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