fs/nfs/pnfs.c

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