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