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 static bool
 744 pnfs_layout_returning(const struct pnfs_layout_hdr *lo,
 745                       struct pnfs_layout_range *range)
 746 {
 747         return test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 748                 (lo->plh_return_iomode == IOMODE_ANY ||
 749                  lo->plh_return_iomode == range->iomode);
 750 }
 751
 752 /* lget is set to 1 if called from inside send_layoutget call chain */
 753 static bool
 754 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo,
 755                         struct pnfs_layout_range *range, int lget)
 756 {
 757         return lo->plh_block_lgets ||
 758                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 759                 (list_empty(&lo->plh_segs) &&
 760                  (atomic_read(&lo->plh_outstanding) > lget)) ||
 761                 pnfs_layout_returning(lo, range);
 762 }
 763
 764 int
 765 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 766                               struct pnfs_layout_range *range,
 767                               struct nfs4_state *open_state)
 768 {
 769         int status = 0;
 770
 771         dprintk("--> %s\n", __func__);
 772         spin_lock(&lo->plh_inode->i_lock);
 773         if (pnfs_layoutgets_blocked(lo, range, 1)) {
 774                 status = -EAGAIN;
 775         } else if (!nfs4_valid_open_stateid(open_state)) {
 776                 status = -EBADF;
 777         } else if (list_empty(&lo->plh_segs) ||
 778                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 779                 int seq;
 780
 781                 do {
 782                         seq = read_seqbegin(&open_state->seqlock);
 783                         nfs4_stateid_copy(dst, &open_state->stateid);
 784                 } while (read_seqretry(&open_state->seqlock, seq));
 785         } else
 786                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 787         spin_unlock(&lo->plh_inode->i_lock);
 788         dprintk("<-- %s\n", __func__);
 789         return status;
 790 }
 791
 792 /*
 793 * Get layout from server.
 794 *    for now, assume that whole file layouts are requested.
 795 *    arg->offset: 0
 796 *    arg->length: all ones
 797 */
 798 static struct pnfs_layout_segment *
 799 send_layoutget(struct pnfs_layout_hdr *lo,
 800            struct nfs_open_context *ctx,
 801            struct pnfs_layout_range *range,
 802            gfp_t gfp_flags)
 803 {
 804         struct inode *ino = lo->plh_inode;
 805         struct nfs_server *server = NFS_SERVER(ino);
 806         struct nfs4_layoutget *lgp;
 807         struct pnfs_layout_segment *lseg;
 808
 809         dprintk("--> %s\n", __func__);
 810
 811         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 812         if (lgp == NULL)
 813                 return NULL;
 814
 815         lgp->args.minlength = PAGE_CACHE_SIZE;
 816         if (lgp->args.minlength > range->length)
 817                 lgp->args.minlength = range->length;
 818         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 819         lgp->args.range = *range;
 820         lgp->args.type = server->pnfs_curr_ld->id;
 821         lgp->args.inode = ino;
 822         lgp->args.ctx = get_nfs_open_context(ctx);
 823         lgp->gfp_flags = gfp_flags;
 824         lgp->cred = lo->plh_lc_cred;
 825
 826         /* Synchronously retrieve layout information from server and
 827          * store in lseg.
 828          */
 829         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 830         if (IS_ERR(lseg)) {
 831                 switch (PTR_ERR(lseg)) {
 832                 case -ENOMEM:
 833                 case -ERESTARTSYS:
 834                         break;
 835                 default:
 836                         /* remember that LAYOUTGET failed and suspend trying */
 837                         pnfs_layout_io_set_failed(lo, range->iomode);
 838                 }
 839                 return NULL;
 840         }
 841
 842         return lseg;
 843 }
 844
 845 static void pnfs_clear_layoutcommit(struct inode *inode,
 846                 struct list_head *head)
 847 {
 848         struct nfs_inode *nfsi = NFS_I(inode);
 849         struct pnfs_layout_segment *lseg, *tmp;
 850
 851         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 852                 return;
 853         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 854                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 855                         continue;
 856                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 857         }
 858 }
 859
 860 static int
 861 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 862                        enum pnfs_iomode iomode)
 863 {
 864         struct inode *ino = lo->plh_inode;
 865         struct nfs4_layoutreturn *lrp;
 866         int status = 0;
 867
 868         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 869         if (unlikely(lrp == NULL)) {
 870                 status = -ENOMEM;
 871                 spin_lock(&ino->i_lock);
 872                 lo->plh_block_lgets--;
 873                 spin_unlock(&ino->i_lock);
 874                 pnfs_put_layout_hdr(lo);
 875                 goto out;
 876         }
 877
 878         lrp->args.stateid = stateid;
 879         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 880         lrp->args.inode = ino;
 881         lrp->args.iomode = iomode;
 882         lrp->args.layout = lo;
 883         lrp->clp = NFS_SERVER(ino)->nfs_client;
 884         lrp->cred = lo->plh_lc_cred;
 885
 886         status = nfs4_proc_layoutreturn(lrp);
 887 out:
 888         dprintk("<-- %s status: %d\n", __func__, status);
 889         return status;
 890 }
 891
 892 /*
 893  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 894  * when the layout segment list is empty.
 895  *
 896  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 897  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 898  * deviceid is marked invalid.
 899  */
 900 int
 901 _pnfs_return_layout(struct inode *ino)
 902 {
 903         struct pnfs_layout_hdr *lo = NULL;
 904         struct nfs_inode *nfsi = NFS_I(ino);
 905         LIST_HEAD(tmp_list);
 906         nfs4_stateid stateid;
 907         int status = 0, empty;
 908
 909         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 910
 911         spin_lock(&ino->i_lock);
 912         lo = nfsi->layout;
 913         if (!lo) {
 914                 spin_unlock(&ino->i_lock);
 915                 dprintk("NFS: %s no layout to return\n", __func__);
 916                 goto out;
 917         }
 918         stateid = nfsi->layout->plh_stateid;
 919         /* Reference matched in nfs4_layoutreturn_release */
 920         pnfs_get_layout_hdr(lo);
 921         empty = list_empty(&lo->plh_segs);
 922         pnfs_clear_layoutcommit(ino, &tmp_list);
 923         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 924
 925         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 926                 struct pnfs_layout_range range = {
 927                         .iomode         = IOMODE_ANY,
 928                         .offset         = 0,
 929                         .length         = NFS4_MAX_UINT64,
 930                 };
 931                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
 932         }
 933
 934         /* Don't send a LAYOUTRETURN if list was initially empty */
 935         if (empty) {
 936                 spin_unlock(&ino->i_lock);
 937                 pnfs_put_layout_hdr(lo);
 938                 dprintk("NFS: %s no layout segments to return\n", __func__);
 939                 goto out;
 940         }
 941
 942         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 943         lo->plh_block_lgets++;
 944         spin_unlock(&ino->i_lock);
 945         pnfs_free_lseg_list(&tmp_list);
 946
 947         status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY);
 948 out:
 949         dprintk("<-- %s status: %d\n", __func__, status);
 950         return status;
 951 }
 952 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 953
 954 int
 955 pnfs_commit_and_return_layout(struct inode *inode)
 956 {
 957         struct pnfs_layout_hdr *lo;
 958         int ret;
 959
 960         spin_lock(&inode->i_lock);
 961         lo = NFS_I(inode)->layout;
 962         if (lo == NULL) {
 963                 spin_unlock(&inode->i_lock);
 964                 return 0;
 965         }
 966         pnfs_get_layout_hdr(lo);
 967         /* Block new layoutgets and read/write to ds */
 968         lo->plh_block_lgets++;
 969         spin_unlock(&inode->i_lock);
 970         filemap_fdatawait(inode->i_mapping);
 971         ret = pnfs_layoutcommit_inode(inode, true);
 972         if (ret == 0)
 973                 ret = _pnfs_return_layout(inode);
 974         spin_lock(&inode->i_lock);
 975         lo->plh_block_lgets--;
 976         spin_unlock(&inode->i_lock);
 977         pnfs_put_layout_hdr(lo);
 978         return ret;
 979 }
 980
 981 bool pnfs_roc(struct inode *ino)
 982 {
 983         struct pnfs_layout_hdr *lo;
 984         struct pnfs_layout_segment *lseg, *tmp;
 985         LIST_HEAD(tmp_list);
 986         bool found = false;
 987
 988         spin_lock(&ino->i_lock);
 989         lo = NFS_I(ino)->layout;
 990         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 991             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 992                 goto out_nolayout;
 993         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 994                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 995                         mark_lseg_invalid(lseg, &tmp_list);
 996                         found = true;
 997                 }
 998         if (!found)
 999                 goto out_nolayout;
1000         lo->plh_block_lgets++;
1001         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
1002         spin_unlock(&ino->i_lock);
1003         pnfs_free_lseg_list(&tmp_list);
1004         return true;
1005
1006 out_nolayout:
1007         spin_unlock(&ino->i_lock);
1008         return false;
1009 }
1010
1011 void pnfs_roc_release(struct inode *ino)
1012 {
1013         struct pnfs_layout_hdr *lo;
1014
1015         spin_lock(&ino->i_lock);
1016         lo = NFS_I(ino)->layout;
1017         lo->plh_block_lgets--;
1018         if (atomic_dec_and_test(&lo->plh_refcount)) {
1019                 pnfs_detach_layout_hdr(lo);
1020                 spin_unlock(&ino->i_lock);
1021                 pnfs_free_layout_hdr(lo);
1022         } else
1023                 spin_unlock(&ino->i_lock);
1024 }
1025
1026 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1027 {
1028         struct pnfs_layout_hdr *lo;
1029
1030         spin_lock(&ino->i_lock);
1031         lo = NFS_I(ino)->layout;
1032         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1033                 lo->plh_barrier = barrier;
1034         spin_unlock(&ino->i_lock);
1035 }
1036
1037 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
1038 {
1039         struct nfs_inode *nfsi = NFS_I(ino);
1040         struct pnfs_layout_hdr *lo;
1041         struct pnfs_layout_segment *lseg;
1042         u32 current_seqid;
1043         bool found = false;
1044
1045         spin_lock(&ino->i_lock);
1046         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1047                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1048                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1049                         found = true;
1050                         goto out;
1051                 }
1052         lo = nfsi->layout;
1053         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1054
1055         /* Since close does not return a layout stateid for use as
1056          * a barrier, we choose the worst-case barrier.
1057          */
1058         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1059 out:
1060         spin_unlock(&ino->i_lock);
1061         return found;
1062 }
1063
1064 /*
1065  * Compare two layout segments for sorting into layout cache.
1066  * We want to preferentially return RW over RO layouts, so ensure those
1067  * are seen first.
1068  */
1069 static s64
1070 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1071            const struct pnfs_layout_range *l2)
1072 {
1073         s64 d;
1074
1075         /* high offset > low offset */
1076         d = l1->offset - l2->offset;
1077         if (d)
1078                 return d;
1079
1080         /* short length > long length */
1081         d = l2->length - l1->length;
1082         if (d)
1083                 return d;
1084
1085         /* read > read/write */
1086         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1087 }
1088
1089 static void
1090 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1091                    struct pnfs_layout_segment *lseg)
1092 {
1093         struct pnfs_layout_segment *lp;
1094
1095         dprintk("%s:Begin\n", __func__);
1096
1097         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1098                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1099                         continue;
1100                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1101                 dprintk("%s: inserted lseg %p "
1102                         "iomode %d offset %llu length %llu before "
1103                         "lp %p iomode %d offset %llu length %llu\n",
1104                         __func__, lseg, lseg->pls_range.iomode,
1105                         lseg->pls_range.offset, lseg->pls_range.length,
1106                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1107                         lp->pls_range.length);
1108                 goto out;
1109         }
1110         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1111         dprintk("%s: inserted lseg %p "
1112                 "iomode %d offset %llu length %llu at tail\n",
1113                 __func__, lseg, lseg->pls_range.iomode,
1114                 lseg->pls_range.offset, lseg->pls_range.length);
1115 out:
1116         pnfs_get_layout_hdr(lo);
1117
1118         dprintk("%s:Return\n", __func__);
1119 }
1120
1121 static struct pnfs_layout_hdr *
1122 alloc_init_layout_hdr(struct inode *ino,
1123                       struct nfs_open_context *ctx,
1124                       gfp_t gfp_flags)
1125 {
1126         struct pnfs_layout_hdr *lo;
1127
1128         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1129         if (!lo)
1130                 return NULL;
1131         atomic_set(&lo->plh_refcount, 1);
1132         INIT_LIST_HEAD(&lo->plh_layouts);
1133         INIT_LIST_HEAD(&lo->plh_segs);
1134         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1135         lo->plh_inode = ino;
1136         lo->plh_lc_cred = get_rpccred(ctx->cred);
1137         return lo;
1138 }
1139
1140 static struct pnfs_layout_hdr *
1141 pnfs_find_alloc_layout(struct inode *ino,
1142                        struct nfs_open_context *ctx,
1143                        gfp_t gfp_flags)
1144 {
1145         struct nfs_inode *nfsi = NFS_I(ino);
1146         struct pnfs_layout_hdr *new = NULL;
1147
1148         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1149
1150         if (nfsi->layout != NULL)
1151                 goto out_existing;
1152         spin_unlock(&ino->i_lock);
1153         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1154         spin_lock(&ino->i_lock);
1155
1156         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1157                 nfsi->layout = new;
1158                 return new;
1159         } else if (new != NULL)
1160                 pnfs_free_layout_hdr(new);
1161 out_existing:
1162         pnfs_get_layout_hdr(nfsi->layout);
1163         return nfsi->layout;
1164 }
1165
1166 /*
1167  * iomode matching rules:
1168  * iomode       lseg    match
1169  * -----        -----   -----
1170  * ANY          READ    true
1171  * ANY          RW      true
1172  * RW           READ    false
1173  * RW           RW      true
1174  * READ         READ    true
1175  * READ         RW      true
1176  */
1177 static bool
1178 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1179                  const struct pnfs_layout_range *range)
1180 {
1181         struct pnfs_layout_range range1;
1182
1183         if ((range->iomode == IOMODE_RW &&
1184              ls_range->iomode != IOMODE_RW) ||
1185             !pnfs_lseg_range_intersecting(ls_range, range))
1186                 return 0;
1187
1188         /* range1 covers only the first byte in the range */
1189         range1 = *range;
1190         range1.length = 1;
1191         return pnfs_lseg_range_contained(ls_range, &range1);
1192 }
1193
1194 /*
1195  * lookup range in layout
1196  */
1197 static struct pnfs_layout_segment *
1198 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1199                 struct pnfs_layout_range *range)
1200 {
1201         struct pnfs_layout_segment *lseg, *ret = NULL;
1202
1203         dprintk("%s:Begin\n", __func__);
1204
1205         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1206                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1207                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1208                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1209                         ret = pnfs_get_lseg(lseg);
1210                         break;
1211                 }
1212                 if (lseg->pls_range.offset > range->offset)
1213                         break;
1214         }
1215
1216         dprintk("%s:Return lseg %p ref %d\n",
1217                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1218         return ret;
1219 }
1220
1221 /*
1222  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1223  * to the MDS or over pNFS
1224  *
1225  * The nfs_inode read_io and write_io fields are cumulative counters reset
1226  * when there are no layout segments. Note that in pnfs_update_layout iomode
1227  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1228  * WRITE request.
1229  *
1230  * A return of true means use MDS I/O.
1231  *
1232  * From rfc 5661:
1233  * If a file's size is smaller than the file size threshold, data accesses
1234  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1235  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1236  * server.  If both file size and I/O size are provided, the client SHOULD
1237  * reach or exceed  both thresholds before sending its read or write
1238  * requests to the data server.
1239  */
1240 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1241                                      struct inode *ino, int iomode)
1242 {
1243         struct nfs4_threshold *t = ctx->mdsthreshold;
1244         struct nfs_inode *nfsi = NFS_I(ino);
1245         loff_t fsize = i_size_read(ino);
1246         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1247
1248         if (t == NULL)
1249                 return ret;
1250
1251         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1252                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1253
1254         switch (iomode) {
1255         case IOMODE_READ:
1256                 if (t->bm & THRESHOLD_RD) {
1257                         dprintk("%s fsize %llu\n", __func__, fsize);
1258                         size_set = true;
1259                         if (fsize < t->rd_sz)
1260                                 size = true;
1261                 }
1262                 if (t->bm & THRESHOLD_RD_IO) {
1263                         dprintk("%s nfsi->read_io %llu\n", __func__,
1264                                 nfsi->read_io);
1265                         io_set = true;
1266                         if (nfsi->read_io < t->rd_io_sz)
1267                                 io = true;
1268                 }
1269                 break;
1270         case IOMODE_RW:
1271                 if (t->bm & THRESHOLD_WR) {
1272                         dprintk("%s fsize %llu\n", __func__, fsize);
1273                         size_set = true;
1274                         if (fsize < t->wr_sz)
1275                                 size = true;
1276                 }
1277                 if (t->bm & THRESHOLD_WR_IO) {
1278                         dprintk("%s nfsi->write_io %llu\n", __func__,
1279                                 nfsi->write_io);
1280                         io_set = true;
1281                         if (nfsi->write_io < t->wr_io_sz)
1282                                 io = true;
1283                 }
1284                 break;
1285         }
1286         if (size_set && io_set) {
1287                 if (size && io)
1288                         ret = true;
1289         } else if (size || io)
1290                 ret = true;
1291
1292         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1293         return ret;
1294 }
1295
1296 /*
1297  * Layout segment is retreived from the server if not cached.
1298  * The appropriate layout segment is referenced and returned to the caller.
1299  */
1300 struct pnfs_layout_segment *
1301 pnfs_update_layout(struct inode *ino,
1302                    struct nfs_open_context *ctx,
1303                    loff_t pos,
1304                    u64 count,
1305                    enum pnfs_iomode iomode,
1306                    gfp_t gfp_flags)
1307 {
1308         struct pnfs_layout_range arg = {
1309                 .iomode = iomode,
1310                 .offset = pos,
1311                 .length = count,
1312         };
1313         unsigned pg_offset;
1314         struct nfs_server *server = NFS_SERVER(ino);
1315         struct nfs_client *clp = server->nfs_client;
1316         struct pnfs_layout_hdr *lo;
1317         struct pnfs_layout_segment *lseg = NULL;
1318         bool first;
1319
1320         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1321                 goto out;
1322
1323         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1324                 goto out;
1325
1326 lookup_again:
1327         first = false;
1328         spin_lock(&ino->i_lock);
1329         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1330         if (lo == NULL) {
1331                 spin_unlock(&ino->i_lock);
1332                 goto out;
1333         }
1334
1335         /* Do we even need to bother with this? */
1336         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1337                 dprintk("%s matches recall, use MDS\n", __func__);
1338                 goto out_unlock;
1339         }
1340
1341         /* if LAYOUTGET already failed once we don't try again */
1342         if (pnfs_layout_io_test_failed(lo, iomode))
1343                 goto out_unlock;
1344
1345         first = list_empty(&lo->plh_segs);
1346         if (first) {
1347                 /* The first layoutget for the file. Need to serialize per
1348                  * RFC 5661 Errata 3208.
1349                  */
1350                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1351                                      &lo->plh_flags)) {
1352                         spin_unlock(&ino->i_lock);
1353                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1354                                     TASK_UNINTERRUPTIBLE);
1355                         pnfs_put_layout_hdr(lo);
1356                         goto lookup_again;
1357                 }
1358         } else {
1359                 /* Check to see if the layout for the given range
1360                  * already exists
1361                  */
1362                 lseg = pnfs_find_lseg(lo, &arg);
1363                 if (lseg)
1364                         goto out_unlock;
1365         }
1366
1367         if (pnfs_layoutgets_blocked(lo, &arg, 0))
1368                 goto out_unlock;
1369         atomic_inc(&lo->plh_outstanding);
1370         spin_unlock(&ino->i_lock);
1371
1372         if (list_empty(&lo->plh_layouts)) {
1373                 /* The lo must be on the clp list if there is any
1374                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1375                  */
1376                 spin_lock(&clp->cl_lock);
1377                 if (list_empty(&lo->plh_layouts))
1378                         list_add_tail(&lo->plh_layouts, &server->layouts);
1379                 spin_unlock(&clp->cl_lock);
1380         }
1381
1382         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1383         if (pg_offset) {
1384                 arg.offset -= pg_offset;
1385                 arg.length += pg_offset;
1386         }
1387         if (arg.length != NFS4_MAX_UINT64)
1388                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1389
1390         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1391         atomic_dec(&lo->plh_outstanding);
1392 out_put_layout_hdr:
1393         if (first) {
1394                 unsigned long *bitlock = &lo->plh_flags;
1395
1396                 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1397                 smp_mb__after_atomic();
1398                 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1399         }
1400         pnfs_put_layout_hdr(lo);
1401 out:
1402         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1403                         "(%s, offset: %llu, length: %llu)\n",
1404                         __func__, ino->i_sb->s_id,
1405                         (unsigned long long)NFS_FILEID(ino),
1406                         lseg == NULL ? "not found" : "found",
1407                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1408                         (unsigned long long)pos,
1409                         (unsigned long long)count);
1410         return lseg;
1411 out_unlock:
1412         spin_unlock(&ino->i_lock);
1413         goto out_put_layout_hdr;
1414 }
1415 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1416
1417 struct pnfs_layout_segment *
1418 pnfs_layout_process(struct nfs4_layoutget *lgp)
1419 {
1420         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1421         struct nfs4_layoutget_res *res = &lgp->res;
1422         struct pnfs_layout_segment *lseg;
1423         struct inode *ino = lo->plh_inode;
1424         LIST_HEAD(free_me);
1425         int status = 0;
1426
1427         /* Inject layout blob into I/O device driver */
1428         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1429         if (!lseg || IS_ERR(lseg)) {
1430                 if (!lseg)
1431                         status = -ENOMEM;
1432                 else
1433                         status = PTR_ERR(lseg);
1434                 dprintk("%s: Could not allocate layout: error %d\n",
1435                        __func__, status);
1436                 goto out;
1437         }
1438
1439         init_lseg(lo, lseg);
1440         lseg->pls_range = res->range;
1441
1442         spin_lock(&ino->i_lock);
1443         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1444                 dprintk("%s forget reply due to recall\n", __func__);
1445                 goto out_forget_reply;
1446         }
1447
1448         if (pnfs_layoutgets_blocked(lo, &lgp->args.range, 1)) {
1449                 dprintk("%s forget reply due to state\n", __func__);
1450                 goto out_forget_reply;
1451         }
1452
1453         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1454                 /* existing state ID, make sure the sequence number matches. */
1455                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1456                         dprintk("%s forget reply due to sequence\n", __func__);
1457                         goto out_forget_reply;
1458                 }
1459                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1460         } else {
1461                 /*
1462                  * We got an entirely new state ID.  Mark all segments for the
1463                  * inode invalid, and don't bother validating the stateid
1464                  * sequence number.
1465                  */
1466                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1467
1468                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1469                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1470         }
1471
1472         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1473
1474         pnfs_get_lseg(lseg);
1475         pnfs_layout_insert_lseg(lo, lseg);
1476
1477         if (res->return_on_close) {
1478                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1479                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1480         }
1481
1482         spin_unlock(&ino->i_lock);
1483         pnfs_free_lseg_list(&free_me);
1484         return lseg;
1485 out:
1486         return ERR_PTR(status);
1487
1488 out_forget_reply:
1489         spin_unlock(&ino->i_lock);
1490         lseg->pls_layout = lo;
1491         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1492         goto out;
1493 }
1494
1495 static void
1496 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1497                                 struct list_head *tmp_list,
1498                                 struct pnfs_layout_range *return_range)
1499 {
1500         struct pnfs_layout_segment *lseg, *next;
1501
1502         dprintk("%s:Begin lo %p\n", __func__, lo);
1503
1504         if (list_empty(&lo->plh_segs))
1505                 return;
1506
1507         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1508                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1509                         dprintk("%s: marking lseg %p iomode %d "
1510                                 "offset %llu length %llu\n", __func__,
1511                                 lseg, lseg->pls_range.iomode,
1512                                 lseg->pls_range.offset,
1513                                 lseg->pls_range.length);
1514                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1515                         mark_lseg_invalid(lseg, tmp_list);
1516                 }
1517 }
1518
1519 void pnfs_error_mark_layout_for_return(struct inode *inode,
1520                                        struct pnfs_layout_segment *lseg)
1521 {
1522         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1523         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1524         struct pnfs_layout_range range = {
1525                 .iomode = lseg->pls_range.iomode,
1526                 .offset = 0,
1527                 .length = NFS4_MAX_UINT64,
1528         };
1529         LIST_HEAD(free_me);
1530
1531         spin_lock(&inode->i_lock);
1532         /* set failure bit so that pnfs path will be retried later */
1533         pnfs_layout_set_fail_bit(lo, iomode);
1534         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1535         if (lo->plh_return_iomode == 0)
1536                 lo->plh_return_iomode = range.iomode;
1537         else if (lo->plh_return_iomode != range.iomode)
1538                 lo->plh_return_iomode = IOMODE_ANY;
1539         /*
1540          * mark all matching lsegs so that we are sure to have no live
1541          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1542          * for how it works.
1543          */
1544         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1545         spin_unlock(&inode->i_lock);
1546         pnfs_free_lseg_list(&free_me);
1547 }
1548 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1549
1550 void
1551 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1552 {
1553         u64 rd_size = req->wb_bytes;
1554
1555         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1556
1557         if (pgio->pg_dreq == NULL)
1558                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1559         else
1560                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1561
1562         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1563                                            req->wb_context,
1564                                            req_offset(req),
1565                                            rd_size,
1566                                            IOMODE_READ,
1567                                            GFP_KERNEL);
1568         /* If no lseg, fall back to read through mds */
1569         if (pgio->pg_lseg == NULL)
1570                 nfs_pageio_reset_read_mds(pgio);
1571
1572 }
1573 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1574
1575 void
1576 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1577                            struct nfs_page *req, u64 wb_size)
1578 {
1579         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1580
1581         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1582                                            req->wb_context,
1583                                            req_offset(req),
1584                                            wb_size,
1585                                            IOMODE_RW,
1586                                            GFP_NOFS);
1587         /* If no lseg, fall back to write through mds */
1588         if (pgio->pg_lseg == NULL)
1589                 nfs_pageio_reset_write_mds(pgio);
1590 }
1591 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1592
1593 /*
1594  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1595  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1596  */
1597 size_t
1598 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1599                      struct nfs_page *req)
1600 {
1601         unsigned int size;
1602         u64 seg_end, req_start, seg_left;
1603
1604         size = nfs_generic_pg_test(pgio, prev, req);
1605         if (!size)
1606                 return 0;
1607
1608         /*
1609          * 'size' contains the number of bytes left in the current page (up
1610          * to the original size asked for in @req->wb_bytes).
1611          *
1612          * Calculate how many bytes are left in the layout segment
1613          * and if there are less bytes than 'size', return that instead.
1614          *
1615          * Please also note that 'end_offset' is actually the offset of the
1616          * first byte that lies outside the pnfs_layout_range. FIXME?
1617          *
1618          */
1619         if (pgio->pg_lseg) {
1620                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1621                                      pgio->pg_lseg->pls_range.length);
1622                 req_start = req_offset(req);
1623                 WARN_ON_ONCE(req_start > seg_end);
1624                 /* start of request is past the last byte of this segment */
1625                 if (req_start >= seg_end)
1626                         return 0;
1627
1628                 /* adjust 'size' iff there are fewer bytes left in the
1629                  * segment than what nfs_generic_pg_test returned */
1630                 seg_left = seg_end - req_start;
1631                 if (seg_left < size)
1632                         size = (unsigned int)seg_left;
1633         }
1634
1635         return size;
1636 }
1637 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1638
1639 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1640 {
1641         struct nfs_pageio_descriptor pgio;
1642
1643         /* Resend all requests through the MDS */
1644         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1645                               hdr->completion_ops);
1646         return nfs_pageio_resend(&pgio, hdr);
1647 }
1648 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1649
1650 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1651 {
1652
1653         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1654         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1655             PNFS_LAYOUTRET_ON_ERROR) {
1656                 pnfs_return_layout(hdr->inode);
1657         }
1658         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1659                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1660 }
1661
1662 /*
1663  * Called by non rpc-based layout drivers
1664  */
1665 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1666 {
1667         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1668         if (!hdr->pnfs_error) {
1669                 pnfs_set_layoutcommit(hdr);
1670                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1671         } else
1672                 pnfs_ld_handle_write_error(hdr);
1673         hdr->mds_ops->rpc_release(hdr);
1674 }
1675 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1676
1677 static void
1678 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1679                 struct nfs_pgio_header *hdr)
1680 {
1681         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1682                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1683                 nfs_pageio_reset_write_mds(desc);
1684                 desc->pg_recoalesce = 1;
1685         }
1686         nfs_pgio_data_destroy(hdr);
1687 }
1688
1689 static enum pnfs_try_status
1690 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1691                         const struct rpc_call_ops *call_ops,
1692                         struct pnfs_layout_segment *lseg,
1693                         int how)
1694 {
1695         struct inode *inode = hdr->inode;
1696         enum pnfs_try_status trypnfs;
1697         struct nfs_server *nfss = NFS_SERVER(inode);
1698
1699         hdr->mds_ops = call_ops;
1700
1701         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1702                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1703         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1704         if (trypnfs != PNFS_NOT_ATTEMPTED)
1705                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1706         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1707         return trypnfs;
1708 }
1709
1710 static void
1711 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1712               struct nfs_pgio_header *hdr, int how)
1713 {
1714         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1715         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1716         enum pnfs_try_status trypnfs;
1717
1718         desc->pg_lseg = NULL;
1719         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1720         if (trypnfs == PNFS_NOT_ATTEMPTED)
1721                 pnfs_write_through_mds(desc, hdr);
1722         pnfs_put_lseg(lseg);
1723 }
1724
1725 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1726 {
1727         pnfs_put_lseg(hdr->lseg);
1728         nfs_pgio_header_free(hdr);
1729 }
1730 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1731
1732 int
1733 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1734 {
1735         struct nfs_pgio_header *hdr;
1736         int ret;
1737
1738         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1739         if (!hdr) {
1740                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1741                 pnfs_put_lseg(desc->pg_lseg);
1742                 desc->pg_lseg = NULL;
1743                 return -ENOMEM;
1744         }
1745         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1746         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1747         ret = nfs_generic_pgio(desc, hdr);
1748         if (ret != 0) {
1749                 pnfs_put_lseg(desc->pg_lseg);
1750                 desc->pg_lseg = NULL;
1751         } else
1752                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1753         return ret;
1754 }
1755 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1756
1757 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1758 {
1759         struct nfs_pageio_descriptor pgio;
1760
1761         /* Resend all requests through the MDS */
1762         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1763         return nfs_pageio_resend(&pgio, hdr);
1764 }
1765 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1766
1767 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1768 {
1769         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1770         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1771             PNFS_LAYOUTRET_ON_ERROR) {
1772                 pnfs_return_layout(hdr->inode);
1773         }
1774         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1775                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1776 }
1777
1778 /*
1779  * Called by non rpc-based layout drivers
1780  */
1781 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1782 {
1783         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1784         if (likely(!hdr->pnfs_error)) {
1785                 __nfs4_read_done_cb(hdr);
1786                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1787         } else
1788                 pnfs_ld_handle_read_error(hdr);
1789         hdr->mds_ops->rpc_release(hdr);
1790 }
1791 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1792
1793 static void
1794 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1795                 struct nfs_pgio_header *hdr)
1796 {
1797         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1798                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1799                 nfs_pageio_reset_read_mds(desc);
1800                 desc->pg_recoalesce = 1;
1801         }
1802         nfs_pgio_data_destroy(hdr);
1803 }
1804
1805 /*
1806  * Call the appropriate parallel I/O subsystem read function.
1807  */
1808 static enum pnfs_try_status
1809 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1810                        const struct rpc_call_ops *call_ops,
1811                        struct pnfs_layout_segment *lseg)
1812 {
1813         struct inode *inode = hdr->inode;
1814         struct nfs_server *nfss = NFS_SERVER(inode);
1815         enum pnfs_try_status trypnfs;
1816
1817         hdr->mds_ops = call_ops;
1818
1819         dprintk("%s: Reading ino:%lu %u@%llu\n",
1820                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1821
1822         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1823         if (trypnfs != PNFS_NOT_ATTEMPTED)
1824                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1825         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1826         return trypnfs;
1827 }
1828
1829 static void
1830 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1831 {
1832         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1833         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1834         enum pnfs_try_status trypnfs;
1835
1836         desc->pg_lseg = NULL;
1837         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1838         if (trypnfs == PNFS_NOT_ATTEMPTED)
1839                 pnfs_read_through_mds(desc, hdr);
1840         pnfs_put_lseg(lseg);
1841 }
1842
1843 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1844 {
1845         pnfs_put_lseg(hdr->lseg);
1846         nfs_pgio_header_free(hdr);
1847 }
1848 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1849
1850 int
1851 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1852 {
1853         struct nfs_pgio_header *hdr;
1854         int ret;
1855
1856         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1857         if (!hdr) {
1858                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1859                 ret = -ENOMEM;
1860                 pnfs_put_lseg(desc->pg_lseg);
1861                 desc->pg_lseg = NULL;
1862                 return ret;
1863         }
1864         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1865         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1866         ret = nfs_generic_pgio(desc, hdr);
1867         if (ret != 0) {
1868                 pnfs_put_lseg(desc->pg_lseg);
1869                 desc->pg_lseg = NULL;
1870         } else
1871                 pnfs_do_read(desc, hdr);
1872         return ret;
1873 }
1874 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1875
1876 static void pnfs_clear_layoutcommitting(struct inode *inode)
1877 {
1878         unsigned long *bitlock = &NFS_I(inode)->flags;
1879
1880         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1881         smp_mb__after_atomic();
1882         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1883 }
1884
1885 /*
1886  * There can be multiple RW segments.
1887  */
1888 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1889 {
1890         struct pnfs_layout_segment *lseg;
1891
1892         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1893                 if (lseg->pls_range.iomode == IOMODE_RW &&
1894                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1895                         list_add(&lseg->pls_lc_list, listp);
1896         }
1897 }
1898
1899 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1900 {
1901         struct pnfs_layout_segment *lseg, *tmp;
1902
1903         /* Matched by references in pnfs_set_layoutcommit */
1904         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1905                 list_del_init(&lseg->pls_lc_list);
1906                 pnfs_put_lseg(lseg);
1907         }
1908
1909         pnfs_clear_layoutcommitting(inode);
1910 }
1911
1912 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1913 {
1914         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1915 }
1916 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1917
1918 void
1919 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1920 {
1921         struct inode *inode = hdr->inode;
1922         struct nfs_inode *nfsi = NFS_I(inode);
1923         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1924         bool mark_as_dirty = false;
1925
1926         spin_lock(&inode->i_lock);
1927         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1928                 mark_as_dirty = true;
1929                 dprintk("%s: Set layoutcommit for inode %lu ",
1930                         __func__, inode->i_ino);
1931         }
1932         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1933                 /* references matched in nfs4_layoutcommit_release */
1934                 pnfs_get_lseg(hdr->lseg);
1935         }
1936         if (end_pos > nfsi->layout->plh_lwb)
1937                 nfsi->layout->plh_lwb = end_pos;
1938         spin_unlock(&inode->i_lock);
1939         dprintk("%s: lseg %p end_pos %llu\n",
1940                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1941
1942         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1943          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1944         if (mark_as_dirty)
1945                 mark_inode_dirty_sync(inode);
1946 }
1947 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1948
1949 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
1950 {
1951         struct inode *inode = data->inode;
1952         struct nfs_inode *nfsi = NFS_I(inode);
1953         bool mark_as_dirty = false;
1954
1955         spin_lock(&inode->i_lock);
1956         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1957                 mark_as_dirty = true;
1958                 dprintk("%s: Set layoutcommit for inode %lu ",
1959                         __func__, inode->i_ino);
1960         }
1961         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
1962                 /* references matched in nfs4_layoutcommit_release */
1963                 pnfs_get_lseg(data->lseg);
1964         }
1965         if (data->lwb > nfsi->layout->plh_lwb)
1966                 nfsi->layout->plh_lwb = data->lwb;
1967         spin_unlock(&inode->i_lock);
1968         dprintk("%s: lseg %p end_pos %llu\n",
1969                 __func__, data->lseg, nfsi->layout->plh_lwb);
1970
1971         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1972          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1973         if (mark_as_dirty)
1974                 mark_inode_dirty_sync(inode);
1975 }
1976 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
1977
1978 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1979 {
1980         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1981
1982         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1983                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1984         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1985 }
1986
1987 /*
1988  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1989  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1990  * data to disk to allow the server to recover the data if it crashes.
1991  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1992  * is off, and a COMMIT is sent to a data server, or
1993  * if WRITEs to a data server return NFS_DATA_SYNC.
1994  */
1995 int
1996 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1997 {
1998         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1999         struct nfs4_layoutcommit_data *data;
2000         struct nfs_inode *nfsi = NFS_I(inode);
2001         loff_t end_pos;
2002         int status;
2003
2004         if (!pnfs_layoutcommit_outstanding(inode))
2005                 return 0;
2006
2007         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2008
2009         status = -EAGAIN;
2010         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2011                 if (!sync)
2012                         goto out;
2013                 status = wait_on_bit_lock_action(&nfsi->flags,
2014                                 NFS_INO_LAYOUTCOMMITTING,
2015                                 nfs_wait_bit_killable,
2016                                 TASK_KILLABLE);
2017                 if (status)
2018                         goto out;
2019         }
2020
2021         status = -ENOMEM;
2022         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2023         data = kzalloc(sizeof(*data), GFP_NOFS);
2024         if (!data)
2025                 goto clear_layoutcommitting;
2026
2027         status = 0;
2028         spin_lock(&inode->i_lock);
2029         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2030                 goto out_unlock;
2031
2032         INIT_LIST_HEAD(&data->lseg_list);
2033         pnfs_list_write_lseg(inode, &data->lseg_list);
2034
2035         end_pos = nfsi->layout->plh_lwb;
2036         nfsi->layout->plh_lwb = 0;
2037
2038         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2039         spin_unlock(&inode->i_lock);
2040
2041         data->args.inode = inode;
2042         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2043         nfs_fattr_init(&data->fattr);
2044         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2045         data->res.fattr = &data->fattr;
2046         data->args.lastbytewritten = end_pos - 1;
2047         data->res.server = NFS_SERVER(inode);
2048
2049         if (ld->prepare_layoutcommit) {
2050                 status = ld->prepare_layoutcommit(&data->args);
2051                 if (status) {
2052                         spin_lock(&inode->i_lock);
2053                         if (end_pos < nfsi->layout->plh_lwb)
2054                                 nfsi->layout->plh_lwb = end_pos;
2055                         spin_unlock(&inode->i_lock);
2056                         put_rpccred(data->cred);
2057                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2058                         goto clear_layoutcommitting;
2059                 }
2060         }
2061
2062
2063         status = nfs4_proc_layoutcommit(data, sync);
2064 out:
2065         if (status)
2066                 mark_inode_dirty_sync(inode);
2067         dprintk("<-- %s status %d\n", __func__, status);
2068         return status;
2069 out_unlock:
2070         spin_unlock(&inode->i_lock);
2071         kfree(data);
2072 clear_layoutcommitting:
2073         pnfs_clear_layoutcommitting(inode);
2074         goto out;
2075 }
2076 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2077
2078 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2079 {
2080         struct nfs4_threshold *thp;
2081
2082         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2083         if (!thp) {
2084                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2085                 return NULL;
2086         }
2087         return thp;
2088 }