fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36 #include "nfs4trace.h"
  37
  38 #define NFSDBG_FACILITY         NFSDBG_PNFS
  39 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  40
  41 /* Locking:
  42  *
  43  * pnfs_spinlock:
  44  *      protects pnfs_modules_tbl.
  45  */
  46 static DEFINE_SPINLOCK(pnfs_spinlock);
  47
  48 /*
  49  * pnfs_modules_tbl holds all pnfs modules
  50  */
  51 static LIST_HEAD(pnfs_modules_tbl);
  52
  53 static int
  54 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
  55                        enum pnfs_iomode iomode);
  56
  57 /* Return the registered pnfs layout driver module matching given id */
  58 static struct pnfs_layoutdriver_type *
  59 find_pnfs_driver_locked(u32 id)
  60 {
  61         struct pnfs_layoutdriver_type *local;
  62
  63         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  64                 if (local->id == id)
  65                         goto out;
  66         local = NULL;
  67 out:
  68         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  69         return local;
  70 }
  71
  72 static struct pnfs_layoutdriver_type *
  73 find_pnfs_driver(u32 id)
  74 {
  75         struct pnfs_layoutdriver_type *local;
  76
  77         spin_lock(&pnfs_spinlock);
  78         local = find_pnfs_driver_locked(id);
  79         if (local != NULL && !try_module_get(local->owner)) {
  80                 dprintk("%s: Could not grab reference on module\n", __func__);
  81                 local = NULL;
  82         }
  83         spin_unlock(&pnfs_spinlock);
  84         return local;
  85 }
  86
  87 void
  88 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  89 {
  90         if (nfss->pnfs_curr_ld) {
  91                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  92                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  93                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  94                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  95                         nfs4_deviceid_purge_client(nfss->nfs_client);
  96                 module_put(nfss->pnfs_curr_ld->owner);
  97         }
  98         nfss->pnfs_curr_ld = NULL;
  99 }
 100
 101 /*
 102  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 103  * Currently only one pNFS layout driver per filesystem is supported.
 104  *
 105  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 106  */
 107 void
 108 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 109                       u32 id)
 110 {
 111         struct pnfs_layoutdriver_type *ld_type = NULL;
 112
 113         if (id == 0)
 114                 goto out_no_driver;
 115         if (!(server->nfs_client->cl_exchange_flags &
 116                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 117                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 118                         __func__, id, server->nfs_client->cl_exchange_flags);
 119                 goto out_no_driver;
 120         }
 121         ld_type = find_pnfs_driver(id);
 122         if (!ld_type) {
 123                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 124                 ld_type = find_pnfs_driver(id);
 125                 if (!ld_type) {
 126                         dprintk("%s: No pNFS module found for %u.\n",
 127                                 __func__, id);
 128                         goto out_no_driver;
 129                 }
 130         }
 131         server->pnfs_curr_ld = ld_type;
 132         if (ld_type->set_layoutdriver
 133             && ld_type->set_layoutdriver(server, mntfh)) {
 134                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 135                         "driver %u.\n", __func__, id);
 136                 module_put(ld_type->owner);
 137                 goto out_no_driver;
 138         }
 139         /* Bump the MDS count */
 140         atomic_inc(&server->nfs_client->cl_mds_count);
 141
 142         dprintk("%s: pNFS module for %u set\n", __func__, id);
 143         return;
 144
 145 out_no_driver:
 146         dprintk("%s: Using NFSv4 I/O\n", __func__);
 147         server->pnfs_curr_ld = NULL;
 148 }
 149
 150 int
 151 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 152 {
 153         int status = -EINVAL;
 154         struct pnfs_layoutdriver_type *tmp;
 155
 156         if (ld_type->id == 0) {
 157                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 158                 return status;
 159         }
 160         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 161                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 162                        "alloc_lseg and free_lseg.\n", __func__);
 163                 return status;
 164         }
 165
 166         spin_lock(&pnfs_spinlock);
 167         tmp = find_pnfs_driver_locked(ld_type->id);
 168         if (!tmp) {
 169                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 170                 status = 0;
 171                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 172                         ld_type->name);
 173         } else {
 174                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 175                         __func__, ld_type->id);
 176         }
 177         spin_unlock(&pnfs_spinlock);
 178
 179         return status;
 180 }
 181 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 182
 183 void
 184 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 185 {
 186         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 187         spin_lock(&pnfs_spinlock);
 188         list_del(&ld_type->pnfs_tblid);
 189         spin_unlock(&pnfs_spinlock);
 190 }
 191 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 192
 193 /*
 194  * pNFS client layout cache
 195  */
 196
 197 /* Need to hold i_lock if caller does not already hold reference */
 198 void
 199 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 200 {
 201         atomic_inc(&lo->plh_refcount);
 202 }
 203
 204 static struct pnfs_layout_hdr *
 205 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 206 {
 207         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 208         return ld->alloc_layout_hdr(ino, gfp_flags);
 209 }
 210
 211 static void
 212 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 213 {
 214         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 215         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 216
 217         if (!list_empty(&lo->plh_layouts)) {
 218                 struct nfs_client *clp = server->nfs_client;
 219
 220                 spin_lock(&clp->cl_lock);
 221                 list_del_init(&lo->plh_layouts);
 222                 spin_unlock(&clp->cl_lock);
 223         }
 224         put_rpccred(lo->plh_lc_cred);
 225         return ld->free_layout_hdr(lo);
 226 }
 227
 228 static void
 229 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 230 {
 231         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 232         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 233         nfsi->layout = NULL;
 234         /* Reset MDS Threshold I/O counters */
 235         nfsi->write_io = 0;
 236         nfsi->read_io = 0;
 237 }
 238
 239 void
 240 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 241 {
 242         struct inode *inode = lo->plh_inode;
 243
 244         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 245                 if (!list_empty(&lo->plh_segs))
 246                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 247                 pnfs_detach_layout_hdr(lo);
 248                 spin_unlock(&inode->i_lock);
 249                 pnfs_free_layout_hdr(lo);
 250         }
 251 }
 252
 253 static int
 254 pnfs_iomode_to_fail_bit(u32 iomode)
 255 {
 256         return iomode == IOMODE_RW ?
 257                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 258 }
 259
 260 static void
 261 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 262 {
 263         lo->plh_retry_timestamp = jiffies;
 264         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 265                 atomic_inc(&lo->plh_refcount);
 266 }
 267
 268 static void
 269 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 270 {
 271         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 272                 atomic_dec(&lo->plh_refcount);
 273 }
 274
 275 static void
 276 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 277 {
 278         struct inode *inode = lo->plh_inode;
 279         struct pnfs_layout_range range = {
 280                 .iomode = iomode,
 281                 .offset = 0,
 282                 .length = NFS4_MAX_UINT64,
 283         };
 284         LIST_HEAD(head);
 285
 286         spin_lock(&inode->i_lock);
 287         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 288         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 289         spin_unlock(&inode->i_lock);
 290         pnfs_free_lseg_list(&head);
 291         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 292                         iomode == IOMODE_RW ?  "RW" : "READ");
 293 }
 294
 295 static bool
 296 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 297 {
 298         unsigned long start, end;
 299         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 300
 301         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 302                 return false;
 303         end = jiffies;
 304         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 305         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 306                 /* It is time to retry the failed layoutgets */
 307                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 308                 return false;
 309         }
 310         return true;
 311 }
 312
 313 static void
 314 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 315 {
 316         INIT_LIST_HEAD(&lseg->pls_list);
 317         INIT_LIST_HEAD(&lseg->pls_lc_list);
 318         atomic_set(&lseg->pls_refcount, 1);
 319         smp_mb();
 320         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 321         lseg->pls_layout = lo;
 322 }
 323
 324 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 325 {
 326         struct inode *ino = lseg->pls_layout->plh_inode;
 327
 328         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 329 }
 330
 331 static void
 332 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 333                 struct pnfs_layout_segment *lseg)
 334 {
 335         struct inode *inode = lo->plh_inode;
 336
 337         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 338         list_del_init(&lseg->pls_list);
 339         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 340         atomic_dec(&lo->plh_refcount);
 341         if (list_empty(&lo->plh_segs))
 342                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 343         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 344 }
 345
 346 /* Return true if layoutreturn is needed */
 347 static bool
 348 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 349                         struct pnfs_layout_segment *lseg,
 350                         nfs4_stateid *stateid, enum pnfs_iomode *iomode)
 351 {
 352         struct pnfs_layout_segment *s;
 353
 354         if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 355                 return false;
 356
 357         list_for_each_entry(s, &lo->plh_segs, pls_list)
 358                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 359                         return false;
 360
 361         *stateid = lo->plh_stateid;
 362         *iomode = lo->plh_return_iomode;
 363         /* decreased in pnfs_send_layoutreturn() */
 364         lo->plh_block_lgets++;
 365         lo->plh_return_iomode = 0;
 366         return true;
 367 }
 368
 369 void
 370 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 371 {
 372         struct pnfs_layout_hdr *lo;
 373         struct inode *inode;
 374
 375         if (!lseg)
 376                 return;
 377
 378         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 379                 atomic_read(&lseg->pls_refcount),
 380                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 381         lo = lseg->pls_layout;
 382         inode = lo->plh_inode;
 383         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 384                 bool need_return;
 385                 nfs4_stateid stateid;
 386                 enum pnfs_iomode iomode;
 387
 388                 pnfs_get_layout_hdr(lo);
 389                 pnfs_layout_remove_lseg(lo, lseg);
 390                 need_return = pnfs_layout_need_return(lo, lseg,
 391                                                       &stateid, &iomode);
 392                 spin_unlock(&inode->i_lock);
 393                 pnfs_free_lseg(lseg);
 394                 if (need_return)
 395                         pnfs_send_layoutreturn(lo, stateid, iomode);
 396                 else
 397                         pnfs_put_layout_hdr(lo);
 398         }
 399 }
 400 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 401
 402 static void pnfs_free_lseg_async_work(struct work_struct *work)
 403 {
 404         struct pnfs_layout_segment *lseg;
 405         struct pnfs_layout_hdr *lo;
 406
 407         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 408         lo = lseg->pls_layout;
 409
 410         pnfs_free_lseg(lseg);
 411         pnfs_put_layout_hdr(lo);
 412 }
 413
 414 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 415 {
 416         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 417         schedule_work(&lseg->pls_work);
 418 }
 419
 420 void
 421 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 422 {
 423         if (!lseg)
 424                 return;
 425
 426         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 427
 428         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 429                 atomic_read(&lseg->pls_refcount),
 430                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 431         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 432                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 433                 pnfs_get_layout_hdr(lo);
 434                 pnfs_layout_remove_lseg(lo, lseg);
 435                 pnfs_free_lseg_async(lseg);
 436         }
 437 }
 438 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 439
 440 static u64
 441 end_offset(u64 start, u64 len)
 442 {
 443         u64 end;
 444
 445         end = start + len;
 446         return end >= start ? end : NFS4_MAX_UINT64;
 447 }
 448
 449 /*
 450  * is l2 fully contained in l1?
 451  *   start1                             end1
 452  *   [----------------------------------)
 453  *           start2           end2
 454  *           [----------------)
 455  */
 456 static bool
 457 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 458                  const struct pnfs_layout_range *l2)
 459 {
 460         u64 start1 = l1->offset;
 461         u64 end1 = end_offset(start1, l1->length);
 462         u64 start2 = l2->offset;
 463         u64 end2 = end_offset(start2, l2->length);
 464
 465         return (start1 <= start2) && (end1 >= end2);
 466 }
 467
 468 /*
 469  * is l1 and l2 intersecting?
 470  *   start1                             end1
 471  *   [----------------------------------)
 472  *                              start2           end2
 473  *                              [----------------)
 474  */
 475 static bool
 476 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 477                     const struct pnfs_layout_range *l2)
 478 {
 479         u64 start1 = l1->offset;
 480         u64 end1 = end_offset(start1, l1->length);
 481         u64 start2 = l2->offset;
 482         u64 end2 = end_offset(start2, l2->length);
 483
 484         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 485                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 486 }
 487
 488 static bool
 489 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 490                  const struct pnfs_layout_range *recall_range)
 491 {
 492         return (recall_range->iomode == IOMODE_ANY ||
 493                 lseg_range->iomode == recall_range->iomode) &&
 494                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 495 }
 496
 497 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 498                 struct list_head *tmp_list)
 499 {
 500         if (!atomic_dec_and_test(&lseg->pls_refcount))
 501                 return false;
 502         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 503         list_add(&lseg->pls_list, tmp_list);
 504         return true;
 505 }
 506
 507 /* Returns 1 if lseg is removed from list, 0 otherwise */
 508 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 509                              struct list_head *tmp_list)
 510 {
 511         int rv = 0;
 512
 513         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 514                 /* Remove the reference keeping the lseg in the
 515                  * list.  It will now be removed when all
 516                  * outstanding io is finished.
 517                  */
 518                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 519                         atomic_read(&lseg->pls_refcount));
 520                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 521                         rv = 1;
 522         }
 523         return rv;
 524 }
 525
 526 /* Returns count of number of matching invalid lsegs remaining in list
 527  * after call.
 528  */
 529 int
 530 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 531                             struct list_head *tmp_list,
 532                             struct pnfs_layout_range *recall_range)
 533 {
 534         struct pnfs_layout_segment *lseg, *next;
 535         int invalid = 0, removed = 0;
 536
 537         dprintk("%s:Begin lo %p\n", __func__, lo);
 538
 539         if (list_empty(&lo->plh_segs))
 540                 return 0;
 541         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 542                 if (!recall_range ||
 543                     should_free_lseg(&lseg->pls_range, recall_range)) {
 544                         dprintk("%s: freeing lseg %p iomode %d "
 545                                 "offset %llu length %llu\n", __func__,
 546                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 547                                 lseg->pls_range.length);
 548                         invalid++;
 549                         removed += mark_lseg_invalid(lseg, tmp_list);
 550                 }
 551         dprintk("%s:Return %i\n", __func__, invalid - removed);
 552         return invalid - removed;
 553 }
 554
 555 /* note free_me must contain lsegs from a single layout_hdr */
 556 void
 557 pnfs_free_lseg_list(struct list_head *free_me)
 558 {
 559         struct pnfs_layout_segment *lseg, *tmp;
 560
 561         if (list_empty(free_me))
 562                 return;
 563
 564         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 565                 list_del(&lseg->pls_list);
 566                 pnfs_free_lseg(lseg);
 567         }
 568 }
 569
 570 void
 571 pnfs_destroy_layout(struct nfs_inode *nfsi)
 572 {
 573         struct pnfs_layout_hdr *lo;
 574         LIST_HEAD(tmp_list);
 575
 576         spin_lock(&nfsi->vfs_inode.i_lock);
 577         lo = nfsi->layout;
 578         if (lo) {
 579                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 580                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 581                 pnfs_get_layout_hdr(lo);
 582                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 583                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 584                 spin_unlock(&nfsi->vfs_inode.i_lock);
 585                 pnfs_free_lseg_list(&tmp_list);
 586                 pnfs_put_layout_hdr(lo);
 587         } else
 588                 spin_unlock(&nfsi->vfs_inode.i_lock);
 589 }
 590 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 591
 592 static bool
 593 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 594                 struct list_head *layout_list)
 595 {
 596         struct pnfs_layout_hdr *lo;
 597         bool ret = false;
 598
 599         spin_lock(&inode->i_lock);
 600         lo = NFS_I(inode)->layout;
 601         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 602                 pnfs_get_layout_hdr(lo);
 603                 list_add(&lo->plh_bulk_destroy, layout_list);
 604                 ret = true;
 605         }
 606         spin_unlock(&inode->i_lock);
 607         return ret;
 608 }
 609
 610 /* Caller must hold rcu_read_lock and clp->cl_lock */
 611 static int
 612 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 613                 struct nfs_server *server,
 614                 struct list_head *layout_list)
 615 {
 616         struct pnfs_layout_hdr *lo, *next;
 617         struct inode *inode;
 618
 619         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 620                 inode = igrab(lo->plh_inode);
 621                 if (inode == NULL)
 622                         continue;
 623                 list_del_init(&lo->plh_layouts);
 624                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 625                         continue;
 626                 rcu_read_unlock();
 627                 spin_unlock(&clp->cl_lock);
 628                 iput(inode);
 629                 spin_lock(&clp->cl_lock);
 630                 rcu_read_lock();
 631                 return -EAGAIN;
 632         }
 633         return 0;
 634 }
 635
 636 static int
 637 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 638                 bool is_bulk_recall)
 639 {
 640         struct pnfs_layout_hdr *lo;
 641         struct inode *inode;
 642         struct pnfs_layout_range range = {
 643                 .iomode = IOMODE_ANY,
 644                 .offset = 0,
 645                 .length = NFS4_MAX_UINT64,
 646         };
 647         LIST_HEAD(lseg_list);
 648         int ret = 0;
 649
 650         while (!list_empty(layout_list)) {
 651                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 652                                 plh_bulk_destroy);
 653                 dprintk("%s freeing layout for inode %lu\n", __func__,
 654                         lo->plh_inode->i_ino);
 655                 inode = lo->plh_inode;
 656
 657                 pnfs_layoutcommit_inode(inode, false);
 658
 659                 spin_lock(&inode->i_lock);
 660                 list_del_init(&lo->plh_bulk_destroy);
 661                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 662                 if (is_bulk_recall)
 663                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 664                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 665                         ret = -EAGAIN;
 666                 spin_unlock(&inode->i_lock);
 667                 pnfs_free_lseg_list(&lseg_list);
 668                 pnfs_put_layout_hdr(lo);
 669                 iput(inode);
 670         }
 671         return ret;
 672 }
 673
 674 int
 675 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 676                 struct nfs_fsid *fsid,
 677                 bool is_recall)
 678 {
 679         struct nfs_server *server;
 680         LIST_HEAD(layout_list);
 681
 682         spin_lock(&clp->cl_lock);
 683         rcu_read_lock();
 684 restart:
 685         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 686                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 687                         continue;
 688                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 689                                 server,
 690                                 &layout_list) != 0)
 691                         goto restart;
 692         }
 693         rcu_read_unlock();
 694         spin_unlock(&clp->cl_lock);
 695
 696         if (list_empty(&layout_list))
 697                 return 0;
 698         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 699 }
 700
 701 int
 702 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 703                 bool is_recall)
 704 {
 705         struct nfs_server *server;
 706         LIST_HEAD(layout_list);
 707
 708         spin_lock(&clp->cl_lock);
 709         rcu_read_lock();
 710 restart:
 711         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 712                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 713                                         server,
 714                                         &layout_list) != 0)
 715                         goto restart;
 716         }
 717         rcu_read_unlock();
 718         spin_unlock(&clp->cl_lock);
 719
 720         if (list_empty(&layout_list))
 721                 return 0;
 722         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 723 }
 724
 725 /*
 726  * Called by the state manger to remove all layouts established under an
 727  * expired lease.
 728  */
 729 void
 730 pnfs_destroy_all_layouts(struct nfs_client *clp)
 731 {
 732         nfs4_deviceid_mark_client_invalid(clp);
 733         nfs4_deviceid_purge_client(clp);
 734
 735         pnfs_destroy_layouts_byclid(clp, false);
 736 }
 737
 738 /*
 739  * Compare 2 layout stateid sequence ids, to see which is newer,
 740  * taking into account wraparound issues.
 741  */
 742 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 743 {
 744         return (s32)(s1 - s2) > 0;
 745 }
 746
 747 /* update lo->plh_stateid with new if is more recent */
 748 void
 749 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 750                         bool update_barrier)
 751 {
 752         u32 oldseq, newseq, new_barrier;
 753         int empty = list_empty(&lo->plh_segs);
 754
 755         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 756         newseq = be32_to_cpu(new->seqid);
 757         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 758                 nfs4_stateid_copy(&lo->plh_stateid, new);
 759                 if (update_barrier) {
 760                         new_barrier = be32_to_cpu(new->seqid);
 761                 } else {
 762                         /* Because of wraparound, we want to keep the barrier
 763                          * "close" to the current seqids.
 764                          */
 765                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 766                 }
 767                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 768                         lo->plh_barrier = new_barrier;
 769         }
 770 }
 771
 772 static bool
 773 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 774                 const nfs4_stateid *stateid)
 775 {
 776         u32 seqid = be32_to_cpu(stateid->seqid);
 777
 778         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 779 }
 780
 781 static bool
 782 pnfs_layout_returning(const struct pnfs_layout_hdr *lo,
 783                       struct pnfs_layout_range *range)
 784 {
 785         return test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 786                 (lo->plh_return_iomode == IOMODE_ANY ||
 787                  lo->plh_return_iomode == range->iomode);
 788 }
 789
 790 /* lget is set to 1 if called from inside send_layoutget call chain */
 791 static bool
 792 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo,
 793                         struct pnfs_layout_range *range, int lget)
 794 {
 795         return lo->plh_block_lgets ||
 796                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 797                 (list_empty(&lo->plh_segs) &&
 798                  (atomic_read(&lo->plh_outstanding) > lget)) ||
 799                 pnfs_layout_returning(lo, range);
 800 }
 801
 802 int
 803 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 804                               struct pnfs_layout_range *range,
 805                               struct nfs4_state *open_state)
 806 {
 807         int status = 0;
 808
 809         dprintk("--> %s\n", __func__);
 810         spin_lock(&lo->plh_inode->i_lock);
 811         if (pnfs_layoutgets_blocked(lo, range, 1)) {
 812                 status = -EAGAIN;
 813         } else if (!nfs4_valid_open_stateid(open_state)) {
 814                 status = -EBADF;
 815         } else if (list_empty(&lo->plh_segs) ||
 816                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 817                 int seq;
 818
 819                 do {
 820                         seq = read_seqbegin(&open_state->seqlock);
 821                         nfs4_stateid_copy(dst, &open_state->stateid);
 822                 } while (read_seqretry(&open_state->seqlock, seq));
 823         } else
 824                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 825         spin_unlock(&lo->plh_inode->i_lock);
 826         dprintk("<-- %s\n", __func__);
 827         return status;
 828 }
 829
 830 /*
 831 * Get layout from server.
 832 *    for now, assume that whole file layouts are requested.
 833 *    arg->offset: 0
 834 *    arg->length: all ones
 835 */
 836 static struct pnfs_layout_segment *
 837 send_layoutget(struct pnfs_layout_hdr *lo,
 838            struct nfs_open_context *ctx,
 839            struct pnfs_layout_range *range,
 840            gfp_t gfp_flags)
 841 {
 842         struct inode *ino = lo->plh_inode;
 843         struct nfs_server *server = NFS_SERVER(ino);
 844         struct nfs4_layoutget *lgp;
 845         struct pnfs_layout_segment *lseg;
 846
 847         dprintk("--> %s\n", __func__);
 848
 849         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 850         if (lgp == NULL)
 851                 return NULL;
 852
 853         lgp->args.minlength = PAGE_CACHE_SIZE;
 854         if (lgp->args.minlength > range->length)
 855                 lgp->args.minlength = range->length;
 856         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 857         lgp->args.range = *range;
 858         lgp->args.type = server->pnfs_curr_ld->id;
 859         lgp->args.inode = ino;
 860         lgp->args.ctx = get_nfs_open_context(ctx);
 861         lgp->gfp_flags = gfp_flags;
 862         lgp->cred = lo->plh_lc_cred;
 863
 864         /* Synchronously retrieve layout information from server and
 865          * store in lseg.
 866          */
 867         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 868         if (IS_ERR(lseg)) {
 869                 switch (PTR_ERR(lseg)) {
 870                 case -ENOMEM:
 871                 case -ERESTARTSYS:
 872                         break;
 873                 default:
 874                         /* remember that LAYOUTGET failed and suspend trying */
 875                         pnfs_layout_io_set_failed(lo, range->iomode);
 876                 }
 877                 return NULL;
 878         }
 879
 880         return lseg;
 881 }
 882
 883 static void pnfs_clear_layoutcommit(struct inode *inode,
 884                 struct list_head *head)
 885 {
 886         struct nfs_inode *nfsi = NFS_I(inode);
 887         struct pnfs_layout_segment *lseg, *tmp;
 888
 889         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 890                 return;
 891         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 892                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 893                         continue;
 894                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 895         }
 896 }
 897
 898 static int
 899 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 900                        enum pnfs_iomode iomode)
 901 {
 902         struct inode *ino = lo->plh_inode;
 903         struct nfs4_layoutreturn *lrp;
 904         int status = 0;
 905
 906         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 907         if (unlikely(lrp == NULL)) {
 908                 status = -ENOMEM;
 909                 spin_lock(&ino->i_lock);
 910                 lo->plh_block_lgets--;
 911                 spin_unlock(&ino->i_lock);
 912                 pnfs_put_layout_hdr(lo);
 913                 goto out;
 914         }
 915
 916         lrp->args.stateid = stateid;
 917         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 918         lrp->args.inode = ino;
 919         lrp->args.range.iomode = iomode;
 920         lrp->args.range.offset = 0;
 921         lrp->args.range.length = NFS4_MAX_UINT64;
 922         lrp->args.layout = lo;
 923         lrp->clp = NFS_SERVER(ino)->nfs_client;
 924         lrp->cred = lo->plh_lc_cred;
 925
 926         status = nfs4_proc_layoutreturn(lrp);
 927 out:
 928         if (status) {
 929                 spin_lock(&ino->i_lock);
 930                 clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 931                 spin_unlock(&ino->i_lock);
 932         }
 933         dprintk("<-- %s status: %d\n", __func__, status);
 934         return status;
 935 }
 936
 937 /*
 938  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 939  * when the layout segment list is empty.
 940  *
 941  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 942  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 943  * deviceid is marked invalid.
 944  */
 945 int
 946 _pnfs_return_layout(struct inode *ino)
 947 {
 948         struct pnfs_layout_hdr *lo = NULL;
 949         struct nfs_inode *nfsi = NFS_I(ino);
 950         LIST_HEAD(tmp_list);
 951         nfs4_stateid stateid;
 952         int status = 0, empty;
 953
 954         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 955
 956         spin_lock(&ino->i_lock);
 957         lo = nfsi->layout;
 958         if (!lo) {
 959                 spin_unlock(&ino->i_lock);
 960                 dprintk("NFS: %s no layout to return\n", __func__);
 961                 goto out;
 962         }
 963         stateid = nfsi->layout->plh_stateid;
 964         /* Reference matched in nfs4_layoutreturn_release */
 965         pnfs_get_layout_hdr(lo);
 966         empty = list_empty(&lo->plh_segs);
 967         pnfs_clear_layoutcommit(ino, &tmp_list);
 968         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 969
 970         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 971                 struct pnfs_layout_range range = {
 972                         .iomode         = IOMODE_ANY,
 973                         .offset         = 0,
 974                         .length         = NFS4_MAX_UINT64,
 975                 };
 976                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
 977         }
 978
 979         /* Don't send a LAYOUTRETURN if list was initially empty */
 980         if (empty) {
 981                 spin_unlock(&ino->i_lock);
 982                 pnfs_put_layout_hdr(lo);
 983                 dprintk("NFS: %s no layout segments to return\n", __func__);
 984                 goto out;
 985         }
 986
 987         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 988         lo->plh_block_lgets++;
 989         spin_unlock(&ino->i_lock);
 990         pnfs_free_lseg_list(&tmp_list);
 991
 992         status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY);
 993 out:
 994         dprintk("<-- %s status: %d\n", __func__, status);
 995         return status;
 996 }
 997 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 998
 999 int
1000 pnfs_commit_and_return_layout(struct inode *inode)
1001 {
1002         struct pnfs_layout_hdr *lo;
1003         int ret;
1004
1005         spin_lock(&inode->i_lock);
1006         lo = NFS_I(inode)->layout;
1007         if (lo == NULL) {
1008                 spin_unlock(&inode->i_lock);
1009                 return 0;
1010         }
1011         pnfs_get_layout_hdr(lo);
1012         /* Block new layoutgets and read/write to ds */
1013         lo->plh_block_lgets++;
1014         spin_unlock(&inode->i_lock);
1015         filemap_fdatawait(inode->i_mapping);
1016         ret = pnfs_layoutcommit_inode(inode, true);
1017         if (ret == 0)
1018                 ret = _pnfs_return_layout(inode);
1019         spin_lock(&inode->i_lock);
1020         lo->plh_block_lgets--;
1021         spin_unlock(&inode->i_lock);
1022         pnfs_put_layout_hdr(lo);
1023         return ret;
1024 }
1025
1026 bool pnfs_roc(struct inode *ino)
1027 {
1028         struct pnfs_layout_hdr *lo;
1029         struct pnfs_layout_segment *lseg, *tmp;
1030         LIST_HEAD(tmp_list);
1031         bool found = false;
1032
1033         spin_lock(&ino->i_lock);
1034         lo = NFS_I(ino)->layout;
1035         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
1036             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1037                 goto out_nolayout;
1038         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1039                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1040                         mark_lseg_invalid(lseg, &tmp_list);
1041                         found = true;
1042                 }
1043         if (!found)
1044                 goto out_nolayout;
1045         lo->plh_block_lgets++;
1046         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
1047         spin_unlock(&ino->i_lock);
1048         pnfs_free_lseg_list(&tmp_list);
1049         return true;
1050
1051 out_nolayout:
1052         spin_unlock(&ino->i_lock);
1053         return false;
1054 }
1055
1056 void pnfs_roc_release(struct inode *ino)
1057 {
1058         struct pnfs_layout_hdr *lo;
1059
1060         spin_lock(&ino->i_lock);
1061         lo = NFS_I(ino)->layout;
1062         lo->plh_block_lgets--;
1063         if (atomic_dec_and_test(&lo->plh_refcount)) {
1064                 pnfs_detach_layout_hdr(lo);
1065                 spin_unlock(&ino->i_lock);
1066                 pnfs_free_layout_hdr(lo);
1067         } else
1068                 spin_unlock(&ino->i_lock);
1069 }
1070
1071 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1072 {
1073         struct pnfs_layout_hdr *lo;
1074
1075         spin_lock(&ino->i_lock);
1076         lo = NFS_I(ino)->layout;
1077         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1078                 lo->plh_barrier = barrier;
1079         spin_unlock(&ino->i_lock);
1080 }
1081
1082 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
1083 {
1084         struct nfs_inode *nfsi = NFS_I(ino);
1085         struct pnfs_layout_hdr *lo;
1086         struct pnfs_layout_segment *lseg;
1087         u32 current_seqid;
1088         bool found = false;
1089
1090         spin_lock(&ino->i_lock);
1091         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1092                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1093                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1094                         found = true;
1095                         goto out;
1096                 }
1097         lo = nfsi->layout;
1098         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1099
1100         /* Since close does not return a layout stateid for use as
1101          * a barrier, we choose the worst-case barrier.
1102          */
1103         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1104 out:
1105         spin_unlock(&ino->i_lock);
1106         return found;
1107 }
1108
1109 /*
1110  * Compare two layout segments for sorting into layout cache.
1111  * We want to preferentially return RW over RO layouts, so ensure those
1112  * are seen first.
1113  */
1114 static s64
1115 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1116            const struct pnfs_layout_range *l2)
1117 {
1118         s64 d;
1119
1120         /* high offset > low offset */
1121         d = l1->offset - l2->offset;
1122         if (d)
1123                 return d;
1124
1125         /* short length > long length */
1126         d = l2->length - l1->length;
1127         if (d)
1128                 return d;
1129
1130         /* read > read/write */
1131         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1132 }
1133
1134 static void
1135 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1136                    struct pnfs_layout_segment *lseg)
1137 {
1138         struct pnfs_layout_segment *lp;
1139
1140         dprintk("%s:Begin\n", __func__);
1141
1142         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1143                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1144                         continue;
1145                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1146                 dprintk("%s: inserted lseg %p "
1147                         "iomode %d offset %llu length %llu before "
1148                         "lp %p iomode %d offset %llu length %llu\n",
1149                         __func__, lseg, lseg->pls_range.iomode,
1150                         lseg->pls_range.offset, lseg->pls_range.length,
1151                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1152                         lp->pls_range.length);
1153                 goto out;
1154         }
1155         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1156         dprintk("%s: inserted lseg %p "
1157                 "iomode %d offset %llu length %llu at tail\n",
1158                 __func__, lseg, lseg->pls_range.iomode,
1159                 lseg->pls_range.offset, lseg->pls_range.length);
1160 out:
1161         pnfs_get_layout_hdr(lo);
1162
1163         dprintk("%s:Return\n", __func__);
1164 }
1165
1166 static struct pnfs_layout_hdr *
1167 alloc_init_layout_hdr(struct inode *ino,
1168                       struct nfs_open_context *ctx,
1169                       gfp_t gfp_flags)
1170 {
1171         struct pnfs_layout_hdr *lo;
1172
1173         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1174         if (!lo)
1175                 return NULL;
1176         atomic_set(&lo->plh_refcount, 1);
1177         INIT_LIST_HEAD(&lo->plh_layouts);
1178         INIT_LIST_HEAD(&lo->plh_segs);
1179         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1180         lo->plh_inode = ino;
1181         lo->plh_lc_cred = get_rpccred(ctx->cred);
1182         return lo;
1183 }
1184
1185 static struct pnfs_layout_hdr *
1186 pnfs_find_alloc_layout(struct inode *ino,
1187                        struct nfs_open_context *ctx,
1188                        gfp_t gfp_flags)
1189 {
1190         struct nfs_inode *nfsi = NFS_I(ino);
1191         struct pnfs_layout_hdr *new = NULL;
1192
1193         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1194
1195         if (nfsi->layout != NULL)
1196                 goto out_existing;
1197         spin_unlock(&ino->i_lock);
1198         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1199         spin_lock(&ino->i_lock);
1200
1201         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1202                 nfsi->layout = new;
1203                 return new;
1204         } else if (new != NULL)
1205                 pnfs_free_layout_hdr(new);
1206 out_existing:
1207         pnfs_get_layout_hdr(nfsi->layout);
1208         return nfsi->layout;
1209 }
1210
1211 /*
1212  * iomode matching rules:
1213  * iomode       lseg    match
1214  * -----        -----   -----
1215  * ANY          READ    true
1216  * ANY          RW      true
1217  * RW           READ    false
1218  * RW           RW      true
1219  * READ         READ    true
1220  * READ         RW      true
1221  */
1222 static bool
1223 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1224                  const struct pnfs_layout_range *range)
1225 {
1226         struct pnfs_layout_range range1;
1227
1228         if ((range->iomode == IOMODE_RW &&
1229              ls_range->iomode != IOMODE_RW) ||
1230             !pnfs_lseg_range_intersecting(ls_range, range))
1231                 return 0;
1232
1233         /* range1 covers only the first byte in the range */
1234         range1 = *range;
1235         range1.length = 1;
1236         return pnfs_lseg_range_contained(ls_range, &range1);
1237 }
1238
1239 /*
1240  * lookup range in layout
1241  */
1242 static struct pnfs_layout_segment *
1243 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1244                 struct pnfs_layout_range *range)
1245 {
1246         struct pnfs_layout_segment *lseg, *ret = NULL;
1247
1248         dprintk("%s:Begin\n", __func__);
1249
1250         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1251                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1252                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1253                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1254                         ret = pnfs_get_lseg(lseg);
1255                         break;
1256                 }
1257                 if (lseg->pls_range.offset > range->offset)
1258                         break;
1259         }
1260
1261         dprintk("%s:Return lseg %p ref %d\n",
1262                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1263         return ret;
1264 }
1265
1266 /*
1267  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1268  * to the MDS or over pNFS
1269  *
1270  * The nfs_inode read_io and write_io fields are cumulative counters reset
1271  * when there are no layout segments. Note that in pnfs_update_layout iomode
1272  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1273  * WRITE request.
1274  *
1275  * A return of true means use MDS I/O.
1276  *
1277  * From rfc 5661:
1278  * If a file's size is smaller than the file size threshold, data accesses
1279  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1280  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1281  * server.  If both file size and I/O size are provided, the client SHOULD
1282  * reach or exceed  both thresholds before sending its read or write
1283  * requests to the data server.
1284  */
1285 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1286                                      struct inode *ino, int iomode)
1287 {
1288         struct nfs4_threshold *t = ctx->mdsthreshold;
1289         struct nfs_inode *nfsi = NFS_I(ino);
1290         loff_t fsize = i_size_read(ino);
1291         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1292
1293         if (t == NULL)
1294                 return ret;
1295
1296         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1297                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1298
1299         switch (iomode) {
1300         case IOMODE_READ:
1301                 if (t->bm & THRESHOLD_RD) {
1302                         dprintk("%s fsize %llu\n", __func__, fsize);
1303                         size_set = true;
1304                         if (fsize < t->rd_sz)
1305                                 size = true;
1306                 }
1307                 if (t->bm & THRESHOLD_RD_IO) {
1308                         dprintk("%s nfsi->read_io %llu\n", __func__,
1309                                 nfsi->read_io);
1310                         io_set = true;
1311                         if (nfsi->read_io < t->rd_io_sz)
1312                                 io = true;
1313                 }
1314                 break;
1315         case IOMODE_RW:
1316                 if (t->bm & THRESHOLD_WR) {
1317                         dprintk("%s fsize %llu\n", __func__, fsize);
1318                         size_set = true;
1319                         if (fsize < t->wr_sz)
1320                                 size = true;
1321                 }
1322                 if (t->bm & THRESHOLD_WR_IO) {
1323                         dprintk("%s nfsi->write_io %llu\n", __func__,
1324                                 nfsi->write_io);
1325                         io_set = true;
1326                         if (nfsi->write_io < t->wr_io_sz)
1327                                 io = true;
1328                 }
1329                 break;
1330         }
1331         if (size_set && io_set) {
1332                 if (size && io)
1333                         ret = true;
1334         } else if (size || io)
1335                 ret = true;
1336
1337         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1338         return ret;
1339 }
1340
1341 /*
1342  * Layout segment is retreived from the server if not cached.
1343  * The appropriate layout segment is referenced and returned to the caller.
1344  */
1345 struct pnfs_layout_segment *
1346 pnfs_update_layout(struct inode *ino,
1347                    struct nfs_open_context *ctx,
1348                    loff_t pos,
1349                    u64 count,
1350                    enum pnfs_iomode iomode,
1351                    gfp_t gfp_flags)
1352 {
1353         struct pnfs_layout_range arg = {
1354                 .iomode = iomode,
1355                 .offset = pos,
1356                 .length = count,
1357         };
1358         unsigned pg_offset;
1359         struct nfs_server *server = NFS_SERVER(ino);
1360         struct nfs_client *clp = server->nfs_client;
1361         struct pnfs_layout_hdr *lo;
1362         struct pnfs_layout_segment *lseg = NULL;
1363         bool first;
1364
1365         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1366                 goto out;
1367
1368         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1369                 goto out;
1370
1371 lookup_again:
1372         first = false;
1373         spin_lock(&ino->i_lock);
1374         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1375         if (lo == NULL) {
1376                 spin_unlock(&ino->i_lock);
1377                 goto out;
1378         }
1379
1380         /* Do we even need to bother with this? */
1381         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1382                 dprintk("%s matches recall, use MDS\n", __func__);
1383                 goto out_unlock;
1384         }
1385
1386         /* if LAYOUTGET already failed once we don't try again */
1387         if (pnfs_layout_io_test_failed(lo, iomode))
1388                 goto out_unlock;
1389
1390         first = list_empty(&lo->plh_segs);
1391         if (first) {
1392                 /* The first layoutget for the file. Need to serialize per
1393                  * RFC 5661 Errata 3208.
1394                  */
1395                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1396                                      &lo->plh_flags)) {
1397                         spin_unlock(&ino->i_lock);
1398                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1399                                     TASK_UNINTERRUPTIBLE);
1400                         pnfs_put_layout_hdr(lo);
1401                         goto lookup_again;
1402                 }
1403         } else {
1404                 /* Check to see if the layout for the given range
1405                  * already exists
1406                  */
1407                 lseg = pnfs_find_lseg(lo, &arg);
1408                 if (lseg)
1409                         goto out_unlock;
1410         }
1411
1412         if (pnfs_layoutgets_blocked(lo, &arg, 0))
1413                 goto out_unlock;
1414         atomic_inc(&lo->plh_outstanding);
1415         spin_unlock(&ino->i_lock);
1416
1417         if (list_empty(&lo->plh_layouts)) {
1418                 /* The lo must be on the clp list if there is any
1419                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1420                  */
1421                 spin_lock(&clp->cl_lock);
1422                 if (list_empty(&lo->plh_layouts))
1423                         list_add_tail(&lo->plh_layouts, &server->layouts);
1424                 spin_unlock(&clp->cl_lock);
1425         }
1426
1427         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1428         if (pg_offset) {
1429                 arg.offset -= pg_offset;
1430                 arg.length += pg_offset;
1431         }
1432         if (arg.length != NFS4_MAX_UINT64)
1433                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1434
1435         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1436         atomic_dec(&lo->plh_outstanding);
1437 out_put_layout_hdr:
1438         if (first) {
1439                 unsigned long *bitlock = &lo->plh_flags;
1440
1441                 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1442                 smp_mb__after_atomic();
1443                 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1444         }
1445         pnfs_put_layout_hdr(lo);
1446 out:
1447         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1448                         "(%s, offset: %llu, length: %llu)\n",
1449                         __func__, ino->i_sb->s_id,
1450                         (unsigned long long)NFS_FILEID(ino),
1451                         lseg == NULL ? "not found" : "found",
1452                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1453                         (unsigned long long)pos,
1454                         (unsigned long long)count);
1455         return lseg;
1456 out_unlock:
1457         spin_unlock(&ino->i_lock);
1458         goto out_put_layout_hdr;
1459 }
1460 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1461
1462 struct pnfs_layout_segment *
1463 pnfs_layout_process(struct nfs4_layoutget *lgp)
1464 {
1465         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1466         struct nfs4_layoutget_res *res = &lgp->res;
1467         struct pnfs_layout_segment *lseg;
1468         struct inode *ino = lo->plh_inode;
1469         LIST_HEAD(free_me);
1470         int status = 0;
1471
1472         /* Inject layout blob into I/O device driver */
1473         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1474         if (!lseg || IS_ERR(lseg)) {
1475                 if (!lseg)
1476                         status = -ENOMEM;
1477                 else
1478                         status = PTR_ERR(lseg);
1479                 dprintk("%s: Could not allocate layout: error %d\n",
1480                        __func__, status);
1481                 goto out;
1482         }
1483
1484         init_lseg(lo, lseg);
1485         lseg->pls_range = res->range;
1486
1487         spin_lock(&ino->i_lock);
1488         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1489                 dprintk("%s forget reply due to recall\n", __func__);
1490                 goto out_forget_reply;
1491         }
1492
1493         if (pnfs_layoutgets_blocked(lo, &lgp->args.range, 1)) {
1494                 dprintk("%s forget reply due to state\n", __func__);
1495                 goto out_forget_reply;
1496         }
1497
1498         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1499                 /* existing state ID, make sure the sequence number matches. */
1500                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1501                         dprintk("%s forget reply due to sequence\n", __func__);
1502                         goto out_forget_reply;
1503                 }
1504                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1505         } else {
1506                 /*
1507                  * We got an entirely new state ID.  Mark all segments for the
1508                  * inode invalid, and don't bother validating the stateid
1509                  * sequence number.
1510                  */
1511                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1512
1513                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1514                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1515         }
1516
1517         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1518
1519         pnfs_get_lseg(lseg);
1520         pnfs_layout_insert_lseg(lo, lseg);
1521
1522         if (res->return_on_close) {
1523                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1524                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1525         }
1526
1527         spin_unlock(&ino->i_lock);
1528         pnfs_free_lseg_list(&free_me);
1529         return lseg;
1530 out:
1531         return ERR_PTR(status);
1532
1533 out_forget_reply:
1534         spin_unlock(&ino->i_lock);
1535         lseg->pls_layout = lo;
1536         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1537         goto out;
1538 }
1539
1540 static void
1541 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1542                                 struct list_head *tmp_list,
1543                                 struct pnfs_layout_range *return_range)
1544 {
1545         struct pnfs_layout_segment *lseg, *next;
1546
1547         dprintk("%s:Begin lo %p\n", __func__, lo);
1548
1549         if (list_empty(&lo->plh_segs))
1550                 return;
1551
1552         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1553                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1554                         dprintk("%s: marking lseg %p iomode %d "
1555                                 "offset %llu length %llu\n", __func__,
1556                                 lseg, lseg->pls_range.iomode,
1557                                 lseg->pls_range.offset,
1558                                 lseg->pls_range.length);
1559                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1560                         mark_lseg_invalid(lseg, tmp_list);
1561                 }
1562 }
1563
1564 void pnfs_error_mark_layout_for_return(struct inode *inode,
1565                                        struct pnfs_layout_segment *lseg)
1566 {
1567         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1568         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1569         struct pnfs_layout_range range = {
1570                 .iomode = lseg->pls_range.iomode,
1571                 .offset = 0,
1572                 .length = NFS4_MAX_UINT64,
1573         };
1574         LIST_HEAD(free_me);
1575
1576         spin_lock(&inode->i_lock);
1577         /* set failure bit so that pnfs path will be retried later */
1578         pnfs_layout_set_fail_bit(lo, iomode);
1579         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1580         if (lo->plh_return_iomode == 0)
1581                 lo->plh_return_iomode = range.iomode;
1582         else if (lo->plh_return_iomode != range.iomode)
1583                 lo->plh_return_iomode = IOMODE_ANY;
1584         /*
1585          * mark all matching lsegs so that we are sure to have no live
1586          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1587          * for how it works.
1588          */
1589         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1590         spin_unlock(&inode->i_lock);
1591         pnfs_free_lseg_list(&free_me);
1592 }
1593 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1594
1595 void
1596 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1597 {
1598         u64 rd_size = req->wb_bytes;
1599
1600         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1601
1602         if (pgio->pg_dreq == NULL)
1603                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1604         else
1605                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1606
1607         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1608                                            req->wb_context,
1609                                            req_offset(req),
1610                                            rd_size,
1611                                            IOMODE_READ,
1612                                            GFP_KERNEL);
1613         /* If no lseg, fall back to read through mds */
1614         if (pgio->pg_lseg == NULL)
1615                 nfs_pageio_reset_read_mds(pgio);
1616
1617 }
1618 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1619
1620 void
1621 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1622                            struct nfs_page *req, u64 wb_size)
1623 {
1624         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1625
1626         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1627                                            req->wb_context,
1628                                            req_offset(req),
1629                                            wb_size,
1630                                            IOMODE_RW,
1631                                            GFP_NOFS);
1632         /* If no lseg, fall back to write through mds */
1633         if (pgio->pg_lseg == NULL)
1634                 nfs_pageio_reset_write_mds(pgio);
1635 }
1636 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1637
1638 void
1639 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1640 {
1641         if (desc->pg_lseg) {
1642                 pnfs_put_lseg(desc->pg_lseg);
1643                 desc->pg_lseg = NULL;
1644         }
1645 }
1646 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1647
1648 /*
1649  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1650  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1651  */
1652 size_t
1653 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1654                      struct nfs_page *prev, struct nfs_page *req)
1655 {
1656         unsigned int size;
1657         u64 seg_end, req_start, seg_left;
1658
1659         size = nfs_generic_pg_test(pgio, prev, req);
1660         if (!size)
1661                 return 0;
1662
1663         /*
1664          * 'size' contains the number of bytes left in the current page (up
1665          * to the original size asked for in @req->wb_bytes).
1666          *
1667          * Calculate how many bytes are left in the layout segment
1668          * and if there are less bytes than 'size', return that instead.
1669          *
1670          * Please also note that 'end_offset' is actually the offset of the
1671          * first byte that lies outside the pnfs_layout_range. FIXME?
1672          *
1673          */
1674         if (pgio->pg_lseg) {
1675                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1676                                      pgio->pg_lseg->pls_range.length);
1677                 req_start = req_offset(req);
1678                 WARN_ON_ONCE(req_start > seg_end);
1679                 /* start of request is past the last byte of this segment */
1680                 if (req_start >= seg_end)
1681                         return 0;
1682
1683                 /* adjust 'size' iff there are fewer bytes left in the
1684                  * segment than what nfs_generic_pg_test returned */
1685                 seg_left = seg_end - req_start;
1686                 if (seg_left < size)
1687                         size = (unsigned int)seg_left;
1688         }
1689
1690         return size;
1691 }
1692 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1693
1694 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1695 {
1696         struct nfs_pageio_descriptor pgio;
1697
1698         /* Resend all requests through the MDS */
1699         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1700                               hdr->completion_ops);
1701         return nfs_pageio_resend(&pgio, hdr);
1702 }
1703 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1704
1705 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1706 {
1707
1708         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1709         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1710             PNFS_LAYOUTRET_ON_ERROR) {
1711                 pnfs_return_layout(hdr->inode);
1712         }
1713         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1714                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1715 }
1716
1717 /*
1718  * Called by non rpc-based layout drivers
1719  */
1720 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1721 {
1722         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1723         if (!hdr->pnfs_error) {
1724                 pnfs_set_layoutcommit(hdr);
1725                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1726         } else
1727                 pnfs_ld_handle_write_error(hdr);
1728         hdr->mds_ops->rpc_release(hdr);
1729 }
1730 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1731
1732 static void
1733 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1734                 struct nfs_pgio_header *hdr)
1735 {
1736         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1737
1738         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1739                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1740                 nfs_pageio_reset_write_mds(desc);
1741                 mirror->pg_recoalesce = 1;
1742         }
1743         nfs_pgio_data_destroy(hdr);
1744 }
1745
1746 static enum pnfs_try_status
1747 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1748                         const struct rpc_call_ops *call_ops,
1749                         struct pnfs_layout_segment *lseg,
1750                         int how)
1751 {
1752         struct inode *inode = hdr->inode;
1753         enum pnfs_try_status trypnfs;
1754         struct nfs_server *nfss = NFS_SERVER(inode);
1755
1756         hdr->mds_ops = call_ops;
1757
1758         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1759                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1760         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1761         if (trypnfs != PNFS_NOT_ATTEMPTED)
1762                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1763         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1764         return trypnfs;
1765 }
1766
1767 static void
1768 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1769               struct nfs_pgio_header *hdr, int how)
1770 {
1771         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1772         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1773         enum pnfs_try_status trypnfs;
1774
1775         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1776         if (trypnfs == PNFS_NOT_ATTEMPTED)
1777                 pnfs_write_through_mds(desc, hdr);
1778 }
1779
1780 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1781 {
1782         pnfs_put_lseg(hdr->lseg);
1783         nfs_pgio_header_free(hdr);
1784 }
1785 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1786
1787 int
1788 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1789 {
1790         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1791
1792         struct nfs_pgio_header *hdr;
1793         int ret;
1794
1795         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1796         if (!hdr) {
1797                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1798                 return -ENOMEM;
1799         }
1800         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1801
1802         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1803         ret = nfs_generic_pgio(desc, hdr);
1804         if (!ret)
1805                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1806
1807         return ret;
1808 }
1809 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1810
1811 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1812 {
1813         struct nfs_pageio_descriptor pgio;
1814
1815         /* Resend all requests through the MDS */
1816         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1817         return nfs_pageio_resend(&pgio, hdr);
1818 }
1819 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1820
1821 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1822 {
1823         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1824         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1825             PNFS_LAYOUTRET_ON_ERROR) {
1826                 pnfs_return_layout(hdr->inode);
1827         }
1828         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1829                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1830 }
1831
1832 /*
1833  * Called by non rpc-based layout drivers
1834  */
1835 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1836 {
1837         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1838         if (likely(!hdr->pnfs_error)) {
1839                 __nfs4_read_done_cb(hdr);
1840                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1841         } else
1842                 pnfs_ld_handle_read_error(hdr);
1843         hdr->mds_ops->rpc_release(hdr);
1844 }
1845 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1846
1847 static void
1848 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1849                 struct nfs_pgio_header *hdr)
1850 {
1851         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1852
1853         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1854                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1855                 nfs_pageio_reset_read_mds(desc);
1856                 mirror->pg_recoalesce = 1;
1857         }
1858         nfs_pgio_data_destroy(hdr);
1859 }
1860
1861 /*
1862  * Call the appropriate parallel I/O subsystem read function.
1863  */
1864 static enum pnfs_try_status
1865 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1866                        const struct rpc_call_ops *call_ops,
1867                        struct pnfs_layout_segment *lseg)
1868 {
1869         struct inode *inode = hdr->inode;
1870         struct nfs_server *nfss = NFS_SERVER(inode);
1871         enum pnfs_try_status trypnfs;
1872
1873         hdr->mds_ops = call_ops;
1874
1875         dprintk("%s: Reading ino:%lu %u@%llu\n",
1876                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1877
1878         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1879         if (trypnfs != PNFS_NOT_ATTEMPTED)
1880                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1881         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1882         return trypnfs;
1883 }
1884
1885 /* Resend all requests through pnfs. */
1886 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
1887 {
1888         struct nfs_pageio_descriptor pgio;
1889
1890         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
1891         return nfs_pageio_resend(&pgio, hdr);
1892 }
1893 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
1894
1895 static void
1896 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1897 {
1898         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1899         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1900         enum pnfs_try_status trypnfs;
1901         int err = 0;
1902
1903         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1904         if (trypnfs == PNFS_TRY_AGAIN)
1905                 err = pnfs_read_resend_pnfs(hdr);
1906         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
1907                 pnfs_read_through_mds(desc, hdr);
1908 }
1909
1910 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1911 {
1912         pnfs_put_lseg(hdr->lseg);
1913         nfs_pgio_header_free(hdr);
1914 }
1915 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1916
1917 int
1918 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1919 {
1920         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1921
1922         struct nfs_pgio_header *hdr;
1923         int ret;
1924
1925         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1926         if (!hdr) {
1927                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1928                 return -ENOMEM;
1929         }
1930         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1931         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1932         ret = nfs_generic_pgio(desc, hdr);
1933         if (!ret)
1934                 pnfs_do_read(desc, hdr);
1935         return ret;
1936 }
1937 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1938
1939 static void pnfs_clear_layoutcommitting(struct inode *inode)
1940 {
1941         unsigned long *bitlock = &NFS_I(inode)->flags;
1942
1943         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1944         smp_mb__after_atomic();
1945         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1946 }
1947
1948 /*
1949  * There can be multiple RW segments.
1950  */
1951 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1952 {
1953         struct pnfs_layout_segment *lseg;
1954
1955         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1956                 if (lseg->pls_range.iomode == IOMODE_RW &&
1957                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1958                         list_add(&lseg->pls_lc_list, listp);
1959         }
1960 }
1961
1962 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1963 {
1964         struct pnfs_layout_segment *lseg, *tmp;
1965
1966         /* Matched by references in pnfs_set_layoutcommit */
1967         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1968                 list_del_init(&lseg->pls_lc_list);
1969                 pnfs_put_lseg(lseg);
1970         }
1971
1972         pnfs_clear_layoutcommitting(inode);
1973 }
1974
1975 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1976 {
1977         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1978 }
1979 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1980
1981 void
1982 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1983 {
1984         struct inode *inode = hdr->inode;
1985         struct nfs_inode *nfsi = NFS_I(inode);
1986         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1987         bool mark_as_dirty = false;
1988
1989         spin_lock(&inode->i_lock);
1990         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1991                 mark_as_dirty = true;
1992                 dprintk("%s: Set layoutcommit for inode %lu ",
1993                         __func__, inode->i_ino);
1994         }
1995         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1996                 /* references matched in nfs4_layoutcommit_release */
1997                 pnfs_get_lseg(hdr->lseg);
1998         }
1999         if (end_pos > nfsi->layout->plh_lwb)
2000                 nfsi->layout->plh_lwb = end_pos;
2001         spin_unlock(&inode->i_lock);
2002         dprintk("%s: lseg %p end_pos %llu\n",
2003                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
2004
2005         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2006          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2007         if (mark_as_dirty)
2008                 mark_inode_dirty_sync(inode);
2009 }
2010 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2011
2012 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
2013 {
2014         struct inode *inode = data->inode;
2015         struct nfs_inode *nfsi = NFS_I(inode);
2016         bool mark_as_dirty = false;
2017
2018         spin_lock(&inode->i_lock);
2019         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2020                 mark_as_dirty = true;
2021                 dprintk("%s: Set layoutcommit for inode %lu ",
2022                         __func__, inode->i_ino);
2023         }
2024         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
2025                 /* references matched in nfs4_layoutcommit_release */
2026                 pnfs_get_lseg(data->lseg);
2027         }
2028         if (data->lwb > nfsi->layout->plh_lwb)
2029                 nfsi->layout->plh_lwb = data->lwb;
2030         spin_unlock(&inode->i_lock);
2031         dprintk("%s: lseg %p end_pos %llu\n",
2032                 __func__, data->lseg, nfsi->layout->plh_lwb);
2033
2034         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2035          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2036         if (mark_as_dirty)
2037                 mark_inode_dirty_sync(inode);
2038 }
2039 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
2040
2041 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2042 {
2043         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2044
2045         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2046                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2047         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2048 }
2049
2050 /*
2051  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2052  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2053  * data to disk to allow the server to recover the data if it crashes.
2054  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2055  * is off, and a COMMIT is sent to a data server, or
2056  * if WRITEs to a data server return NFS_DATA_SYNC.
2057  */
2058 int
2059 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2060 {
2061         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2062         struct nfs4_layoutcommit_data *data;
2063         struct nfs_inode *nfsi = NFS_I(inode);
2064         loff_t end_pos;
2065         int status;
2066
2067         if (!pnfs_layoutcommit_outstanding(inode))
2068                 return 0;
2069
2070         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2071
2072         status = -EAGAIN;
2073         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2074                 if (!sync)
2075                         goto out;
2076                 status = wait_on_bit_lock_action(&nfsi->flags,
2077                                 NFS_INO_LAYOUTCOMMITTING,
2078                                 nfs_wait_bit_killable,
2079                                 TASK_KILLABLE);
2080                 if (status)
2081                         goto out;
2082         }
2083
2084         status = -ENOMEM;
2085         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2086         data = kzalloc(sizeof(*data), GFP_NOFS);
2087         if (!data)
2088                 goto clear_layoutcommitting;
2089
2090         status = 0;
2091         spin_lock(&inode->i_lock);
2092         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2093                 goto out_unlock;
2094
2095         INIT_LIST_HEAD(&data->lseg_list);
2096         pnfs_list_write_lseg(inode, &data->lseg_list);
2097
2098         end_pos = nfsi->layout->plh_lwb;
2099         nfsi->layout->plh_lwb = 0;
2100
2101         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2102         spin_unlock(&inode->i_lock);
2103
2104         data->args.inode = inode;
2105         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2106         nfs_fattr_init(&data->fattr);
2107         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2108         data->res.fattr = &data->fattr;
2109         data->args.lastbytewritten = end_pos - 1;
2110         data->res.server = NFS_SERVER(inode);
2111
2112         if (ld->prepare_layoutcommit) {
2113                 status = ld->prepare_layoutcommit(&data->args);
2114                 if (status) {
2115                         spin_lock(&inode->i_lock);
2116                         if (end_pos < nfsi->layout->plh_lwb)
2117                                 nfsi->layout->plh_lwb = end_pos;
2118                         spin_unlock(&inode->i_lock);
2119                         put_rpccred(data->cred);
2120                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2121                         goto clear_layoutcommitting;
2122                 }
2123         }
2124
2125
2126         status = nfs4_proc_layoutcommit(data, sync);
2127 out:
2128         if (status)
2129                 mark_inode_dirty_sync(inode);
2130         dprintk("<-- %s status %d\n", __func__, status);
2131         return status;
2132 out_unlock:
2133         spin_unlock(&inode->i_lock);
2134         kfree(data);
2135 clear_layoutcommitting:
2136         pnfs_clear_layoutcommitting(inode);
2137         goto out;
2138 }
2139 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2140
2141 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2142 {
2143         struct nfs4_threshold *thp;
2144
2145         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2146         if (!thp) {
2147                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2148                 return NULL;
2149         }
2150         return thp;
2151 }