fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36 #include "nfs4trace.h"
  37 #include "delegation.h"
  38 #include "nfs42.h"
  39
  40 #define NFSDBG_FACILITY         NFSDBG_PNFS
  41 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  42
  43 /* Locking:
  44  *
  45  * pnfs_spinlock:
  46  *      protects pnfs_modules_tbl.
  47  */
  48 static DEFINE_SPINLOCK(pnfs_spinlock);
  49
  50 /*
  51  * pnfs_modules_tbl holds all pnfs modules
  52  */
  53 static LIST_HEAD(pnfs_modules_tbl);
  54
  55 static int
  56 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
  57                        enum pnfs_iomode iomode, bool sync);
  58
  59 /* Return the registered pnfs layout driver module matching given id */
  60 static struct pnfs_layoutdriver_type *
  61 find_pnfs_driver_locked(u32 id)
  62 {
  63         struct pnfs_layoutdriver_type *local;
  64
  65         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  66                 if (local->id == id)
  67                         goto out;
  68         local = NULL;
  69 out:
  70         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  71         return local;
  72 }
  73
  74 static struct pnfs_layoutdriver_type *
  75 find_pnfs_driver(u32 id)
  76 {
  77         struct pnfs_layoutdriver_type *local;
  78
  79         spin_lock(&pnfs_spinlock);
  80         local = find_pnfs_driver_locked(id);
  81         if (local != NULL && !try_module_get(local->owner)) {
  82                 dprintk("%s: Could not grab reference on module\n", __func__);
  83                 local = NULL;
  84         }
  85         spin_unlock(&pnfs_spinlock);
  86         return local;
  87 }
  88
  89 void
  90 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  91 {
  92         if (nfss->pnfs_curr_ld) {
  93                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  94                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  95                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  96                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  97                         nfs4_deviceid_purge_client(nfss->nfs_client);
  98                 module_put(nfss->pnfs_curr_ld->owner);
  99         }
 100         nfss->pnfs_curr_ld = NULL;
 101 }
 102
 103 /*
 104  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 105  * Currently only one pNFS layout driver per filesystem is supported.
 106  *
 107  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 108  */
 109 void
 110 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 111                       u32 id)
 112 {
 113         struct pnfs_layoutdriver_type *ld_type = NULL;
 114
 115         if (id == 0)
 116                 goto out_no_driver;
 117         if (!(server->nfs_client->cl_exchange_flags &
 118                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 119                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 120                         __func__, id, server->nfs_client->cl_exchange_flags);
 121                 goto out_no_driver;
 122         }
 123         ld_type = find_pnfs_driver(id);
 124         if (!ld_type) {
 125                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 126                 ld_type = find_pnfs_driver(id);
 127                 if (!ld_type) {
 128                         dprintk("%s: No pNFS module found for %u.\n",
 129                                 __func__, id);
 130                         goto out_no_driver;
 131                 }
 132         }
 133         server->pnfs_curr_ld = ld_type;
 134         if (ld_type->set_layoutdriver
 135             && ld_type->set_layoutdriver(server, mntfh)) {
 136                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 137                         "driver %u.\n", __func__, id);
 138                 module_put(ld_type->owner);
 139                 goto out_no_driver;
 140         }
 141         /* Bump the MDS count */
 142         atomic_inc(&server->nfs_client->cl_mds_count);
 143
 144         dprintk("%s: pNFS module for %u set\n", __func__, id);
 145         return;
 146
 147 out_no_driver:
 148         dprintk("%s: Using NFSv4 I/O\n", __func__);
 149         server->pnfs_curr_ld = NULL;
 150 }
 151
 152 int
 153 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 154 {
 155         int status = -EINVAL;
 156         struct pnfs_layoutdriver_type *tmp;
 157
 158         if (ld_type->id == 0) {
 159                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 160                 return status;
 161         }
 162         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 163                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 164                        "alloc_lseg and free_lseg.\n", __func__);
 165                 return status;
 166         }
 167
 168         spin_lock(&pnfs_spinlock);
 169         tmp = find_pnfs_driver_locked(ld_type->id);
 170         if (!tmp) {
 171                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 172                 status = 0;
 173                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 174                         ld_type->name);
 175         } else {
 176                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 177                         __func__, ld_type->id);
 178         }
 179         spin_unlock(&pnfs_spinlock);
 180
 181         return status;
 182 }
 183 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 184
 185 void
 186 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 187 {
 188         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 189         spin_lock(&pnfs_spinlock);
 190         list_del(&ld_type->pnfs_tblid);
 191         spin_unlock(&pnfs_spinlock);
 192 }
 193 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 194
 195 /*
 196  * pNFS client layout cache
 197  */
 198
 199 /* Need to hold i_lock if caller does not already hold reference */
 200 void
 201 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 202 {
 203         atomic_inc(&lo->plh_refcount);
 204 }
 205
 206 static struct pnfs_layout_hdr *
 207 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 208 {
 209         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 210         return ld->alloc_layout_hdr(ino, gfp_flags);
 211 }
 212
 213 static void
 214 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 215 {
 216         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 217         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 218
 219         if (!list_empty(&lo->plh_layouts)) {
 220                 struct nfs_client *clp = server->nfs_client;
 221
 222                 spin_lock(&clp->cl_lock);
 223                 list_del_init(&lo->plh_layouts);
 224                 spin_unlock(&clp->cl_lock);
 225         }
 226         put_rpccred(lo->plh_lc_cred);
 227         return ld->free_layout_hdr(lo);
 228 }
 229
 230 static void
 231 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 232 {
 233         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 234         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 235         nfsi->layout = NULL;
 236         /* Reset MDS Threshold I/O counters */
 237         nfsi->write_io = 0;
 238         nfsi->read_io = 0;
 239 }
 240
 241 void
 242 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 243 {
 244         struct inode *inode = lo->plh_inode;
 245
 246         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 247                 if (!list_empty(&lo->plh_segs))
 248                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 249                 pnfs_detach_layout_hdr(lo);
 250                 spin_unlock(&inode->i_lock);
 251                 pnfs_free_layout_hdr(lo);
 252         }
 253 }
 254
 255 static int
 256 pnfs_iomode_to_fail_bit(u32 iomode)
 257 {
 258         return iomode == IOMODE_RW ?
 259                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 260 }
 261
 262 static void
 263 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 264 {
 265         lo->plh_retry_timestamp = jiffies;
 266         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 267                 atomic_inc(&lo->plh_refcount);
 268 }
 269
 270 static void
 271 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 272 {
 273         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 274                 atomic_dec(&lo->plh_refcount);
 275 }
 276
 277 static void
 278 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 279 {
 280         struct inode *inode = lo->plh_inode;
 281         struct pnfs_layout_range range = {
 282                 .iomode = iomode,
 283                 .offset = 0,
 284                 .length = NFS4_MAX_UINT64,
 285         };
 286         LIST_HEAD(head);
 287
 288         spin_lock(&inode->i_lock);
 289         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 290         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 291         spin_unlock(&inode->i_lock);
 292         pnfs_free_lseg_list(&head);
 293         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 294                         iomode == IOMODE_RW ?  "RW" : "READ");
 295 }
 296
 297 static bool
 298 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 299 {
 300         unsigned long start, end;
 301         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 302
 303         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 304                 return false;
 305         end = jiffies;
 306         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 307         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 308                 /* It is time to retry the failed layoutgets */
 309                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 310                 return false;
 311         }
 312         return true;
 313 }
 314
 315 static void
 316 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 317 {
 318         INIT_LIST_HEAD(&lseg->pls_list);
 319         INIT_LIST_HEAD(&lseg->pls_lc_list);
 320         atomic_set(&lseg->pls_refcount, 1);
 321         smp_mb();
 322         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 323         lseg->pls_layout = lo;
 324 }
 325
 326 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 327 {
 328         struct inode *ino = lseg->pls_layout->plh_inode;
 329
 330         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 331 }
 332
 333 static void
 334 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 335                 struct pnfs_layout_segment *lseg)
 336 {
 337         struct inode *inode = lo->plh_inode;
 338
 339         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 340         list_del_init(&lseg->pls_list);
 341         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 342         atomic_dec(&lo->plh_refcount);
 343         if (list_empty(&lo->plh_segs))
 344                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 345         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 346 }
 347
 348 /* Return true if layoutreturn is needed */
 349 static bool
 350 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 351                         struct pnfs_layout_segment *lseg)
 352 {
 353         struct pnfs_layout_segment *s;
 354
 355         if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 356                 return false;
 357
 358         list_for_each_entry(s, &lo->plh_segs, pls_list)
 359                 if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 360                         return false;
 361
 362         return true;
 363 }
 364
 365 static bool
 366 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 367 {
 368         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 369                 return false;
 370         lo->plh_return_iomode = 0;
 371         pnfs_get_layout_hdr(lo);
 372         clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
 373         return true;
 374 }
 375
 376 static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
 377                 struct pnfs_layout_hdr *lo, struct inode *inode)
 378 {
 379         lo = lseg->pls_layout;
 380         inode = lo->plh_inode;
 381
 382         spin_lock(&inode->i_lock);
 383         if (pnfs_layout_need_return(lo, lseg)) {
 384                 nfs4_stateid stateid;
 385                 enum pnfs_iomode iomode;
 386                 bool send;
 387
 388                 stateid = lo->plh_stateid;
 389                 iomode = lo->plh_return_iomode;
 390                 send = pnfs_prepare_layoutreturn(lo);
 391                 spin_unlock(&inode->i_lock);
 392                 if (send) {
 393                         /* Send an async layoutreturn so we dont deadlock */
 394                         pnfs_send_layoutreturn(lo, stateid, iomode, false);
 395                 }
 396         } else
 397                 spin_unlock(&inode->i_lock);
 398 }
 399
 400 void
 401 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 402 {
 403         struct pnfs_layout_hdr *lo;
 404         struct inode *inode;
 405
 406         if (!lseg)
 407                 return;
 408
 409         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 410                 atomic_read(&lseg->pls_refcount),
 411                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 412
 413         /* Handle the case where refcount != 1 */
 414         if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
 415                 return;
 416
 417         lo = lseg->pls_layout;
 418         inode = lo->plh_inode;
 419         /* Do we need a layoutreturn? */
 420         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 421                 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
 422
 423         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 424                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 425                         spin_unlock(&inode->i_lock);
 426                         return;
 427                 }
 428                 pnfs_get_layout_hdr(lo);
 429                 pnfs_layout_remove_lseg(lo, lseg);
 430                 spin_unlock(&inode->i_lock);
 431                 pnfs_free_lseg(lseg);
 432                 pnfs_put_layout_hdr(lo);
 433         }
 434 }
 435 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 436
 437 static void pnfs_free_lseg_async_work(struct work_struct *work)
 438 {
 439         struct pnfs_layout_segment *lseg;
 440         struct pnfs_layout_hdr *lo;
 441
 442         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 443         lo = lseg->pls_layout;
 444
 445         pnfs_free_lseg(lseg);
 446         pnfs_put_layout_hdr(lo);
 447 }
 448
 449 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 450 {
 451         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 452         schedule_work(&lseg->pls_work);
 453 }
 454
 455 void
 456 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 457 {
 458         if (!lseg)
 459                 return;
 460
 461         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 462
 463         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 464                 atomic_read(&lseg->pls_refcount),
 465                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 466         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 467                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 468                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 469                         return;
 470                 pnfs_get_layout_hdr(lo);
 471                 pnfs_layout_remove_lseg(lo, lseg);
 472                 pnfs_free_lseg_async(lseg);
 473         }
 474 }
 475 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 476
 477 static u64
 478 end_offset(u64 start, u64 len)
 479 {
 480         u64 end;
 481
 482         end = start + len;
 483         return end >= start ? end : NFS4_MAX_UINT64;
 484 }
 485
 486 /*
 487  * is l2 fully contained in l1?
 488  *   start1                             end1
 489  *   [----------------------------------)
 490  *           start2           end2
 491  *           [----------------)
 492  */
 493 static bool
 494 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 495                  const struct pnfs_layout_range *l2)
 496 {
 497         u64 start1 = l1->offset;
 498         u64 end1 = end_offset(start1, l1->length);
 499         u64 start2 = l2->offset;
 500         u64 end2 = end_offset(start2, l2->length);
 501
 502         return (start1 <= start2) && (end1 >= end2);
 503 }
 504
 505 /*
 506  * is l1 and l2 intersecting?
 507  *   start1                             end1
 508  *   [----------------------------------)
 509  *                              start2           end2
 510  *                              [----------------)
 511  */
 512 static bool
 513 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 514                     const struct pnfs_layout_range *l2)
 515 {
 516         u64 start1 = l1->offset;
 517         u64 end1 = end_offset(start1, l1->length);
 518         u64 start2 = l2->offset;
 519         u64 end2 = end_offset(start2, l2->length);
 520
 521         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 522                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 523 }
 524
 525 static bool
 526 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 527                  const struct pnfs_layout_range *recall_range)
 528 {
 529         return (recall_range->iomode == IOMODE_ANY ||
 530                 lseg_range->iomode == recall_range->iomode) &&
 531                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 532 }
 533
 534 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 535                 struct list_head *tmp_list)
 536 {
 537         if (!atomic_dec_and_test(&lseg->pls_refcount))
 538                 return false;
 539         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 540         list_add(&lseg->pls_list, tmp_list);
 541         return true;
 542 }
 543
 544 /* Returns 1 if lseg is removed from list, 0 otherwise */
 545 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 546                              struct list_head *tmp_list)
 547 {
 548         int rv = 0;
 549
 550         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 551                 /* Remove the reference keeping the lseg in the
 552                  * list.  It will now be removed when all
 553                  * outstanding io is finished.
 554                  */
 555                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 556                         atomic_read(&lseg->pls_refcount));
 557                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 558                         rv = 1;
 559         }
 560         return rv;
 561 }
 562
 563 /* Returns count of number of matching invalid lsegs remaining in list
 564  * after call.
 565  */
 566 int
 567 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 568                             struct list_head *tmp_list,
 569                             struct pnfs_layout_range *recall_range)
 570 {
 571         struct pnfs_layout_segment *lseg, *next;
 572         int invalid = 0, removed = 0;
 573
 574         dprintk("%s:Begin lo %p\n", __func__, lo);
 575
 576         if (list_empty(&lo->plh_segs))
 577                 return 0;
 578         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 579                 if (!recall_range ||
 580                     should_free_lseg(&lseg->pls_range, recall_range)) {
 581                         dprintk("%s: freeing lseg %p iomode %d "
 582                                 "offset %llu length %llu\n", __func__,
 583                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 584                                 lseg->pls_range.length);
 585                         invalid++;
 586                         removed += mark_lseg_invalid(lseg, tmp_list);
 587                 }
 588         dprintk("%s:Return %i\n", __func__, invalid - removed);
 589         return invalid - removed;
 590 }
 591
 592 /* note free_me must contain lsegs from a single layout_hdr */
 593 void
 594 pnfs_free_lseg_list(struct list_head *free_me)
 595 {
 596         struct pnfs_layout_segment *lseg, *tmp;
 597
 598         if (list_empty(free_me))
 599                 return;
 600
 601         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 602                 list_del(&lseg->pls_list);
 603                 pnfs_free_lseg(lseg);
 604         }
 605 }
 606
 607 void
 608 pnfs_destroy_layout(struct nfs_inode *nfsi)
 609 {
 610         struct pnfs_layout_hdr *lo;
 611         LIST_HEAD(tmp_list);
 612
 613         spin_lock(&nfsi->vfs_inode.i_lock);
 614         lo = nfsi->layout;
 615         if (lo) {
 616                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 617                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 618                 pnfs_get_layout_hdr(lo);
 619                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 620                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 621                 pnfs_clear_retry_layoutget(lo);
 622                 spin_unlock(&nfsi->vfs_inode.i_lock);
 623                 pnfs_free_lseg_list(&tmp_list);
 624                 pnfs_put_layout_hdr(lo);
 625         } else
 626                 spin_unlock(&nfsi->vfs_inode.i_lock);
 627 }
 628 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 629
 630 static bool
 631 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 632                 struct list_head *layout_list)
 633 {
 634         struct pnfs_layout_hdr *lo;
 635         bool ret = false;
 636
 637         spin_lock(&inode->i_lock);
 638         lo = NFS_I(inode)->layout;
 639         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 640                 pnfs_get_layout_hdr(lo);
 641                 list_add(&lo->plh_bulk_destroy, layout_list);
 642                 ret = true;
 643         }
 644         spin_unlock(&inode->i_lock);
 645         return ret;
 646 }
 647
 648 /* Caller must hold rcu_read_lock and clp->cl_lock */
 649 static int
 650 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 651                 struct nfs_server *server,
 652                 struct list_head *layout_list)
 653 {
 654         struct pnfs_layout_hdr *lo, *next;
 655         struct inode *inode;
 656
 657         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 658                 inode = igrab(lo->plh_inode);
 659                 if (inode == NULL)
 660                         continue;
 661                 list_del_init(&lo->plh_layouts);
 662                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 663                         continue;
 664                 rcu_read_unlock();
 665                 spin_unlock(&clp->cl_lock);
 666                 iput(inode);
 667                 spin_lock(&clp->cl_lock);
 668                 rcu_read_lock();
 669                 return -EAGAIN;
 670         }
 671         return 0;
 672 }
 673
 674 static int
 675 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 676                 bool is_bulk_recall)
 677 {
 678         struct pnfs_layout_hdr *lo;
 679         struct inode *inode;
 680         struct pnfs_layout_range range = {
 681                 .iomode = IOMODE_ANY,
 682                 .offset = 0,
 683                 .length = NFS4_MAX_UINT64,
 684         };
 685         LIST_HEAD(lseg_list);
 686         int ret = 0;
 687
 688         while (!list_empty(layout_list)) {
 689                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 690                                 plh_bulk_destroy);
 691                 dprintk("%s freeing layout for inode %lu\n", __func__,
 692                         lo->plh_inode->i_ino);
 693                 inode = lo->plh_inode;
 694
 695                 pnfs_layoutcommit_inode(inode, false);
 696
 697                 spin_lock(&inode->i_lock);
 698                 list_del_init(&lo->plh_bulk_destroy);
 699                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 700                 if (is_bulk_recall)
 701                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 702                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 703                         ret = -EAGAIN;
 704                 spin_unlock(&inode->i_lock);
 705                 pnfs_free_lseg_list(&lseg_list);
 706                 pnfs_put_layout_hdr(lo);
 707                 iput(inode);
 708         }
 709         return ret;
 710 }
 711
 712 int
 713 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 714                 struct nfs_fsid *fsid,
 715                 bool is_recall)
 716 {
 717         struct nfs_server *server;
 718         LIST_HEAD(layout_list);
 719
 720         spin_lock(&clp->cl_lock);
 721         rcu_read_lock();
 722 restart:
 723         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 724                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 725                         continue;
 726                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 727                                 server,
 728                                 &layout_list) != 0)
 729                         goto restart;
 730         }
 731         rcu_read_unlock();
 732         spin_unlock(&clp->cl_lock);
 733
 734         if (list_empty(&layout_list))
 735                 return 0;
 736         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 737 }
 738
 739 int
 740 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 741                 bool is_recall)
 742 {
 743         struct nfs_server *server;
 744         LIST_HEAD(layout_list);
 745
 746         spin_lock(&clp->cl_lock);
 747         rcu_read_lock();
 748 restart:
 749         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 750                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 751                                         server,
 752                                         &layout_list) != 0)
 753                         goto restart;
 754         }
 755         rcu_read_unlock();
 756         spin_unlock(&clp->cl_lock);
 757
 758         if (list_empty(&layout_list))
 759                 return 0;
 760         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 761 }
 762
 763 /*
 764  * Called by the state manger to remove all layouts established under an
 765  * expired lease.
 766  */
 767 void
 768 pnfs_destroy_all_layouts(struct nfs_client *clp)
 769 {
 770         nfs4_deviceid_mark_client_invalid(clp);
 771         nfs4_deviceid_purge_client(clp);
 772
 773         pnfs_destroy_layouts_byclid(clp, false);
 774 }
 775
 776 /*
 777  * Compare 2 layout stateid sequence ids, to see which is newer,
 778  * taking into account wraparound issues.
 779  */
 780 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 781 {
 782         return (s32)(s1 - s2) > 0;
 783 }
 784
 785 /* update lo->plh_stateid with new if is more recent */
 786 void
 787 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 788                         bool update_barrier)
 789 {
 790         u32 oldseq, newseq, new_barrier;
 791         int empty = list_empty(&lo->plh_segs);
 792
 793         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 794         newseq = be32_to_cpu(new->seqid);
 795         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 796                 nfs4_stateid_copy(&lo->plh_stateid, new);
 797                 if (update_barrier) {
 798                         new_barrier = be32_to_cpu(new->seqid);
 799                 } else {
 800                         /* Because of wraparound, we want to keep the barrier
 801                          * "close" to the current seqids.
 802                          */
 803                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 804                 }
 805                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 806                         lo->plh_barrier = new_barrier;
 807         }
 808 }
 809
 810 static bool
 811 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 812                 const nfs4_stateid *stateid)
 813 {
 814         u32 seqid = be32_to_cpu(stateid->seqid);
 815
 816         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 817 }
 818
 819 /* lget is set to 1 if called from inside send_layoutget call chain */
 820 static bool
 821 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 822 {
 823         return lo->plh_block_lgets ||
 824                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 825 }
 826
 827 int
 828 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 829                               struct pnfs_layout_range *range,
 830                               struct nfs4_state *open_state)
 831 {
 832         int status = 0;
 833
 834         dprintk("--> %s\n", __func__);
 835         spin_lock(&lo->plh_inode->i_lock);
 836         if (pnfs_layoutgets_blocked(lo)) {
 837                 status = -EAGAIN;
 838         } else if (!nfs4_valid_open_stateid(open_state)) {
 839                 status = -EBADF;
 840         } else if (list_empty(&lo->plh_segs) ||
 841                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 842                 int seq;
 843
 844                 do {
 845                         seq = read_seqbegin(&open_state->seqlock);
 846                         nfs4_stateid_copy(dst, &open_state->stateid);
 847                 } while (read_seqretry(&open_state->seqlock, seq));
 848         } else
 849                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 850         spin_unlock(&lo->plh_inode->i_lock);
 851         dprintk("<-- %s\n", __func__);
 852         return status;
 853 }
 854
 855 /*
 856 * Get layout from server.
 857 *    for now, assume that whole file layouts are requested.
 858 *    arg->offset: 0
 859 *    arg->length: all ones
 860 */
 861 static struct pnfs_layout_segment *
 862 send_layoutget(struct pnfs_layout_hdr *lo,
 863            struct nfs_open_context *ctx,
 864            struct pnfs_layout_range *range,
 865            gfp_t gfp_flags)
 866 {
 867         struct inode *ino = lo->plh_inode;
 868         struct nfs_server *server = NFS_SERVER(ino);
 869         struct nfs4_layoutget *lgp;
 870         struct pnfs_layout_segment *lseg;
 871
 872         dprintk("--> %s\n", __func__);
 873
 874         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 875         if (lgp == NULL)
 876                 return NULL;
 877
 878         lgp->args.minlength = PAGE_CACHE_SIZE;
 879         if (lgp->args.minlength > range->length)
 880                 lgp->args.minlength = range->length;
 881         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 882         lgp->args.range = *range;
 883         lgp->args.type = server->pnfs_curr_ld->id;
 884         lgp->args.inode = ino;
 885         lgp->args.ctx = get_nfs_open_context(ctx);
 886         lgp->gfp_flags = gfp_flags;
 887         lgp->cred = lo->plh_lc_cred;
 888
 889         /* Synchronously retrieve layout information from server and
 890          * store in lseg.
 891          */
 892         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 893         if (IS_ERR(lseg)) {
 894                 switch (PTR_ERR(lseg)) {
 895                 case -ENOMEM:
 896                 case -ERESTARTSYS:
 897                         break;
 898                 default:
 899                         /* remember that LAYOUTGET failed and suspend trying */
 900                         pnfs_layout_io_set_failed(lo, range->iomode);
 901                 }
 902                 return NULL;
 903         } else
 904                 pnfs_layout_clear_fail_bit(lo,
 905                                 pnfs_iomode_to_fail_bit(range->iomode));
 906
 907         return lseg;
 908 }
 909
 910 static void pnfs_clear_layoutcommit(struct inode *inode,
 911                 struct list_head *head)
 912 {
 913         struct nfs_inode *nfsi = NFS_I(inode);
 914         struct pnfs_layout_segment *lseg, *tmp;
 915
 916         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 917                 return;
 918         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 919                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 920                         continue;
 921                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 922         }
 923 }
 924
 925 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 926 {
 927         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 928         smp_mb__after_atomic();
 929         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 930         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 931 }
 932
 933 static int
 934 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 935                        enum pnfs_iomode iomode, bool sync)
 936 {
 937         struct inode *ino = lo->plh_inode;
 938         struct nfs4_layoutreturn *lrp;
 939         int status = 0;
 940
 941         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 942         if (unlikely(lrp == NULL)) {
 943                 status = -ENOMEM;
 944                 spin_lock(&ino->i_lock);
 945                 pnfs_clear_layoutreturn_waitbit(lo);
 946                 spin_unlock(&ino->i_lock);
 947                 pnfs_put_layout_hdr(lo);
 948                 goto out;
 949         }
 950
 951         lrp->args.stateid = stateid;
 952         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 953         lrp->args.inode = ino;
 954         lrp->args.range.iomode = iomode;
 955         lrp->args.range.offset = 0;
 956         lrp->args.range.length = NFS4_MAX_UINT64;
 957         lrp->args.layout = lo;
 958         lrp->clp = NFS_SERVER(ino)->nfs_client;
 959         lrp->cred = lo->plh_lc_cred;
 960
 961         status = nfs4_proc_layoutreturn(lrp, sync);
 962 out:
 963         dprintk("<-- %s status: %d\n", __func__, status);
 964         return status;
 965 }
 966
 967 /*
 968  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 969  * when the layout segment list is empty.
 970  *
 971  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 972  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 973  * deviceid is marked invalid.
 974  */
 975 int
 976 _pnfs_return_layout(struct inode *ino)
 977 {
 978         struct pnfs_layout_hdr *lo = NULL;
 979         struct nfs_inode *nfsi = NFS_I(ino);
 980         LIST_HEAD(tmp_list);
 981         nfs4_stateid stateid;
 982         int status = 0, empty;
 983         bool send;
 984
 985         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 986
 987         spin_lock(&ino->i_lock);
 988         lo = nfsi->layout;
 989         if (!lo) {
 990                 spin_unlock(&ino->i_lock);
 991                 dprintk("NFS: %s no layout to return\n", __func__);
 992                 goto out;
 993         }
 994         stateid = nfsi->layout->plh_stateid;
 995         /* Reference matched in nfs4_layoutreturn_release */
 996         pnfs_get_layout_hdr(lo);
 997         empty = list_empty(&lo->plh_segs);
 998         pnfs_clear_layoutcommit(ino, &tmp_list);
 999         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1000
1001         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1002                 struct pnfs_layout_range range = {
1003                         .iomode         = IOMODE_ANY,
1004                         .offset         = 0,
1005                         .length         = NFS4_MAX_UINT64,
1006                 };
1007                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1008         }
1009
1010         /* Don't send a LAYOUTRETURN if list was initially empty */
1011         if (empty) {
1012                 spin_unlock(&ino->i_lock);
1013                 dprintk("NFS: %s no layout segments to return\n", __func__);
1014                 goto out_put_layout_hdr;
1015         }
1016
1017         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1018         send = pnfs_prepare_layoutreturn(lo);
1019         spin_unlock(&ino->i_lock);
1020         pnfs_free_lseg_list(&tmp_list);
1021         if (send)
1022                 status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1023 out_put_layout_hdr:
1024         pnfs_put_layout_hdr(lo);
1025 out:
1026         dprintk("<-- %s status: %d\n", __func__, status);
1027         return status;
1028 }
1029 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1030
1031 int
1032 pnfs_commit_and_return_layout(struct inode *inode)
1033 {
1034         struct pnfs_layout_hdr *lo;
1035         int ret;
1036
1037         spin_lock(&inode->i_lock);
1038         lo = NFS_I(inode)->layout;
1039         if (lo == NULL) {
1040                 spin_unlock(&inode->i_lock);
1041                 return 0;
1042         }
1043         pnfs_get_layout_hdr(lo);
1044         /* Block new layoutgets and read/write to ds */
1045         lo->plh_block_lgets++;
1046         spin_unlock(&inode->i_lock);
1047         filemap_fdatawait(inode->i_mapping);
1048         ret = pnfs_layoutcommit_inode(inode, true);
1049         if (ret == 0)
1050                 ret = _pnfs_return_layout(inode);
1051         spin_lock(&inode->i_lock);
1052         lo->plh_block_lgets--;
1053         spin_unlock(&inode->i_lock);
1054         pnfs_put_layout_hdr(lo);
1055         return ret;
1056 }
1057
1058 bool pnfs_roc(struct inode *ino)
1059 {
1060         struct nfs_inode *nfsi = NFS_I(ino);
1061         struct nfs_open_context *ctx;
1062         struct nfs4_state *state;
1063         struct pnfs_layout_hdr *lo;
1064         struct pnfs_layout_segment *lseg, *tmp;
1065         nfs4_stateid stateid;
1066         LIST_HEAD(tmp_list);
1067         bool found = false, layoutreturn = false, roc = false;
1068
1069         spin_lock(&ino->i_lock);
1070         lo = nfsi->layout;
1071         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1072                 goto out_noroc;
1073
1074         /* no roc if we hold a delegation */
1075         if (nfs4_check_delegation(ino, FMODE_READ))
1076                 goto out_noroc;
1077
1078         list_for_each_entry(ctx, &nfsi->open_files, list) {
1079                 state = ctx->state;
1080                 /* Don't return layout if there is open file state */
1081                 if (state != NULL && state->state != 0)
1082                         goto out_noroc;
1083         }
1084
1085         stateid = lo->plh_stateid;
1086         /* always send layoutreturn if being marked so */
1087         if (test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1088                                    &lo->plh_flags))
1089                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1090
1091         pnfs_clear_retry_layoutget(lo);
1092         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1093                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1094                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1095                         mark_lseg_invalid(lseg, &tmp_list);
1096                         found = true;
1097                 }
1098         /* pnfs_prepare_layoutreturn() grabs lo ref and it will be put
1099          * in pnfs_roc_release(). We don't really send a layoutreturn but
1100          * still want others to view us like we are sending one!
1101          *
1102          * If pnfs_prepare_layoutreturn() fails, it means someone else is doing
1103          * LAYOUTRETURN, so we proceed like there are no layouts to return.
1104          *
1105          * ROC in three conditions:
1106          * 1. there are ROC lsegs
1107          * 2. we don't send layoutreturn
1108          * 3. no others are sending layoutreturn
1109          */
1110         if (found && !layoutreturn && pnfs_prepare_layoutreturn(lo))
1111                 roc = true;
1112
1113 out_noroc:
1114         spin_unlock(&ino->i_lock);
1115         pnfs_free_lseg_list(&tmp_list);
1116         pnfs_layoutcommit_inode(ino, true);
1117         if (layoutreturn)
1118                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1119         return roc;
1120 }
1121
1122 void pnfs_roc_release(struct inode *ino)
1123 {
1124         struct pnfs_layout_hdr *lo;
1125
1126         spin_lock(&ino->i_lock);
1127         lo = NFS_I(ino)->layout;
1128         pnfs_clear_layoutreturn_waitbit(lo);
1129         if (atomic_dec_and_test(&lo->plh_refcount)) {
1130                 pnfs_detach_layout_hdr(lo);
1131                 spin_unlock(&ino->i_lock);
1132                 pnfs_free_layout_hdr(lo);
1133         } else
1134                 spin_unlock(&ino->i_lock);
1135 }
1136
1137 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1138 {
1139         struct pnfs_layout_hdr *lo;
1140
1141         spin_lock(&ino->i_lock);
1142         lo = NFS_I(ino)->layout;
1143         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1144                 lo->plh_barrier = barrier;
1145         spin_unlock(&ino->i_lock);
1146         trace_nfs4_layoutreturn_on_close(ino, 0);
1147 }
1148
1149 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1150 {
1151         struct nfs_inode *nfsi = NFS_I(ino);
1152         struct pnfs_layout_hdr *lo;
1153         u32 current_seqid;
1154
1155         spin_lock(&ino->i_lock);
1156         lo = nfsi->layout;
1157         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1158
1159         /* Since close does not return a layout stateid for use as
1160          * a barrier, we choose the worst-case barrier.
1161          */
1162         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1163         spin_unlock(&ino->i_lock);
1164 }
1165
1166 /*
1167  * Compare two layout segments for sorting into layout cache.
1168  * We want to preferentially return RW over RO layouts, so ensure those
1169  * are seen first.
1170  */
1171 static s64
1172 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1173            const struct pnfs_layout_range *l2)
1174 {
1175         s64 d;
1176
1177         /* high offset > low offset */
1178         d = l1->offset - l2->offset;
1179         if (d)
1180                 return d;
1181
1182         /* short length > long length */
1183         d = l2->length - l1->length;
1184         if (d)
1185                 return d;
1186
1187         /* read > read/write */
1188         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1189 }
1190
1191 static void
1192 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1193                    struct pnfs_layout_segment *lseg)
1194 {
1195         struct pnfs_layout_segment *lp;
1196
1197         dprintk("%s:Begin\n", __func__);
1198
1199         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1200                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1201                         continue;
1202                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1203                 dprintk("%s: inserted lseg %p "
1204                         "iomode %d offset %llu length %llu before "
1205                         "lp %p iomode %d offset %llu length %llu\n",
1206                         __func__, lseg, lseg->pls_range.iomode,
1207                         lseg->pls_range.offset, lseg->pls_range.length,
1208                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1209                         lp->pls_range.length);
1210                 goto out;
1211         }
1212         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1213         dprintk("%s: inserted lseg %p "
1214                 "iomode %d offset %llu length %llu at tail\n",
1215                 __func__, lseg, lseg->pls_range.iomode,
1216                 lseg->pls_range.offset, lseg->pls_range.length);
1217 out:
1218         pnfs_get_layout_hdr(lo);
1219
1220         dprintk("%s:Return\n", __func__);
1221 }
1222
1223 static struct pnfs_layout_hdr *
1224 alloc_init_layout_hdr(struct inode *ino,
1225                       struct nfs_open_context *ctx,
1226                       gfp_t gfp_flags)
1227 {
1228         struct pnfs_layout_hdr *lo;
1229
1230         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1231         if (!lo)
1232                 return NULL;
1233         atomic_set(&lo->plh_refcount, 1);
1234         INIT_LIST_HEAD(&lo->plh_layouts);
1235         INIT_LIST_HEAD(&lo->plh_segs);
1236         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1237         lo->plh_inode = ino;
1238         lo->plh_lc_cred = get_rpccred(ctx->cred);
1239         return lo;
1240 }
1241
1242 static struct pnfs_layout_hdr *
1243 pnfs_find_alloc_layout(struct inode *ino,
1244                        struct nfs_open_context *ctx,
1245                        gfp_t gfp_flags)
1246 {
1247         struct nfs_inode *nfsi = NFS_I(ino);
1248         struct pnfs_layout_hdr *new = NULL;
1249
1250         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1251
1252         if (nfsi->layout != NULL)
1253                 goto out_existing;
1254         spin_unlock(&ino->i_lock);
1255         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1256         spin_lock(&ino->i_lock);
1257
1258         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1259                 nfsi->layout = new;
1260                 return new;
1261         } else if (new != NULL)
1262                 pnfs_free_layout_hdr(new);
1263 out_existing:
1264         pnfs_get_layout_hdr(nfsi->layout);
1265         return nfsi->layout;
1266 }
1267
1268 /*
1269  * iomode matching rules:
1270  * iomode       lseg    match
1271  * -----        -----   -----
1272  * ANY          READ    true
1273  * ANY          RW      true
1274  * RW           READ    false
1275  * RW           RW      true
1276  * READ         READ    true
1277  * READ         RW      true
1278  */
1279 static bool
1280 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1281                  const struct pnfs_layout_range *range)
1282 {
1283         struct pnfs_layout_range range1;
1284
1285         if ((range->iomode == IOMODE_RW &&
1286              ls_range->iomode != IOMODE_RW) ||
1287             !pnfs_lseg_range_intersecting(ls_range, range))
1288                 return 0;
1289
1290         /* range1 covers only the first byte in the range */
1291         range1 = *range;
1292         range1.length = 1;
1293         return pnfs_lseg_range_contained(ls_range, &range1);
1294 }
1295
1296 /*
1297  * lookup range in layout
1298  */
1299 static struct pnfs_layout_segment *
1300 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1301                 struct pnfs_layout_range *range)
1302 {
1303         struct pnfs_layout_segment *lseg, *ret = NULL;
1304
1305         dprintk("%s:Begin\n", __func__);
1306
1307         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1308                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1309                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1310                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1311                         ret = pnfs_get_lseg(lseg);
1312                         break;
1313                 }
1314                 if (lseg->pls_range.offset > range->offset)
1315                         break;
1316         }
1317
1318         dprintk("%s:Return lseg %p ref %d\n",
1319                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1320         return ret;
1321 }
1322
1323 /*
1324  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1325  * to the MDS or over pNFS
1326  *
1327  * The nfs_inode read_io and write_io fields are cumulative counters reset
1328  * when there are no layout segments. Note that in pnfs_update_layout iomode
1329  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1330  * WRITE request.
1331  *
1332  * A return of true means use MDS I/O.
1333  *
1334  * From rfc 5661:
1335  * If a file's size is smaller than the file size threshold, data accesses
1336  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1337  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1338  * server.  If both file size and I/O size are provided, the client SHOULD
1339  * reach or exceed  both thresholds before sending its read or write
1340  * requests to the data server.
1341  */
1342 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1343                                      struct inode *ino, int iomode)
1344 {
1345         struct nfs4_threshold *t = ctx->mdsthreshold;
1346         struct nfs_inode *nfsi = NFS_I(ino);
1347         loff_t fsize = i_size_read(ino);
1348         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1349
1350         if (t == NULL)
1351                 return ret;
1352
1353         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1354                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1355
1356         switch (iomode) {
1357         case IOMODE_READ:
1358                 if (t->bm & THRESHOLD_RD) {
1359                         dprintk("%s fsize %llu\n", __func__, fsize);
1360                         size_set = true;
1361                         if (fsize < t->rd_sz)
1362                                 size = true;
1363                 }
1364                 if (t->bm & THRESHOLD_RD_IO) {
1365                         dprintk("%s nfsi->read_io %llu\n", __func__,
1366                                 nfsi->read_io);
1367                         io_set = true;
1368                         if (nfsi->read_io < t->rd_io_sz)
1369                                 io = true;
1370                 }
1371                 break;
1372         case IOMODE_RW:
1373                 if (t->bm & THRESHOLD_WR) {
1374                         dprintk("%s fsize %llu\n", __func__, fsize);
1375                         size_set = true;
1376                         if (fsize < t->wr_sz)
1377                                 size = true;
1378                 }
1379                 if (t->bm & THRESHOLD_WR_IO) {
1380                         dprintk("%s nfsi->write_io %llu\n", __func__,
1381                                 nfsi->write_io);
1382                         io_set = true;
1383                         if (nfsi->write_io < t->wr_io_sz)
1384                                 io = true;
1385                 }
1386                 break;
1387         }
1388         if (size_set && io_set) {
1389                 if (size && io)
1390                         ret = true;
1391         } else if (size || io)
1392                 ret = true;
1393
1394         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1395         return ret;
1396 }
1397
1398 /* stop waiting if someone clears NFS_LAYOUT_RETRY_LAYOUTGET bit. */
1399 static int pnfs_layoutget_retry_bit_wait(struct wait_bit_key *key)
1400 {
1401         if (!test_bit(NFS_LAYOUT_RETRY_LAYOUTGET, key->flags))
1402                 return 1;
1403         return nfs_wait_bit_killable(key);
1404 }
1405
1406 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1407 {
1408         if (!pnfs_should_retry_layoutget(lo))
1409                 return false;
1410         /*
1411          * send layoutcommit as it can hold up layoutreturn due to lseg
1412          * reference
1413          */
1414         pnfs_layoutcommit_inode(lo->plh_inode, false);
1415         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1416                                    pnfs_layoutget_retry_bit_wait,
1417                                    TASK_UNINTERRUPTIBLE);
1418 }
1419
1420 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1421 {
1422         unsigned long *bitlock = &lo->plh_flags;
1423
1424         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1425         smp_mb__after_atomic();
1426         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1427 }
1428
1429 /*
1430  * Layout segment is retreived from the server if not cached.
1431  * The appropriate layout segment is referenced and returned to the caller.
1432  */
1433 struct pnfs_layout_segment *
1434 pnfs_update_layout(struct inode *ino,
1435                    struct nfs_open_context *ctx,
1436                    loff_t pos,
1437                    u64 count,
1438                    enum pnfs_iomode iomode,
1439                    gfp_t gfp_flags)
1440 {
1441         struct pnfs_layout_range arg = {
1442                 .iomode = iomode,
1443                 .offset = pos,
1444                 .length = count,
1445         };
1446         unsigned pg_offset;
1447         struct nfs_server *server = NFS_SERVER(ino);
1448         struct nfs_client *clp = server->nfs_client;
1449         struct pnfs_layout_hdr *lo;
1450         struct pnfs_layout_segment *lseg = NULL;
1451         bool first;
1452
1453         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1454                 goto out;
1455
1456         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1457                 goto out;
1458
1459 lookup_again:
1460         first = false;
1461         spin_lock(&ino->i_lock);
1462         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1463         if (lo == NULL) {
1464                 spin_unlock(&ino->i_lock);
1465                 goto out;
1466         }
1467
1468         /* Do we even need to bother with this? */
1469         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1470                 dprintk("%s matches recall, use MDS\n", __func__);
1471                 goto out_unlock;
1472         }
1473
1474         /* if LAYOUTGET already failed once we don't try again */
1475         if (pnfs_layout_io_test_failed(lo, iomode) &&
1476             !pnfs_should_retry_layoutget(lo))
1477                 goto out_unlock;
1478
1479         first = list_empty(&lo->plh_segs);
1480         if (first) {
1481                 /* The first layoutget for the file. Need to serialize per
1482                  * RFC 5661 Errata 3208.
1483                  */
1484                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1485                                      &lo->plh_flags)) {
1486                         spin_unlock(&ino->i_lock);
1487                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1488                                     TASK_UNINTERRUPTIBLE);
1489                         pnfs_put_layout_hdr(lo);
1490                         goto lookup_again;
1491                 }
1492         } else {
1493                 /* Check to see if the layout for the given range
1494                  * already exists
1495                  */
1496                 lseg = pnfs_find_lseg(lo, &arg);
1497                 if (lseg)
1498                         goto out_unlock;
1499         }
1500
1501         /*
1502          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1503          * for LAYOUTRETURN even if first is true.
1504          */
1505         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1506                 spin_unlock(&ino->i_lock);
1507                 dprintk("%s wait for layoutreturn\n", __func__);
1508                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1509                         if (first)
1510                                 pnfs_clear_first_layoutget(lo);
1511                         pnfs_put_layout_hdr(lo);
1512                         dprintk("%s retrying\n", __func__);
1513                         goto lookup_again;
1514                 }
1515                 goto out_put_layout_hdr;
1516         }
1517
1518         if (pnfs_layoutgets_blocked(lo))
1519                 goto out_unlock;
1520         atomic_inc(&lo->plh_outstanding);
1521         spin_unlock(&ino->i_lock);
1522
1523         if (list_empty(&lo->plh_layouts)) {
1524                 /* The lo must be on the clp list if there is any
1525                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1526                  */
1527                 spin_lock(&clp->cl_lock);
1528                 if (list_empty(&lo->plh_layouts))
1529                         list_add_tail(&lo->plh_layouts, &server->layouts);
1530                 spin_unlock(&clp->cl_lock);
1531         }
1532
1533         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1534         if (pg_offset) {
1535                 arg.offset -= pg_offset;
1536                 arg.length += pg_offset;
1537         }
1538         if (arg.length != NFS4_MAX_UINT64)
1539                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1540
1541         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1542         pnfs_clear_retry_layoutget(lo);
1543         atomic_dec(&lo->plh_outstanding);
1544 out_put_layout_hdr:
1545         if (first)
1546                 pnfs_clear_first_layoutget(lo);
1547         pnfs_put_layout_hdr(lo);
1548 out:
1549         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1550                         "(%s, offset: %llu, length: %llu)\n",
1551                         __func__, ino->i_sb->s_id,
1552                         (unsigned long long)NFS_FILEID(ino),
1553                         lseg == NULL ? "not found" : "found",
1554                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1555                         (unsigned long long)pos,
1556                         (unsigned long long)count);
1557         return lseg;
1558 out_unlock:
1559         spin_unlock(&ino->i_lock);
1560         goto out_put_layout_hdr;
1561 }
1562 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1563
1564 struct pnfs_layout_segment *
1565 pnfs_layout_process(struct nfs4_layoutget *lgp)
1566 {
1567         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1568         struct nfs4_layoutget_res *res = &lgp->res;
1569         struct pnfs_layout_segment *lseg;
1570         struct inode *ino = lo->plh_inode;
1571         LIST_HEAD(free_me);
1572         int status = 0;
1573
1574         /* Inject layout blob into I/O device driver */
1575         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1576         if (!lseg || IS_ERR(lseg)) {
1577                 if (!lseg)
1578                         status = -ENOMEM;
1579                 else
1580                         status = PTR_ERR(lseg);
1581                 dprintk("%s: Could not allocate layout: error %d\n",
1582                        __func__, status);
1583                 goto out;
1584         }
1585
1586         init_lseg(lo, lseg);
1587         lseg->pls_range = res->range;
1588
1589         spin_lock(&ino->i_lock);
1590         if (pnfs_layoutgets_blocked(lo)) {
1591                 dprintk("%s forget reply due to state\n", __func__);
1592                 goto out_forget_reply;
1593         }
1594
1595         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1596                 /* existing state ID, make sure the sequence number matches. */
1597                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1598                         dprintk("%s forget reply due to sequence\n", __func__);
1599                         goto out_forget_reply;
1600                 }
1601                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1602         } else {
1603                 /*
1604                  * We got an entirely new state ID.  Mark all segments for the
1605                  * inode invalid, and don't bother validating the stateid
1606                  * sequence number.
1607                  */
1608                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1609
1610                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1611                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1612         }
1613
1614         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1615
1616         pnfs_get_lseg(lseg);
1617         pnfs_layout_insert_lseg(lo, lseg);
1618
1619         if (res->return_on_close)
1620                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1621
1622         spin_unlock(&ino->i_lock);
1623         pnfs_free_lseg_list(&free_me);
1624         return lseg;
1625 out:
1626         return ERR_PTR(status);
1627
1628 out_forget_reply:
1629         spin_unlock(&ino->i_lock);
1630         lseg->pls_layout = lo;
1631         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1632         goto out;
1633 }
1634
1635 static void
1636 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1637                                 struct list_head *tmp_list,
1638                                 struct pnfs_layout_range *return_range)
1639 {
1640         struct pnfs_layout_segment *lseg, *next;
1641
1642         dprintk("%s:Begin lo %p\n", __func__, lo);
1643
1644         if (list_empty(&lo->plh_segs))
1645                 return;
1646
1647         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1648                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1649                         dprintk("%s: marking lseg %p iomode %d "
1650                                 "offset %llu length %llu\n", __func__,
1651                                 lseg, lseg->pls_range.iomode,
1652                                 lseg->pls_range.offset,
1653                                 lseg->pls_range.length);
1654                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1655                         mark_lseg_invalid(lseg, tmp_list);
1656                 }
1657 }
1658
1659 void pnfs_error_mark_layout_for_return(struct inode *inode,
1660                                        struct pnfs_layout_segment *lseg)
1661 {
1662         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1663         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1664         struct pnfs_layout_range range = {
1665                 .iomode = lseg->pls_range.iomode,
1666                 .offset = 0,
1667                 .length = NFS4_MAX_UINT64,
1668         };
1669         LIST_HEAD(free_me);
1670
1671         spin_lock(&inode->i_lock);
1672         /* set failure bit so that pnfs path will be retried later */
1673         pnfs_layout_set_fail_bit(lo, iomode);
1674         if (lo->plh_return_iomode == 0)
1675                 lo->plh_return_iomode = range.iomode;
1676         else if (lo->plh_return_iomode != range.iomode)
1677                 lo->plh_return_iomode = IOMODE_ANY;
1678         /*
1679          * mark all matching lsegs so that we are sure to have no live
1680          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1681          * for how it works.
1682          */
1683         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1684         spin_unlock(&inode->i_lock);
1685         pnfs_free_lseg_list(&free_me);
1686 }
1687 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1688
1689 void
1690 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1691 {
1692         u64 rd_size = req->wb_bytes;
1693
1694         if (pgio->pg_lseg == NULL) {
1695                 if (pgio->pg_dreq == NULL)
1696                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1697                 else
1698                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1699
1700                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1701                                                    req->wb_context,
1702                                                    req_offset(req),
1703                                                    rd_size,
1704                                                    IOMODE_READ,
1705                                                    GFP_KERNEL);
1706         }
1707         /* If no lseg, fall back to read through mds */
1708         if (pgio->pg_lseg == NULL)
1709                 nfs_pageio_reset_read_mds(pgio);
1710
1711 }
1712 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1713
1714 void
1715 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1716                            struct nfs_page *req, u64 wb_size)
1717 {
1718         if (pgio->pg_lseg == NULL)
1719                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1720                                                    req->wb_context,
1721                                                    req_offset(req),
1722                                                    wb_size,
1723                                                    IOMODE_RW,
1724                                                    GFP_NOFS);
1725         /* If no lseg, fall back to write through mds */
1726         if (pgio->pg_lseg == NULL)
1727                 nfs_pageio_reset_write_mds(pgio);
1728 }
1729 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1730
1731 void
1732 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1733 {
1734         if (desc->pg_lseg) {
1735                 pnfs_put_lseg(desc->pg_lseg);
1736                 desc->pg_lseg = NULL;
1737         }
1738 }
1739 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1740
1741 /*
1742  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1743  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1744  */
1745 size_t
1746 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1747                      struct nfs_page *prev, struct nfs_page *req)
1748 {
1749         unsigned int size;
1750         u64 seg_end, req_start, seg_left;
1751
1752         size = nfs_generic_pg_test(pgio, prev, req);
1753         if (!size)
1754                 return 0;
1755
1756         /*
1757          * 'size' contains the number of bytes left in the current page (up
1758          * to the original size asked for in @req->wb_bytes).
1759          *
1760          * Calculate how many bytes are left in the layout segment
1761          * and if there are less bytes than 'size', return that instead.
1762          *
1763          * Please also note that 'end_offset' is actually the offset of the
1764          * first byte that lies outside the pnfs_layout_range. FIXME?
1765          *
1766          */
1767         if (pgio->pg_lseg) {
1768                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1769                                      pgio->pg_lseg->pls_range.length);
1770                 req_start = req_offset(req);
1771                 WARN_ON_ONCE(req_start >= seg_end);
1772                 /* start of request is past the last byte of this segment */
1773                 if (req_start >= seg_end) {
1774                         /* reference the new lseg */
1775                         if (pgio->pg_ops->pg_cleanup)
1776                                 pgio->pg_ops->pg_cleanup(pgio);
1777                         if (pgio->pg_ops->pg_init)
1778                                 pgio->pg_ops->pg_init(pgio, req);
1779                         return 0;
1780                 }
1781
1782                 /* adjust 'size' iff there are fewer bytes left in the
1783                  * segment than what nfs_generic_pg_test returned */
1784                 seg_left = seg_end - req_start;
1785                 if (seg_left < size)
1786                         size = (unsigned int)seg_left;
1787         }
1788
1789         return size;
1790 }
1791 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1792
1793 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1794 {
1795         struct nfs_pageio_descriptor pgio;
1796
1797         /* Resend all requests through the MDS */
1798         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1799                               hdr->completion_ops);
1800         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1801         return nfs_pageio_resend(&pgio, hdr);
1802 }
1803 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1804
1805 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1806 {
1807
1808         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1809         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1810             PNFS_LAYOUTRET_ON_ERROR) {
1811                 pnfs_return_layout(hdr->inode);
1812         }
1813         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1814                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1815 }
1816
1817 /*
1818  * Called by non rpc-based layout drivers
1819  */
1820 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1821 {
1822         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1823         if (!hdr->pnfs_error) {
1824                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1825                                 hdr->mds_offset + hdr->res.count);
1826                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1827         } else
1828                 pnfs_ld_handle_write_error(hdr);
1829         hdr->mds_ops->rpc_release(hdr);
1830 }
1831 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1832
1833 static void
1834 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1835                 struct nfs_pgio_header *hdr)
1836 {
1837         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1838
1839         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1840                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1841                 nfs_pageio_reset_write_mds(desc);
1842                 mirror->pg_recoalesce = 1;
1843         }
1844         nfs_pgio_data_destroy(hdr);
1845         hdr->release(hdr);
1846 }
1847
1848 static enum pnfs_try_status
1849 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1850                         const struct rpc_call_ops *call_ops,
1851                         struct pnfs_layout_segment *lseg,
1852                         int how)
1853 {
1854         struct inode *inode = hdr->inode;
1855         enum pnfs_try_status trypnfs;
1856         struct nfs_server *nfss = NFS_SERVER(inode);
1857
1858         hdr->mds_ops = call_ops;
1859
1860         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1861                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1862         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1863         if (trypnfs != PNFS_NOT_ATTEMPTED)
1864                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1865         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1866         return trypnfs;
1867 }
1868
1869 static void
1870 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1871               struct nfs_pgio_header *hdr, int how)
1872 {
1873         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1874         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1875         enum pnfs_try_status trypnfs;
1876
1877         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1878         if (trypnfs == PNFS_NOT_ATTEMPTED)
1879                 pnfs_write_through_mds(desc, hdr);
1880 }
1881
1882 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1883 {
1884         pnfs_put_lseg(hdr->lseg);
1885         nfs_pgio_header_free(hdr);
1886 }
1887
1888 int
1889 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1890 {
1891         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1892
1893         struct nfs_pgio_header *hdr;
1894         int ret;
1895
1896         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1897         if (!hdr) {
1898                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1899                 return -ENOMEM;
1900         }
1901         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1902
1903         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1904         ret = nfs_generic_pgio(desc, hdr);
1905         if (!ret)
1906                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1907
1908         return ret;
1909 }
1910 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1911
1912 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1913 {
1914         struct nfs_pageio_descriptor pgio;
1915
1916         /* Resend all requests through the MDS */
1917         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1918         return nfs_pageio_resend(&pgio, hdr);
1919 }
1920 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1921
1922 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1923 {
1924         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1925         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1926             PNFS_LAYOUTRET_ON_ERROR) {
1927                 pnfs_return_layout(hdr->inode);
1928         }
1929         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1930                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1931 }
1932
1933 /*
1934  * Called by non rpc-based layout drivers
1935  */
1936 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1937 {
1938         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1939         if (likely(!hdr->pnfs_error)) {
1940                 __nfs4_read_done_cb(hdr);
1941                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1942         } else
1943                 pnfs_ld_handle_read_error(hdr);
1944         hdr->mds_ops->rpc_release(hdr);
1945 }
1946 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1947
1948 static void
1949 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1950                 struct nfs_pgio_header *hdr)
1951 {
1952         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1953
1954         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1955                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1956                 nfs_pageio_reset_read_mds(desc);
1957                 mirror->pg_recoalesce = 1;
1958         }
1959         nfs_pgio_data_destroy(hdr);
1960         hdr->release(hdr);
1961 }
1962
1963 /*
1964  * Call the appropriate parallel I/O subsystem read function.
1965  */
1966 static enum pnfs_try_status
1967 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1968                        const struct rpc_call_ops *call_ops,
1969                        struct pnfs_layout_segment *lseg)
1970 {
1971         struct inode *inode = hdr->inode;
1972         struct nfs_server *nfss = NFS_SERVER(inode);
1973         enum pnfs_try_status trypnfs;
1974
1975         hdr->mds_ops = call_ops;
1976
1977         dprintk("%s: Reading ino:%lu %u@%llu\n",
1978                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1979
1980         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1981         if (trypnfs != PNFS_NOT_ATTEMPTED)
1982                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1983         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1984         return trypnfs;
1985 }
1986
1987 /* Resend all requests through pnfs. */
1988 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
1989 {
1990         struct nfs_pageio_descriptor pgio;
1991
1992         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
1993         return nfs_pageio_resend(&pgio, hdr);
1994 }
1995 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
1996
1997 static void
1998 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1999 {
2000         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2001         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2002         enum pnfs_try_status trypnfs;
2003         int err = 0;
2004
2005         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2006         if (trypnfs == PNFS_TRY_AGAIN)
2007                 err = pnfs_read_resend_pnfs(hdr);
2008         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2009                 pnfs_read_through_mds(desc, hdr);
2010 }
2011
2012 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2013 {
2014         pnfs_put_lseg(hdr->lseg);
2015         nfs_pgio_header_free(hdr);
2016 }
2017
2018 int
2019 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2020 {
2021         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2022
2023         struct nfs_pgio_header *hdr;
2024         int ret;
2025
2026         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2027         if (!hdr) {
2028                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2029                 return -ENOMEM;
2030         }
2031         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2032         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2033         ret = nfs_generic_pgio(desc, hdr);
2034         if (!ret)
2035                 pnfs_do_read(desc, hdr);
2036         return ret;
2037 }
2038 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2039
2040 static void pnfs_clear_layoutcommitting(struct inode *inode)
2041 {
2042         unsigned long *bitlock = &NFS_I(inode)->flags;
2043
2044         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2045         smp_mb__after_atomic();
2046         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2047 }
2048
2049 /*
2050  * There can be multiple RW segments.
2051  */
2052 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2053 {
2054         struct pnfs_layout_segment *lseg;
2055
2056         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2057                 if (lseg->pls_range.iomode == IOMODE_RW &&
2058                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2059                         list_add(&lseg->pls_lc_list, listp);
2060         }
2061 }
2062
2063 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2064 {
2065         struct pnfs_layout_segment *lseg, *tmp;
2066
2067         /* Matched by references in pnfs_set_layoutcommit */
2068         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2069                 list_del_init(&lseg->pls_lc_list);
2070                 pnfs_put_lseg(lseg);
2071         }
2072
2073         pnfs_clear_layoutcommitting(inode);
2074 }
2075
2076 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2077 {
2078         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2079 }
2080 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2081
2082 void
2083 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2084                 loff_t end_pos)
2085 {
2086         struct nfs_inode *nfsi = NFS_I(inode);
2087         bool mark_as_dirty = false;
2088
2089         spin_lock(&inode->i_lock);
2090         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2091                 nfsi->layout->plh_lwb = end_pos;
2092                 mark_as_dirty = true;
2093                 dprintk("%s: Set layoutcommit for inode %lu ",
2094                         __func__, inode->i_ino);
2095         } else if (end_pos > nfsi->layout->plh_lwb)
2096                 nfsi->layout->plh_lwb = end_pos;
2097         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2098                 /* references matched in nfs4_layoutcommit_release */
2099                 pnfs_get_lseg(lseg);
2100         }
2101         spin_unlock(&inode->i_lock);
2102         dprintk("%s: lseg %p end_pos %llu\n",
2103                 __func__, lseg, nfsi->layout->plh_lwb);
2104
2105         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2106          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2107         if (mark_as_dirty)
2108                 mark_inode_dirty_sync(inode);
2109 }
2110 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2111
2112 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2113 {
2114         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2115
2116         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2117                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2118         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2119 }
2120
2121 /*
2122  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2123  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2124  * data to disk to allow the server to recover the data if it crashes.
2125  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2126  * is off, and a COMMIT is sent to a data server, or
2127  * if WRITEs to a data server return NFS_DATA_SYNC.
2128  */
2129 int
2130 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2131 {
2132         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2133         struct nfs4_layoutcommit_data *data;
2134         struct nfs_inode *nfsi = NFS_I(inode);
2135         loff_t end_pos;
2136         int status;
2137
2138         if (!pnfs_layoutcommit_outstanding(inode))
2139                 return 0;
2140
2141         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2142
2143         status = -EAGAIN;
2144         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2145                 if (!sync)
2146                         goto out;
2147                 status = wait_on_bit_lock_action(&nfsi->flags,
2148                                 NFS_INO_LAYOUTCOMMITTING,
2149                                 nfs_wait_bit_killable,
2150                                 TASK_KILLABLE);
2151                 if (status)
2152                         goto out;
2153         }
2154
2155         status = -ENOMEM;
2156         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2157         data = kzalloc(sizeof(*data), GFP_NOFS);
2158         if (!data)
2159                 goto clear_layoutcommitting;
2160
2161         status = 0;
2162         spin_lock(&inode->i_lock);
2163         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2164                 goto out_unlock;
2165
2166         INIT_LIST_HEAD(&data->lseg_list);
2167         pnfs_list_write_lseg(inode, &data->lseg_list);
2168
2169         end_pos = nfsi->layout->plh_lwb;
2170
2171         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2172         spin_unlock(&inode->i_lock);
2173
2174         data->args.inode = inode;
2175         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2176         nfs_fattr_init(&data->fattr);
2177         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2178         data->res.fattr = &data->fattr;
2179         data->args.lastbytewritten = end_pos - 1;
2180         data->res.server = NFS_SERVER(inode);
2181
2182         if (ld->prepare_layoutcommit) {
2183                 status = ld->prepare_layoutcommit(&data->args);
2184                 if (status) {
2185                         put_rpccred(data->cred);
2186                         spin_lock(&inode->i_lock);
2187                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2188                         if (end_pos > nfsi->layout->plh_lwb)
2189                                 nfsi->layout->plh_lwb = end_pos;
2190                         goto out_unlock;
2191                 }
2192         }
2193
2194
2195         status = nfs4_proc_layoutcommit(data, sync);
2196 out:
2197         if (status)
2198                 mark_inode_dirty_sync(inode);
2199         dprintk("<-- %s status %d\n", __func__, status);
2200         return status;
2201 out_unlock:
2202         spin_unlock(&inode->i_lock);
2203         kfree(data);
2204 clear_layoutcommitting:
2205         pnfs_clear_layoutcommitting(inode);
2206         goto out;
2207 }
2208 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2209
2210 int
2211 pnfs_generic_sync(struct inode *inode, bool datasync)
2212 {
2213         return pnfs_layoutcommit_inode(inode, true);
2214 }
2215 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2216
2217 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2218 {
2219         struct nfs4_threshold *thp;
2220
2221         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2222         if (!thp) {
2223                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2224                 return NULL;
2225         }
2226         return thp;
2227 }
2228
2229 #if IS_ENABLED(CONFIG_NFS_V4_2)
2230 int
2231 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2232 {
2233         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2234         struct nfs_server *server = NFS_SERVER(inode);
2235         struct nfs_inode *nfsi = NFS_I(inode);
2236         struct nfs42_layoutstat_data *data;
2237         struct pnfs_layout_hdr *hdr;
2238         int status = 0;
2239
2240         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2241                 goto out;
2242
2243         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2244                 goto out;
2245
2246         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2247                 goto out;
2248
2249         spin_lock(&inode->i_lock);
2250         if (!NFS_I(inode)->layout) {
2251                 spin_unlock(&inode->i_lock);
2252                 goto out;
2253         }
2254         hdr = NFS_I(inode)->layout;
2255         pnfs_get_layout_hdr(hdr);
2256         spin_unlock(&inode->i_lock);
2257
2258         data = kzalloc(sizeof(*data), gfp_flags);
2259         if (!data) {
2260                 status = -ENOMEM;
2261                 goto out_put;
2262         }
2263
2264         data->args.fh = NFS_FH(inode);
2265         data->args.inode = inode;
2266         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2267         status = ld->prepare_layoutstats(&data->args);
2268         if (status)
2269                 goto out_free;
2270
2271         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2272
2273 out:
2274         dprintk("%s returns %d\n", __func__, status);
2275         return status;
2276
2277 out_free:
2278         kfree(data);
2279 out_put:
2280         pnfs_put_layout_hdr(hdr);
2281         smp_mb__before_atomic();
2282         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2283         smp_mb__after_atomic();
2284         goto out;
2285 }
2286 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2287 #endif