lo->plh_lc_cred;
 884
 885         /* Synchronously retrieve layout information from server and
 886          * store in lseg.
 887          */
 888         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 889         if (IS_ERR(lseg)) {
 890                 switch (PTR_ERR(lseg)) {
 891                 case -ENOMEM:
 892                 case -ERESTARTSYS:
 893                         break;
 894                 default:
 895                         /* remember that LAYOUTGET failed and suspend trying */
 896                         pnfs_layout_io_set_failed(lo, range->iomode);
 897                 }
 898                 return NULL;
 899         } else
 900                 pnfs_layout_clear_fail_bit(lo,
 901                                 pnfs_iomode_to_fail_bit(range->iomode));
 902
 903         return lseg;
 904 }
 905
 906 static void pnfs_clear_layoutcommit(struct inode *inode,
 907                 struct list_head *head)
 908 {
 909         struct nfs_inode *nfsi = NFS_I(inode);
 910         struct pnfs_layout_segment *lseg, *tmp;
 911
 912         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 913                 return;
 914         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 915                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 916                         continue;
 917                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 918         }
 919 }
 920
 921 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 922 {
 923         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 924         smp_mb__after_atomic();
 925         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 926 }
 927
 928 static int
 929 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 930                        enum pnfs_iomode iomode, bool sync)
 931 {
 932         struct inode *ino = lo->plh_inode;
 933         struct nfs4_layoutreturn *lrp;
 934         int status = 0;
 935
 936         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 937         if (unlikely(lrp == NULL)) {
 938                 status = -ENOMEM;
 939                 spin_lock(&ino->i_lock);
 940                 lo->plh_block_lgets--;
 941                 pnfs_clear_layoutreturn_waitbit(lo);
 942                 rpc_wake_up(&NFS_SERVER(ino)->roc_rpcwaitq);
 943                 spin_unlock(&ino->i_lock);
 944                 pnfs_put_layout_hdr(lo);
 945                 goto out;
 946         }
 947
 948         lrp->args.stateid = stateid;
 949         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 950         lrp->args.inode = ino;
 951         lrp->args.range.iomode = iomode;
 952         lrp->args.range.offset = 0;
 953         lrp->args.range.length = NFS4_MAX_UINT64;
 954         lrp->args.layout = lo;
 955         lrp->clp = NFS_SERVER(ino)->nfs_client;
 956         lrp->cred = lo->plh_lc_cred;
 957
 958         status = nfs4_proc_layoutreturn(lrp, sync);
 959 out:
 960         dprintk("<-- %s status: %d\n", __func__, status);
 961         return status;
 962 }
 963
 964 /*
 965  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 966  * when the layout segment list is empty.
 967  *
 968  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 969  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 970  * deviceid is marked invalid.
 971  */
 972 int
 973 _pnfs_return_layout(struct inode *ino)
 974 {
 975         struct pnfs_layout_hdr *lo = NULL;
 976         struct nfs_inode *nfsi = NFS_I(ino);
 977         LIST_HEAD(tmp_list);
 978         nfs4_stateid stateid;
 979         int status = 0, empty;
 980
 981         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 982
 983         spin_lock(&ino->i_lock);
 984         lo = nfsi->layout;
 985         if (!lo) {
 986                 spin_unlock(&ino->i_lock);
 987                 dprintk("NFS: %s no layout to return\n", __func__);
 988                 goto out;
 989         }
 990         stateid = nfsi->layout->plh_stateid;
 991         /* Reference matched in nfs4_layoutreturn_release */
 992         pnfs_get_layout_hdr(lo);
 993         empty = list_empty(&lo->plh_segs);
 994         pnfs_clear_layoutcommit(ino, &tmp_list);
 995         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 996
 997         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 998                 struct pnfs_layout_range range = {
 999                         .iomode         = IOMODE_ANY,
1000                         .offset         = 0,
1001                         .length         = NFS4_MAX_UINT64,
1002                 };
1003                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1004         }
1005
1006         /* Don't send a LAYOUTRETURN if list was initially empty */
1007         if (empty) {
1008                 spin_unlock(&ino->i_lock);
1009                 pnfs_put_layout_hdr(lo);
1010                 dprintk("NFS: %s no layout segments to return\n", __func__);
1011                 goto out;
1012         }
1013
1014         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1015         lo->plh_block_lgets++;
1016         spin_unlock(&ino->i_lock);
1017         pnfs_free_lseg_list(&tmp_list);
1018
1019         status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1020 out:
1021         dprintk("<-- %s status: %d\n", __func__, status);
1022         return status;
1023 }
1024 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1025
1026 int
1027 pnfs_commit_and_return_layout(struct inode *inode)
1028 {
1029         struct pnfs_layout_hdr *lo;
1030         int ret;
1031
1032         spin_lock(&inode->i_lock);
1033         lo = NFS_I(inode)->layout;
1034         if (lo == NULL) {
1035                 spin_unlock(&inode->i_lock);
1036                 return 0;
1037         }
1038         pnfs_get_layout_hdr(lo);
1039         /* Block new layoutgets and read/write to ds */
1040         lo->plh_block_lgets++;
1041         spin_unlock(&inode->i_lock);
1042         filemap_fdatawait(inode->i_mapping);
1043         ret = pnfs_layoutcommit_inode(inode, true);
1044         if (ret == 0)
1045                 ret = _pnfs_return_layout(inode);
1046         spin_lock(&inode->i_lock);
1047         lo->plh_block_lgets--;
1048         spin_unlock(&inode->i_lock);
1049         pnfs_put_layout_hdr(lo);
1050         return ret;
1051 }
1052
1053 bool pnfs_roc(struct inode *ino)
1054 {
1055         struct nfs_inode *nfsi = NFS_I(ino);
1056         struct nfs_open_context *ctx;
1057         struct nfs4_state *state;
1058         struct pnfs_layout_hdr *lo;
1059         struct pnfs_layout_segment *lseg, *tmp;
1060         nfs4_stateid stateid;
1061         LIST_HEAD(tmp_list);
1062         bool found = false, layoutreturn = false;
1063
1064         spin_lock(&ino->i_lock);
1065         lo = nfsi->layout;
1066         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
1067             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1068                 goto out_noroc;
1069
1070         /* Don't return layout if we hold a delegation */
1071         if (nfs4_check_delegation(ino, FMODE_READ))
1072                 goto out_noroc;
1073
1074         list_for_each_entry(ctx, &nfsi->open_files, list) {
1075                 state = ctx->state;
1076                 /* Don't return layout if there is open file state */
1077                 if (state != NULL && state->state != 0)
1078                         goto out_noroc;
1079         }
1080
1081         pnfs_clear_retry_layoutget(lo);
1082         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1083                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1084                         mark_lseg_invalid(lseg, &tmp_list);
1085                         found = true;
1086                 }
1087         if (!found)
1088                 goto out_noroc;
1089         lo->plh_block_lgets++;
1090         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
1091         spin_unlock(&ino->i_lock);
1092         pnfs_free_lseg_list(&tmp_list);
1093         return true;
1094
1095 out_noroc:
1096         if (lo) {
1097                 stateid = lo->plh_stateid;
1098                 layoutreturn =
1099                         test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1100                                            &lo->plh_flags);
1101                 if (layoutreturn) {
1102                         lo->plh_block_lgets++;
1103                         pnfs_get_layout_hdr(lo);
1104                 }
1105         }
1106         spin_unlock(&ino->i_lock);
1107         if (layoutreturn)
1108                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1109         return false;
1110 }
1111
1112 void pnfs_roc_release(struct inode *ino)
1113 {
1114         struct pnfs_layout_hdr *lo;
1115
1116         spin_lock(&ino->i_lock);
1117         lo = NFS_I(ino)->layout;
1118         lo->plh_block_lgets--;
1119         if (atomic_dec_and_test(&lo->plh_refcount)) {
1120                 pnfs_detach_layout_hdr(lo);
1121                 spin_unlock(&ino->i_lock);
1122                 pnfs_free_layout_hdr(lo);
1123         } else
1124                 spin_unlock(&ino->i_lock);
1125 }
1126
1127 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1128 {
1129         struct pnfs_layout_hdr *lo;
1130
1131         spin_lock(&ino->i_lock);
1132         lo = NFS_I(ino)->layout;
1133         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1134                 lo->plh_barrier = barrier;
1135         spin_unlock(&ino->i_lock);
1136 }
1137
1138 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
1139 {
1140         struct nfs_inode *nfsi = NFS_I(ino);
1141         struct pnfs_layout_hdr *lo;
1142         struct pnfs_layout_segment *lseg;
1143         nfs4_stateid stateid;
1144         u32 current_seqid;
1145         bool found = false, layoutreturn = false;
1146
1147         spin_lock(&ino->i_lock);
1148         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1149                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1150                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1151                         found = true;
1152                         goto out;
1153                 }
1154         lo = nfsi->layout;
1155         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1156
1157         /* Since close does not return a layout stateid for use as
1158          * a barrier, we choose the worst-case barrier.
1159          */
1160         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1161 out:
1162         if (!found) {
1163                 stateid = lo->plh_stateid;
1164                 layoutreturn =
1165                         test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1166                                            &lo->plh_flags);
1167                 if (layoutreturn) {
1168                         lo->plh_block_lgets++;
1169                         pnfs_get_layout_hdr(lo);
1170                 }
1171         }
1172         spin_unlock(&ino->i_lock);
1173         if (layoutreturn) {
1174                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1175                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, false);
1176         }
1177         return found;
1178 }
1179
1180 /*
1181  * Compare two layout segments for sorting into layout cache.
1182  * We want to preferentially return RW over RO layouts, so ensure those
1183  * are seen first.
1184  */
1185 static s64
1186 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1187            const struct pnfs_layout_range *l2)
1188 {
1189         s64 d;
1190
1191         /* high offset > low offset */
1192         d = l1->offset - l2->offset;
1193         if (d)
1194                 return d;
1195
1196         /* short length > long length */
1197         d = l2->length - l1->length;
1198         if (d)
1199                 return d;
1200
1201         /* read > read/write */
1202         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1203 }
1204
1205 static void
1206 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1207                    struct pnfs_layout_segment *lseg)
1208 {
1209         struct pnfs_layout_segment *lp;
1210
1211         dprintk("%s:Begin\n", __func__);
1212
1213         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1214                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1215                         continue;
1216                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1217                 dprintk("%s: inserted lseg %p "
1218                         "iomode %d offset %llu length %llu before "
1219                         "lp %p iomode %d offset %llu length %llu\n",
1220                         __func__, lseg, lseg->pls_range.iomode,
1221                         lseg->pls_range.offset, lseg->pls_range.length,
1222                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1223                         lp->pls_range.length);
1224                 goto out;
1225         }
1226         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1227         dprintk("%s: inserted lseg %p "
1228                 "iomode %d offset %llu length %llu at tail\n",
1229                 __func__, lseg, lseg->pls_range.iomode,
1230                 lseg->pls_range.offset, lseg->pls_range.length);
1231 out:
1232         pnfs_get_layout_hdr(lo);
1233
1234         dprintk("%s:Return\n", __func__);
1235 }
1236
1237 static struct pnfs_layout_hdr *
1238 alloc_init_layout_hdr(struct inode *ino,
1239                       struct nfs_open_context *ctx,
1240                       gfp_t gfp_flags)
1241 {
1242         struct pnfs_layout_hdr *lo;
1243
1244         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1245         if (!lo)
1246                 return NULL;
1247         atomic_set(&lo->plh_refcount, 1);
1248         INIT_LIST_HEAD(&lo->plh_layouts);
1249         INIT_LIST_HEAD(&lo->plh_segs);
1250         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1251         lo->plh_inode = ino;
1252         lo->plh_lc_cred = get_rpccred(ctx->cred);
1253         return lo;
1254 }
1255
1256 static struct pnfs_layout_hdr *
1257 pnfs_find_alloc_layout(struct inode *ino,
1258                        struct nfs_open_context *ctx,
1259                        gfp_t gfp_flags)
1260 {
1261         struct nfs_inode *nfsi = NFS_I(ino);
1262         struct pnfs_layout_hdr *new = NULL;
1263
1264         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1265
1266         if (nfsi->layout != NULL)
1267                 goto out_existing;
1268         spin_unlock(&ino->i_lock);
1269         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1270         spin_lock(&ino->i_lock);
1271
1272         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1273                 nfsi->layout = new;
1274                 return new;
1275         } else if (new != NULL)
1276                 pnfs_free_layout_hdr(new);
1277 out_existing:
1278         pnfs_get_layout_hdr(nfsi->layout);
1279         return nfsi->layout;
1280 }
1281
1282 /*
1283  * iomode matching rules:
1284  * iomode       lseg    match
1285  * -----        -----   -----
1286  * ANY          READ    true
1287  * ANY          RW      true
1288  * RW           READ    false
1289  * RW           RW      true
1290  * READ         READ    true
1291  * READ         RW      true
1292  */
1293 static bool
1294 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1295                  const struct pnfs_layout_range *range)
1296 {
1297         struct pnfs_layout_range range1;
1298
1299         if ((range->iomode == IOMODE_RW &&
1300              ls_range->iomode != IOMODE_RW) ||
1301             !pnfs_lseg_range_intersecting(ls_range, range))
1302                 return 0;
1303
1304         /* range1 covers only the first byte in the range */
1305         range1 = *range;
1306         range1.length = 1;
1307         return pnfs_lseg_range_contained(ls_range, &range1);
1308 }
1309
1310 /*
1311  * lookup range in layout
1312  */
1313 static struct pnfs_layout_segment *
1314 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1315                 struct pnfs_layout_range *range)
1316 {
1317         struct pnfs_layout_segment *lseg, *ret = NULL;
1318
1319         dprintk("%s:Begin\n", __func__);
1320
1321         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1322                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1323                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1324                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1325                         ret = pnfs_get_lseg(lseg);
1326                         break;
1327                 }
1328                 if (lseg->pls_range.offset > range->offset)
1329                         break;
1330         }
1331
1332         dprintk("%s:Return lseg %p ref %d\n",
1333                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1334         return ret;
1335 }
1336
1337 /*
1338  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1339  * to the MDS or over pNFS
1340  *
1341  * The nfs_inode read_io and write_io fields are cumulative counters reset
1342  * when there are no layout segments. Note that in pnfs_update_layout iomode
1343  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1344  * WRITE request.
1345  *
1346  * A return of true means use MDS I/O.
1347  *
1348  * From rfc 5661:
1349  * If a file's size is smaller than the file size threshold, data accesses
1350  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1351  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1352  * server.  If both file size and I/O size are provided, the client SHOULD
1353  * reach or exceed  both thresholds before sending its read or write
1354  * requests to the data server.
1355  */
1356 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1357                                      struct inode *ino, int iomode)
1358 {
1359         struct nfs4_threshold *t = ctx->mdsthreshold;
1360         struct nfs_inode *nfsi = NFS_I(ino);
1361         loff_t fsize = i_size_read(ino);
1362         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1363
1364         if (t == NULL)
1365                 return ret;
1366
1367         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1368                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1369
1370         switch (iomode) {
1371         case IOMODE_READ:
1372                 if (t->bm & THRESHOLD_RD) {
1373                         dprintk("%s fsize %llu\n", __func__, fsize);
1374                         size_set = true;
1375                         if (fsize < t->rd_sz)
1376                                 size = true;
1377                 }
1378                 if (t->bm & THRESHOLD_RD_IO) {
1379                         dprintk("%s nfsi->read_io %llu\n", __func__,
1380                                 nfsi->read_io);
1381                         io_set = true;
1382                         if (nfsi->read_io < t->rd_io_sz)
1383                                 io = true;
1384                 }
1385                 break;
1386         case IOMODE_RW:
1387                 if (t->bm & THRESHOLD_WR) {
1388                         dprintk("%s fsize %llu\n", __func__, fsize);
1389                         size_set = true;
1390                         if (fsize < t->wr_sz)
1391                                 size = true;
1392                 }
1393                 if (t->bm & THRESHOLD_WR_IO) {
1394                         dprintk("%s nfsi->write_io %llu\n", __func__,
1395                                 nfsi->write_io);
1396                         io_set = true;
1397                         if (nfsi->write_io < t->wr_io_sz)
1398                                 io = true;
1399                 }
1400                 break;
1401         }
1402         if (size_set && io_set) {
1403                 if (size && io)
1404                         ret = true;
1405         } else if (size || io)
1406                 ret = true;
1407
1408         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1409         return ret;
1410 }
1411
1412 /* stop waiting if someone clears NFS_LAYOUT_RETRY_LAYOUTGET bit. */
1413 static int pnfs_layoutget_retry_bit_wait(struct wait_bit_key *key)
1414 {
1415         if (!test_bit(NFS_LAYOUT_RETRY_LAYOUTGET, key->flags))
1416                 return 1;
1417         return nfs_wait_bit_killable(key);
1418 }
1419
1420 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1421 {
1422         /*
1423          * send layoutcommit as it can hold up layoutreturn due to lseg
1424          * reference
1425          */
1426         pnfs_layoutcommit_inode(lo->plh_inode, false);
1427         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1428                                    pnfs_layoutget_retry_bit_wait,
1429                                    TASK_UNINTERRUPTIBLE);
1430 }
1431
1432 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1433 {
1434         unsigned long *bitlock = &lo->plh_flags;
1435
1436         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1437         smp_mb__after_atomic();
1438         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1439 }
1440
1441 /*
1442  * Layout segment is retreived from the server if not cached.
1443  * The appropriate layout segment is referenced and returned to the caller.
1444  */
1445 struct pnfs_layout_segment *
1446 pnfs_update_layout(struct inode *ino,
1447                    struct nfs_open_context *ctx,
1448                    loff_t pos,
1449                    u64 count,
1450                    enum pnfs_iomode iomode,
1451                    gfp_t gfp_flags)
1452 {
1453         struct pnfs_layout_range arg = {
1454                 .iomode = iomode,
1455                 .offset = pos,
1456                 .length = count,
1457         };
1458         unsigned pg_offset;
1459         struct nfs_server *server = NFS_SERVER(ino);
1460         struct nfs_client *clp = server->nfs_client;
1461         struct pnfs_layout_hdr *lo;
1462         struct pnfs_layout_segment *lseg = NULL;
1463         bool first;
1464
1465         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1466                 goto out;
1467
1468         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1469                 goto out;
1470
1471 lookup_again:
1472         first = false;
1473         spin_lock(&ino->i_lock);
1474         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1475         if (lo == NULL) {
1476                 spin_unlock(&ino->i_lock);
1477                 goto out;
1478         }
1479
1480         /* Do we even need to bother with this? */
1481         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1482                 dprintk("%s matches recall, use MDS\n", __func__);
1483                 goto out_unlock;
1484         }
1485
1486         /* if LAYOUTGET already failed once we don't try again */
1487         if (pnfs_layout_io_test_failed(lo, iomode) &&
1488             !pnfs_should_retry_layoutget(lo))
1489                 goto out_unlock;
1490
1491         first = list_empty(&lo->plh_segs);
1492         if (first) {
1493                 /* The first layoutget for the file. Need to serialize per
1494                  * RFC 5661 Errata 3208.
1495                  */
1496                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1497                                      &lo->plh_flags)) {
1498                         spin_unlock(&ino->i_lock);
1499                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1500                                     TASK_UNINTERRUPTIBLE);
1501                         pnfs_put_layout_hdr(lo);
1502                         goto lookup_again;
1503                 }
1504         } else {
1505                 /* Check to see if the layout for the given range
1506                  * already exists
1507                  */
1508                 lseg = pnfs_find_lseg(lo, &arg);
1509                 if (lseg)
1510                         goto out_unlock;
1511         }
1512
1513         /*
1514          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1515          * for LAYOUTRETURN even if first is true.
1516          */
1517         if (!lseg && pnfs_should_retry_layoutget(lo) &&
1518             test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1519                 spin_unlock(&ino->i_lock);
1520                 dprintk("%s wait for layoutreturn\n", __func__);
1521                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1522                         if (first)
1523                                 pnfs_clear_first_layoutget(lo);
1524                         pnfs_put_layout_hdr(lo);
1525                         dprintk("%s retrying\n", __func__);
1526                         goto lookup_again;
1527                 }
1528                 goto out_put_layout_hdr;
1529         }
1530
1531         if (pnfs_layoutgets_blocked(lo, &arg, 0))
1532                 goto out_unlock;
1533         atomic_inc(&lo->plh_outstanding);
1534         spin_unlock(&ino->i_lock);
1535
1536         if (list_empty(&lo->plh_layouts)) {
1537                 /* The lo must be on the clp list if there is any
1538                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1539                  */
1540                 spin_lock(&clp->cl_lock);
1541                 if (list_empty(&lo->plh_layouts))
1542                         list_add_tail(&lo->plh_layouts, &server->layouts);
1543                 spin_unlock(&clp->cl_lock);
1544         }
1545
1546         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1547         if (pg_offset) {
1548                 arg.offset -= pg_offset;
1549                 arg.length += pg_offset;
1550         }
1551         if (arg.length != NFS4_MAX_UINT64)
1552                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1553
1554         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1555         pnfs_clear_retry_layoutget(lo);
1556         atomic_dec(&lo->plh_outstanding);
1557 out_put_layout_hdr:
1558         if (first)
1559                 pnfs_clear_first_layoutget(lo);
1560         pnfs_put_layout_hdr(lo);
1561 out:
1562         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1563                         "(%s, offset: %llu, length: %llu)\n",
1564                         __func__, ino->i_sb->s_id,
1565                         (unsigned long long)NFS_FILEID(ino),
1566                         lseg == NULL ? "not found" : "found",
1567                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1568                         (unsigned long long)pos,
1569                         (unsigned long long)count);
1570         return lseg;
1571 out_unlock:
1572         spin_unlock(&ino->i_lock);
1573         goto out_put_layout_hdr;
1574 }
1575 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1576
1577 struct pnfs_layout_segment *
1578 pnfs_layout_process(struct nfs4_layoutget *lgp)
1579 {
1580         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1581         struct nfs4_layoutget_res *res = &lgp->res;
1582         struct pnfs_layout_segment *lseg;
1583         struct inode *ino = lo->plh_inode;
1584         LIST_HEAD(free_me);
1585         int status = 0;
1586
1587         /* Inject layout blob into I/O device driver */
1588         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1589         if (!lseg || IS_ERR(lseg)) {
1590                 if (!lseg)
1591                         status = -ENOMEM;
1592                 else
1593                         status = PTR_ERR(lseg);
1594                 dprintk("%s: Could not allocate layout: error %d\n",
1595                        __func__, status);
1596                 goto out;
1597         }
1598
1599         init_lseg(lo, lseg);
1600         lseg->pls_range = res->range;
1601
1602         spin_lock(&ino->i_lock);
1603         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1604                 dprintk("%s forget reply due to recall\n", __func__);
1605                 goto out_forget_reply;
1606         }
1607
1608         if (pnfs_layoutgets_blocked(lo, &lgp->args.range, 1)) {
1609                 dprintk("%s forget reply due to state\n", __func__);
1610                 goto out_forget_reply;
1611         }
1612
1613         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1614                 /* existing state ID, make sure the sequence number matches. */
1615                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1616                         dprintk("%s forget reply due to sequence\n", __func__);
1617                         goto out_forget_reply;
1618                 }
1619                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1620         } else {
1621                 /*
1622                  * We got an entirely new state ID.  Mark all segments for the
1623                  * inode invalid, and don't bother validating the stateid
1624                  * sequence number.
1625                  */
1626                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1627
1628                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1629                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1630         }
1631
1632         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1633
1634         pnfs_get_lseg(lseg);
1635         pnfs_layout_insert_lseg(lo, lseg);
1636
1637         if (res->return_on_close) {
1638                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1639                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1640         }
1641
1642         spin_unlock(&ino->i_lock);
1643         pnfs_free_lseg_list(&free_me);
1644         return lseg;
1645 out:
1646         return ERR_PTR(status);
1647
1648 out_forget_reply:
1649         spin_unlock(&ino->i_lock);
1650         lseg->pls_layout = lo;
1651         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1652         goto out;
1653 }
1654
1655 static void
1656 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1657                                 struct list_head *tmp_list,
1658                                 struct pnfs_layout_range *return_range)
1659 {
1660         struct pnfs_layout_segment *lseg, *next;
1661
1662         dprintk("%s:Begin lo %p\n", __func__, lo);
1663
1664         if (list_empty(&lo->plh_segs))
1665                 return;
1666
1667         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1668                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1669                         dprintk("%s: marking lseg %p iomode %d "
1670                                 "offset %llu length %llu\n", __func__,
1671                                 lseg, lseg->pls_range.iomode,
1672                                 lseg->pls_range.offset,
1673                                 lseg->pls_range.length);
1674                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1675                         mark_lseg_invalid(lseg, tmp_list);
1676                 }
1677 }
1678
1679 void pnfs_error_mark_layout_for_return(struct inode *inode,
1680                                        struct pnfs_layout_segment *lseg)
1681 {
1682         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1683         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1684         struct pnfs_layout_range range = {
1685                 .iomode = lseg->pls_range.iomode,
1686                 .offset = 0,
1687                 .length = NFS4_MAX_UINT64,
1688         };
1689         LIST_HEAD(free_me);
1690
1691         spin_lock(&inode->i_lock);
1692         /* set failure bit so that pnfs path will be retried later */
1693         pnfs_layout_set_fail_bit(lo, iomode);
1694         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1695         if (lo->plh_return_iomode == 0)
1696                 lo->plh_return_iomode = range.iomode;
1697         else if (lo->plh_return_iomode != range.iomode)
1698                 lo->plh_return_iomode = IOMODE_ANY;
1699         /*
1700          * mark all matching lsegs so that we are sure to have no live
1701          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1702          * for how it works.
1703          */
1704         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1705         spin_unlock(&inode->i_lock);
1706         pnfs_free_lseg_list(&free_me);
1707 }
1708 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1709
1710 void
1711 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1712 {
1713         u64 rd_size = req->wb_bytes;
1714
1715         if (pgio->pg_lseg == NULL) {
1716                 if (pgio->pg_dreq == NULL)
1717                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1718                 else
1719                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1720
1721                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1722                                                    req->wb_context,
1723                                                    req_offset(req),
1724                                                    rd_size,
1725                                                    IOMODE_READ,
1726                                                    GFP_KERNEL);
1727         }
1728         /* If no lseg, fall back to read through mds */
1729         if (pgio->pg_lseg == NULL)
1730                 nfs_pageio_reset_read_mds(pgio);
1731
1732 }
1733 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1734
1735 void
1736 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1737                            struct nfs_page *req, u64 wb_size)
1738 {
1739         if (pgio->pg_lseg == NULL)
1740                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1741                                                    req->wb_context,
1742                                                    req_offset(req),
1743                                                    wb_size,
1744                                                    IOMODE_RW,
1745                                                    GFP_NOFS);
1746         /* If no lseg, fall back to write through mds */
1747         if (pgio->pg_lseg == NULL)
1748                 nfs_pageio_reset_write_mds(pgio);
1749 }
1750 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1751
1752 void
1753 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1754 {
1755         if (desc->pg_lseg) {
1756                 pnfs_put_lseg(desc->pg_lseg);
1757                 desc->pg_lseg = NULL;
1758         }
1759 }
1760 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1761
1762 /*
1763  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1764  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1765  */
1766 size_t
1767 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1768                      struct nfs_page *prev, struct nfs_page *req)
1769 {
1770         unsigned int size;
1771         u64 seg_end, req_start, seg_left;
1772
1773         size = nfs_generic_pg_test(pgio, prev, req);
1774         if (!size)
1775                 return 0;
1776
1777         /*
1778          * 'size' contains the number of bytes left in the current page (up
1779          * to the original size asked for in @req->wb_bytes).
1780          *
1781          * Calculate how many bytes are left in the layout segment
1782          * and if there are less bytes than 'size', return that instead.
1783          *
1784          * Please also note that 'end_offset' is actually the offset of the
1785          * first byte that lies outside the pnfs_layout_range. FIXME?
1786          *
1787          */
1788         if (pgio->pg_lseg) {
1789                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1790                                      pgio->pg_lseg->pls_range.length);
1791                 req_start = req_offset(req);
1792                 WARN_ON_ONCE(req_start >= seg_end);
1793                 /* start of request is past the last byte of this segment */
1794                 if (req_start >= seg_end) {
1795                         /* reference the new lseg */
1796                         if (pgio->pg_ops->pg_cleanup)
1797                                 pgio->pg_ops->pg_cleanup(pgio);
1798                         if (pgio->pg_ops->pg_init)
1799                                 pgio->pg_ops->pg_init(pgio, req);
1800                         return 0;
1801                 }
1802
1803                 /* adjust 'size' iff there are fewer bytes left in the
1804                  * segment than what nfs_generic_pg_test returned */
1805                 seg_left = seg_end - req_start;
1806                 if (seg_left < size)
1807                         size = (unsigned int)seg_left;
1808         }
1809
1810         return size;
1811 }
1812 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1813
1814 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1815 {
1816         struct nfs_pageio_descriptor pgio;
1817
1818         /* Resend all requests through the MDS */
1819         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1820                               hdr->completion_ops);
1821         return nfs_pageio_resend(&pgio, hdr);
1822 }
1823 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1824
1825 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1826 {
1827
1828         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1829         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1830             PNFS_LAYOUTRET_ON_ERROR) {
1831                 pnfs_return_layout(hdr->inode);
1832         }
1833         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1834                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1835 }
1836
1837 /*
1838  * Called by non rpc-based layout drivers
1839  */
1840 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1841 {
1842         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1843         if (!hdr->pnfs_error) {
1844                 pnfs_set_layoutcommit(hdr);
1845                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1846         } else
1847                 pnfs_ld_handle_write_error(hdr);
1848         hdr->mds_ops->rpc_release(hdr);
1849 }
1850 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1851
1852 static void
1853 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1854                 struct nfs_pgio_header *hdr)
1855 {
1856         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1857
1858         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1859                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1860                 nfs_pageio_reset_write_mds(desc);
1861                 mirror->pg_recoalesce = 1;
1862         }
1863         nfs_pgio_data_destroy(hdr);
1864 }
1865
1866 static enum pnfs_try_status
1867 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1868                         const struct rpc_call_ops *call_ops,
1869                         struct pnfs_layout_segment *lseg,
1870                         int how)
1871 {
1872         struct inode *inode = hdr->inode;
1873         enum pnfs_try_status trypnfs;
1874         struct nfs_server *nfss = NFS_SERVER(inode);
1875
1876         hdr->mds_ops = call_ops;
1877
1878         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1879                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1880         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1881         if (trypnfs != PNFS_NOT_ATTEMPTED)
1882                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1883         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1884         return trypnfs;
1885 }
1886
1887 static void
1888 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1889               struct nfs_pgio_header *hdr, int how)
1890 {
1891         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1892         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1893         enum pnfs_try_status trypnfs;
1894
1895         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1896         if (trypnfs == PNFS_NOT_ATTEMPTED)
1897                 pnfs_write_through_mds(desc, hdr);
1898 }
1899
1900 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1901 {
1902         pnfs_put_lseg(hdr->lseg);
1903         nfs_pgio_header_free(hdr);
1904 }
1905 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1906
1907 int
1908 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1909 {
1910         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1911
1912         struct nfs_pgio_header *hdr;
1913         int ret;
1914
1915         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1916         if (!hdr) {
1917                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1918                 return -ENOMEM;
1919         }
1920         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1921
1922         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1923         ret = nfs_generic_pgio(desc, hdr);
1924         if (!ret)
1925                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1926
1927         return ret;
1928 }
1929 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1930
1931 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1932 {
1933         struct nfs_pageio_descriptor pgio;
1934
1935         /* Resend all requests through the MDS */
1936         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1937         return nfs_pageio_resend(&pgio, hdr);
1938 }
1939 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1940
1941 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1942 {
1943         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1944         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1945             PNFS_LAYOUTRET_ON_ERROR) {
1946                 pnfs_return_layout(hdr->inode);
1947         }
1948         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1949                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1950 }
1951
1952 /*
1953  * Called by non rpc-based layout drivers
1954  */
1955 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1956 {
1957         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1958         if (likely(!hdr->pnfs_error)) {
1959                 __nfs4_read_done_cb(hdr);
1960                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1961         } else
1962                 pnfs_ld_handle_read_error(hdr);
1963         hdr->mds_ops->rpc_release(hdr);
1964 }
1965 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1966
1967 static void
1968 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1969                 struct nfs_pgio_header *hdr)
1970 {
1971         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1972
1973         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1974                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1975                 nfs_pageio_reset_read_mds(desc);
1976                 mirror->pg_recoalesce = 1;
1977         }
1978         nfs_pgio_data_destroy(hdr);
1979 }
1980
1981 /*
1982  * Call the appropriate parallel I/O subsystem read function.
1983  */
1984 static enum pnfs_try_status
1985 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1986                        const struct rpc_call_ops *call_ops,
1987                        struct pnfs_layout_segment *lseg)
1988 {
1989         struct inode *inode = hdr->inode;
1990         struct nfs_server *nfss = NFS_SERVER(inode);
1991         enum pnfs_try_status trypnfs;
1992
1993         hdr->mds_ops = call_ops;
1994
1995         dprintk("%s: Reading ino:%lu %u@%llu\n",
1996                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1997
1998         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1999         if (trypnfs != PNFS_NOT_ATTEMPTED)
2000                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2001         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2002         return trypnfs;
2003 }
2004
2005 /* Resend all requests through pnfs. */
2006 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2007 {
2008         struct nfs_pageio_descriptor pgio;
2009
2010         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2011         return nfs_pageio_resend(&pgio, hdr);
2012 }
2013 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2014
2015 static void
2016 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2017 {
2018         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2019         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2020         enum pnfs_try_status trypnfs;
2021         int err = 0;
2022
2023         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2024         if (trypnfs == PNFS_TRY_AGAIN)
2025                 err = pnfs_read_resend_pnfs(hdr);
2026         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2027                 pnfs_read_through_mds(desc, hdr);
2028 }
2029
2030 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2031 {
2032         pnfs_put_lseg(hdr->lseg);
2033         nfs_pgio_header_free(hdr);
2034 }
2035 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
2036
2037 int
2038 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2039 {
2040         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2041
2042         struct nfs_pgio_header *hdr;
2043         int ret;
2044
2045         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2046         if (!hdr) {
2047                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2048                 return -ENOMEM;
2049         }
2050         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2051         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2052         ret = nfs_generic_pgio(desc, hdr);
2053         if (!ret)
2054                 pnfs_do_read(desc, hdr);
2055         return ret;
2056 }
2057 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2058
2059 static void pnfs_clear_layoutcommitting(struct inode *inode)
2060 {
2061         unsigned long *bitlock = &NFS_I(inode)->flags;
2062
2063         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2064         smp_mb__after_atomic();
2065         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2066 }
2067
2068 /*
2069  * There can be multiple RW segments.
2070  */
2071 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2072 {
2073         struct pnfs_layout_segment *lseg;
2074
2075         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2076                 if (lseg->pls_range.iomode == IOMODE_RW &&
2077                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2078                         list_add(&lseg->pls_lc_list, listp);
2079         }
2080 }
2081
2082 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2083 {
2084         struct pnfs_layout_segment *lseg, *tmp;
2085
2086         /* Matched by references in pnfs_set_layoutcommit */
2087         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2088                 list_del_init(&lseg->pls_lc_list);
2089                 pnfs_put_lseg(lseg);
2090         }
2091
2092         pnfs_clear_layoutcommitting(inode);
2093 }
2094
2095 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2096 {
2097         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2098 }
2099 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2100
2101 void
2102 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
2103 {
2104         struct inode *inode = hdr->inode;
2105         struct nfs_inode *nfsi = NFS_I(inode);
2106         loff_t end_pos = hdr->mds_offset + hdr->res.count;
2107         bool mark_as_dirty = false;
2108
2109         spin_lock(&inode->i_lock);
2110         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2111                 mark_as_dirty = true;
2112                 dprintk("%s: Set layoutcommit for inode %lu ",
2113                         __func__, inode->i_ino);
2114         }
2115         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
2116                 /* references matched in nfs4_layoutcommit_release */
2117                 pnfs_get_lseg(hdr->lseg);
2118         }
2119         if (end_pos > nfsi->layout->plh_lwb)
2120                 nfsi->layout->plh_lwb = end_pos;
2121         spin_unlock(&inode->i_lock);
2122         dprintk("%s: lseg %p end_pos %llu\n",
2123                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
2124
2125         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2126          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2127         if (mark_as_dirty)
2128                 mark_inode_dirty_sync(inode);
2129 }
2130 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2131
2132 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
2133 {
2134         struct inode *inode = data->inode;
2135         struct nfs_inode *nfsi = NFS_I(inode);
2136         bool mark_as_dirty = false;
2137
2138         spin_lock(&inode->i_lock);
2139         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2140                 mark_as_dirty = true;
2141                 dprintk("%s: Set layoutcommit for inode %lu ",
2142                         __func__, inode->i_ino);
2143         }
2144         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
2145                 /* references matched in nfs4_layoutcommit_release */
2146                 pnfs_get_lseg(data->lseg);
2147         }
2148         if (data->lwb > nfsi->layout->plh_lwb)
2149                 nfsi->layout->plh_lwb = data->lwb;
2150         spin_unlock(&inode->i_lock);
2151         dprintk("%s: lseg %p end_pos %llu\n",
2152                 __func__, data->lseg, nfsi->layout->plh_lwb);
2153
2154         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2155          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2156         if (mark_as_dirty)
2157                 mark_inode_dirty_sync(inode);
2158 }
2159 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
2160
2161 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2162 {
2163         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2164
2165         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2166                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2167         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2168 }
2169
2170 /*
2171  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2172  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2173  * data to disk to allow the server to recover the data if it crashes.
2174  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2175  * is off, and a COMMIT is sent to a data server, or
2176  * if WRITEs to a data server return NFS_DATA_SYNC.
2177  */
2178 int
2179 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2180 {
2181         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2182         struct nfs4_layoutcommit_data *data;
2183         struct nfs_inode *nfsi = NFS_I(inode);
2184         loff_t end_pos;
2185         int status;
2186
2187         if (!pnfs_layoutcommit_outstanding(inode))
2188                 return 0;
2189
2190         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2191
2192         status = -EAGAIN;
2193         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2194                 if (!sync)
2195                         goto out;
2196                 status = wait_on_bit_lock_action(&nfsi->flags,
2197                                 NFS_INO_LAYOUTCOMMITTING,
2198                                 nfs_wait_bit_killable,
2199                                 TASK_KILLABLE);
2200                 if (status)
2201                         goto out;
2202         }
2203
2204         status = -ENOMEM;
2205         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2206         data = kzalloc(sizeof(*data), GFP_NOFS);
2207         if (!data)
2208                 goto clear_layoutcommitting;
2209
2210         status = 0;
2211         spin_lock(&inode->i_lock);
2212         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2213                 goto out_unlock;
2214
2215         INIT_LIST_HEAD(&data->lseg_list);
2216         pnfs_list_write_lseg(inode, &data->lseg_list);
2217
2218         end_pos = nfsi->layout->plh_lwb;
2219         nfsi->layout->plh_lwb = 0;
2220
2221         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2222         spin_unlock(&inode->i_lock);
2223
2224         data->args.inode = inode;
2225         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2226         nfs_fattr_init(&data->fattr);
2227         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2228         data->res.fattr = &data->fattr;
2229         data->args.lastbytewritten = end_pos - 1;
2230         data->res.server = NFS_SERVER(inode);
2231
2232         if (ld->prepare_layoutcommit) {
2233                 status = ld->prepare_layoutcommit(&data->args);
2234                 if (status) {
2235                         spin_lock(&inode->i_lock);
2236                         if (end_pos < nfsi->layout->plh_lwb)
2237                                 nfsi->layout->plh_lwb = end_pos;
2238                         spin_unlock(&inode->i_lock);
2239                         put_rpccred(data->cred);
2240                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2241                         goto clear_layoutcommitting;
2242                 }
2243         }
2244
2245
2246         status = nfs4_proc_layoutcommit(data, sync);
2247 out:
2248         if (status)
2249                 mark_inode_dirty_sync(inode);
2250         dprintk("<-- %s status %d\n", __func__, status);
2251         return status;
2252 out_unlock:
2253         spin_unlock(&inode->i_lock);
2254         kfree(data);
2255 clear_layoutcommitting:
2256         pnfs_clear_layoutcommitting(inode);
2257         goto out;
2258 }
2259 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2260
2261 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2262 {
2263         struct nfs4_threshold *thp;
2264
2265         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2266         if (!thp) {
2267                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2268                 return NULL;
2269         }
2270         return thp;
2271 }