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 /* update lo->plh_stateid with new if is more recent */
 553 void
 554 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 555                         bool update_barrier)
 556 {
 557         u32 oldseq, newseq;
 558
 559         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 560         newseq = be32_to_cpu(new->seqid);
 561         if ((int)(newseq - oldseq) > 0) {
 562                 nfs4_stateid_copy(&lo->plh_stateid, new);
 563                 if (update_barrier) {
 564                         u32 new_barrier = be32_to_cpu(new->seqid);
 565
 566                         if ((int)(new_barrier - lo->plh_barrier))
 567                                 lo->plh_barrier = new_barrier;
 568                 } else {
 569                         /* Because of wraparound, we want to keep the barrier
 570                          * "close" to the current seqids.  It needs to be
 571                          * within 2**31 to count as "behind", so if it
 572                          * gets too near that limit, give us a litle leeway
 573                          * and bring it to within 2**30.
 574                          * NOTE - and yes, this is all unsigned arithmetic.
 575                          */
 576                         if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
 577                                 lo->plh_barrier = newseq - (1 << 30);
 578                 }
 579         }
 580 }
 581
 582 /* lget is set to 1 if called from inside send_layoutget call chain */
 583 static bool
 584 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
 585                         int lget)
 586 {
 587         if ((stateid) &&
 588             (int)(lo->plh_barrier - be32_to_cpu(stateid->seqid)) >= 0)
 589                 return true;
 590         return lo->plh_block_lgets ||
 591                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 592                 (list_empty(&lo->plh_segs) &&
 593                  (atomic_read(&lo->plh_outstanding) > lget));
 594 }
 595
 596 int
 597 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 598                               struct nfs4_state *open_state)
 599 {
 600         int status = 0;
 601
 602         dprintk("--> %s\n", __func__);
 603         spin_lock(&lo->plh_inode->i_lock);
 604         if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
 605                 status = -EAGAIN;
 606         } else if (list_empty(&lo->plh_segs)) {
 607                 int seq;
 608
 609                 do {
 610                         seq = read_seqbegin(&open_state->seqlock);
 611                         nfs4_stateid_copy(dst, &open_state->stateid);
 612                 } while (read_seqretry(&open_state->seqlock, seq));
 613         } else
 614                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 615         spin_unlock(&lo->plh_inode->i_lock);
 616         dprintk("<-- %s\n", __func__);
 617         return status;
 618 }
 619
 620 /*
 621 * Get layout from server.
 622 *    for now, assume that whole file layouts are requested.
 623 *    arg->offset: 0
 624 *    arg->length: all ones
 625 */
 626 static struct pnfs_layout_segment *
 627 send_layoutget(struct pnfs_layout_hdr *lo,
 628            struct nfs_open_context *ctx,
 629            struct pnfs_layout_range *range,
 630            gfp_t gfp_flags)
 631 {
 632         struct inode *ino = lo->plh_inode;
 633         struct nfs_server *server = NFS_SERVER(ino);
 634         struct nfs4_layoutget *lgp;
 635         struct pnfs_layout_segment *lseg;
 636
 637         dprintk("--> %s\n", __func__);
 638
 639         BUG_ON(ctx == NULL);
 640         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 641         if (lgp == NULL)
 642                 return NULL;
 643
 644         lgp->args.minlength = PAGE_CACHE_SIZE;
 645         if (lgp->args.minlength > range->length)
 646                 lgp->args.minlength = range->length;
 647         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 648         lgp->args.range = *range;
 649         lgp->args.type = server->pnfs_curr_ld->id;
 650         lgp->args.inode = ino;
 651         lgp->args.ctx = get_nfs_open_context(ctx);
 652         lgp->gfp_flags = gfp_flags;
 653
 654         /* Synchronously retrieve layout information from server and
 655          * store in lseg.
 656          */
 657         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 658         if (IS_ERR(lseg)) {
 659                 switch (PTR_ERR(lseg)) {
 660                 case -ENOMEM:
 661                 case -ERESTARTSYS:
 662                         break;
 663                 default:
 664                         /* remember that LAYOUTGET failed and suspend trying */
 665                         pnfs_layout_io_set_failed(lo, range->iomode);
 666                 }
 667                 return NULL;
 668         }
 669
 670         return lseg;
 671 }
 672
 673 /*
 674  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 675  * when the layout segment list is empty.
 676  *
 677  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 678  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 679  * deviceid is marked invalid.
 680  */
 681 int
 682 _pnfs_return_layout(struct inode *ino)
 683 {
 684         struct pnfs_layout_hdr *lo = NULL;
 685         struct nfs_inode *nfsi = NFS_I(ino);
 686         LIST_HEAD(tmp_list);
 687         struct nfs4_layoutreturn *lrp;
 688         nfs4_stateid stateid;
 689         int status = 0, empty;
 690
 691         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 692
 693         spin_lock(&ino->i_lock);
 694         lo = nfsi->layout;
 695         if (!lo || pnfs_test_layout_returned(lo)) {
 696                 spin_unlock(&ino->i_lock);
 697                 dprintk("NFS: %s no layout to return\n", __func__);
 698                 goto out;
 699         }
 700         stateid = nfsi->layout->plh_stateid;
 701         /* Reference matched in nfs4_layoutreturn_release */
 702         pnfs_get_layout_hdr(lo);
 703         empty = list_empty(&lo->plh_segs);
 704         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 705         /* Don't send a LAYOUTRETURN if list was initially empty */
 706         if (empty) {
 707                 spin_unlock(&ino->i_lock);
 708                 pnfs_put_layout_hdr(lo);
 709                 dprintk("NFS: %s no layout segments to return\n", __func__);
 710                 goto out;
 711         }
 712         lo->plh_block_lgets++;
 713         pnfs_mark_layout_returned(lo);
 714         spin_unlock(&ino->i_lock);
 715         pnfs_free_lseg_list(&tmp_list);
 716
 717         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 718
 719         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 720         if (unlikely(lrp == NULL)) {
 721                 status = -ENOMEM;
 722                 pnfs_layout_io_set_failed(lo, IOMODE_RW);
 723                 pnfs_layout_io_set_failed(lo, IOMODE_READ);
 724                 pnfs_clear_layout_returned(lo);
 725                 pnfs_put_layout_hdr(lo);
 726                 goto out;
 727         }
 728
 729         lrp->args.stateid = stateid;
 730         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 731         lrp->args.inode = ino;
 732         lrp->args.layout = lo;
 733         lrp->clp = NFS_SERVER(ino)->nfs_client;
 734
 735         status = nfs4_proc_layoutreturn(lrp);
 736 out:
 737         dprintk("<-- %s status: %d\n", __func__, status);
 738         return status;
 739 }
 740 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 741
 742 bool pnfs_roc(struct inode *ino)
 743 {
 744         struct pnfs_layout_hdr *lo;
 745         struct pnfs_layout_segment *lseg, *tmp;
 746         LIST_HEAD(tmp_list);
 747         bool found = false;
 748
 749         spin_lock(&ino->i_lock);
 750         lo = NFS_I(ino)->layout;
 751         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 752             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 753                 goto out_nolayout;
 754         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 755                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 756                         mark_lseg_invalid(lseg, &tmp_list);
 757                         found = true;
 758                 }
 759         if (!found)
 760                 goto out_nolayout;
 761         lo->plh_block_lgets++;
 762         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 763         spin_unlock(&ino->i_lock);
 764         pnfs_free_lseg_list(&tmp_list);
 765         return true;
 766
 767 out_nolayout:
 768         spin_unlock(&ino->i_lock);
 769         return false;
 770 }
 771
 772 void pnfs_roc_release(struct inode *ino)
 773 {
 774         struct pnfs_layout_hdr *lo;
 775
 776         spin_lock(&ino->i_lock);
 777         lo = NFS_I(ino)->layout;
 778         lo->plh_block_lgets--;
 779         if (atomic_dec_and_test(&lo->plh_refcount)) {
 780                 pnfs_detach_layout_hdr(lo);
 781                 spin_unlock(&ino->i_lock);
 782                 pnfs_free_layout_hdr(lo);
 783         } else
 784                 spin_unlock(&ino->i_lock);
 785 }
 786
 787 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 788 {
 789         struct pnfs_layout_hdr *lo;
 790
 791         spin_lock(&ino->i_lock);
 792         lo = NFS_I(ino)->layout;
 793         if ((int)(barrier - lo->plh_barrier) > 0)
 794                 lo->plh_barrier = barrier;
 795         spin_unlock(&ino->i_lock);
 796 }
 797
 798 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 799 {
 800         struct nfs_inode *nfsi = NFS_I(ino);
 801         struct pnfs_layout_hdr *lo;
 802         struct pnfs_layout_segment *lseg;
 803         u32 current_seqid;
 804         bool found = false;
 805
 806         spin_lock(&ino->i_lock);
 807         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 808                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 809                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 810                         found = true;
 811                         goto out;
 812                 }
 813         lo = nfsi->layout;
 814         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 815
 816         /* Since close does not return a layout stateid for use as
 817          * a barrier, we choose the worst-case barrier.
 818          */
 819         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 820 out:
 821         spin_unlock(&ino->i_lock);
 822         return found;
 823 }
 824
 825 /*
 826  * Compare two layout segments for sorting into layout cache.
 827  * We want to preferentially return RW over RO layouts, so ensure those
 828  * are seen first.
 829  */
 830 static s64
 831 cmp_layout(struct pnfs_layout_range *l1,
 832            struct pnfs_layout_range *l2)
 833 {
 834         s64 d;
 835
 836         /* high offset > low offset */
 837         d = l1->offset - l2->offset;
 838         if (d)
 839                 return d;
 840
 841         /* short length > long length */
 842         d = l2->length - l1->length;
 843         if (d)
 844                 return d;
 845
 846         /* read > read/write */
 847         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 848 }
 849
 850 static void
 851 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
 852                    struct pnfs_layout_segment *lseg)
 853 {
 854         struct pnfs_layout_segment *lp;
 855
 856         dprintk("%s:Begin\n", __func__);
 857
 858         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 859                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 860                         continue;
 861                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 862                 dprintk("%s: inserted lseg %p "
 863                         "iomode %d offset %llu length %llu before "
 864                         "lp %p iomode %d offset %llu length %llu\n",
 865                         __func__, lseg, lseg->pls_range.iomode,
 866                         lseg->pls_range.offset, lseg->pls_range.length,
 867                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 868                         lp->pls_range.length);
 869                 goto out;
 870         }
 871         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 872         dprintk("%s: inserted lseg %p "
 873                 "iomode %d offset %llu length %llu at tail\n",
 874                 __func__, lseg, lseg->pls_range.iomode,
 875                 lseg->pls_range.offset, lseg->pls_range.length);
 876 out:
 877         pnfs_get_layout_hdr(lo);
 878
 879         dprintk("%s:Return\n", __func__);
 880 }
 881
 882 static struct pnfs_layout_hdr *
 883 alloc_init_layout_hdr(struct inode *ino,
 884                       struct nfs_open_context *ctx,
 885                       gfp_t gfp_flags)
 886 {
 887         struct pnfs_layout_hdr *lo;
 888
 889         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 890         if (!lo)
 891                 return NULL;
 892         atomic_set(&lo->plh_refcount, 1);
 893         INIT_LIST_HEAD(&lo->plh_layouts);
 894         INIT_LIST_HEAD(&lo->plh_segs);
 895         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 896         lo->plh_inode = ino;
 897         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 898         return lo;
 899 }
 900
 901 static struct pnfs_layout_hdr *
 902 pnfs_find_alloc_layout(struct inode *ino,
 903                        struct nfs_open_context *ctx,
 904                        gfp_t gfp_flags)
 905 {
 906         struct nfs_inode *nfsi = NFS_I(ino);
 907         struct pnfs_layout_hdr *new = NULL;
 908
 909         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 910
 911         if (nfsi->layout) {
 912                 pnfs_get_layout_hdr(nfsi->layout);
 913                 return nfsi->layout;
 914         }
 915         spin_unlock(&ino->i_lock);
 916         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 917         spin_lock(&ino->i_lock);
 918
 919         if (likely(nfsi->layout == NULL))       /* Won the race? */
 920                 nfsi->layout = new;
 921         else
 922                 pnfs_free_layout_hdr(new);
 923         return nfsi->layout;
 924 }
 925
 926 /*
 927  * iomode matching rules:
 928  * iomode       lseg    match
 929  * -----        -----   -----
 930  * ANY          READ    true
 931  * ANY          RW      true
 932  * RW           READ    false
 933  * RW           RW      true
 934  * READ         READ    true
 935  * READ         RW      true
 936  */
 937 static int
 938 is_matching_lseg(struct pnfs_layout_range *ls_range,
 939                  struct pnfs_layout_range *range)
 940 {
 941         struct pnfs_layout_range range1;
 942
 943         if ((range->iomode == IOMODE_RW &&
 944              ls_range->iomode != IOMODE_RW) ||
 945             !lo_seg_intersecting(ls_range, range))
 946                 return 0;
 947
 948         /* range1 covers only the first byte in the range */
 949         range1 = *range;
 950         range1.length = 1;
 951         return lo_seg_contained(ls_range, &range1);
 952 }
 953
 954 /*
 955  * lookup range in layout
 956  */
 957 static struct pnfs_layout_segment *
 958 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 959                 struct pnfs_layout_range *range)
 960 {
 961         struct pnfs_layout_segment *lseg, *ret = NULL;
 962
 963         dprintk("%s:Begin\n", __func__);
 964
 965         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 966                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 967                     is_matching_lseg(&lseg->pls_range, range)) {
 968                         ret = pnfs_get_lseg(lseg);
 969                         break;
 970                 }
 971                 if (lseg->pls_range.offset > range->offset)
 972                         break;
 973         }
 974
 975         dprintk("%s:Return lseg %p ref %d\n",
 976                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 977         return ret;
 978 }
 979
 980 /*
 981  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
 982  * to the MDS or over pNFS
 983  *
 984  * The nfs_inode read_io and write_io fields are cumulative counters reset
 985  * when there are no layout segments. Note that in pnfs_update_layout iomode
 986  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
 987  * WRITE request.
 988  *
 989  * A return of true means use MDS I/O.
 990  *
 991  * From rfc 5661:
 992  * If a file's size is smaller than the file size threshold, data accesses
 993  * SHOULD be sent to the metadata server.  If an I/O request has a length that
 994  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
 995  * server.  If both file size and I/O size are provided, the client SHOULD
 996  * reach or exceed  both thresholds before sending its read or write
 997  * requests to the data server.
 998  */
 999 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1000                                      struct inode *ino, int iomode)
1001 {
1002         struct nfs4_threshold *t = ctx->mdsthreshold;
1003         struct nfs_inode *nfsi = NFS_I(ino);
1004         loff_t fsize = i_size_read(ino);
1005         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1006
1007         if (t == NULL)
1008                 return ret;
1009
1010         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1011                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1012
1013         switch (iomode) {
1014         case IOMODE_READ:
1015                 if (t->bm & THRESHOLD_RD) {
1016                         dprintk("%s fsize %llu\n", __func__, fsize);
1017                         size_set = true;
1018                         if (fsize < t->rd_sz)
1019                                 size = true;
1020                 }
1021                 if (t->bm & THRESHOLD_RD_IO) {
1022                         dprintk("%s nfsi->read_io %llu\n", __func__,
1023                                 nfsi->read_io);
1024                         io_set = true;
1025                         if (nfsi->read_io < t->rd_io_sz)
1026                                 io = true;
1027                 }
1028                 break;
1029         case IOMODE_RW:
1030                 if (t->bm & THRESHOLD_WR) {
1031                         dprintk("%s fsize %llu\n", __func__, fsize);
1032                         size_set = true;
1033                         if (fsize < t->wr_sz)
1034                                 size = true;
1035                 }
1036                 if (t->bm & THRESHOLD_WR_IO) {
1037                         dprintk("%s nfsi->write_io %llu\n", __func__,
1038                                 nfsi->write_io);
1039                         io_set = true;
1040                         if (nfsi->write_io < t->wr_io_sz)
1041                                 io = true;
1042                 }
1043                 break;
1044         }
1045         if (size_set && io_set) {
1046                 if (size && io)
1047                         ret = true;
1048         } else if (size || io)
1049                 ret = true;
1050
1051         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1052         return ret;
1053 }
1054
1055 /*
1056  * Layout segment is retreived from the server if not cached.
1057  * The appropriate layout segment is referenced and returned to the caller.
1058  */
1059 struct pnfs_layout_segment *
1060 pnfs_update_layout(struct inode *ino,
1061                    struct nfs_open_context *ctx,
1062                    loff_t pos,
1063                    u64 count,
1064                    enum pnfs_iomode iomode,
1065                    gfp_t gfp_flags)
1066 {
1067         struct pnfs_layout_range arg = {
1068                 .iomode = iomode,
1069                 .offset = pos,
1070                 .length = count,
1071         };
1072         unsigned pg_offset;
1073         struct nfs_server *server = NFS_SERVER(ino);
1074         struct nfs_client *clp = server->nfs_client;
1075         struct pnfs_layout_hdr *lo;
1076         struct pnfs_layout_segment *lseg = NULL;
1077         bool first = false;
1078
1079         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1080                 goto out;
1081
1082         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1083                 goto out;
1084
1085         spin_lock(&ino->i_lock);
1086         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1087         if (lo == NULL) {
1088                 spin_unlock(&ino->i_lock);
1089                 goto out;
1090         }
1091
1092         /* Do we even need to bother with this? */
1093         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1094                 dprintk("%s matches recall, use MDS\n", __func__);
1095                 goto out_unlock;
1096         }
1097
1098         /* if LAYOUTGET already failed once we don't try again */
1099         if (pnfs_layout_io_test_failed(lo, iomode))
1100                 goto out_unlock;
1101
1102         /* Check to see if the layout for the given range already exists */
1103         lseg = pnfs_find_lseg(lo, &arg);
1104         if (lseg)
1105                 goto out_unlock;
1106
1107         if (pnfs_layoutgets_blocked(lo, NULL, 0))
1108                 goto out_unlock;
1109         atomic_inc(&lo->plh_outstanding);
1110
1111         if (list_empty(&lo->plh_segs))
1112                 first = true;
1113
1114         /* Enable LAYOUTRETURNs */
1115         pnfs_clear_layout_returned(lo);
1116
1117         spin_unlock(&ino->i_lock);
1118         if (first) {
1119                 /* The lo must be on the clp list if there is any
1120                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1121                  */
1122                 spin_lock(&clp->cl_lock);
1123                 BUG_ON(!list_empty(&lo->plh_layouts));
1124                 list_add_tail(&lo->plh_layouts, &server->layouts);
1125                 spin_unlock(&clp->cl_lock);
1126         }
1127
1128         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1129         if (pg_offset) {
1130                 arg.offset -= pg_offset;
1131                 arg.length += pg_offset;
1132         }
1133         if (arg.length != NFS4_MAX_UINT64)
1134                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1135
1136         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1137         atomic_dec(&lo->plh_outstanding);
1138 out_put_layout_hdr:
1139         pnfs_put_layout_hdr(lo);
1140 out:
1141         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1142                         "(%s, offset: %llu, length: %llu)\n",
1143                         __func__, ino->i_sb->s_id,
1144                         (unsigned long long)NFS_FILEID(ino),
1145                         lseg == NULL ? "not found" : "found",
1146                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1147                         (unsigned long long)pos,
1148                         (unsigned long long)count);
1149         return lseg;
1150 out_unlock:
1151         spin_unlock(&ino->i_lock);
1152         goto out_put_layout_hdr;
1153 }
1154 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1155
1156 struct pnfs_layout_segment *
1157 pnfs_layout_process(struct nfs4_layoutget *lgp)
1158 {
1159         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1160         struct nfs4_layoutget_res *res = &lgp->res;
1161         struct pnfs_layout_segment *lseg;
1162         struct inode *ino = lo->plh_inode;
1163         int status = 0;
1164
1165         /* Inject layout blob into I/O device driver */
1166         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1167         if (!lseg || IS_ERR(lseg)) {
1168                 if (!lseg)
1169                         status = -ENOMEM;
1170                 else
1171                         status = PTR_ERR(lseg);
1172                 dprintk("%s: Could not allocate layout: error %d\n",
1173                        __func__, status);
1174                 goto out;
1175         }
1176
1177         spin_lock(&ino->i_lock);
1178         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1179                 dprintk("%s forget reply due to recall\n", __func__);
1180                 goto out_forget_reply;
1181         }
1182
1183         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1184                 dprintk("%s forget reply due to state\n", __func__);
1185                 goto out_forget_reply;
1186         }
1187         init_lseg(lo, lseg);
1188         lseg->pls_range = res->range;
1189         pnfs_get_lseg(lseg);
1190         pnfs_layout_insert_lseg(lo, lseg);
1191
1192         if (res->return_on_close) {
1193                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1194                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1195         }
1196
1197         /* Done processing layoutget. Set the layout stateid */
1198         pnfs_set_layout_stateid(lo, &res->stateid, false);
1199         spin_unlock(&ino->i_lock);
1200         return lseg;
1201 out:
1202         return ERR_PTR(status);
1203
1204 out_forget_reply:
1205         spin_unlock(&ino->i_lock);
1206         lseg->pls_layout = lo;
1207         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1208         goto out;
1209 }
1210
1211 void
1212 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1213 {
1214         BUG_ON(pgio->pg_lseg != NULL);
1215
1216         if (req->wb_offset != req->wb_pgbase) {
1217                 nfs_pageio_reset_read_mds(pgio);
1218                 return;
1219         }
1220         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1221                                            req->wb_context,
1222                                            req_offset(req),
1223                                            req->wb_bytes,
1224                                            IOMODE_READ,
1225                                            GFP_KERNEL);
1226         /* If no lseg, fall back to read through mds */
1227         if (pgio->pg_lseg == NULL)
1228                 nfs_pageio_reset_read_mds(pgio);
1229
1230 }
1231 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1232
1233 void
1234 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1235 {
1236         BUG_ON(pgio->pg_lseg != NULL);
1237
1238         if (req->wb_offset != req->wb_pgbase) {
1239                 nfs_pageio_reset_write_mds(pgio);
1240                 return;
1241         }
1242         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1243                                            req->wb_context,
1244                                            req_offset(req),
1245                                            req->wb_bytes,
1246                                            IOMODE_RW,
1247                                            GFP_NOFS);
1248         /* If no lseg, fall back to write through mds */
1249         if (pgio->pg_lseg == NULL)
1250                 nfs_pageio_reset_write_mds(pgio);
1251 }
1252 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1253
1254 void
1255 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1256                       const struct nfs_pgio_completion_ops *compl_ops)
1257 {
1258         struct nfs_server *server = NFS_SERVER(inode);
1259         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1260
1261         if (ld == NULL)
1262                 nfs_pageio_init_read(pgio, inode, compl_ops);
1263         else
1264                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1265 }
1266
1267 void
1268 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1269                        int ioflags,
1270                        const struct nfs_pgio_completion_ops *compl_ops)
1271 {
1272         struct nfs_server *server = NFS_SERVER(inode);
1273         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1274
1275         if (ld == NULL)
1276                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1277         else
1278                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1279 }
1280
1281 bool
1282 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1283                      struct nfs_page *req)
1284 {
1285         if (pgio->pg_lseg == NULL)
1286                 return nfs_generic_pg_test(pgio, prev, req);
1287
1288         /*
1289          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1290          * Note that this test makes several assumptions:
1291          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1292          *   is known to lie within the range.
1293          *   - that the nfs_page being tested is known to be contiguous with the
1294          *   previous nfs_page.
1295          *   - Layout ranges are page aligned, so we only have to test the
1296          *   start offset of the request.
1297          *
1298          * Please also note that 'end_offset' is actually the offset of the
1299          * first byte that lies outside the pnfs_layout_range. FIXME?
1300          *
1301          */
1302         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1303                                          pgio->pg_lseg->pls_range.length);
1304 }
1305 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1306
1307 int pnfs_write_done_resend_to_mds(struct inode *inode,
1308                                 struct list_head *head,
1309                                 const struct nfs_pgio_completion_ops *compl_ops)
1310 {
1311         struct nfs_pageio_descriptor pgio;
1312         LIST_HEAD(failed);
1313
1314         /* Resend all requests through the MDS */
1315         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1316         while (!list_empty(head)) {
1317                 struct nfs_page *req = nfs_list_entry(head->next);
1318
1319                 nfs_list_remove_request(req);
1320                 if (!nfs_pageio_add_request(&pgio, req))
1321                         nfs_list_add_request(req, &failed);
1322         }
1323         nfs_pageio_complete(&pgio);
1324
1325         if (!list_empty(&failed)) {
1326                 /* For some reason our attempt to resend pages. Mark the
1327                  * overall send request as having failed, and let
1328                  * nfs_writeback_release_full deal with the error.
1329                  */
1330                 list_move(&failed, head);
1331                 return -EIO;
1332         }
1333         return 0;
1334 }
1335 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1336
1337 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1338 {
1339         struct nfs_pgio_header *hdr = data->header;
1340
1341         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1342         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1343             PNFS_LAYOUTRET_ON_ERROR) {
1344                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1345                 pnfs_return_layout(hdr->inode);
1346         }
1347         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1348                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1349                                                         &hdr->pages,
1350                                                         hdr->completion_ops);
1351 }
1352
1353 /*
1354  * Called by non rpc-based layout drivers
1355  */
1356 void pnfs_ld_write_done(struct nfs_write_data *data)
1357 {
1358         struct nfs_pgio_header *hdr = data->header;
1359
1360         if (!hdr->pnfs_error) {
1361                 pnfs_set_layoutcommit(data);
1362                 hdr->mds_ops->rpc_call_done(&data->task, data);
1363         } else
1364                 pnfs_ld_handle_write_error(data);
1365         hdr->mds_ops->rpc_release(data);
1366 }
1367 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1368
1369 static void
1370 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1371                 struct nfs_write_data *data)
1372 {
1373         struct nfs_pgio_header *hdr = data->header;
1374
1375         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1376                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1377                 nfs_pageio_reset_write_mds(desc);
1378                 desc->pg_recoalesce = 1;
1379         }
1380         nfs_writedata_release(data);
1381 }
1382
1383 static enum pnfs_try_status
1384 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1385                         const struct rpc_call_ops *call_ops,
1386                         struct pnfs_layout_segment *lseg,
1387                         int how)
1388 {
1389         struct nfs_pgio_header *hdr = wdata->header;
1390         struct inode *inode = hdr->inode;
1391         enum pnfs_try_status trypnfs;
1392         struct nfs_server *nfss = NFS_SERVER(inode);
1393
1394         hdr->mds_ops = call_ops;
1395
1396         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1397                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1398         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1399         if (trypnfs != PNFS_NOT_ATTEMPTED)
1400                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1401         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1402         return trypnfs;
1403 }
1404
1405 static void
1406 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1407 {
1408         struct nfs_write_data *data;
1409         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1410         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1411
1412         desc->pg_lseg = NULL;
1413         while (!list_empty(head)) {
1414                 enum pnfs_try_status trypnfs;
1415
1416                 data = list_first_entry(head, struct nfs_write_data, list);
1417                 list_del_init(&data->list);
1418
1419                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1420                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1421                         pnfs_write_through_mds(desc, data);
1422         }
1423         pnfs_put_lseg(lseg);
1424 }
1425
1426 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1427 {
1428         pnfs_put_lseg(hdr->lseg);
1429         nfs_writehdr_free(hdr);
1430 }
1431 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1432
1433 int
1434 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1435 {
1436         struct nfs_write_header *whdr;
1437         struct nfs_pgio_header *hdr;
1438         int ret;
1439
1440         whdr = nfs_writehdr_alloc();
1441         if (!whdr) {
1442                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1443                 pnfs_put_lseg(desc->pg_lseg);
1444                 desc->pg_lseg = NULL;
1445                 return -ENOMEM;
1446         }
1447         hdr = &whdr->header;
1448         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1449         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1450         atomic_inc(&hdr->refcnt);
1451         ret = nfs_generic_flush(desc, hdr);
1452         if (ret != 0) {
1453                 pnfs_put_lseg(desc->pg_lseg);
1454                 desc->pg_lseg = NULL;
1455         } else
1456                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1457         if (atomic_dec_and_test(&hdr->refcnt))
1458                 hdr->completion_ops->completion(hdr);
1459         return ret;
1460 }
1461 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1462
1463 int pnfs_read_done_resend_to_mds(struct inode *inode,
1464                                 struct list_head *head,
1465                                 const struct nfs_pgio_completion_ops *compl_ops)
1466 {
1467         struct nfs_pageio_descriptor pgio;
1468         LIST_HEAD(failed);
1469
1470         /* Resend all requests through the MDS */
1471         nfs_pageio_init_read(&pgio, inode, compl_ops);
1472         while (!list_empty(head)) {
1473                 struct nfs_page *req = nfs_list_entry(head->next);
1474
1475                 nfs_list_remove_request(req);
1476                 if (!nfs_pageio_add_request(&pgio, req))
1477                         nfs_list_add_request(req, &failed);
1478         }
1479         nfs_pageio_complete(&pgio);
1480
1481         if (!list_empty(&failed)) {
1482                 list_move(&failed, head);
1483                 return -EIO;
1484         }
1485         return 0;
1486 }
1487 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1488
1489 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1490 {
1491         struct nfs_pgio_header *hdr = data->header;
1492
1493         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1494         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1495             PNFS_LAYOUTRET_ON_ERROR) {
1496                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1497                 pnfs_return_layout(hdr->inode);
1498         }
1499         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1500                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1501                                                         &hdr->pages,
1502                                                         hdr->completion_ops);
1503 }
1504
1505 /*
1506  * Called by non rpc-based layout drivers
1507  */
1508 void pnfs_ld_read_done(struct nfs_read_data *data)
1509 {
1510         struct nfs_pgio_header *hdr = data->header;
1511
1512         if (likely(!hdr->pnfs_error)) {
1513                 __nfs4_read_done_cb(data);
1514                 hdr->mds_ops->rpc_call_done(&data->task, data);
1515         } else
1516                 pnfs_ld_handle_read_error(data);
1517         hdr->mds_ops->rpc_release(data);
1518 }
1519 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1520
1521 static void
1522 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1523                 struct nfs_read_data *data)
1524 {
1525         struct nfs_pgio_header *hdr = data->header;
1526
1527         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1528                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1529                 nfs_pageio_reset_read_mds(desc);
1530                 desc->pg_recoalesce = 1;
1531         }
1532         nfs_readdata_release(data);
1533 }
1534
1535 /*
1536  * Call the appropriate parallel I/O subsystem read function.
1537  */
1538 static enum pnfs_try_status
1539 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1540                        const struct rpc_call_ops *call_ops,
1541                        struct pnfs_layout_segment *lseg)
1542 {
1543         struct nfs_pgio_header *hdr = rdata->header;
1544         struct inode *inode = hdr->inode;
1545         struct nfs_server *nfss = NFS_SERVER(inode);
1546         enum pnfs_try_status trypnfs;
1547
1548         hdr->mds_ops = call_ops;
1549
1550         dprintk("%s: Reading ino:%lu %u@%llu\n",
1551                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1552
1553         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1554         if (trypnfs != PNFS_NOT_ATTEMPTED)
1555                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1556         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1557         return trypnfs;
1558 }
1559
1560 static void
1561 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1562 {
1563         struct nfs_read_data *data;
1564         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1565         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1566
1567         desc->pg_lseg = NULL;
1568         while (!list_empty(head)) {
1569                 enum pnfs_try_status trypnfs;
1570
1571                 data = list_first_entry(head, struct nfs_read_data, list);
1572                 list_del_init(&data->list);
1573
1574                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1575                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1576                         pnfs_read_through_mds(desc, data);
1577         }
1578         pnfs_put_lseg(lseg);
1579 }
1580
1581 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1582 {
1583         pnfs_put_lseg(hdr->lseg);
1584         nfs_readhdr_free(hdr);
1585 }
1586 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1587
1588 int
1589 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1590 {
1591         struct nfs_read_header *rhdr;
1592         struct nfs_pgio_header *hdr;
1593         int ret;
1594
1595         rhdr = nfs_readhdr_alloc();
1596         if (!rhdr) {
1597                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1598                 ret = -ENOMEM;
1599                 pnfs_put_lseg(desc->pg_lseg);
1600                 desc->pg_lseg = NULL;
1601                 return ret;
1602         }
1603         hdr = &rhdr->header;
1604         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1605         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1606         atomic_inc(&hdr->refcnt);
1607         ret = nfs_generic_pagein(desc, hdr);
1608         if (ret != 0) {
1609                 pnfs_put_lseg(desc->pg_lseg);
1610                 desc->pg_lseg = NULL;
1611         } else
1612                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1613         if (atomic_dec_and_test(&hdr->refcnt))
1614                 hdr->completion_ops->completion(hdr);
1615         return ret;
1616 }
1617 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1618
1619 /*
1620  * There can be multiple RW segments.
1621  */
1622 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1623 {
1624         struct pnfs_layout_segment *lseg;
1625
1626         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1627                 if (lseg->pls_range.iomode == IOMODE_RW &&
1628                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1629                         list_add(&lseg->pls_lc_list, listp);
1630         }
1631 }
1632
1633 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1634 {
1635         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1636 }
1637 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1638
1639 void
1640 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1641 {
1642         struct nfs_pgio_header *hdr = wdata->header;
1643         struct inode *inode = hdr->inode;
1644         struct nfs_inode *nfsi = NFS_I(inode);
1645         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1646         bool mark_as_dirty = false;
1647
1648         spin_lock(&inode->i_lock);
1649         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1650                 mark_as_dirty = true;
1651                 dprintk("%s: Set layoutcommit for inode %lu ",
1652                         __func__, inode->i_ino);
1653         }
1654         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1655                 /* references matched in nfs4_layoutcommit_release */
1656                 pnfs_get_lseg(hdr->lseg);
1657         }
1658         if (end_pos > nfsi->layout->plh_lwb)
1659                 nfsi->layout->plh_lwb = end_pos;
1660         spin_unlock(&inode->i_lock);
1661         dprintk("%s: lseg %p end_pos %llu\n",
1662                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1663
1664         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1665          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1666         if (mark_as_dirty)
1667                 mark_inode_dirty_sync(inode);
1668 }
1669 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1670
1671 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1672 {
1673         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1674
1675         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1676                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1677 }
1678
1679 /*
1680  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1681  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1682  * data to disk to allow the server to recover the data if it crashes.
1683  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1684  * is off, and a COMMIT is sent to a data server, or
1685  * if WRITEs to a data server return NFS_DATA_SYNC.
1686  */
1687 int
1688 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1689 {
1690         struct nfs4_layoutcommit_data *data;
1691         struct nfs_inode *nfsi = NFS_I(inode);
1692         loff_t end_pos;
1693         int status = 0;
1694
1695         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1696
1697         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1698                 return 0;
1699
1700         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1701         data = kzalloc(sizeof(*data), GFP_NOFS);
1702         if (!data) {
1703                 status = -ENOMEM;
1704                 goto out;
1705         }
1706
1707         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1708                 goto out_free;
1709
1710         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1711                 if (!sync) {
1712                         status = -EAGAIN;
1713                         goto out_free;
1714                 }
1715                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1716                                         nfs_wait_bit_killable, TASK_KILLABLE);
1717                 if (status)
1718                         goto out_free;
1719         }
1720
1721         INIT_LIST_HEAD(&data->lseg_list);
1722         spin_lock(&inode->i_lock);
1723         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1724                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1725                 spin_unlock(&inode->i_lock);
1726                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1727                 goto out_free;
1728         }
1729
1730         pnfs_list_write_lseg(inode, &data->lseg_list);
1731
1732         end_pos = nfsi->layout->plh_lwb;
1733         nfsi->layout->plh_lwb = 0;
1734
1735         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1736         spin_unlock(&inode->i_lock);
1737
1738         data->args.inode = inode;
1739         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1740         nfs_fattr_init(&data->fattr);
1741         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1742         data->res.fattr = &data->fattr;
1743         data->args.lastbytewritten = end_pos - 1;
1744         data->res.server = NFS_SERVER(inode);
1745
1746         status = nfs4_proc_layoutcommit(data, sync);
1747 out:
1748         if (status)
1749                 mark_inode_dirty_sync(inode);
1750         dprintk("<-- %s status %d\n", __func__, status);
1751         return status;
1752 out_free:
1753         kfree(data);
1754         goto out;
1755 }
1756
1757 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1758 {
1759         struct nfs4_threshold *thp;
1760
1761         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1762         if (!thp) {
1763                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1764                 return NULL;
1765         }
1766         return thp;
1767 }