fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36 #include "nfs4trace.h"
  37 #include "delegation.h"
  38
  39 #define NFSDBG_FACILITY         NFSDBG_PNFS
  40 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  41
  42 /* Locking:
  43  *
  44  * pnfs_spinlock:
  45  *      protects pnfs_modules_tbl.
  46  */
  47 static DEFINE_SPINLOCK(pnfs_spinlock);
  48
  49 /*
  50  * pnfs_modules_tbl holds all pnfs modules
  51  */
  52 static LIST_HEAD(pnfs_modules_tbl);
  53
  54 /* Return the registered pnfs layout driver module matching given id */
  55 static struct pnfs_layoutdriver_type *
  56 find_pnfs_driver_locked(u32 id)
  57 {
  58         struct pnfs_layoutdriver_type *local;
  59
  60         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  61                 if (local->id == id)
  62                         goto out;
  63         local = NULL;
  64 out:
  65         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  66         return local;
  67 }
  68
  69 static struct pnfs_layoutdriver_type *
  70 find_pnfs_driver(u32 id)
  71 {
  72         struct pnfs_layoutdriver_type *local;
  73
  74         spin_lock(&pnfs_spinlock);
  75         local = find_pnfs_driver_locked(id);
  76         if (local != NULL && !try_module_get(local->owner)) {
  77                 dprintk("%s: Could not grab reference on module\n", __func__);
  78                 local = NULL;
  79         }
  80         spin_unlock(&pnfs_spinlock);
  81         return local;
  82 }
  83
  84 void
  85 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  86 {
  87         if (nfss->pnfs_curr_ld) {
  88                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  89                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  90                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  91                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  92                         nfs4_deviceid_purge_client(nfss->nfs_client);
  93                 module_put(nfss->pnfs_curr_ld->owner);
  94         }
  95         nfss->pnfs_curr_ld = NULL;
  96 }
  97
  98 /*
  99  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 100  * Currently only one pNFS layout driver per filesystem is supported.
 101  *
 102  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 103  */
 104 void
 105 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 106                       u32 id)
 107 {
 108         struct pnfs_layoutdriver_type *ld_type = NULL;
 109
 110         if (id == 0)
 111                 goto out_no_driver;
 112         if (!(server->nfs_client->cl_exchange_flags &
 113                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 114                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 115                         __func__, id, server->nfs_client->cl_exchange_flags);
 116                 goto out_no_driver;
 117         }
 118         ld_type = find_pnfs_driver(id);
 119         if (!ld_type) {
 120                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 121                 ld_type = find_pnfs_driver(id);
 122                 if (!ld_type) {
 123                         dprintk("%s: No pNFS module found for %u.\n",
 124                                 __func__, id);
 125                         goto out_no_driver;
 126                 }
 127         }
 128         server->pnfs_curr_ld = ld_type;
 129         if (ld_type->set_layoutdriver
 130             && ld_type->set_layoutdriver(server, mntfh)) {
 131                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 132                         "driver %u.\n", __func__, id);
 133                 module_put(ld_type->owner);
 134                 goto out_no_driver;
 135         }
 136         /* Bump the MDS count */
 137         atomic_inc(&server->nfs_client->cl_mds_count);
 138
 139         dprintk("%s: pNFS module for %u set\n", __func__, id);
 140         return;
 141
 142 out_no_driver:
 143         dprintk("%s: Using NFSv4 I/O\n", __func__);
 144         server->pnfs_curr_ld = NULL;
 145 }
 146
 147 int
 148 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 149 {
 150         int status = -EINVAL;
 151         struct pnfs_layoutdriver_type *tmp;
 152
 153         if (ld_type->id == 0) {
 154                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 155                 return status;
 156         }
 157         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 158                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 159                        "alloc_lseg and free_lseg.\n", __func__);
 160                 return status;
 161         }
 162
 163         spin_lock(&pnfs_spinlock);
 164         tmp = find_pnfs_driver_locked(ld_type->id);
 165         if (!tmp) {
 166                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 167                 status = 0;
 168                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 169                         ld_type->name);
 170         } else {
 171                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 172                         __func__, ld_type->id);
 173         }
 174         spin_unlock(&pnfs_spinlock);
 175
 176         return status;
 177 }
 178 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 179
 180 void
 181 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 182 {
 183         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 184         spin_lock(&pnfs_spinlock);
 185         list_del(&ld_type->pnfs_tblid);
 186         spin_unlock(&pnfs_spinlock);
 187 }
 188 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 189
 190 /*
 191  * pNFS client layout cache
 192  */
 193
 194 /* Need to hold i_lock if caller does not already hold reference */
 195 void
 196 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 197 {
 198         atomic_inc(&lo->plh_refcount);
 199 }
 200
 201 static struct pnfs_layout_hdr *
 202 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 203 {
 204         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 205         return ld->alloc_layout_hdr(ino, gfp_flags);
 206 }
 207
 208 static void
 209 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 210 {
 211         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 212         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 213
 214         if (!list_empty(&lo->plh_layouts)) {
 215                 struct nfs_client *clp = server->nfs_client;
 216
 217                 spin_lock(&clp->cl_lock);
 218                 list_del_init(&lo->plh_layouts);
 219                 spin_unlock(&clp->cl_lock);
 220         }
 221         put_rpccred(lo->plh_lc_cred);
 222         return ld->free_layout_hdr(lo);
 223 }
 224
 225 static void
 226 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 227 {
 228         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 229         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 230         nfsi->layout = NULL;
 231         /* Reset MDS Threshold I/O counters */
 232         nfsi->write_io = 0;
 233         nfsi->read_io = 0;
 234 }
 235
 236 void
 237 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 238 {
 239         struct inode *inode = lo->plh_inode;
 240
 241         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 242                 pnfs_detach_layout_hdr(lo);
 243                 spin_unlock(&inode->i_lock);
 244                 pnfs_free_layout_hdr(lo);
 245         }
 246 }
 247
 248 static int
 249 pnfs_iomode_to_fail_bit(u32 iomode)
 250 {
 251         return iomode == IOMODE_RW ?
 252                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 253 }
 254
 255 static void
 256 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 257 {
 258         lo->plh_retry_timestamp = jiffies;
 259         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 260                 atomic_inc(&lo->plh_refcount);
 261 }
 262
 263 static void
 264 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 265 {
 266         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 267                 atomic_dec(&lo->plh_refcount);
 268 }
 269
 270 static void
 271 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 272 {
 273         struct inode *inode = lo->plh_inode;
 274         struct pnfs_layout_range range = {
 275                 .iomode = iomode,
 276                 .offset = 0,
 277                 .length = NFS4_MAX_UINT64,
 278         };
 279         LIST_HEAD(head);
 280
 281         spin_lock(&inode->i_lock);
 282         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 283         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 284         spin_unlock(&inode->i_lock);
 285         pnfs_free_lseg_list(&head);
 286         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 287                         iomode == IOMODE_RW ?  "RW" : "READ");
 288 }
 289
 290 static bool
 291 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 292 {
 293         unsigned long start, end;
 294         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 295
 296         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 297                 return false;
 298         end = jiffies;
 299         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 300         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 301                 /* It is time to retry the failed layoutgets */
 302                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 303                 return false;
 304         }
 305         return true;
 306 }
 307
 308 static void
 309 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 310 {
 311         INIT_LIST_HEAD(&lseg->pls_list);
 312         INIT_LIST_HEAD(&lseg->pls_lc_list);
 313         atomic_set(&lseg->pls_refcount, 1);
 314         smp_mb();
 315         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 316         lseg->pls_layout = lo;
 317 }
 318
 319 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 320 {
 321         struct inode *ino = lseg->pls_layout->plh_inode;
 322
 323         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 324 }
 325
 326 static void
 327 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 328                 struct pnfs_layout_segment *lseg)
 329 {
 330         struct inode *inode = lo->plh_inode;
 331
 332         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 333         list_del_init(&lseg->pls_list);
 334         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 335         atomic_dec(&lo->plh_refcount);
 336         if (list_empty(&lo->plh_segs))
 337                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 338         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 339 }
 340
 341 void
 342 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 343 {
 344         struct pnfs_layout_hdr *lo;
 345         struct inode *inode;
 346
 347         if (!lseg)
 348                 return;
 349
 350         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 351                 atomic_read(&lseg->pls_refcount),
 352                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 353         lo = lseg->pls_layout;
 354         inode = lo->plh_inode;
 355         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 356                 pnfs_get_layout_hdr(lo);
 357                 pnfs_layout_remove_lseg(lo, lseg);
 358                 spin_unlock(&inode->i_lock);
 359                 pnfs_free_lseg(lseg);
 360                 pnfs_put_layout_hdr(lo);
 361         }
 362 }
 363 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 364
 365 static void pnfs_free_lseg_async_work(struct work_struct *work)
 366 {
 367         struct pnfs_layout_segment *lseg;
 368         struct pnfs_layout_hdr *lo;
 369
 370         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 371         lo = lseg->pls_layout;
 372
 373         pnfs_free_lseg(lseg);
 374         pnfs_put_layout_hdr(lo);
 375 }
 376
 377 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 378 {
 379         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 380         schedule_work(&lseg->pls_work);
 381 }
 382
 383 void
 384 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 385 {
 386         if (!lseg)
 387                 return;
 388
 389         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 390
 391         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 392                 atomic_read(&lseg->pls_refcount),
 393                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 394         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 395                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 396                 pnfs_get_layout_hdr(lo);
 397                 pnfs_layout_remove_lseg(lo, lseg);
 398                 pnfs_free_lseg_async(lseg);
 399         }
 400 }
 401 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 402
 403 static u64
 404 end_offset(u64 start, u64 len)
 405 {
 406         u64 end;
 407
 408         end = start + len;
 409         return end >= start ? end : NFS4_MAX_UINT64;
 410 }
 411
 412 /*
 413  * is l2 fully contained in l1?
 414  *   start1                             end1
 415  *   [----------------------------------)
 416  *           start2           end2
 417  *           [----------------)
 418  */
 419 static bool
 420 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 421                  const struct pnfs_layout_range *l2)
 422 {
 423         u64 start1 = l1->offset;
 424         u64 end1 = end_offset(start1, l1->length);
 425         u64 start2 = l2->offset;
 426         u64 end2 = end_offset(start2, l2->length);
 427
 428         return (start1 <= start2) && (end1 >= end2);
 429 }
 430
 431 /*
 432  * is l1 and l2 intersecting?
 433  *   start1                             end1
 434  *   [----------------------------------)
 435  *                              start2           end2
 436  *                              [----------------)
 437  */
 438 static bool
 439 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 440                     const struct pnfs_layout_range *l2)
 441 {
 442         u64 start1 = l1->offset;
 443         u64 end1 = end_offset(start1, l1->length);
 444         u64 start2 = l2->offset;
 445         u64 end2 = end_offset(start2, l2->length);
 446
 447         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 448                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 449 }
 450
 451 static bool
 452 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 453                  const struct pnfs_layout_range *recall_range)
 454 {
 455         return (recall_range->iomode == IOMODE_ANY ||
 456                 lseg_range->iomode == recall_range->iomode) &&
 457                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 458 }
 459
 460 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 461                 struct list_head *tmp_list)
 462 {
 463         if (!atomic_dec_and_test(&lseg->pls_refcount))
 464                 return false;
 465         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 466         list_add(&lseg->pls_list, tmp_list);
 467         return true;
 468 }
 469
 470 /* Returns 1 if lseg is removed from list, 0 otherwise */
 471 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 472                              struct list_head *tmp_list)
 473 {
 474         int rv = 0;
 475
 476         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 477                 /* Remove the reference keeping the lseg in the
 478                  * list.  It will now be removed when all
 479                  * outstanding io is finished.
 480                  */
 481                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 482                         atomic_read(&lseg->pls_refcount));
 483                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 484                         rv = 1;
 485         }
 486         return rv;
 487 }
 488
 489 /* Returns count of number of matching invalid lsegs remaining in list
 490  * after call.
 491  */
 492 int
 493 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 494                             struct list_head *tmp_list,
 495                             struct pnfs_layout_range *recall_range)
 496 {
 497         struct pnfs_layout_segment *lseg, *next;
 498         int invalid = 0, removed = 0;
 499
 500         dprintk("%s:Begin lo %p\n", __func__, lo);
 501
 502         if (list_empty(&lo->plh_segs))
 503                 return 0;
 504         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 505                 if (!recall_range ||
 506                     should_free_lseg(&lseg->pls_range, recall_range)) {
 507                         dprintk("%s: freeing lseg %p iomode %d "
 508                                 "offset %llu length %llu\n", __func__,
 509                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 510                                 lseg->pls_range.length);
 511                         invalid++;
 512                         removed += mark_lseg_invalid(lseg, tmp_list);
 513                 }
 514         dprintk("%s:Return %i\n", __func__, invalid - removed);
 515         return invalid - removed;
 516 }
 517
 518 /* note free_me must contain lsegs from a single layout_hdr */
 519 void
 520 pnfs_free_lseg_list(struct list_head *free_me)
 521 {
 522         struct pnfs_layout_segment *lseg, *tmp;
 523
 524         if (list_empty(free_me))
 525                 return;
 526
 527         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 528                 list_del(&lseg->pls_list);
 529                 pnfs_free_lseg(lseg);
 530         }
 531 }
 532
 533 void
 534 pnfs_destroy_layout(struct nfs_inode *nfsi)
 535 {
 536         struct pnfs_layout_hdr *lo;
 537         LIST_HEAD(tmp_list);
 538
 539         spin_lock(&nfsi->vfs_inode.i_lock);
 540         lo = nfsi->layout;
 541         if (lo) {
 542                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 543                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 544                 pnfs_get_layout_hdr(lo);
 545                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 546                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 547                 spin_unlock(&nfsi->vfs_inode.i_lock);
 548                 pnfs_free_lseg_list(&tmp_list);
 549                 pnfs_put_layout_hdr(lo);
 550         } else
 551                 spin_unlock(&nfsi->vfs_inode.i_lock);
 552 }
 553 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 554
 555 static bool
 556 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 557                 struct list_head *layout_list)
 558 {
 559         struct pnfs_layout_hdr *lo;
 560         bool ret = false;
 561
 562         spin_lock(&inode->i_lock);
 563         lo = NFS_I(inode)->layout;
 564         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 565                 pnfs_get_layout_hdr(lo);
 566                 list_add(&lo->plh_bulk_destroy, layout_list);
 567                 ret = true;
 568         }
 569         spin_unlock(&inode->i_lock);
 570         return ret;
 571 }
 572
 573 /* Caller must hold rcu_read_lock and clp->cl_lock */
 574 static int
 575 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 576                 struct nfs_server *server,
 577                 struct list_head *layout_list)
 578 {
 579         struct pnfs_layout_hdr *lo, *next;
 580         struct inode *inode;
 581
 582         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 583                 inode = igrab(lo->plh_inode);
 584                 if (inode == NULL)
 585                         continue;
 586                 list_del_init(&lo->plh_layouts);
 587                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 588                         continue;
 589                 rcu_read_unlock();
 590                 spin_unlock(&clp->cl_lock);
 591                 iput(inode);
 592                 spin_lock(&clp->cl_lock);
 593                 rcu_read_lock();
 594                 return -EAGAIN;
 595         }
 596         return 0;
 597 }
 598
 599 static int
 600 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 601                 bool is_bulk_recall)
 602 {
 603         struct pnfs_layout_hdr *lo;
 604         struct inode *inode;
 605         struct pnfs_layout_range range = {
 606                 .iomode = IOMODE_ANY,
 607                 .offset = 0,
 608                 .length = NFS4_MAX_UINT64,
 609         };
 610         LIST_HEAD(lseg_list);
 611         int ret = 0;
 612
 613         while (!list_empty(layout_list)) {
 614                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 615                                 plh_bulk_destroy);
 616                 dprintk("%s freeing layout for inode %lu\n", __func__,
 617                         lo->plh_inode->i_ino);
 618                 inode = lo->plh_inode;
 619
 620                 pnfs_layoutcommit_inode(inode, false);
 621
 622                 spin_lock(&inode->i_lock);
 623                 list_del_init(&lo->plh_bulk_destroy);
 624                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 625                 if (is_bulk_recall)
 626                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 627                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 628                         ret = -EAGAIN;
 629                 spin_unlock(&inode->i_lock);
 630                 pnfs_free_lseg_list(&lseg_list);
 631                 pnfs_put_layout_hdr(lo);
 632                 iput(inode);
 633         }
 634         return ret;
 635 }
 636
 637 int
 638 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 639                 struct nfs_fsid *fsid,
 640                 bool is_recall)
 641 {
 642         struct nfs_server *server;
 643         LIST_HEAD(layout_list);
 644
 645         spin_lock(&clp->cl_lock);
 646         rcu_read_lock();
 647 restart:
 648         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 649                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 650                         continue;
 651                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 652                                 server,
 653                                 &layout_list) != 0)
 654                         goto restart;
 655         }
 656         rcu_read_unlock();
 657         spin_unlock(&clp->cl_lock);
 658
 659         if (list_empty(&layout_list))
 660                 return 0;
 661         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 662 }
 663
 664 int
 665 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 666                 bool is_recall)
 667 {
 668         struct nfs_server *server;
 669         LIST_HEAD(layout_list);
 670
 671         spin_lock(&clp->cl_lock);
 672         rcu_read_lock();
 673 restart:
 674         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 675                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 676                                         server,
 677                                         &layout_list) != 0)
 678                         goto restart;
 679         }
 680         rcu_read_unlock();
 681         spin_unlock(&clp->cl_lock);
 682
 683         if (list_empty(&layout_list))
 684                 return 0;
 685         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 686 }
 687
 688 /*
 689  * Called by the state manger to remove all layouts established under an
 690  * expired lease.
 691  */
 692 void
 693 pnfs_destroy_all_layouts(struct nfs_client *clp)
 694 {
 695         nfs4_deviceid_mark_client_invalid(clp);
 696         nfs4_deviceid_purge_client(clp);
 697
 698         pnfs_destroy_layouts_byclid(clp, false);
 699 }
 700
 701 /*
 702  * Compare 2 layout stateid sequence ids, to see which is newer,
 703  * taking into account wraparound issues.
 704  */
 705 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 706 {
 707         return (s32)(s1 - s2) > 0;
 708 }
 709
 710 /* update lo->plh_stateid with new if is more recent */
 711 void
 712 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 713                         bool update_barrier)
 714 {
 715         u32 oldseq, newseq, new_barrier;
 716         int empty = list_empty(&lo->plh_segs);
 717
 718         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 719         newseq = be32_to_cpu(new->seqid);
 720         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 721                 nfs4_stateid_copy(&lo->plh_stateid, new);
 722                 if (update_barrier) {
 723                         new_barrier = be32_to_cpu(new->seqid);
 724                 } else {
 725                         /* Because of wraparound, we want to keep the barrier
 726                          * "close" to the current seqids.
 727                          */
 728                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 729                 }
 730                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 731                         lo->plh_barrier = new_barrier;
 732         }
 733 }
 734
 735 static bool
 736 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 737                 const nfs4_stateid *stateid)
 738 {
 739         u32 seqid = be32_to_cpu(stateid->seqid);
 740
 741         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 742 }
 743
 744 /* lget is set to 1 if called from inside send_layoutget call chain */
 745 static bool
 746 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 747 {
 748         return lo->plh_block_lgets ||
 749                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 750                 (list_empty(&lo->plh_segs) &&
 751                  (atomic_read(&lo->plh_outstanding) > lget));
 752 }
 753
 754 int
 755 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 756                               struct nfs4_state *open_state)
 757 {
 758         int status = 0;
 759
 760         dprintk("--> %s\n", __func__);
 761         spin_lock(&lo->plh_inode->i_lock);
 762         if (pnfs_layoutgets_blocked(lo, 1)) {
 763                 status = -EAGAIN;
 764         } else if (!nfs4_valid_open_stateid(open_state)) {
 765                 status = -EBADF;
 766         } else if (list_empty(&lo->plh_segs) ||
 767                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 768                 int seq;
 769
 770                 do {
 771                         seq = read_seqbegin(&open_state->seqlock);
 772                         nfs4_stateid_copy(dst, &open_state->stateid);
 773                 } while (read_seqretry(&open_state->seqlock, seq));
 774         } else
 775                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 776         spin_unlock(&lo->plh_inode->i_lock);
 777         dprintk("<-- %s\n", __func__);
 778         return status;
 779 }
 780
 781 /*
 782 * Get layout from server.
 783 *    for now, assume that whole file layouts are requested.
 784 *    arg->offset: 0
 785 *    arg->length: all ones
 786 */
 787 static struct pnfs_layout_segment *
 788 send_layoutget(struct pnfs_layout_hdr *lo,
 789            struct nfs_open_context *ctx,
 790            struct pnfs_layout_range *range,
 791            gfp_t gfp_flags)
 792 {
 793         struct inode *ino = lo->plh_inode;
 794         struct nfs_server *server = NFS_SERVER(ino);
 795         struct nfs4_layoutget *lgp;
 796         struct pnfs_layout_segment *lseg;
 797
 798         dprintk("--> %s\n", __func__);
 799
 800         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 801         if (lgp == NULL)
 802                 return NULL;
 803
 804         lgp->args.minlength = PAGE_CACHE_SIZE;
 805         if (lgp->args.minlength > range->length)
 806                 lgp->args.minlength = range->length;
 807         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 808         lgp->args.range = *range;
 809         lgp->args.type = server->pnfs_curr_ld->id;
 810         lgp->args.inode = ino;
 811         lgp->args.ctx = get_nfs_open_context(ctx);
 812         lgp->gfp_flags = gfp_flags;
 813         lgp->cred = lo->plh_lc_cred;
 814
 815         /* Synchronously retrieve layout information from server and
 816          * store in lseg.
 817          */
 818         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 819         if (IS_ERR(lseg)) {
 820                 switch (PTR_ERR(lseg)) {
 821                 case -ENOMEM:
 822                 case -ERESTARTSYS:
 823                         break;
 824                 default:
 825                         /* remember that LAYOUTGET failed and suspend trying */
 826                         pnfs_layout_io_set_failed(lo, range->iomode);
 827                 }
 828                 return NULL;
 829         }
 830
 831         return lseg;
 832 }
 833
 834 static void pnfs_clear_layoutcommit(struct inode *inode,
 835                 struct list_head *head)
 836 {
 837         struct nfs_inode *nfsi = NFS_I(inode);
 838         struct pnfs_layout_segment *lseg, *tmp;
 839
 840         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 841                 return;
 842         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 843                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 844                         continue;
 845                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 846         }
 847 }
 848
 849 /*
 850  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 851  * when the layout segment list is empty.
 852  *
 853  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 854  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 855  * deviceid is marked invalid.
 856  */
 857 int
 858 _pnfs_return_layout(struct inode *ino)
 859 {
 860         struct pnfs_layout_hdr *lo = NULL;
 861         struct nfs_inode *nfsi = NFS_I(ino);
 862         LIST_HEAD(tmp_list);
 863         struct nfs4_layoutreturn *lrp;
 864         nfs4_stateid stateid;
 865         int status = 0, empty;
 866
 867         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 868
 869         spin_lock(&ino->i_lock);
 870         lo = nfsi->layout;
 871         if (!lo) {
 872                 spin_unlock(&ino->i_lock);
 873                 dprintk("NFS: %s no layout to return\n", __func__);
 874                 goto out;
 875         }
 876         stateid = nfsi->layout->plh_stateid;
 877         /* Reference matched in nfs4_layoutreturn_release */
 878         pnfs_get_layout_hdr(lo);
 879         empty = list_empty(&lo->plh_segs);
 880         pnfs_clear_layoutcommit(ino, &tmp_list);
 881         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 882
 883         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 884                 struct pnfs_layout_range range = {
 885                         .iomode         = IOMODE_ANY,
 886                         .offset         = 0,
 887                         .length         = NFS4_MAX_UINT64,
 888                 };
 889                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
 890         }
 891
 892         /* Don't send a LAYOUTRETURN if list was initially empty */
 893         if (empty) {
 894                 spin_unlock(&ino->i_lock);
 895                 pnfs_put_layout_hdr(lo);
 896                 dprintk("NFS: %s no layout segments to return\n", __func__);
 897                 goto out;
 898         }
 899
 900         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 901         lo->plh_block_lgets++;
 902         spin_unlock(&ino->i_lock);
 903         pnfs_free_lseg_list(&tmp_list);
 904
 905         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 906         if (unlikely(lrp == NULL)) {
 907                 status = -ENOMEM;
 908                 spin_lock(&ino->i_lock);
 909                 lo->plh_block_lgets--;
 910                 spin_unlock(&ino->i_lock);
 911                 pnfs_put_layout_hdr(lo);
 912                 goto out;
 913         }
 914
 915         lrp->args.stateid = stateid;
 916         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 917         lrp->args.inode = ino;
 918         lrp->args.layout = lo;
 919         lrp->clp = NFS_SERVER(ino)->nfs_client;
 920         lrp->cred = lo->plh_lc_cred;
 921
 922         status = nfs4_proc_layoutreturn(lrp);
 923 out:
 924         dprintk("<-- %s status: %d\n", __func__, status);
 925         return status;
 926 }
 927 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 928
 929 int
 930 pnfs_commit_and_return_layout(struct inode *inode)
 931 {
 932         struct pnfs_layout_hdr *lo;
 933         int ret;
 934
 935         spin_lock(&inode->i_lock);
 936         lo = NFS_I(inode)->layout;
 937         if (lo == NULL) {
 938                 spin_unlock(&inode->i_lock);
 939                 return 0;
 940         }
 941         pnfs_get_layout_hdr(lo);
 942         /* Block new layoutgets and read/write to ds */
 943         lo->plh_block_lgets++;
 944         spin_unlock(&inode->i_lock);
 945         filemap_fdatawait(inode->i_mapping);
 946         ret = pnfs_layoutcommit_inode(inode, true);
 947         if (ret == 0)
 948                 ret = _pnfs_return_layout(inode);
 949         spin_lock(&inode->i_lock);
 950         lo->plh_block_lgets--;
 951         spin_unlock(&inode->i_lock);
 952         pnfs_put_layout_hdr(lo);
 953         return ret;
 954 }
 955
 956 bool pnfs_roc(struct inode *ino)
 957 {
 958         struct nfs_inode *nfsi = NFS_I(ino);
 959         struct nfs_open_context *ctx;
 960         struct nfs4_state *state;
 961         struct pnfs_layout_hdr *lo;
 962         struct pnfs_layout_segment *lseg, *tmp;
 963         LIST_HEAD(tmp_list);
 964         bool found = false;
 965
 966         spin_lock(&ino->i_lock);
 967         lo = nfsi->layout;
 968         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 969             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 970                 goto out_noroc;
 971
 972         /* Don't return layout if we hold a delegation */
 973         if (nfs4_check_delegation(ino, FMODE_READ))
 974                 goto out_noroc;
 975
 976         list_for_each_entry(ctx, &nfsi->open_files, list) {
 977                 state = ctx->state;
 978                 /* Don't return layout if there is open file state */
 979                 if (state != NULL && state->state != 0)
 980                         goto out_noroc;
 981         }
 982
 983         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 984                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 985                         mark_lseg_invalid(lseg, &tmp_list);
 986                         found = true;
 987                 }
 988         if (!found)
 989                 goto out_noroc;
 990         lo->plh_block_lgets++;
 991         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 992         spin_unlock(&ino->i_lock);
 993         pnfs_free_lseg_list(&tmp_list);
 994         return true;
 995
 996 out_noroc:
 997         spin_unlock(&ino->i_lock);
 998         return false;
 999 }
1000
1001 void pnfs_roc_release(struct inode *ino)
1002 {
1003         struct pnfs_layout_hdr *lo;
1004
1005         spin_lock(&ino->i_lock);
1006         lo = NFS_I(ino)->layout;
1007         lo->plh_block_lgets--;
1008         if (atomic_dec_and_test(&lo->plh_refcount)) {
1009                 pnfs_detach_layout_hdr(lo);
1010                 spin_unlock(&ino->i_lock);
1011                 pnfs_free_layout_hdr(lo);
1012         } else
1013                 spin_unlock(&ino->i_lock);
1014 }
1015
1016 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1017 {
1018         struct pnfs_layout_hdr *lo;
1019
1020         spin_lock(&ino->i_lock);
1021         lo = NFS_I(ino)->layout;
1022         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1023                 lo->plh_barrier = barrier;
1024         spin_unlock(&ino->i_lock);
1025 }
1026
1027 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
1028 {
1029         struct nfs_inode *nfsi = NFS_I(ino);
1030         struct pnfs_layout_hdr *lo;
1031         struct pnfs_layout_segment *lseg;
1032         u32 current_seqid;
1033         bool found = false;
1034
1035         spin_lock(&ino->i_lock);
1036         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1037                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1038                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1039                         found = true;
1040                         goto out;
1041                 }
1042         lo = nfsi->layout;
1043         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1044
1045         /* Since close does not return a layout stateid for use as
1046          * a barrier, we choose the worst-case barrier.
1047          */
1048         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1049 out:
1050         spin_unlock(&ino->i_lock);
1051         return found;
1052 }
1053
1054 /*
1055  * Compare two layout segments for sorting into layout cache.
1056  * We want to preferentially return RW over RO layouts, so ensure those
1057  * are seen first.
1058  */
1059 static s64
1060 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1061            const struct pnfs_layout_range *l2)
1062 {
1063         s64 d;
1064
1065         /* high offset > low offset */
1066         d = l1->offset - l2->offset;
1067         if (d)
1068                 return d;
1069
1070         /* short length > long length */
1071         d = l2->length - l1->length;
1072         if (d)
1073                 return d;
1074
1075         /* read > read/write */
1076         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1077 }
1078
1079 static void
1080 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1081                    struct pnfs_layout_segment *lseg)
1082 {
1083         struct pnfs_layout_segment *lp;
1084
1085         dprintk("%s:Begin\n", __func__);
1086
1087         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1088                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1089                         continue;
1090                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1091                 dprintk("%s: inserted lseg %p "
1092                         "iomode %d offset %llu length %llu before "
1093                         "lp %p iomode %d offset %llu length %llu\n",
1094                         __func__, lseg, lseg->pls_range.iomode,
1095                         lseg->pls_range.offset, lseg->pls_range.length,
1096                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1097                         lp->pls_range.length);
1098                 goto out;
1099         }
1100         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1101         dprintk("%s: inserted lseg %p "
1102                 "iomode %d offset %llu length %llu at tail\n",
1103                 __func__, lseg, lseg->pls_range.iomode,
1104                 lseg->pls_range.offset, lseg->pls_range.length);
1105 out:
1106         pnfs_get_layout_hdr(lo);
1107
1108         dprintk("%s:Return\n", __func__);
1109 }
1110
1111 static struct pnfs_layout_hdr *
1112 alloc_init_layout_hdr(struct inode *ino,
1113                       struct nfs_open_context *ctx,
1114                       gfp_t gfp_flags)
1115 {
1116         struct pnfs_layout_hdr *lo;
1117
1118         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1119         if (!lo)
1120                 return NULL;
1121         atomic_set(&lo->plh_refcount, 1);
1122         INIT_LIST_HEAD(&lo->plh_layouts);
1123         INIT_LIST_HEAD(&lo->plh_segs);
1124         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1125         lo->plh_inode = ino;
1126         lo->plh_lc_cred = get_rpccred(ctx->cred);
1127         return lo;
1128 }
1129
1130 static struct pnfs_layout_hdr *
1131 pnfs_find_alloc_layout(struct inode *ino,
1132                        struct nfs_open_context *ctx,
1133                        gfp_t gfp_flags)
1134 {
1135         struct nfs_inode *nfsi = NFS_I(ino);
1136         struct pnfs_layout_hdr *new = NULL;
1137
1138         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1139
1140         if (nfsi->layout != NULL)
1141                 goto out_existing;
1142         spin_unlock(&ino->i_lock);
1143         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1144         spin_lock(&ino->i_lock);
1145
1146         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1147                 nfsi->layout = new;
1148                 return new;
1149         } else if (new != NULL)
1150                 pnfs_free_layout_hdr(new);
1151 out_existing:
1152         pnfs_get_layout_hdr(nfsi->layout);
1153         return nfsi->layout;
1154 }
1155
1156 /*
1157  * iomode matching rules:
1158  * iomode       lseg    match
1159  * -----        -----   -----
1160  * ANY          READ    true
1161  * ANY          RW      true
1162  * RW           READ    false
1163  * RW           RW      true
1164  * READ         READ    true
1165  * READ         RW      true
1166  */
1167 static bool
1168 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1169                  const struct pnfs_layout_range *range)
1170 {
1171         struct pnfs_layout_range range1;
1172
1173         if ((range->iomode == IOMODE_RW &&
1174              ls_range->iomode != IOMODE_RW) ||
1175             !pnfs_lseg_range_intersecting(ls_range, range))
1176                 return 0;
1177
1178         /* range1 covers only the first byte in the range */
1179         range1 = *range;
1180         range1.length = 1;
1181         return pnfs_lseg_range_contained(ls_range, &range1);
1182 }
1183
1184 /*
1185  * lookup range in layout
1186  */
1187 static struct pnfs_layout_segment *
1188 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1189                 struct pnfs_layout_range *range)
1190 {
1191         struct pnfs_layout_segment *lseg, *ret = NULL;
1192
1193         dprintk("%s:Begin\n", __func__);
1194
1195         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1196                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1197                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1198                         ret = pnfs_get_lseg(lseg);
1199                         break;
1200                 }
1201                 if (lseg->pls_range.offset > range->offset)
1202                         break;
1203         }
1204
1205         dprintk("%s:Return lseg %p ref %d\n",
1206                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1207         return ret;
1208 }
1209
1210 /*
1211  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1212  * to the MDS or over pNFS
1213  *
1214  * The nfs_inode read_io and write_io fields are cumulative counters reset
1215  * when there are no layout segments. Note that in pnfs_update_layout iomode
1216  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1217  * WRITE request.
1218  *
1219  * A return of true means use MDS I/O.
1220  *
1221  * From rfc 5661:
1222  * If a file's size is smaller than the file size threshold, data accesses
1223  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1224  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1225  * server.  If both file size and I/O size are provided, the client SHOULD
1226  * reach or exceed  both thresholds before sending its read or write
1227  * requests to the data server.
1228  */
1229 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1230                                      struct inode *ino, int iomode)
1231 {
1232         struct nfs4_threshold *t = ctx->mdsthreshold;
1233         struct nfs_inode *nfsi = NFS_I(ino);
1234         loff_t fsize = i_size_read(ino);
1235         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1236
1237         if (t == NULL)
1238                 return ret;
1239
1240         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1241                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1242
1243         switch (iomode) {
1244         case IOMODE_READ:
1245                 if (t->bm & THRESHOLD_RD) {
1246                         dprintk("%s fsize %llu\n", __func__, fsize);
1247                         size_set = true;
1248                         if (fsize < t->rd_sz)
1249                                 size = true;
1250                 }
1251                 if (t->bm & THRESHOLD_RD_IO) {
1252                         dprintk("%s nfsi->read_io %llu\n", __func__,
1253                                 nfsi->read_io);
1254                         io_set = true;
1255                         if (nfsi->read_io < t->rd_io_sz)
1256                                 io = true;
1257                 }
1258                 break;
1259         case IOMODE_RW:
1260                 if (t->bm & THRESHOLD_WR) {
1261                         dprintk("%s fsize %llu\n", __func__, fsize);
1262                         size_set = true;
1263                         if (fsize < t->wr_sz)
1264                                 size = true;
1265                 }
1266                 if (t->bm & THRESHOLD_WR_IO) {
1267                         dprintk("%s nfsi->write_io %llu\n", __func__,
1268                                 nfsi->write_io);
1269                         io_set = true;
1270                         if (nfsi->write_io < t->wr_io_sz)
1271                                 io = true;
1272                 }
1273                 break;
1274         }
1275         if (size_set && io_set) {
1276                 if (size && io)
1277                         ret = true;
1278         } else if (size || io)
1279                 ret = true;
1280
1281         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1282         return ret;
1283 }
1284
1285 /*
1286  * Layout segment is retreived from the server if not cached.
1287  * The appropriate layout segment is referenced and returned to the caller.
1288  */
1289 struct pnfs_layout_segment *
1290 pnfs_update_layout(struct inode *ino,
1291                    struct nfs_open_context *ctx,
1292                    loff_t pos,
1293                    u64 count,
1294                    enum pnfs_iomode iomode,
1295                    gfp_t gfp_flags)
1296 {
1297         struct pnfs_layout_range arg = {
1298                 .iomode = iomode,
1299                 .offset = pos,
1300                 .length = count,
1301         };
1302         unsigned pg_offset;
1303         struct nfs_server *server = NFS_SERVER(ino);
1304         struct nfs_client *clp = server->nfs_client;
1305         struct pnfs_layout_hdr *lo;
1306         struct pnfs_layout_segment *lseg = NULL;
1307         bool first;
1308
1309         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1310                 goto out;
1311
1312         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1313                 goto out;
1314
1315         spin_lock(&ino->i_lock);
1316         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1317         if (lo == NULL) {
1318                 spin_unlock(&ino->i_lock);
1319                 goto out;
1320         }
1321
1322         /* Do we even need to bother with this? */
1323         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1324                 dprintk("%s matches recall, use MDS\n", __func__);
1325                 goto out_unlock;
1326         }
1327
1328         /* if LAYOUTGET already failed once we don't try again */
1329         if (pnfs_layout_io_test_failed(lo, iomode))
1330                 goto out_unlock;
1331
1332         /* Check to see if the layout for the given range already exists */
1333         lseg = pnfs_find_lseg(lo, &arg);
1334         if (lseg)
1335                 goto out_unlock;
1336
1337         if (pnfs_layoutgets_blocked(lo, 0))
1338                 goto out_unlock;
1339         atomic_inc(&lo->plh_outstanding);
1340
1341         first = list_empty(&lo->plh_layouts) ? true : false;
1342         spin_unlock(&ino->i_lock);
1343
1344         if (first) {
1345                 /* The lo must be on the clp list if there is any
1346                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1347                  */
1348                 spin_lock(&clp->cl_lock);
1349                 list_add_tail(&lo->plh_layouts, &server->layouts);
1350                 spin_unlock(&clp->cl_lock);
1351         }
1352
1353         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1354         if (pg_offset) {
1355                 arg.offset -= pg_offset;
1356                 arg.length += pg_offset;
1357         }
1358         if (arg.length != NFS4_MAX_UINT64)
1359                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1360
1361         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1362         atomic_dec(&lo->plh_outstanding);
1363 out_put_layout_hdr:
1364         pnfs_put_layout_hdr(lo);
1365 out:
1366         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1367                         "(%s, offset: %llu, length: %llu)\n",
1368                         __func__, ino->i_sb->s_id,
1369                         (unsigned long long)NFS_FILEID(ino),
1370                         lseg == NULL ? "not found" : "found",
1371                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1372                         (unsigned long long)pos,
1373                         (unsigned long long)count);
1374         return lseg;
1375 out_unlock:
1376         spin_unlock(&ino->i_lock);
1377         goto out_put_layout_hdr;
1378 }
1379 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1380
1381 struct pnfs_layout_segment *
1382 pnfs_layout_process(struct nfs4_layoutget *lgp)
1383 {
1384         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1385         struct nfs4_layoutget_res *res = &lgp->res;
1386         struct pnfs_layout_segment *lseg;
1387         struct inode *ino = lo->plh_inode;
1388         LIST_HEAD(free_me);
1389         int status = 0;
1390
1391         /* Inject layout blob into I/O device driver */
1392         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1393         if (!lseg || IS_ERR(lseg)) {
1394                 if (!lseg)
1395                         status = -ENOMEM;
1396                 else
1397                         status = PTR_ERR(lseg);
1398                 dprintk("%s: Could not allocate layout: error %d\n",
1399                        __func__, status);
1400                 goto out;
1401         }
1402
1403         init_lseg(lo, lseg);
1404         lseg->pls_range = res->range;
1405
1406         spin_lock(&ino->i_lock);
1407         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1408                 dprintk("%s forget reply due to recall\n", __func__);
1409                 goto out_forget_reply;
1410         }
1411
1412         if (pnfs_layoutgets_blocked(lo, 1)) {
1413                 dprintk("%s forget reply due to state\n", __func__);
1414                 goto out_forget_reply;
1415         }
1416
1417         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1418                 /* existing state ID, make sure the sequence number matches. */
1419                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1420                         dprintk("%s forget reply due to sequence\n", __func__);
1421                         goto out_forget_reply;
1422                 }
1423                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1424         } else {
1425                 /*
1426                  * We got an entirely new state ID.  Mark all segments for the
1427                  * inode invalid, and don't bother validating the stateid
1428                  * sequence number.
1429                  */
1430                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1431
1432                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1433                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1434         }
1435
1436         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1437
1438         pnfs_get_lseg(lseg);
1439         pnfs_layout_insert_lseg(lo, lseg);
1440
1441         if (res->return_on_close) {
1442                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1443                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1444         }
1445
1446         spin_unlock(&ino->i_lock);
1447         pnfs_free_lseg_list(&free_me);
1448         return lseg;
1449 out:
1450         return ERR_PTR(status);
1451
1452 out_forget_reply:
1453         spin_unlock(&ino->i_lock);
1454         lseg->pls_layout = lo;
1455         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1456         goto out;
1457 }
1458
1459 void
1460 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1461 {
1462         u64 rd_size = req->wb_bytes;
1463
1464         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1465
1466         if (pgio->pg_dreq == NULL)
1467                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1468         else
1469                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1470
1471         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1472                                            req->wb_context,
1473                                            req_offset(req),
1474                                            rd_size,
1475                                            IOMODE_READ,
1476                                            GFP_KERNEL);
1477         /* If no lseg, fall back to read through mds */
1478         if (pgio->pg_lseg == NULL)
1479                 nfs_pageio_reset_read_mds(pgio);
1480
1481 }
1482 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1483
1484 void
1485 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1486                            struct nfs_page *req, u64 wb_size)
1487 {
1488         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1489
1490         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1491                                            req->wb_context,
1492                                            req_offset(req),
1493                                            wb_size,
1494                                            IOMODE_RW,
1495                                            GFP_NOFS);
1496         /* If no lseg, fall back to write through mds */
1497         if (pgio->pg_lseg == NULL)
1498                 nfs_pageio_reset_write_mds(pgio);
1499 }
1500 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1501
1502 /*
1503  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1504  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1505  */
1506 size_t
1507 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1508                      struct nfs_page *req)
1509 {
1510         unsigned int size;
1511         u64 seg_end, req_start, seg_left;
1512
1513         size = nfs_generic_pg_test(pgio, prev, req);
1514         if (!size)
1515                 return 0;
1516
1517         /*
1518          * 'size' contains the number of bytes left in the current page (up
1519          * to the original size asked for in @req->wb_bytes).
1520          *
1521          * Calculate how many bytes are left in the layout segment
1522          * and if there are less bytes than 'size', return that instead.
1523          *
1524          * Please also note that 'end_offset' is actually the offset of the
1525          * first byte that lies outside the pnfs_layout_range. FIXME?
1526          *
1527          */
1528         if (pgio->pg_lseg) {
1529                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1530                                      pgio->pg_lseg->pls_range.length);
1531                 req_start = req_offset(req);
1532                 WARN_ON_ONCE(req_start > seg_end);
1533                 /* start of request is past the last byte of this segment */
1534                 if (req_start >= seg_end)
1535                         return 0;
1536
1537                 /* adjust 'size' iff there are fewer bytes left in the
1538                  * segment than what nfs_generic_pg_test returned */
1539                 seg_left = seg_end - req_start;
1540                 if (seg_left < size)
1541                         size = (unsigned int)seg_left;
1542         }
1543
1544         return size;
1545 }
1546 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1547
1548 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1549 {
1550         struct nfs_pageio_descriptor pgio;
1551
1552         /* Resend all requests through the MDS */
1553         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1554                               hdr->completion_ops);
1555         return nfs_pageio_resend(&pgio, hdr);
1556 }
1557 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1558
1559 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1560 {
1561
1562         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1563         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1564             PNFS_LAYOUTRET_ON_ERROR) {
1565                 pnfs_return_layout(hdr->inode);
1566         }
1567         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1568                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1569 }
1570
1571 /*
1572  * Called by non rpc-based layout drivers
1573  */
1574 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1575 {
1576         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1577         if (!hdr->pnfs_error) {
1578                 pnfs_set_layoutcommit(hdr);
1579                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1580         } else
1581                 pnfs_ld_handle_write_error(hdr);
1582         hdr->mds_ops->rpc_release(hdr);
1583 }
1584 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1585
1586 static void
1587 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1588                 struct nfs_pgio_header *hdr)
1589 {
1590         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1591                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1592                 nfs_pageio_reset_write_mds(desc);
1593                 desc->pg_recoalesce = 1;
1594         }
1595         nfs_pgio_data_destroy(hdr);
1596 }
1597
1598 static enum pnfs_try_status
1599 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1600                         const struct rpc_call_ops *call_ops,
1601                         struct pnfs_layout_segment *lseg,
1602                         int how)
1603 {
1604         struct inode *inode = hdr->inode;
1605         enum pnfs_try_status trypnfs;
1606         struct nfs_server *nfss = NFS_SERVER(inode);
1607
1608         hdr->mds_ops = call_ops;
1609
1610         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1611                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1612         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1613         if (trypnfs != PNFS_NOT_ATTEMPTED)
1614                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1615         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1616         return trypnfs;
1617 }
1618
1619 static void
1620 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1621               struct nfs_pgio_header *hdr, int how)
1622 {
1623         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1624         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1625         enum pnfs_try_status trypnfs;
1626
1627         desc->pg_lseg = NULL;
1628         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1629         if (trypnfs == PNFS_NOT_ATTEMPTED)
1630                 pnfs_write_through_mds(desc, hdr);
1631         pnfs_put_lseg(lseg);
1632 }
1633
1634 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1635 {
1636         pnfs_put_lseg(hdr->lseg);
1637         nfs_pgio_header_free(hdr);
1638 }
1639 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1640
1641 int
1642 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1643 {
1644         struct nfs_pgio_header *hdr;
1645         int ret;
1646
1647         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1648         if (!hdr) {
1649                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1650                 pnfs_put_lseg(desc->pg_lseg);
1651                 desc->pg_lseg = NULL;
1652                 return -ENOMEM;
1653         }
1654         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1655         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1656         ret = nfs_generic_pgio(desc, hdr);
1657         if (ret != 0) {
1658                 pnfs_put_lseg(desc->pg_lseg);
1659                 desc->pg_lseg = NULL;
1660         } else
1661                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1662         return ret;
1663 }
1664 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1665
1666 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1667 {
1668         struct nfs_pageio_descriptor pgio;
1669
1670         /* Resend all requests through the MDS */
1671         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1672         return nfs_pageio_resend(&pgio, hdr);
1673 }
1674 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1675
1676 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1677 {
1678         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1679         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1680             PNFS_LAYOUTRET_ON_ERROR) {
1681                 pnfs_return_layout(hdr->inode);
1682         }
1683         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1684                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1685 }
1686
1687 /*
1688  * Called by non rpc-based layout drivers
1689  */
1690 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1691 {
1692         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1693         if (likely(!hdr->pnfs_error)) {
1694                 __nfs4_read_done_cb(hdr);
1695                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1696         } else
1697                 pnfs_ld_handle_read_error(hdr);
1698         hdr->mds_ops->rpc_release(hdr);
1699 }
1700 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1701
1702 static void
1703 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1704                 struct nfs_pgio_header *hdr)
1705 {
1706         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1707                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1708                 nfs_pageio_reset_read_mds(desc);
1709                 desc->pg_recoalesce = 1;
1710         }
1711         nfs_pgio_data_destroy(hdr);
1712 }
1713
1714 /*
1715  * Call the appropriate parallel I/O subsystem read function.
1716  */
1717 static enum pnfs_try_status
1718 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1719                        const struct rpc_call_ops *call_ops,
1720                        struct pnfs_layout_segment *lseg)
1721 {
1722         struct inode *inode = hdr->inode;
1723         struct nfs_server *nfss = NFS_SERVER(inode);
1724         enum pnfs_try_status trypnfs;
1725
1726         hdr->mds_ops = call_ops;
1727
1728         dprintk("%s: Reading ino:%lu %u@%llu\n",
1729                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1730
1731         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1732         if (trypnfs != PNFS_NOT_ATTEMPTED)
1733                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1734         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1735         return trypnfs;
1736 }
1737
1738 static void
1739 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1740 {
1741         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1742         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1743         enum pnfs_try_status trypnfs;
1744
1745         desc->pg_lseg = NULL;
1746         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1747         if (trypnfs == PNFS_NOT_ATTEMPTED)
1748                 pnfs_read_through_mds(desc, hdr);
1749         pnfs_put_lseg(lseg);
1750 }
1751
1752 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1753 {
1754         pnfs_put_lseg(hdr->lseg);
1755         nfs_pgio_header_free(hdr);
1756 }
1757 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1758
1759 int
1760 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1761 {
1762         struct nfs_pgio_header *hdr;
1763         int ret;
1764
1765         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1766         if (!hdr) {
1767                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1768                 ret = -ENOMEM;
1769                 pnfs_put_lseg(desc->pg_lseg);
1770                 desc->pg_lseg = NULL;
1771                 return ret;
1772         }
1773         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1774         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1775         ret = nfs_generic_pgio(desc, hdr);
1776         if (ret != 0) {
1777                 pnfs_put_lseg(desc->pg_lseg);
1778                 desc->pg_lseg = NULL;
1779         } else
1780                 pnfs_do_read(desc, hdr);
1781         return ret;
1782 }
1783 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1784
1785 static void pnfs_clear_layoutcommitting(struct inode *inode)
1786 {
1787         unsigned long *bitlock = &NFS_I(inode)->flags;
1788
1789         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1790         smp_mb__after_atomic();
1791         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1792 }
1793
1794 /*
1795  * There can be multiple RW segments.
1796  */
1797 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1798 {
1799         struct pnfs_layout_segment *lseg;
1800
1801         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1802                 if (lseg->pls_range.iomode == IOMODE_RW &&
1803                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1804                         list_add(&lseg->pls_lc_list, listp);
1805         }
1806 }
1807
1808 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1809 {
1810         struct pnfs_layout_segment *lseg, *tmp;
1811
1812         /* Matched by references in pnfs_set_layoutcommit */
1813         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1814                 list_del_init(&lseg->pls_lc_list);
1815                 pnfs_put_lseg(lseg);
1816         }
1817
1818         pnfs_clear_layoutcommitting(inode);
1819 }
1820
1821 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1822 {
1823         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1824 }
1825 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1826
1827 void
1828 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1829 {
1830         struct inode *inode = hdr->inode;
1831         struct nfs_inode *nfsi = NFS_I(inode);
1832         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1833         bool mark_as_dirty = false;
1834
1835         spin_lock(&inode->i_lock);
1836         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1837                 mark_as_dirty = true;
1838                 dprintk("%s: Set layoutcommit for inode %lu ",
1839                         __func__, inode->i_ino);
1840         }
1841         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1842                 /* references matched in nfs4_layoutcommit_release */
1843                 pnfs_get_lseg(hdr->lseg);
1844         }
1845         if (end_pos > nfsi->layout->plh_lwb)
1846                 nfsi->layout->plh_lwb = end_pos;
1847         spin_unlock(&inode->i_lock);
1848         dprintk("%s: lseg %p end_pos %llu\n",
1849                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1850
1851         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1852          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1853         if (mark_as_dirty)
1854                 mark_inode_dirty_sync(inode);
1855 }
1856 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1857
1858 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
1859 {
1860         struct inode *inode = data->inode;
1861         struct nfs_inode *nfsi = NFS_I(inode);
1862         bool mark_as_dirty = false;
1863
1864         spin_lock(&inode->i_lock);
1865         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1866                 mark_as_dirty = true;
1867                 dprintk("%s: Set layoutcommit for inode %lu ",
1868                         __func__, inode->i_ino);
1869         }
1870         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
1871                 /* references matched in nfs4_layoutcommit_release */
1872                 pnfs_get_lseg(data->lseg);
1873         }
1874         if (data->lwb > nfsi->layout->plh_lwb)
1875                 nfsi->layout->plh_lwb = data->lwb;
1876         spin_unlock(&inode->i_lock);
1877         dprintk("%s: lseg %p end_pos %llu\n",
1878                 __func__, data->lseg, nfsi->layout->plh_lwb);
1879
1880         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1881          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1882         if (mark_as_dirty)
1883                 mark_inode_dirty_sync(inode);
1884 }
1885 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
1886
1887 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1888 {
1889         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1890
1891         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1892                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1893         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1894 }
1895
1896 /*
1897  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1898  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1899  * data to disk to allow the server to recover the data if it crashes.
1900  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1901  * is off, and a COMMIT is sent to a data server, or
1902  * if WRITEs to a data server return NFS_DATA_SYNC.
1903  */
1904 int
1905 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1906 {
1907         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1908         struct nfs4_layoutcommit_data *data;
1909         struct nfs_inode *nfsi = NFS_I(inode);
1910         loff_t end_pos;
1911         int status;
1912
1913         if (!pnfs_layoutcommit_outstanding(inode))
1914                 return 0;
1915
1916         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1917
1918         status = -EAGAIN;
1919         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1920                 if (!sync)
1921                         goto out;
1922                 status = wait_on_bit_lock_action(&nfsi->flags,
1923                                 NFS_INO_LAYOUTCOMMITTING,
1924                                 nfs_wait_bit_killable,
1925                                 TASK_KILLABLE);
1926                 if (status)
1927                         goto out;
1928         }
1929
1930         status = -ENOMEM;
1931         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1932         data = kzalloc(sizeof(*data), GFP_NOFS);
1933         if (!data)
1934                 goto clear_layoutcommitting;
1935
1936         status = 0;
1937         spin_lock(&inode->i_lock);
1938         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1939                 goto out_unlock;
1940
1941         INIT_LIST_HEAD(&data->lseg_list);
1942         pnfs_list_write_lseg(inode, &data->lseg_list);
1943
1944         end_pos = nfsi->layout->plh_lwb;
1945         nfsi->layout->plh_lwb = 0;
1946
1947         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1948         spin_unlock(&inode->i_lock);
1949
1950         data->args.inode = inode;
1951         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1952         nfs_fattr_init(&data->fattr);
1953         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1954         data->res.fattr = &data->fattr;
1955         data->args.lastbytewritten = end_pos - 1;
1956         data->res.server = NFS_SERVER(inode);
1957
1958         if (ld->prepare_layoutcommit) {
1959                 status = ld->prepare_layoutcommit(&data->args);
1960                 if (status) {
1961                         spin_lock(&inode->i_lock);
1962                         if (end_pos < nfsi->layout->plh_lwb)
1963                                 nfsi->layout->plh_lwb = end_pos;
1964                         spin_unlock(&inode->i_lock);
1965                         put_rpccred(data->cred);
1966                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
1967                         goto clear_layoutcommitting;
1968                 }
1969         }
1970
1971
1972         status = nfs4_proc_layoutcommit(data, sync);
1973 out:
1974         if (status)
1975                 mark_inode_dirty_sync(inode);
1976         dprintk("<-- %s status %d\n", __func__, status);
1977         return status;
1978 out_unlock:
1979         spin_unlock(&inode->i_lock);
1980         kfree(data);
1981 clear_layoutcommitting:
1982         pnfs_clear_layoutcommitting(inode);
1983         goto out;
1984 }
1985
1986 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1987 {
1988         struct nfs4_threshold *thp;
1989
1990         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1991         if (!thp) {
1992                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1993                 return NULL;
1994         }
1995         return thp;
1996 }