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_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 void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
 366                 struct pnfs_layout_hdr *lo, struct inode *inode)
 367 {
 368         lo = lseg->pls_layout;
 369         inode = lo->plh_inode;
 370
 371         spin_lock(&inode->i_lock);
 372         if (pnfs_layout_need_return(lo, lseg)) {
 373                 nfs4_stateid stateid;
 374                 enum pnfs_iomode iomode;
 375
 376                 stateid = lo->plh_stateid;
 377                 iomode = lo->plh_return_iomode;
 378                 /* decreased in pnfs_send_layoutreturn() */
 379                 lo->plh_block_lgets++;
 380                 lo->plh_return_iomode = 0;
 381                 spin_unlock(&inode->i_lock);
 382                 pnfs_get_layout_hdr(lo);
 383
 384                 /* Send an async layoutreturn so we dont deadlock */
 385                 pnfs_send_layoutreturn(lo, stateid, iomode, false);
 386         } else
 387                 spin_unlock(&inode->i_lock);
 388 }
 389
 390 void
 391 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 392 {
 393         struct pnfs_layout_hdr *lo;
 394         struct inode *inode;
 395
 396         if (!lseg)
 397                 return;
 398
 399         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 400                 atomic_read(&lseg->pls_refcount),
 401                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 402
 403         /* Handle the case where refcount != 1 */
 404         if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
 405                 return;
 406
 407         lo = lseg->pls_layout;
 408         inode = lo->plh_inode;
 409         /* Do we need a layoutreturn? */
 410         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 411                 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
 412
 413         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 414                 pnfs_get_layout_hdr(lo);
 415                 pnfs_layout_remove_lseg(lo, lseg);
 416                 spin_unlock(&inode->i_lock);
 417                 pnfs_free_lseg(lseg);
 418                 pnfs_put_layout_hdr(lo);
 419         }
 420 }
 421 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 422
 423 static void pnfs_free_lseg_async_work(struct work_struct *work)
 424 {
 425         struct pnfs_layout_segment *lseg;
 426         struct pnfs_layout_hdr *lo;
 427
 428         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 429         lo = lseg->pls_layout;
 430
 431         pnfs_free_lseg(lseg);
 432         pnfs_put_layout_hdr(lo);
 433 }
 434
 435 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 436 {
 437         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 438         schedule_work(&lseg->pls_work);
 439 }
 440
 441 void
 442 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 443 {
 444         if (!lseg)
 445                 return;
 446
 447         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 448
 449         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 450                 atomic_read(&lseg->pls_refcount),
 451                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 452         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 453                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 454                 pnfs_get_layout_hdr(lo);
 455                 pnfs_layout_remove_lseg(lo, lseg);
 456                 pnfs_free_lseg_async(lseg);
 457         }
 458 }
 459 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 460
 461 static u64
 462 end_offset(u64 start, u64 len)
 463 {
 464         u64 end;
 465
 466         end = start + len;
 467         return end >= start ? end : NFS4_MAX_UINT64;
 468 }
 469
 470 /*
 471  * is l2 fully contained in l1?
 472  *   start1                             end1
 473  *   [----------------------------------)
 474  *           start2           end2
 475  *           [----------------)
 476  */
 477 static bool
 478 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 479                  const struct pnfs_layout_range *l2)
 480 {
 481         u64 start1 = l1->offset;
 482         u64 end1 = end_offset(start1, l1->length);
 483         u64 start2 = l2->offset;
 484         u64 end2 = end_offset(start2, l2->length);
 485
 486         return (start1 <= start2) && (end1 >= end2);
 487 }
 488
 489 /*
 490  * is l1 and l2 intersecting?
 491  *   start1                             end1
 492  *   [----------------------------------)
 493  *                              start2           end2
 494  *                              [----------------)
 495  */
 496 static bool
 497 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 498                     const struct pnfs_layout_range *l2)
 499 {
 500         u64 start1 = l1->offset;
 501         u64 end1 = end_offset(start1, l1->length);
 502         u64 start2 = l2->offset;
 503         u64 end2 = end_offset(start2, l2->length);
 504
 505         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 506                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 507 }
 508
 509 static bool
 510 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 511                  const struct pnfs_layout_range *recall_range)
 512 {
 513         return (recall_range->iomode == IOMODE_ANY ||
 514                 lseg_range->iomode == recall_range->iomode) &&
 515                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 516 }
 517
 518 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 519                 struct list_head *tmp_list)
 520 {
 521         if (!atomic_dec_and_test(&lseg->pls_refcount))
 522                 return false;
 523         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 524         list_add(&lseg->pls_list, tmp_list);
 525         return true;
 526 }
 527
 528 /* Returns 1 if lseg is removed from list, 0 otherwise */
 529 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 530                              struct list_head *tmp_list)
 531 {
 532         int rv = 0;
 533
 534         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 535                 /* Remove the reference keeping the lseg in the
 536                  * list.  It will now be removed when all
 537                  * outstanding io is finished.
 538                  */
 539                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 540                         atomic_read(&lseg->pls_refcount));
 541                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 542                         rv = 1;
 543         }
 544         return rv;
 545 }
 546
 547 /* Returns count of number of matching invalid lsegs remaining in list
 548  * after call.
 549  */
 550 int
 551 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 552                             struct list_head *tmp_list,
 553                             struct pnfs_layout_range *recall_range)
 554 {
 555         struct pnfs_layout_segment *lseg, *next;
 556         int invalid = 0, removed = 0;
 557
 558         dprintk("%s:Begin lo %p\n", __func__, lo);
 559
 560         if (list_empty(&lo->plh_segs))
 561                 return 0;
 562         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 563                 if (!recall_range ||
 564                     should_free_lseg(&lseg->pls_range, recall_range)) {
 565                         dprintk("%s: freeing lseg %p iomode %d "
 566                                 "offset %llu length %llu\n", __func__,
 567                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 568                                 lseg->pls_range.length);
 569                         invalid++;
 570                         removed += mark_lseg_invalid(lseg, tmp_list);
 571                 }
 572         dprintk("%s:Return %i\n", __func__, invalid - removed);
 573         return invalid - removed;
 574 }
 575
 576 /* note free_me must contain lsegs from a single layout_hdr */
 577 void
 578 pnfs_free_lseg_list(struct list_head *free_me)
 579 {
 580         struct pnfs_layout_segment *lseg, *tmp;
 581
 582         if (list_empty(free_me))
 583                 return;
 584
 585         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 586                 list_del(&lseg->pls_list);
 587                 pnfs_free_lseg(lseg);
 588         }
 589 }
 590
 591 void
 592 pnfs_destroy_layout(struct nfs_inode *nfsi)
 593 {
 594         struct pnfs_layout_hdr *lo;
 595         LIST_HEAD(tmp_list);
 596
 597         spin_lock(&nfsi->vfs_inode.i_lock);
 598         lo = nfsi->layout;
 599         if (lo) {
 600                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 601                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 602                 pnfs_get_layout_hdr(lo);
 603                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 604                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 605                 pnfs_clear_retry_layoutget(lo);
 606                 spin_unlock(&nfsi->vfs_inode.i_lock);
 607                 pnfs_free_lseg_list(&tmp_list);
 608                 pnfs_put_layout_hdr(lo);
 609         } else
 610                 spin_unlock(&nfsi->vfs_inode.i_lock);
 611 }
 612 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 613
 614 static bool
 615 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 616                 struct list_head *layout_list)
 617 {
 618         struct pnfs_layout_hdr *lo;
 619         bool ret = false;
 620
 621         spin_lock(&inode->i_lock);
 622         lo = NFS_I(inode)->layout;
 623         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 624                 pnfs_get_layout_hdr(lo);
 625                 list_add(&lo->plh_bulk_destroy, layout_list);
 626                 ret = true;
 627         }
 628         spin_unlock(&inode->i_lock);
 629         return ret;
 630 }
 631
 632 /* Caller must hold rcu_read_lock and clp->cl_lock */
 633 static int
 634 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 635                 struct nfs_server *server,
 636                 struct list_head *layout_list)
 637 {
 638         struct pnfs_layout_hdr *lo, *next;
 639         struct inode *inode;
 640
 641         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 642                 inode = igrab(lo->plh_inode);
 643                 if (inode == NULL)
 644                         continue;
 645                 list_del_init(&lo->plh_layouts);
 646                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 647                         continue;
 648                 rcu_read_unlock();
 649                 spin_unlock(&clp->cl_lock);
 650                 iput(inode);
 651                 spin_lock(&clp->cl_lock);
 652                 rcu_read_lock();
 653                 return -EAGAIN;
 654         }
 655         return 0;
 656 }
 657
 658 static int
 659 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 660                 bool is_bulk_recall)
 661 {
 662         struct pnfs_layout_hdr *lo;
 663         struct inode *inode;
 664         struct pnfs_layout_range range = {
 665                 .iomode = IOMODE_ANY,
 666                 .offset = 0,
 667                 .length = NFS4_MAX_UINT64,
 668         };
 669         LIST_HEAD(lseg_list);
 670         int ret = 0;
 671
 672         while (!list_empty(layout_list)) {
 673                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 674                                 plh_bulk_destroy);
 675                 dprintk("%s freeing layout for inode %lu\n", __func__,
 676                         lo->plh_inode->i_ino);
 677                 inode = lo->plh_inode;
 678
 679                 pnfs_layoutcommit_inode(inode, false);
 680
 681                 spin_lock(&inode->i_lock);
 682                 list_del_init(&lo->plh_bulk_destroy);
 683                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 684                 if (is_bulk_recall)
 685                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 686                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 687                         ret = -EAGAIN;
 688                 spin_unlock(&inode->i_lock);
 689                 pnfs_free_lseg_list(&lseg_list);
 690                 pnfs_put_layout_hdr(lo);
 691                 iput(inode);
 692         }
 693         return ret;
 694 }
 695
 696 int
 697 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 698                 struct nfs_fsid *fsid,
 699                 bool is_recall)
 700 {
 701         struct nfs_server *server;
 702         LIST_HEAD(layout_list);
 703
 704         spin_lock(&clp->cl_lock);
 705         rcu_read_lock();
 706 restart:
 707         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 708                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 709                         continue;
 710                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 711                                 server,
 712                                 &layout_list) != 0)
 713                         goto restart;
 714         }
 715         rcu_read_unlock();
 716         spin_unlock(&clp->cl_lock);
 717
 718         if (list_empty(&layout_list))
 719                 return 0;
 720         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 721 }
 722
 723 int
 724 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 725                 bool is_recall)
 726 {
 727         struct nfs_server *server;
 728         LIST_HEAD(layout_list);
 729
 730         spin_lock(&clp->cl_lock);
 731         rcu_read_lock();
 732 restart:
 733         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 734                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 735                                         server,
 736                                         &layout_list) != 0)
 737                         goto restart;
 738         }
 739         rcu_read_unlock();
 740         spin_unlock(&clp->cl_lock);
 741
 742         if (list_empty(&layout_list))
 743                 return 0;
 744         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 745 }
 746
 747 /*
 748  * Called by the state manger to remove all layouts established under an
 749  * expired lease.
 750  */
 751 void
 752 pnfs_destroy_all_layouts(struct nfs_client *clp)
 753 {
 754         nfs4_deviceid_mark_client_invalid(clp);
 755         nfs4_deviceid_purge_client(clp);
 756
 757         pnfs_destroy_layouts_byclid(clp, false);
 758 }
 759
 760 /*
 761  * Compare 2 layout stateid sequence ids, to see which is newer,
 762  * taking into account wraparound issues.
 763  */
 764 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 765 {
 766         return (s32)(s1 - s2) > 0;
 767 }
 768
 769 /* update lo->plh_stateid with new if is more recent */
 770 void
 771 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 772                         bool update_barrier)
 773 {
 774         u32 oldseq, newseq, new_barrier;
 775         int empty = list_empty(&lo->plh_segs);
 776
 777         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 778         newseq = be32_to_cpu(new->seqid);
 779         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 780                 nfs4_stateid_copy(&lo->plh_stateid, new);
 781                 if (update_barrier) {
 782                         new_barrier = be32_to_cpu(new->seqid);
 783                 } else {
 784                         /* Because of wraparound, we want to keep the barrier
 785                          * "close" to the current seqids.
 786                          */
 787                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 788                 }
 789                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 790                         lo->plh_barrier = new_barrier;
 791         }
 792 }
 793
 794 static bool
 795 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 796                 const nfs4_stateid *stateid)
 797 {
 798         u32 seqid = be32_to_cpu(stateid->seqid);
 799
 800         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 801 }
 802
 803 static bool
 804 pnfs_layout_returning(const struct pnfs_layout_hdr *lo,
 805                       struct pnfs_layout_range *range)
 806 {
 807         return test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 808                 (lo->plh_return_iomode == IOMODE_ANY ||
 809                  lo->plh_return_iomode == range->iomode);
 810 }
 811
 812 /* lget is set to 1 if called from inside send_layoutget call chain */
 813 static bool
 814 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo,
 815                         struct pnfs_layout_range *range, int lget)
 816 {
 817         return lo->plh_block_lgets ||
 818                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 819                 (list_empty(&lo->plh_segs) &&
 820                  (atomic_read(&lo->plh_outstanding) > lget)) ||
 821                 pnfs_layout_returning(lo, range);
 822 }
 823
 824 int
 825 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 826                               struct pnfs_layout_range *range,
 827                               struct nfs4_state *open_state)
 828 {
 829         int status = 0;
 830
 831         dprintk("--> %s\n", __func__);
 832         spin_lock(&lo->plh_inode->i_lock);
 833         if (pnfs_layoutgets_blocked(lo, range, 1)) {
 834                 status = -EAGAIN;
 835         } else if (!nfs4_valid_open_stateid(open_state)) {
 836                 status = -EBADF;
 837         } else if (list_empty(&lo->plh_segs) ||
 838                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 839                 int seq;
 840
 841                 do {
 842                         seq = read_seqbegin(&open_state->seqlock);
 843                         nfs4_stateid_copy(dst, &open_state->stateid);
 844                 } while (read_seqretry(&open_state->seqlock, seq));
 845         } else
 846                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 847         spin_unlock(&lo->plh_inode->i_lock);
 848         dprintk("<-- %s\n", __func__);
 849         return status;
 850 }
 851
 852 /*
 853 * Get layout from server.
 854 *    for now, assume that whole file layouts are requested.
 855 *    arg->offset: 0
 856 *    arg->length: all ones
 857 */
 858 static struct pnfs_layout_segment *
 859 send_layoutget(struct pnfs_layout_hdr *lo,
 860            struct nfs_open_context *ctx,
 861            struct pnfs_layout_range *range,
 862            gfp_t gfp_flags)
 863 {
 864         struct inode *ino = lo->plh_inode;
 865         struct nfs_server *server = NFS_SERVER(ino);
 866         struct nfs4_layoutget *lgp;
 867         struct pnfs_layout_segment *lseg;
 868
 869         dprintk("--> %s\n", __func__);
 870
 871         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 872         if (lgp == NULL)
 873                 return NULL;
 874
 875         lgp->args.minlength = PAGE_CACHE_SIZE;
 876         if (lgp->args.minlength > range->length)
 877                 lgp->args.minlength = range->length;
 878         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 879         lgp->args.range = *range;
 880         lgp->args.type = server->pnfs_curr_ld->id;
 881         lgp->args.inode = ino;
 882         lgp->args.ctx = get_nfs_open_context(ctx);
 883         lgp->gfp_flags = gfp_flags;
 884         lgp->cred = lo->plh_lc_cred;
 885
 886         /* Synchronously retrieve layout information from server and
 887          * store in lseg.
 888          */
 889         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 890         if (IS_ERR(lseg)) {
 891                 switch (PTR_ERR(lseg)) {
 892                 case -ENOMEM:
 893                 case -ERESTARTSYS:
 894                         break;
 895                 default:
 896                         /* remember that LAYOUTGET failed and suspend trying */
 897                         pnfs_layout_io_set_failed(lo, range->iomode);
 898                 }
 899                 return NULL;
 900         } else
 901                 pnfs_layout_clear_fail_bit(lo,
 902                                 pnfs_iomode_to_fail_bit(range->iomode));
 903
 904         return lseg;
 905 }
 906
 907 static void pnfs_clear_layoutcommit(struct inode *inode,
 908                 struct list_head *head)
 909 {
 910         struct nfs_inode *nfsi = NFS_I(inode);
 911         struct pnfs_layout_segment *lseg, *tmp;
 912
 913         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 914                 return;
 915         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 916                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 917                         continue;
 918                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 919         }
 920 }
 921
 922 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 923 {
 924         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 925         smp_mb__after_atomic();
 926         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 927 }
 928
 929 static int
 930 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 931                        enum pnfs_iomode iomode, bool sync)
 932 {
 933         struct inode *ino = lo->plh_inode;
 934         struct nfs4_layoutreturn *lrp;
 935         int status = 0;
 936
 937         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 938         if (unlikely(lrp == NULL)) {
 939                 status = -ENOMEM;
 940                 spin_lock(&ino->i_lock);
 941                 lo->plh_block_lgets--;
 942                 pnfs_clear_layoutreturn_waitbit(lo);
 943                 rpc_wake_up(&NFS_SERVER(ino)->roc_rpcwaitq);
 944                 spin_unlock(&ino->i_lock);
 945                 pnfs_put_layout_hdr(lo);
 946                 goto out;
 947         }
 948
 949         lrp->args.stateid = stateid;
 950         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 951         lrp->args.inode = ino;
 952         lrp->args.range.iomode = iomode;
 953         lrp->args.range.offset = 0;
 954         lrp->args.range.length = NFS4_MAX_UINT64;
 955         lrp->args.layout = lo;
 956         lrp->clp = NFS_SERVER(ino)->nfs_client;
 957         lrp->cred = lo->plh_lc_cred;
 958
 959         status = nfs4_proc_layoutreturn(lrp, sync);
 960 out:
 961         dprintk("<-- %s status: %d\n", __func__, status);
 962         return status;
 963 }
 964
 965 /*
 966  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 967  * when the layout segment list is empty.
 968  *
 969  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 970  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 971  * deviceid is marked invalid.
 972  */
 973 int
 974 _pnfs_return_layout(struct inode *ino)
 975 {
 976         struct pnfs_layout_hdr *lo = NULL;
 977         struct nfs_inode *nfsi = NFS_I(ino);
 978         LIST_HEAD(tmp_list);
 979         nfs4_stateid stateid;
 980         int status = 0, empty;
 981
 982         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 983
 984         spin_lock(&ino->i_lock);
 985         lo = nfsi->layout;
 986         if (!lo) {
 987                 spin_unlock(&ino->i_lock);
 988                 dprintk("NFS: %s no layout to return\n", __func__);
 989                 goto out;
 990         }
 991         stateid = nfsi->layout->plh_stateid;
 992         /* Reference matched in nfs4_layoutreturn_release */
 993         pnfs_get_layout_hdr(lo);
 994         empty = list_empty(&lo->plh_segs);
 995         pnfs_clear_layoutcommit(ino, &tmp_list);
 996         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 997
 998         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 999                 struct pnfs_layout_range range = {
1000                         .iomode         = IOMODE_ANY,
1001                         .offset         = 0,
1002                         .length         = NFS4_MAX_UINT64,
1003                 };
1004                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1005         }
1006
1007         /* Don't send a LAYOUTRETURN if list was initially empty */
1008         if (empty) {
1009                 spin_unlock(&ino->i_lock);
1010                 pnfs_put_layout_hdr(lo);
1011                 dprintk("NFS: %s no layout segments to return\n", __func__);
1012                 goto out;
1013         }
1014
1015         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1016         lo->plh_block_lgets++;
1017         spin_unlock(&ino->i_lock);
1018         pnfs_free_lseg_list(&tmp_list);
1019
1020         status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1021 out:
1022         dprintk("<-- %s status: %d\n", __func__, status);
1023         return status;
1024 }
1025 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1026
1027 int
1028 pnfs_commit_and_return_layout(struct inode *inode)
1029 {
1030         struct pnfs_layout_hdr *lo;
1031         int ret;
1032
1033         spin_lock(&inode->i_lock);
1034         lo = NFS_I(inode)->layout;
1035         if (lo == NULL) {
1036                 spin_unlock(&inode->i_lock);
1037                 return 0;
1038         }
1039         pnfs_get_layout_hdr(lo);
1040         /* Block new layoutgets and read/write to ds */
1041         lo->plh_block_lgets++;
1042         spin_unlock(&inode->i_lock);
1043         filemap_fdatawait(inode->i_mapping);
1044         ret = pnfs_layoutcommit_inode(inode, true);
1045         if (ret == 0)
1046                 ret = _pnfs_return_layout(inode);
1047         spin_lock(&inode->i_lock);
1048         lo->plh_block_lgets--;
1049         spin_unlock(&inode->i_lock);
1050         pnfs_put_layout_hdr(lo);
1051         return ret;
1052 }
1053
1054 bool pnfs_roc(struct inode *ino)
1055 {
1056         struct nfs_inode *nfsi = NFS_I(ino);
1057         struct nfs_open_context *ctx;
1058         struct nfs4_state *state;
1059         struct pnfs_layout_hdr *lo;
1060         struct pnfs_layout_segment *lseg, *tmp;
1061         nfs4_stateid stateid;
1062         LIST_HEAD(tmp_list);
1063         bool found = false, layoutreturn = false;
1064
1065         spin_lock(&ino->i_lock);
1066         lo = nfsi->layout;
1067         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
1068             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1069                 goto out_noroc;
1070
1071         /* Don't return layout if we hold a delegation */
1072         if (nfs4_check_delegation(ino, FMODE_READ))
1073                 goto out_noroc;
1074
1075         list_for_each_entry(ctx, &nfsi->open_files, list) {
1076                 state = ctx->state;
1077                 /* Don't return layout if there is open file state */
1078                 if (state != NULL && state->state != 0)
1079                         goto out_noroc;
1080         }
1081
1082         pnfs_clear_retry_layoutget(lo);
1083         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1084                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1085                         mark_lseg_invalid(lseg, &tmp_list);
1086                         found = true;
1087                 }
1088         if (!found)
1089                 goto out_noroc;
1090         lo->plh_block_lgets++;
1091         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
1092         spin_unlock(&ino->i_lock);
1093         pnfs_free_lseg_list(&tmp_list);
1094         pnfs_layoutcommit_inode(ino, true);
1095         return true;
1096
1097 out_noroc:
1098         if (lo) {
1099                 stateid = lo->plh_stateid;
1100                 layoutreturn =
1101                         test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1102                                            &lo->plh_flags);
1103                 if (layoutreturn) {
1104                         lo->plh_block_lgets++;
1105                         pnfs_get_layout_hdr(lo);
1106                 }
1107         }
1108         spin_unlock(&ino->i_lock);
1109         if (layoutreturn) {
1110                 pnfs_layoutcommit_inode(ino, true);
1111                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1112         }
1113         return false;
1114 }
1115
1116 void pnfs_roc_release(struct inode *ino)
1117 {
1118         struct pnfs_layout_hdr *lo;
1119
1120         spin_lock(&ino->i_lock);
1121         lo = NFS_I(ino)->layout;
1122         lo->plh_block_lgets--;
1123         if (atomic_dec_and_test(&lo->plh_refcount)) {
1124                 pnfs_detach_layout_hdr(lo);
1125                 spin_unlock(&ino->i_lock);
1126                 pnfs_free_layout_hdr(lo);
1127         } else
1128                 spin_unlock(&ino->i_lock);
1129 }
1130
1131 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1132 {
1133         struct pnfs_layout_hdr *lo;
1134
1135         spin_lock(&ino->i_lock);
1136         lo = NFS_I(ino)->layout;
1137         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1138                 lo->plh_barrier = barrier;
1139         spin_unlock(&ino->i_lock);
1140 }
1141
1142 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
1143 {
1144         struct nfs_inode *nfsi = NFS_I(ino);
1145         struct pnfs_layout_hdr *lo;
1146         struct pnfs_layout_segment *lseg;
1147         nfs4_stateid stateid;
1148         u32 current_seqid;
1149         bool found = false, layoutreturn = false;
1150
1151         spin_lock(&ino->i_lock);
1152         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1153                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1154                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1155                         found = true;
1156                         goto out;
1157                 }
1158         lo = nfsi->layout;
1159         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1160
1161         /* Since close does not return a layout stateid for use as
1162          * a barrier, we choose the worst-case barrier.
1163          */
1164         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1165 out:
1166         if (!found) {
1167                 stateid = lo->plh_stateid;
1168                 layoutreturn =
1169                         test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1170                                            &lo->plh_flags);
1171                 if (layoutreturn) {
1172                         lo->plh_block_lgets++;
1173                         pnfs_get_layout_hdr(lo);
1174                 }
1175         }
1176         spin_unlock(&ino->i_lock);
1177         if (layoutreturn) {
1178                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1179                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, false);
1180         }
1181         return found;
1182 }
1183
1184 /*
1185  * Compare two layout segments for sorting into layout cache.
1186  * We want to preferentially return RW over RO layouts, so ensure those
1187  * are seen first.
1188  */
1189 static s64
1190 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1191            const struct pnfs_layout_range *l2)
1192 {
1193         s64 d;
1194
1195         /* high offset > low offset */
1196         d = l1->offset - l2->offset;
1197         if (d)
1198                 return d;
1199
1200         /* short length > long length */
1201         d = l2->length - l1->length;
1202         if (d)
1203                 return d;
1204
1205         /* read > read/write */
1206         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1207 }
1208
1209 static void
1210 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1211                    struct pnfs_layout_segment *lseg)
1212 {
1213         struct pnfs_layout_segment *lp;
1214
1215         dprintk("%s:Begin\n", __func__);
1216
1217         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1218                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1219                         continue;
1220                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1221                 dprintk("%s: inserted lseg %p "
1222                         "iomode %d offset %llu length %llu before "
1223                         "lp %p iomode %d offset %llu length %llu\n",
1224                         __func__, lseg, lseg->pls_range.iomode,
1225                         lseg->pls_range.offset, lseg->pls_range.length,
1226                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1227                         lp->pls_range.length);
1228                 goto out;
1229         }
1230         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1231         dprintk("%s: inserted lseg %p "
1232                 "iomode %d offset %llu length %llu at tail\n",
1233                 __func__, lseg, lseg->pls_range.iomode,
1234                 lseg->pls_range.offset, lseg->pls_range.length);
1235 out:
1236         pnfs_get_layout_hdr(lo);
1237
1238         dprintk("%s:Return\n", __func__);
1239 }
1240
1241 static struct pnfs_layout_hdr *
1242 alloc_init_layout_hdr(struct inode *ino,
1243                       struct nfs_open_context *ctx,
1244                       gfp_t gfp_flags)
1245 {
1246         struct pnfs_layout_hdr *lo;
1247
1248         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1249         if (!lo)
1250                 return NULL;
1251         atomic_set(&lo->plh_refcount, 1);
1252         INIT_LIST_HEAD(&lo->plh_layouts);
1253         INIT_LIST_HEAD(&lo->plh_segs);
1254         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1255         lo->plh_inode = ino;
1256         lo->plh_lc_cred = get_rpccred(ctx->cred);
1257         return lo;
1258 }
1259
1260 static struct pnfs_layout_hdr *
1261 pnfs_find_alloc_layout(struct inode *ino,
1262                        struct nfs_open_context *ctx,
1263                        gfp_t gfp_flags)
1264 {
1265         struct nfs_inode *nfsi = NFS_I(ino);
1266         struct pnfs_layout_hdr *new = NULL;
1267
1268         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1269
1270         if (nfsi->layout != NULL)
1271                 goto out_existing;
1272         spin_unlock(&ino->i_lock);
1273         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1274         spin_lock(&ino->i_lock);
1275
1276         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1277                 nfsi->layout = new;
1278                 return new;
1279         } else if (new != NULL)
1280                 pnfs_free_layout_hdr(new);
1281 out_existing:
1282         pnfs_get_layout_hdr(nfsi->layout);
1283         return nfsi->layout;
1284 }
1285
1286 /*
1287  * iomode matching rules:
1288  * iomode       lseg    match
1289  * -----        -----   -----
1290  * ANY          READ    true
1291  * ANY          RW      true
1292  * RW           READ    false
1293  * RW           RW      true
1294  * READ         READ    true
1295  * READ         RW      true
1296  */
1297 static bool
1298 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1299                  const struct pnfs_layout_range *range)
1300 {
1301         struct pnfs_layout_range range1;
1302
1303         if ((range->iomode == IOMODE_RW &&
1304              ls_range->iomode != IOMODE_RW) ||
1305             !pnfs_lseg_range_intersecting(ls_range, range))
1306                 return 0;
1307
1308         /* range1 covers only the first byte in the range */
1309         range1 = *range;
1310         range1.length = 1;
1311         return pnfs_lseg_range_contained(ls_range, &range1);
1312 }
1313
1314 /*
1315  * lookup range in layout
1316  */
1317 static struct pnfs_layout_segment *
1318 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1319                 struct pnfs_layout_range *range)
1320 {
1321         struct pnfs_layout_segment *lseg, *ret = NULL;
1322
1323         dprintk("%s:Begin\n", __func__);
1324
1325         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1326                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1327                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1328                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1329                         ret = pnfs_get_lseg(lseg);
1330                         break;
1331                 }
1332                 if (lseg->pls_range.offset > range->offset)
1333                         break;
1334         }
1335
1336         dprintk("%s:Return lseg %p ref %d\n",
1337                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1338         return ret;
1339 }
1340
1341 /*
1342  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1343  * to the MDS or over pNFS
1344  *
1345  * The nfs_inode read_io and write_io fields are cumulative counters reset
1346  * when there are no layout segments. Note that in pnfs_update_layout iomode
1347  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1348  * WRITE request.
1349  *
1350  * A return of true means use MDS I/O.
1351  *
1352  * From rfc 5661:
1353  * If a file's size is smaller than the file size threshold, data accesses
1354  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1355  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1356  * server.  If both file size and I/O size are provided, the client SHOULD
1357  * reach or exceed  both thresholds before sending its read or write
1358  * requests to the data server.
1359  */
1360 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1361                                      struct inode *ino, int iomode)
1362 {
1363         struct nfs4_threshold *t = ctx->mdsthreshold;
1364         struct nfs_inode *nfsi = NFS_I(ino);
1365         loff_t fsize = i_size_read(ino);
1366         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1367
1368         if (t == NULL)
1369                 return ret;
1370
1371         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1372                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1373
1374         switch (iomode) {
1375         case IOMODE_READ:
1376                 if (t->bm & THRESHOLD_RD) {
1377                         dprintk("%s fsize %llu\n", __func__, fsize);
1378                         size_set = true;
1379                         if (fsize < t->rd_sz)
1380                                 size = true;
1381                 }
1382                 if (t->bm & THRESHOLD_RD_IO) {
1383                         dprintk("%s nfsi->read_io %llu\n", __func__,
1384                                 nfsi->read_io);
1385                         io_set = true;
1386                         if (nfsi->read_io < t->rd_io_sz)
1387                                 io = true;
1388                 }
1389                 break;
1390         case IOMODE_RW:
1391                 if (t->bm & THRESHOLD_WR) {
1392                         dprintk("%s fsize %llu\n", __func__, fsize);
1393                         size_set = true;
1394                         if (fsize < t->wr_sz)
1395                                 size = true;
1396                 }
1397                 if (t->bm & THRESHOLD_WR_IO) {
1398                         dprintk("%s nfsi->write_io %llu\n", __func__,
1399                                 nfsi->write_io);
1400                         io_set = true;
1401                         if (nfsi->write_io < t->wr_io_sz)
1402                                 io = true;
1403                 }
1404                 break;
1405         }
1406         if (size_set && io_set) {
1407                 if (size && io)
1408                         ret = true;
1409         } else if (size || io)
1410                 ret = true;
1411
1412         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1413         return ret;
1414 }
1415
1416 /* stop waiting if someone clears NFS_LAYOUT_RETRY_LAYOUTGET bit. */
1417 static int pnfs_layoutget_retry_bit_wait(struct wait_bit_key *key)
1418 {
1419         if (!test_bit(NFS_LAYOUT_RETRY_LAYOUTGET, key->flags))
1420                 return 1;
1421         return nfs_wait_bit_killable(key);
1422 }
1423
1424 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1425 {
1426         /*
1427          * send layoutcommit as it can hold up layoutreturn due to lseg
1428          * reference
1429          */
1430         pnfs_layoutcommit_inode(lo->plh_inode, false);
1431         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1432                                    pnfs_layoutget_retry_bit_wait,
1433                                    TASK_UNINTERRUPTIBLE);
1434 }
1435
1436 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1437 {
1438         unsigned long *bitlock = &lo->plh_flags;
1439
1440         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1441         smp_mb__after_atomic();
1442         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1443 }
1444
1445 /*
1446  * Layout segment is retreived from the server if not cached.
1447  * The appropriate layout segment is referenced and returned to the caller.
1448  */
1449 struct pnfs_layout_segment *
1450 pnfs_update_layout(struct inode *ino,
1451                    struct nfs_open_context *ctx,
1452                    loff_t pos,
1453                    u64 count,
1454                    enum pnfs_iomode iomode,
1455                    gfp_t gfp_flags)
1456 {
1457         struct pnfs_layout_range arg = {
1458                 .iomode = iomode,
1459                 .offset = pos,
1460                 .length = count,
1461         };
1462         unsigned pg_offset;
1463         struct nfs_server *server = NFS_SERVER(ino);
1464         struct nfs_client *clp = server->nfs_client;
1465         struct pnfs_layout_hdr *lo;
1466         struct pnfs_layout_segment *lseg = NULL;
1467         bool first;
1468
1469         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1470                 goto out;
1471
1472         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1473                 goto out;
1474
1475 lookup_again:
1476         first = false;
1477         spin_lock(&ino->i_lock);
1478         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1479         if (lo == NULL) {
1480                 spin_unlock(&ino->i_lock);
1481                 goto out;
1482         }
1483
1484         /* Do we even need to bother with this? */
1485         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1486                 dprintk("%s matches recall, use MDS\n", __func__);
1487                 goto out_unlock;
1488         }
1489
1490         /* if LAYOUTGET already failed once we don't try again */
1491         if (pnfs_layout_io_test_failed(lo, iomode) &&
1492             !pnfs_should_retry_layoutget(lo))
1493                 goto out_unlock;
1494
1495         first = list_empty(&lo->plh_segs);
1496         if (first) {
1497                 /* The first layoutget for the file. Need to serialize per
1498                  * RFC 5661 Errata 3208.
1499                  */
1500                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1501                                      &lo->plh_flags)) {
1502                         spin_unlock(&ino->i_lock);
1503                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1504                                     TASK_UNINTERRUPTIBLE);
1505                         pnfs_put_layout_hdr(lo);
1506                         goto lookup_again;
1507                 }
1508         } else {
1509                 /* Check to see if the layout for the given range
1510                  * already exists
1511                  */
1512                 lseg = pnfs_find_lseg(lo, &arg);
1513                 if (lseg)
1514                         goto out_unlock;
1515         }
1516
1517         /*
1518          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1519          * for LAYOUTRETURN even if first is true.
1520          */
1521         if (!lseg && pnfs_should_retry_layoutget(lo) &&
1522             test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1523                 spin_unlock(&ino->i_lock);
1524                 dprintk("%s wait for layoutreturn\n", __func__);
1525                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1526                         if (first)
1527                                 pnfs_clear_first_layoutget(lo);
1528                         pnfs_put_layout_hdr(lo);
1529                         dprintk("%s retrying\n", __func__);
1530                         goto lookup_again;
1531                 }
1532                 goto out_put_layout_hdr;
1533         }
1534
1535         if (pnfs_layoutgets_blocked(lo, &arg, 0))
1536                 goto out_unlock;
1537         atomic_inc(&lo->plh_outstanding);
1538         spin_unlock(&ino->i_lock);
1539
1540         if (list_empty(&lo->plh_layouts)) {
1541                 /* The lo must be on the clp list if there is any
1542                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1543                  */
1544                 spin_lock(&clp->cl_lock);
1545                 if (list_empty(&lo->plh_layouts))
1546                         list_add_tail(&lo->plh_layouts, &server->layouts);
1547                 spin_unlock(&clp->cl_lock);
1548         }
1549
1550         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1551         if (pg_offset) {
1552                 arg.offset -= pg_offset;
1553                 arg.length += pg_offset;
1554         }
1555         if (arg.length != NFS4_MAX_UINT64)
1556                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1557
1558         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1559         pnfs_clear_retry_layoutget(lo);
1560         atomic_dec(&lo->plh_outstanding);
1561 out_put_layout_hdr:
1562         if (first)
1563                 pnfs_clear_first_layoutget(lo);
1564         pnfs_put_layout_hdr(lo);
1565 out:
1566         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1567                         "(%s, offset: %llu, length: %llu)\n",
1568                         __func__, ino->i_sb->s_id,
1569                         (unsigned long long)NFS_FILEID(ino),
1570                         lseg == NULL ? "not found" : "found",
1571                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1572                         (unsigned long long)pos,
1573                         (unsigned long long)count);
1574         return lseg;
1575 out_unlock:
1576         spin_unlock(&ino->i_lock);
1577         goto out_put_layout_hdr;
1578 }
1579 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1580
1581 struct pnfs_layout_segment *
1582 pnfs_layout_process(struct nfs4_layoutget *lgp)
1583 {
1584         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1585         struct nfs4_layoutget_res *res = &lgp->res;
1586         struct pnfs_layout_segment *lseg;
1587         struct inode *ino = lo->plh_inode;
1588         LIST_HEAD(free_me);
1589         int status = 0;
1590
1591         /* Inject layout blob into I/O device driver */
1592         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1593         if (!lseg || IS_ERR(lseg)) {
1594                 if (!lseg)
1595                         status = -ENOMEM;
1596                 else
1597                         status = PTR_ERR(lseg);
1598                 dprintk("%s: Could not allocate layout: error %d\n",
1599                        __func__, status);
1600                 goto out;
1601         }
1602
1603         init_lseg(lo, lseg);
1604         lseg->pls_range = res->range;
1605
1606         spin_lock(&ino->i_lock);
1607         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1608                 dprintk("%s forget reply due to recall\n", __func__);
1609                 goto out_forget_reply;
1610         }
1611
1612         if (pnfs_layoutgets_blocked(lo, &lgp->args.range, 1)) {
1613                 dprintk("%s forget reply due to state\n", __func__);
1614                 goto out_forget_reply;
1615         }
1616
1617         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1618                 /* existing state ID, make sure the sequence number matches. */
1619                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1620                         dprintk("%s forget reply due to sequence\n", __func__);
1621                         goto out_forget_reply;
1622                 }
1623                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1624         } else {
1625                 /*
1626                  * We got an entirely new state ID.  Mark all segments for the
1627                  * inode invalid, and don't bother validating the stateid
1628                  * sequence number.
1629                  */
1630                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1631
1632                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1633                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1634         }
1635
1636         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1637
1638         pnfs_get_lseg(lseg);
1639         pnfs_layout_insert_lseg(lo, lseg);
1640
1641         if (res->return_on_close) {
1642                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1643                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1644         }
1645
1646         spin_unlock(&ino->i_lock);
1647         pnfs_free_lseg_list(&free_me);
1648         return lseg;
1649 out:
1650         return ERR_PTR(status);
1651
1652 out_forget_reply:
1653         spin_unlock(&ino->i_lock);
1654         lseg->pls_layout = lo;
1655         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1656         goto out;
1657 }
1658
1659 static void
1660 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1661                                 struct list_head *tmp_list,
1662                                 struct pnfs_layout_range *return_range)
1663 {
1664         struct pnfs_layout_segment *lseg, *next;
1665
1666         dprintk("%s:Begin lo %p\n", __func__, lo);
1667
1668         if (list_empty(&lo->plh_segs))
1669                 return;
1670
1671         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1672                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1673                         dprintk("%s: marking lseg %p iomode %d "
1674                                 "offset %llu length %llu\n", __func__,
1675                                 lseg, lseg->pls_range.iomode,
1676                                 lseg->pls_range.offset,
1677                                 lseg->pls_range.length);
1678                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1679                         mark_lseg_invalid(lseg, tmp_list);
1680                 }
1681 }
1682
1683 void pnfs_error_mark_layout_for_return(struct inode *inode,
1684                                        struct pnfs_layout_segment *lseg)
1685 {
1686         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1687         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1688         struct pnfs_layout_range range = {
1689                 .iomode = lseg->pls_range.iomode,
1690                 .offset = 0,
1691                 .length = NFS4_MAX_UINT64,
1692         };
1693         LIST_HEAD(free_me);
1694
1695         spin_lock(&inode->i_lock);
1696         /* set failure bit so that pnfs path will be retried later */
1697         pnfs_layout_set_fail_bit(lo, iomode);
1698         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1699         if (lo->plh_return_iomode == 0)
1700                 lo->plh_return_iomode = range.iomode;
1701         else if (lo->plh_return_iomode != range.iomode)
1702                 lo->plh_return_iomode = IOMODE_ANY;
1703         /*
1704          * mark all matching lsegs so that we are sure to have no live
1705          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1706          * for how it works.
1707          */
1708         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1709         spin_unlock(&inode->i_lock);
1710         pnfs_free_lseg_list(&free_me);
1711 }
1712 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1713
1714 void
1715 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1716 {
1717         u64 rd_size = req->wb_bytes;
1718
1719         if (pgio->pg_lseg == NULL) {
1720                 if (pgio->pg_dreq == NULL)
1721                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1722                 else
1723                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1724
1725                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1726                                                    req->wb_context,
1727                                                    req_offset(req),
1728                                                    rd_size,
1729                                                    IOMODE_READ,
1730                                                    GFP_KERNEL);
1731         }
1732         /* If no lseg, fall back to read through mds */
1733         if (pgio->pg_lseg == NULL)
1734                 nfs_pageio_reset_read_mds(pgio);
1735
1736 }
1737 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1738
1739 void
1740 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1741                            struct nfs_page *req, u64 wb_size)
1742 {
1743         if (pgio->pg_lseg == NULL)
1744                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1745                                                    req->wb_context,
1746                                                    req_offset(req),
1747                                                    wb_size,
1748                                                    IOMODE_RW,
1749                                                    GFP_NOFS);
1750         /* If no lseg, fall back to write through mds */
1751         if (pgio->pg_lseg == NULL)
1752                 nfs_pageio_reset_write_mds(pgio);
1753 }
1754 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1755
1756 void
1757 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1758 {
1759         if (desc->pg_lseg) {
1760                 pnfs_put_lseg(desc->pg_lseg);
1761                 desc->pg_lseg = NULL;
1762         }
1763 }
1764 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1765
1766 /*
1767  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1768  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1769  */
1770 size_t
1771 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1772                      struct nfs_page *prev, struct nfs_page *req)
1773 {
1774         unsigned int size;
1775         u64 seg_end, req_start, seg_left;
1776
1777         size = nfs_generic_pg_test(pgio, prev, req);
1778         if (!size)
1779                 return 0;
1780
1781         /*
1782          * 'size' contains the number of bytes left in the current page (up
1783          * to the original size asked for in @req->wb_bytes).
1784          *
1785          * Calculate how many bytes are left in the layout segment
1786          * and if there are less bytes than 'size', return that instead.
1787          *
1788          * Please also note that 'end_offset' is actually the offset of the
1789          * first byte that lies outside the pnfs_layout_range. FIXME?
1790          *
1791          */
1792         if (pgio->pg_lseg) {
1793                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1794                                      pgio->pg_lseg->pls_range.length);
1795                 req_start = req_offset(req);
1796                 WARN_ON_ONCE(req_start >= seg_end);
1797                 /* start of request is past the last byte of this segment */
1798                 if (req_start >= seg_end) {
1799                         /* reference the new lseg */
1800                         if (pgio->pg_ops->pg_cleanup)
1801                                 pgio->pg_ops->pg_cleanup(pgio);
1802                         if (pgio->pg_ops->pg_init)
1803                                 pgio->pg_ops->pg_init(pgio, req);
1804                         return 0;
1805                 }
1806
1807                 /* adjust 'size' iff there are fewer bytes left in the
1808                  * segment than what nfs_generic_pg_test returned */
1809                 seg_left = seg_end - req_start;
1810                 if (seg_left < size)
1811                         size = (unsigned int)seg_left;
1812         }
1813
1814         return size;
1815 }
1816 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1817
1818 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1819 {
1820         struct nfs_pageio_descriptor pgio;
1821
1822         /* Resend all requests through the MDS */
1823         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1824                               hdr->completion_ops);
1825         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1826         return nfs_pageio_resend(&pgio, hdr);
1827 }
1828 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1829
1830 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1831 {
1832
1833         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1834         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1835             PNFS_LAYOUTRET_ON_ERROR) {
1836                 pnfs_return_layout(hdr->inode);
1837         }
1838         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1839                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1840 }
1841
1842 /*
1843  * Called by non rpc-based layout drivers
1844  */
1845 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1846 {
1847         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1848         if (!hdr->pnfs_error) {
1849                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1850                                 hdr->mds_offset + hdr->res.count);
1851                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1852         } else
1853                 pnfs_ld_handle_write_error(hdr);
1854         hdr->mds_ops->rpc_release(hdr);
1855 }
1856 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1857
1858 static void
1859 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1860                 struct nfs_pgio_header *hdr)
1861 {
1862         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1863
1864         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1865                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1866                 nfs_pageio_reset_write_mds(desc);
1867                 mirror->pg_recoalesce = 1;
1868         }
1869         nfs_pgio_data_destroy(hdr);
1870         hdr->release(hdr);
1871 }
1872
1873 static enum pnfs_try_status
1874 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1875                         const struct rpc_call_ops *call_ops,
1876                         struct pnfs_layout_segment *lseg,
1877                         int how)
1878 {
1879         struct inode *inode = hdr->inode;
1880         enum pnfs_try_status trypnfs;
1881         struct nfs_server *nfss = NFS_SERVER(inode);
1882
1883         hdr->mds_ops = call_ops;
1884
1885         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1886                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1887         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1888         if (trypnfs != PNFS_NOT_ATTEMPTED)
1889                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1890         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1891         return trypnfs;
1892 }
1893
1894 static void
1895 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1896               struct nfs_pgio_header *hdr, int how)
1897 {
1898         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1899         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1900         enum pnfs_try_status trypnfs;
1901
1902         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1903         if (trypnfs == PNFS_NOT_ATTEMPTED)
1904                 pnfs_write_through_mds(desc, hdr);
1905 }
1906
1907 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1908 {
1909         pnfs_put_lseg(hdr->lseg);
1910         nfs_pgio_header_free(hdr);
1911 }
1912
1913 int
1914 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1915 {
1916         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1917
1918         struct nfs_pgio_header *hdr;
1919         int ret;
1920
1921         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1922         if (!hdr) {
1923                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1924                 return -ENOMEM;
1925         }
1926         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1927
1928         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1929         ret = nfs_generic_pgio(desc, hdr);
1930         if (!ret)
1931                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1932
1933         return ret;
1934 }
1935 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1936
1937 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1938 {
1939         struct nfs_pageio_descriptor pgio;
1940
1941         /* Resend all requests through the MDS */
1942         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1943         return nfs_pageio_resend(&pgio, hdr);
1944 }
1945 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1946
1947 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1948 {
1949         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1950         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1951             PNFS_LAYOUTRET_ON_ERROR) {
1952                 pnfs_return_layout(hdr->inode);
1953         }
1954         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1955                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1956 }
1957
1958 /*
1959  * Called by non rpc-based layout drivers
1960  */
1961 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1962 {
1963         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1964         if (likely(!hdr->pnfs_error)) {
1965                 __nfs4_read_done_cb(hdr);
1966                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1967         } else
1968                 pnfs_ld_handle_read_error(hdr);
1969         hdr->mds_ops->rpc_release(hdr);
1970 }
1971 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1972
1973 static void
1974 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1975                 struct nfs_pgio_header *hdr)
1976 {
1977         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1978
1979         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1980                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1981                 nfs_pageio_reset_read_mds(desc);
1982                 mirror->pg_recoalesce = 1;
1983         }
1984         nfs_pgio_data_destroy(hdr);
1985         hdr->release(hdr);
1986 }
1987
1988 /*
1989  * Call the appropriate parallel I/O subsystem read function.
1990  */
1991 static enum pnfs_try_status
1992 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1993                        const struct rpc_call_ops *call_ops,
1994                        struct pnfs_layout_segment *lseg)
1995 {
1996         struct inode *inode = hdr->inode;
1997         struct nfs_server *nfss = NFS_SERVER(inode);
1998         enum pnfs_try_status trypnfs;
1999
2000         hdr->mds_ops = call_ops;
2001
2002         dprintk("%s: Reading ino:%lu %u@%llu\n",
2003                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2004
2005         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2006         if (trypnfs != PNFS_NOT_ATTEMPTED)
2007                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2008         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2009         return trypnfs;
2010 }
2011
2012 /* Resend all requests through pnfs. */
2013 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2014 {
2015         struct nfs_pageio_descriptor pgio;
2016
2017         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2018         return nfs_pageio_resend(&pgio, hdr);
2019 }
2020 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2021
2022 static void
2023 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2024 {
2025         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2026         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2027         enum pnfs_try_status trypnfs;
2028         int err = 0;
2029
2030         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2031         if (trypnfs == PNFS_TRY_AGAIN)
2032                 err = pnfs_read_resend_pnfs(hdr);
2033         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2034                 pnfs_read_through_mds(desc, hdr);
2035 }
2036
2037 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2038 {
2039         pnfs_put_lseg(hdr->lseg);
2040         nfs_pgio_header_free(hdr);
2041 }
2042
2043 int
2044 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2045 {
2046         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2047
2048         struct nfs_pgio_header *hdr;
2049         int ret;
2050
2051         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2052         if (!hdr) {
2053                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2054                 return -ENOMEM;
2055         }
2056         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2057         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2058         ret = nfs_generic_pgio(desc, hdr);
2059         if (!ret)
2060                 pnfs_do_read(desc, hdr);
2061         return ret;
2062 }
2063 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2064
2065 static void pnfs_clear_layoutcommitting(struct inode *inode)
2066 {
2067         unsigned long *bitlock = &NFS_I(inode)->flags;
2068
2069         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2070         smp_mb__after_atomic();
2071         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2072 }
2073
2074 /*
2075  * There can be multiple RW segments.
2076  */
2077 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2078 {
2079         struct pnfs_layout_segment *lseg;
2080
2081         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2082                 if (lseg->pls_range.iomode == IOMODE_RW &&
2083                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2084                         list_add(&lseg->pls_lc_list, listp);
2085         }
2086 }
2087
2088 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2089 {
2090         struct pnfs_layout_segment *lseg, *tmp;
2091
2092         /* Matched by references in pnfs_set_layoutcommit */
2093         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2094                 list_del_init(&lseg->pls_lc_list);
2095                 pnfs_put_lseg(lseg);
2096         }
2097
2098         pnfs_clear_layoutcommitting(inode);
2099 }
2100
2101 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2102 {
2103         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2104 }
2105 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2106
2107 void
2108 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2109                 loff_t end_pos)
2110 {
2111         struct nfs_inode *nfsi = NFS_I(inode);
2112         bool mark_as_dirty = false;
2113
2114         spin_lock(&inode->i_lock);
2115         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2116                 nfsi->layout->plh_lwb = end_pos;
2117                 mark_as_dirty = true;
2118                 dprintk("%s: Set layoutcommit for inode %lu ",
2119                         __func__, inode->i_ino);
2120         } else if (end_pos > nfsi->layout->plh_lwb)
2121                 nfsi->layout->plh_lwb = end_pos;
2122         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2123                 /* references matched in nfs4_layoutcommit_release */
2124                 pnfs_get_lseg(lseg);
2125         }
2126         spin_unlock(&inode->i_lock);
2127         dprintk("%s: lseg %p end_pos %llu\n",
2128                 __func__, lseg, nfsi->layout->plh_lwb);
2129
2130         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2131          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2132         if (mark_as_dirty)
2133                 mark_inode_dirty_sync(inode);
2134 }
2135 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2136
2137 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2138 {
2139         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2140
2141         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2142                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2143         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2144 }
2145
2146 /*
2147  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2148  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2149  * data to disk to allow the server to recover the data if it crashes.
2150  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2151  * is off, and a COMMIT is sent to a data server, or
2152  * if WRITEs to a data server return NFS_DATA_SYNC.
2153  */
2154 int
2155 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2156 {
2157         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2158         struct nfs4_layoutcommit_data *data;
2159         struct nfs_inode *nfsi = NFS_I(inode);
2160         loff_t end_pos;
2161         int status;
2162
2163         if (!pnfs_layoutcommit_outstanding(inode))
2164                 return 0;
2165
2166         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2167
2168         status = -EAGAIN;
2169         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2170                 if (!sync)
2171                         goto out;
2172                 status = wait_on_bit_lock_action(&nfsi->flags,
2173                                 NFS_INO_LAYOUTCOMMITTING,
2174                                 nfs_wait_bit_killable,
2175                                 TASK_KILLABLE);
2176                 if (status)
2177                         goto out;
2178         }
2179
2180         status = -ENOMEM;
2181         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2182         data = kzalloc(sizeof(*data), GFP_NOFS);
2183         if (!data)
2184                 goto clear_layoutcommitting;
2185
2186         status = 0;
2187         spin_lock(&inode->i_lock);
2188         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2189                 goto out_unlock;
2190
2191         INIT_LIST_HEAD(&data->lseg_list);
2192         pnfs_list_write_lseg(inode, &data->lseg_list);
2193
2194         end_pos = nfsi->layout->plh_lwb;
2195
2196         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2197         spin_unlock(&inode->i_lock);
2198
2199         data->args.inode = inode;
2200         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2201         nfs_fattr_init(&data->fattr);
2202         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2203         data->res.fattr = &data->fattr;
2204         data->args.lastbytewritten = end_pos - 1;
2205         data->res.server = NFS_SERVER(inode);
2206
2207         if (ld->prepare_layoutcommit) {
2208                 status = ld->prepare_layoutcommit(&data->args);
2209                 if (status) {
2210                         spin_lock(&inode->i_lock);
2211                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2212                         if (end_pos > nfsi->layout->plh_lwb)
2213                                 nfsi->layout->plh_lwb = end_pos;
2214                         spin_unlock(&inode->i_lock);
2215                         put_rpccred(data->cred);
2216                         goto clear_layoutcommitting;
2217                 }
2218         }
2219
2220
2221         status = nfs4_proc_layoutcommit(data, sync);
2222 out:
2223         if (status)
2224                 mark_inode_dirty_sync(inode);
2225         dprintk("<-- %s status %d\n", __func__, status);
2226         return status;
2227 out_unlock:
2228         spin_unlock(&inode->i_lock);
2229         kfree(data);
2230 clear_layoutcommitting:
2231         pnfs_clear_layoutcommitting(inode);
2232         goto out;
2233 }
2234 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2235
2236 int
2237 pnfs_generic_sync(struct inode *inode, bool datasync)
2238 {
2239         return pnfs_layoutcommit_inode(inode, true);
2240 }
2241 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2242
2243 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2244 {
2245         struct nfs4_threshold *thp;
2246
2247         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2248         if (!thp) {
2249                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2250                 return NULL;
2251         }
2252         return thp;
2253 }
2254
2255 #if IS_ENABLED(CONFIG_NFS_V4_2)
2256 int
2257 pnfs_report_layoutstat(struct inode *inode)
2258 {
2259         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2260         struct nfs_server *server = NFS_SERVER(inode);
2261         struct nfs_inode *nfsi = NFS_I(inode);
2262         struct nfs42_layoutstat_data *data;
2263         struct pnfs_layout_hdr *hdr;
2264         int status = 0;
2265
2266         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2267                 goto out;
2268
2269         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2270                 goto out;
2271
2272         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2273                 goto out;
2274
2275         spin_lock(&inode->i_lock);
2276         if (!NFS_I(inode)->layout) {
2277                 spin_unlock(&inode->i_lock);
2278                 goto out;
2279         }
2280         hdr = NFS_I(inode)->layout;
2281         pnfs_get_layout_hdr(hdr);
2282         spin_unlock(&inode->i_lock);
2283
2284         data = kzalloc(sizeof(*data), GFP_KERNEL);
2285         if (!data) {
2286                 status = -ENOMEM;
2287                 goto out_put;
2288         }
2289
2290         data->args.fh = NFS_FH(inode);
2291         data->args.inode = inode;
2292         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2293         status = ld->prepare_layoutstats(&data->args);
2294         if (status)
2295                 goto out_free;
2296
2297         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2298
2299 out:
2300         dprintk("%s returns %d\n", __func__, status);
2301         return status;
2302
2303 out_free:
2304         kfree(data);
2305 out_put:
2306         pnfs_put_layout_hdr(hdr);
2307         smp_mb__before_atomic();
2308         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2309         smp_mb__after_atomic();
2310         goto out;
2311 }
2312 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2313 #endif