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
  39 /* Locking:
  40  *
  41  * pnfs_spinlock:
  42  *      protects pnfs_modules_tbl.
  43  */
  44 static DEFINE_SPINLOCK(pnfs_spinlock);
  45
  46 /*
  47  * pnfs_modules_tbl holds all pnfs modules
  48  */
  49 static LIST_HEAD(pnfs_modules_tbl);
  50
  51 /* Return the registered pnfs layout driver module matching given id */
  52 static struct pnfs_layoutdriver_type *
  53 find_pnfs_driver_locked(u32 id)
  54 {
  55         struct pnfs_layoutdriver_type *local;
  56
  57         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  58                 if (local->id == id)
  59                         goto out;
  60         local = NULL;
  61 out:
  62         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  63         return local;
  64 }
  65
  66 static struct pnfs_layoutdriver_type *
  67 find_pnfs_driver(u32 id)
  68 {
  69         struct pnfs_layoutdriver_type *local;
  70
  71         spin_lock(&pnfs_spinlock);
  72         local = find_pnfs_driver_locked(id);
  73         spin_unlock(&pnfs_spinlock);
  74         return local;
  75 }
  76
  77 void
  78 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  79 {
  80         if (nfss->pnfs_curr_ld) {
  81                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  82                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  83                 module_put(nfss->pnfs_curr_ld->owner);
  84         }
  85         nfss->pnfs_curr_ld = NULL;
  86 }
  87
  88 /*
  89  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  90  * Currently only one pNFS layout driver per filesystem is supported.
  91  *
  92  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
  93  */
  94 void
  95 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
  96                       u32 id)
  97 {
  98         struct pnfs_layoutdriver_type *ld_type = NULL;
  99
 100         if (id == 0)
 101                 goto out_no_driver;
 102         if (!(server->nfs_client->cl_exchange_flags &
 103                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 104                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 105                         __func__, id, server->nfs_client->cl_exchange_flags);
 106                 goto out_no_driver;
 107         }
 108         ld_type = find_pnfs_driver(id);
 109         if (!ld_type) {
 110                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 111                 ld_type = find_pnfs_driver(id);
 112                 if (!ld_type) {
 113                         dprintk("%s: No pNFS module found for %u.\n",
 114                                 __func__, id);
 115                         goto out_no_driver;
 116                 }
 117         }
 118         if (!try_module_get(ld_type->owner)) {
 119                 dprintk("%s: Could not grab reference on module\n", __func__);
 120                 goto out_no_driver;
 121         }
 122         server->pnfs_curr_ld = ld_type;
 123         if (ld_type->set_layoutdriver
 124             && ld_type->set_layoutdriver(server, mntfh)) {
 125                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 126                         "driver %u.\n", __func__, id);
 127                 module_put(ld_type->owner);
 128                 goto out_no_driver;
 129         }
 130
 131         dprintk("%s: pNFS module for %u set\n", __func__, id);
 132         return;
 133
 134 out_no_driver:
 135         dprintk("%s: Using NFSv4 I/O\n", __func__);
 136         server->pnfs_curr_ld = NULL;
 137 }
 138
 139 int
 140 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 141 {
 142         int status = -EINVAL;
 143         struct pnfs_layoutdriver_type *tmp;
 144
 145         if (ld_type->id == 0) {
 146                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 147                 return status;
 148         }
 149         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 150                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 151                        "alloc_lseg and free_lseg.\n", __func__);
 152                 return status;
 153         }
 154
 155         spin_lock(&pnfs_spinlock);
 156         tmp = find_pnfs_driver_locked(ld_type->id);
 157         if (!tmp) {
 158                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 159                 status = 0;
 160                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 161                         ld_type->name);
 162         } else {
 163                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 164                         __func__, ld_type->id);
 165         }
 166         spin_unlock(&pnfs_spinlock);
 167
 168         return status;
 169 }
 170 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 171
 172 void
 173 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 174 {
 175         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 176         spin_lock(&pnfs_spinlock);
 177         list_del(&ld_type->pnfs_tblid);
 178         spin_unlock(&pnfs_spinlock);
 179 }
 180 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 181
 182 /*
 183  * pNFS client layout cache
 184  */
 185
 186 /* Need to hold i_lock if caller does not already hold reference */
 187 void
 188 get_layout_hdr(struct pnfs_layout_hdr *lo)
 189 {
 190         atomic_inc(&lo->plh_refcount);
 191 }
 192
 193 static struct pnfs_layout_hdr *
 194 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 195 {
 196         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 197         return ld->alloc_layout_hdr ? ld->alloc_layout_hdr(ino, gfp_flags) :
 198                 kzalloc(sizeof(struct pnfs_layout_hdr), gfp_flags);
 199 }
 200
 201 static void
 202 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 203 {
 204         struct pnfs_layoutdriver_type *ld = NFS_SERVER(lo->plh_inode)->pnfs_curr_ld;
 205         put_rpccred(lo->plh_lc_cred);
 206         return ld->alloc_layout_hdr ? ld->free_layout_hdr(lo) : kfree(lo);
 207 }
 208
 209 static void
 210 destroy_layout_hdr(struct pnfs_layout_hdr *lo)
 211 {
 212         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 213         BUG_ON(!list_empty(&lo->plh_layouts));
 214         NFS_I(lo->plh_inode)->layout = NULL;
 215         pnfs_free_layout_hdr(lo);
 216 }
 217
 218 static void
 219 put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
 220 {
 221         if (atomic_dec_and_test(&lo->plh_refcount))
 222                 destroy_layout_hdr(lo);
 223 }
 224
 225 void
 226 put_layout_hdr(struct pnfs_layout_hdr *lo)
 227 {
 228         struct inode *inode = lo->plh_inode;
 229
 230         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 231                 destroy_layout_hdr(lo);
 232                 spin_unlock(&inode->i_lock);
 233         }
 234 }
 235
 236 static void
 237 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 238 {
 239         INIT_LIST_HEAD(&lseg->pls_list);
 240         INIT_LIST_HEAD(&lseg->pls_lc_list);
 241         atomic_set(&lseg->pls_refcount, 1);
 242         smp_mb();
 243         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 244         lseg->pls_layout = lo;
 245 }
 246
 247 static void free_lseg(struct pnfs_layout_segment *lseg)
 248 {
 249         struct inode *ino = lseg->pls_layout->plh_inode;
 250
 251         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 252         /* Matched by get_layout_hdr in pnfs_insert_layout */
 253         put_layout_hdr(NFS_I(ino)->layout);
 254 }
 255
 256 static void
 257 put_lseg_common(struct pnfs_layout_segment *lseg)
 258 {
 259         struct inode *inode = lseg->pls_layout->plh_inode;
 260
 261         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 262         list_del_init(&lseg->pls_list);
 263         if (list_empty(&lseg->pls_layout->plh_segs)) {
 264                 set_bit(NFS_LAYOUT_DESTROYED, &lseg->pls_layout->plh_flags);
 265                 /* Matched by initial refcount set in alloc_init_layout_hdr */
 266                 put_layout_hdr_locked(lseg->pls_layout);
 267         }
 268         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 269 }
 270
 271 void
 272 put_lseg(struct pnfs_layout_segment *lseg)
 273 {
 274         struct inode *inode;
 275
 276         if (!lseg)
 277                 return;
 278
 279         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 280                 atomic_read(&lseg->pls_refcount),
 281                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 282         inode = lseg->pls_layout->plh_inode;
 283         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 284                 LIST_HEAD(free_me);
 285
 286                 put_lseg_common(lseg);
 287                 list_add(&lseg->pls_list, &free_me);
 288                 spin_unlock(&inode->i_lock);
 289                 pnfs_free_lseg_list(&free_me);
 290         }
 291 }
 292 EXPORT_SYMBOL_GPL(put_lseg);
 293
 294 static inline u64
 295 end_offset(u64 start, u64 len)
 296 {
 297         u64 end;
 298
 299         end = start + len;
 300         return end >= start ? end : NFS4_MAX_UINT64;
 301 }
 302
 303 /* last octet in a range */
 304 static inline u64
 305 last_byte_offset(u64 start, u64 len)
 306 {
 307         u64 end;
 308
 309         BUG_ON(!len);
 310         end = start + len;
 311         return end > start ? end - 1 : NFS4_MAX_UINT64;
 312 }
 313
 314 /*
 315  * is l2 fully contained in l1?
 316  *   start1                             end1
 317  *   [----------------------------------)
 318  *           start2           end2
 319  *           [----------------)
 320  */
 321 static inline int
 322 lo_seg_contained(struct pnfs_layout_range *l1,
 323                  struct pnfs_layout_range *l2)
 324 {
 325         u64 start1 = l1->offset;
 326         u64 end1 = end_offset(start1, l1->length);
 327         u64 start2 = l2->offset;
 328         u64 end2 = end_offset(start2, l2->length);
 329
 330         return (start1 <= start2) && (end1 >= end2);
 331 }
 332
 333 /*
 334  * is l1 and l2 intersecting?
 335  *   start1                             end1
 336  *   [----------------------------------)
 337  *                              start2           end2
 338  *                              [----------------)
 339  */
 340 static inline int
 341 lo_seg_intersecting(struct pnfs_layout_range *l1,
 342                     struct pnfs_layout_range *l2)
 343 {
 344         u64 start1 = l1->offset;
 345         u64 end1 = end_offset(start1, l1->length);
 346         u64 start2 = l2->offset;
 347         u64 end2 = end_offset(start2, l2->length);
 348
 349         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 350                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 351 }
 352
 353 static bool
 354 should_free_lseg(struct pnfs_layout_range *lseg_range,
 355                  struct pnfs_layout_range *recall_range)
 356 {
 357         return (recall_range->iomode == IOMODE_ANY ||
 358                 lseg_range->iomode == recall_range->iomode) &&
 359                lo_seg_intersecting(lseg_range, recall_range);
 360 }
 361
 362 /* Returns 1 if lseg is removed from list, 0 otherwise */
 363 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 364                              struct list_head *tmp_list)
 365 {
 366         int rv = 0;
 367
 368         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 369                 /* Remove the reference keeping the lseg in the
 370                  * list.  It will now be removed when all
 371                  * outstanding io is finished.
 372                  */
 373                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 374                         atomic_read(&lseg->pls_refcount));
 375                 if (atomic_dec_and_test(&lseg->pls_refcount)) {
 376                         put_lseg_common(lseg);
 377                         list_add(&lseg->pls_list, tmp_list);
 378                         rv = 1;
 379                 }
 380         }
 381         return rv;
 382 }
 383
 384 /* Returns count of number of matching invalid lsegs remaining in list
 385  * after call.
 386  */
 387 int
 388 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 389                             struct list_head *tmp_list,
 390                             struct pnfs_layout_range *recall_range)
 391 {
 392         struct pnfs_layout_segment *lseg, *next;
 393         int invalid = 0, removed = 0;
 394
 395         dprintk("%s:Begin lo %p\n", __func__, lo);
 396
 397         if (list_empty(&lo->plh_segs)) {
 398                 /* Reset MDS Threshold I/O counters */
 399                 NFS_I(lo->plh_inode)->write_io = 0;
 400                 NFS_I(lo->plh_inode)->read_io = 0;
 401                 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
 402                         put_layout_hdr_locked(lo);
 403                 return 0;
 404         }
 405         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 406                 if (!recall_range ||
 407                     should_free_lseg(&lseg->pls_range, recall_range)) {
 408                         dprintk("%s: freeing lseg %p iomode %d "
 409                                 "offset %llu length %llu\n", __func__,
 410                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 411                                 lseg->pls_range.length);
 412                         invalid++;
 413                         removed += mark_lseg_invalid(lseg, tmp_list);
 414                 }
 415         dprintk("%s:Return %i\n", __func__, invalid - removed);
 416         return invalid - removed;
 417 }
 418
 419 /* note free_me must contain lsegs from a single layout_hdr */
 420 void
 421 pnfs_free_lseg_list(struct list_head *free_me)
 422 {
 423         struct pnfs_layout_segment *lseg, *tmp;
 424         struct pnfs_layout_hdr *lo;
 425
 426         if (list_empty(free_me))
 427                 return;
 428
 429         lo = list_first_entry(free_me, struct pnfs_layout_segment,
 430                               pls_list)->pls_layout;
 431
 432         if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
 433                 struct nfs_client *clp;
 434
 435                 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
 436                 spin_lock(&clp->cl_lock);
 437                 list_del_init(&lo->plh_layouts);
 438                 spin_unlock(&clp->cl_lock);
 439         }
 440         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 441                 list_del(&lseg->pls_list);
 442                 free_lseg(lseg);
 443         }
 444 }
 445
 446 void
 447 pnfs_destroy_layout(struct nfs_inode *nfsi)
 448 {
 449         struct pnfs_layout_hdr *lo;
 450         LIST_HEAD(tmp_list);
 451
 452         spin_lock(&nfsi->vfs_inode.i_lock);
 453         lo = nfsi->layout;
 454         if (lo) {
 455                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 456                 mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 457         }
 458         spin_unlock(&nfsi->vfs_inode.i_lock);
 459         pnfs_free_lseg_list(&tmp_list);
 460 }
 461 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 462
 463 /*
 464  * Called by the state manger to remove all layouts established under an
 465  * expired lease.
 466  */
 467 void
 468 pnfs_destroy_all_layouts(struct nfs_client *clp)
 469 {
 470         struct nfs_server *server;
 471         struct pnfs_layout_hdr *lo;
 472         LIST_HEAD(tmp_list);
 473
 474         nfs4_deviceid_mark_client_invalid(clp);
 475         nfs4_deviceid_purge_client(clp);
 476
 477         spin_lock(&clp->cl_lock);
 478         rcu_read_lock();
 479         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 480                 if (!list_empty(&server->layouts))
 481                         list_splice_init(&server->layouts, &tmp_list);
 482         }
 483         rcu_read_unlock();
 484         spin_unlock(&clp->cl_lock);
 485
 486         while (!list_empty(&tmp_list)) {
 487                 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
 488                                 plh_layouts);
 489                 dprintk("%s freeing layout for inode %lu\n", __func__,
 490                         lo->plh_inode->i_ino);
 491                 list_del_init(&lo->plh_layouts);
 492                 pnfs_destroy_layout(NFS_I(lo->plh_inode));
 493         }
 494 }
 495
 496 /* update lo->plh_stateid with new if is more recent */
 497 void
 498 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 499                         bool update_barrier)
 500 {
 501         u32 oldseq, newseq;
 502
 503         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 504         newseq = be32_to_cpu(new->seqid);
 505         if ((int)(newseq - oldseq) > 0) {
 506                 nfs4_stateid_copy(&lo->plh_stateid, new);
 507                 if (update_barrier) {
 508                         u32 new_barrier = be32_to_cpu(new->seqid);
 509
 510                         if ((int)(new_barrier - lo->plh_barrier))
 511                                 lo->plh_barrier = new_barrier;
 512                 } else {
 513                         /* Because of wraparound, we want to keep the barrier
 514                          * "close" to the current seqids.  It needs to be
 515                          * within 2**31 to count as "behind", so if it
 516                          * gets too near that limit, give us a litle leeway
 517                          * and bring it to within 2**30.
 518                          * NOTE - and yes, this is all unsigned arithmetic.
 519                          */
 520                         if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
 521                                 lo->plh_barrier = newseq - (1 << 30);
 522                 }
 523         }
 524 }
 525
 526 /* lget is set to 1 if called from inside send_layoutget call chain */
 527 static bool
 528 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
 529                         int lget)
 530 {
 531         if ((stateid) &&
 532             (int)(lo->plh_barrier - be32_to_cpu(stateid->seqid)) >= 0)
 533                 return true;
 534         return lo->plh_block_lgets ||
 535                 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
 536                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 537                 (list_empty(&lo->plh_segs) &&
 538                  (atomic_read(&lo->plh_outstanding) > lget));
 539 }
 540
 541 int
 542 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 543                               struct nfs4_state *open_state)
 544 {
 545         int status = 0;
 546
 547         dprintk("--> %s\n", __func__);
 548         spin_lock(&lo->plh_inode->i_lock);
 549         if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
 550                 status = -EAGAIN;
 551         } else if (list_empty(&lo->plh_segs)) {
 552                 int seq;
 553
 554                 do {
 555                         seq = read_seqbegin(&open_state->seqlock);
 556                         nfs4_stateid_copy(dst, &open_state->stateid);
 557                 } while (read_seqretry(&open_state->seqlock, seq));
 558         } else
 559                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 560         spin_unlock(&lo->plh_inode->i_lock);
 561         dprintk("<-- %s\n", __func__);
 562         return status;
 563 }
 564
 565 /*
 566 * Get layout from server.
 567 *    for now, assume that whole file layouts are requested.
 568 *    arg->offset: 0
 569 *    arg->length: all ones
 570 */
 571 static struct pnfs_layout_segment *
 572 send_layoutget(struct pnfs_layout_hdr *lo,
 573            struct nfs_open_context *ctx,
 574            struct pnfs_layout_range *range,
 575            gfp_t gfp_flags)
 576 {
 577         struct inode *ino = lo->plh_inode;
 578         struct nfs_server *server = NFS_SERVER(ino);
 579         struct nfs4_layoutget *lgp;
 580         struct pnfs_layout_segment *lseg = NULL;
 581         struct page **pages = NULL;
 582         int i;
 583         u32 max_resp_sz, max_pages;
 584
 585         dprintk("--> %s\n", __func__);
 586
 587         BUG_ON(ctx == NULL);
 588         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 589         if (lgp == NULL)
 590                 return NULL;
 591
 592         /* allocate pages for xdr post processing */
 593         max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
 594         max_pages = nfs_page_array_len(0, max_resp_sz);
 595
 596         pages = kcalloc(max_pages, sizeof(struct page *), gfp_flags);
 597         if (!pages)
 598                 goto out_err_free;
 599
 600         for (i = 0; i < max_pages; i++) {
 601                 pages[i] = alloc_page(gfp_flags);
 602                 if (!pages[i])
 603                         goto out_err_free;
 604         }
 605
 606         lgp->args.minlength = PAGE_CACHE_SIZE;
 607         if (lgp->args.minlength > range->length)
 608                 lgp->args.minlength = range->length;
 609         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 610         lgp->args.range = *range;
 611         lgp->args.type = server->pnfs_curr_ld->id;
 612         lgp->args.inode = ino;
 613         lgp->args.ctx = get_nfs_open_context(ctx);
 614         lgp->args.layout.pages = pages;
 615         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
 616         lgp->lsegpp = &lseg;
 617         lgp->gfp_flags = gfp_flags;
 618
 619         /* Synchronously retrieve layout information from server and
 620          * store in lseg.
 621          */
 622         nfs4_proc_layoutget(lgp);
 623         if (!lseg) {
 624                 /* remember that LAYOUTGET failed and suspend trying */
 625                 set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
 626         }
 627
 628         /* free xdr pages */
 629         for (i = 0; i < max_pages; i++)
 630                 __free_page(pages[i]);
 631         kfree(pages);
 632
 633         return lseg;
 634
 635 out_err_free:
 636         /* free any allocated xdr pages, lgp as it's not used */
 637         if (pages) {
 638                 for (i = 0; i < max_pages; i++) {
 639                         if (!pages[i])
 640                                 break;
 641                         __free_page(pages[i]);
 642                 }
 643                 kfree(pages);
 644         }
 645         kfree(lgp);
 646         return NULL;
 647 }
 648
 649 /* Initiates a LAYOUTRETURN(FILE) */
 650 int
 651 _pnfs_return_layout(struct inode *ino)
 652 {
 653         struct pnfs_layout_hdr *lo = NULL;
 654         struct nfs_inode *nfsi = NFS_I(ino);
 655         LIST_HEAD(tmp_list);
 656         struct nfs4_layoutreturn *lrp;
 657         nfs4_stateid stateid;
 658         int status = 0;
 659
 660         dprintk("--> %s\n", __func__);
 661
 662         spin_lock(&ino->i_lock);
 663         lo = nfsi->layout;
 664         if (!lo) {
 665                 spin_unlock(&ino->i_lock);
 666                 dprintk("%s: no layout to return\n", __func__);
 667                 return status;
 668         }
 669         stateid = nfsi->layout->plh_stateid;
 670         /* Reference matched in nfs4_layoutreturn_release */
 671         get_layout_hdr(lo);
 672         mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 673         lo->plh_block_lgets++;
 674         spin_unlock(&ino->i_lock);
 675         pnfs_free_lseg_list(&tmp_list);
 676
 677         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 678
 679         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 680         if (unlikely(lrp == NULL)) {
 681                 status = -ENOMEM;
 682                 set_bit(NFS_LAYOUT_RW_FAILED, &lo->plh_flags);
 683                 set_bit(NFS_LAYOUT_RO_FAILED, &lo->plh_flags);
 684                 put_layout_hdr(lo);
 685                 goto out;
 686         }
 687
 688         lrp->args.stateid = stateid;
 689         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 690         lrp->args.inode = ino;
 691         lrp->args.layout = lo;
 692         lrp->clp = NFS_SERVER(ino)->nfs_client;
 693
 694         status = nfs4_proc_layoutreturn(lrp);
 695 out:
 696         dprintk("<-- %s status: %d\n", __func__, status);
 697         return status;
 698 }
 699 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 700
 701 bool pnfs_roc(struct inode *ino)
 702 {
 703         struct pnfs_layout_hdr *lo;
 704         struct pnfs_layout_segment *lseg, *tmp;
 705         LIST_HEAD(tmp_list);
 706         bool found = false;
 707
 708         spin_lock(&ino->i_lock);
 709         lo = NFS_I(ino)->layout;
 710         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 711             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 712                 goto out_nolayout;
 713         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 714                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 715                         mark_lseg_invalid(lseg, &tmp_list);
 716                         found = true;
 717                 }
 718         if (!found)
 719                 goto out_nolayout;
 720         lo->plh_block_lgets++;
 721         get_layout_hdr(lo); /* matched in pnfs_roc_release */
 722         spin_unlock(&ino->i_lock);
 723         pnfs_free_lseg_list(&tmp_list);
 724         return true;
 725
 726 out_nolayout:
 727         spin_unlock(&ino->i_lock);
 728         return false;
 729 }
 730
 731 void pnfs_roc_release(struct inode *ino)
 732 {
 733         struct pnfs_layout_hdr *lo;
 734
 735         spin_lock(&ino->i_lock);
 736         lo = NFS_I(ino)->layout;
 737         lo->plh_block_lgets--;
 738         put_layout_hdr_locked(lo);
 739         spin_unlock(&ino->i_lock);
 740 }
 741
 742 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 743 {
 744         struct pnfs_layout_hdr *lo;
 745
 746         spin_lock(&ino->i_lock);
 747         lo = NFS_I(ino)->layout;
 748         if ((int)(barrier - lo->plh_barrier) > 0)
 749                 lo->plh_barrier = barrier;
 750         spin_unlock(&ino->i_lock);
 751 }
 752
 753 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
 754 {
 755         struct nfs_inode *nfsi = NFS_I(ino);
 756         struct pnfs_layout_segment *lseg;
 757         bool found = false;
 758
 759         spin_lock(&ino->i_lock);
 760         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 761                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 762                         found = true;
 763                         break;
 764                 }
 765         if (!found) {
 766                 struct pnfs_layout_hdr *lo = nfsi->layout;
 767                 u32 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 768
 769                 /* Since close does not return a layout stateid for use as
 770                  * a barrier, we choose the worst-case barrier.
 771                  */
 772                 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 773         }
 774         spin_unlock(&ino->i_lock);
 775         return found;
 776 }
 777
 778 /*
 779  * Compare two layout segments for sorting into layout cache.
 780  * We want to preferentially return RW over RO layouts, so ensure those
 781  * are seen first.
 782  */
 783 static s64
 784 cmp_layout(struct pnfs_layout_range *l1,
 785            struct pnfs_layout_range *l2)
 786 {
 787         s64 d;
 788
 789         /* high offset > low offset */
 790         d = l1->offset - l2->offset;
 791         if (d)
 792                 return d;
 793
 794         /* short length > long length */
 795         d = l2->length - l1->length;
 796         if (d)
 797                 return d;
 798
 799         /* read > read/write */
 800         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 801 }
 802
 803 static void
 804 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
 805                    struct pnfs_layout_segment *lseg)
 806 {
 807         struct pnfs_layout_segment *lp;
 808
 809         dprintk("%s:Begin\n", __func__);
 810
 811         assert_spin_locked(&lo->plh_inode->i_lock);
 812         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 813                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 814                         continue;
 815                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 816                 dprintk("%s: inserted lseg %p "
 817                         "iomode %d offset %llu length %llu before "
 818                         "lp %p iomode %d offset %llu length %llu\n",
 819                         __func__, lseg, lseg->pls_range.iomode,
 820                         lseg->pls_range.offset, lseg->pls_range.length,
 821                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 822                         lp->pls_range.length);
 823                 goto out;
 824         }
 825         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 826         dprintk("%s: inserted lseg %p "
 827                 "iomode %d offset %llu length %llu at tail\n",
 828                 __func__, lseg, lseg->pls_range.iomode,
 829                 lseg->pls_range.offset, lseg->pls_range.length);
 830 out:
 831         get_layout_hdr(lo);
 832
 833         dprintk("%s:Return\n", __func__);
 834 }
 835
 836 static struct pnfs_layout_hdr *
 837 alloc_init_layout_hdr(struct inode *ino,
 838                       struct nfs_open_context *ctx,
 839                       gfp_t gfp_flags)
 840 {
 841         struct pnfs_layout_hdr *lo;
 842
 843         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 844         if (!lo)
 845                 return NULL;
 846         atomic_set(&lo->plh_refcount, 1);
 847         INIT_LIST_HEAD(&lo->plh_layouts);
 848         INIT_LIST_HEAD(&lo->plh_segs);
 849         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 850         lo->plh_inode = ino;
 851         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 852         return lo;
 853 }
 854
 855 static struct pnfs_layout_hdr *
 856 pnfs_find_alloc_layout(struct inode *ino,
 857                        struct nfs_open_context *ctx,
 858                        gfp_t gfp_flags)
 859 {
 860         struct nfs_inode *nfsi = NFS_I(ino);
 861         struct pnfs_layout_hdr *new = NULL;
 862
 863         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 864
 865         assert_spin_locked(&ino->i_lock);
 866         if (nfsi->layout) {
 867                 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
 868                         return NULL;
 869                 else
 870                         return nfsi->layout;
 871         }
 872         spin_unlock(&ino->i_lock);
 873         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 874         spin_lock(&ino->i_lock);
 875
 876         if (likely(nfsi->layout == NULL))       /* Won the race? */
 877                 nfsi->layout = new;
 878         else
 879                 pnfs_free_layout_hdr(new);
 880         return nfsi->layout;
 881 }
 882
 883 /*
 884  * iomode matching rules:
 885  * iomode       lseg    match
 886  * -----        -----   -----
 887  * ANY          READ    true
 888  * ANY          RW      true
 889  * RW           READ    false
 890  * RW           RW      true
 891  * READ         READ    true
 892  * READ         RW      true
 893  */
 894 static int
 895 is_matching_lseg(struct pnfs_layout_range *ls_range,
 896                  struct pnfs_layout_range *range)
 897 {
 898         struct pnfs_layout_range range1;
 899
 900         if ((range->iomode == IOMODE_RW &&
 901              ls_range->iomode != IOMODE_RW) ||
 902             !lo_seg_intersecting(ls_range, range))
 903                 return 0;
 904
 905         /* range1 covers only the first byte in the range */
 906         range1 = *range;
 907         range1.length = 1;
 908         return lo_seg_contained(ls_range, &range1);
 909 }
 910
 911 /*
 912  * lookup range in layout
 913  */
 914 static struct pnfs_layout_segment *
 915 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 916                 struct pnfs_layout_range *range)
 917 {
 918         struct pnfs_layout_segment *lseg, *ret = NULL;
 919
 920         dprintk("%s:Begin\n", __func__);
 921
 922         assert_spin_locked(&lo->plh_inode->i_lock);
 923         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 924                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 925                     is_matching_lseg(&lseg->pls_range, range)) {
 926                         ret = get_lseg(lseg);
 927                         break;
 928                 }
 929                 if (lseg->pls_range.offset > range->offset)
 930                         break;
 931         }
 932
 933         dprintk("%s:Return lseg %p ref %d\n",
 934                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 935         return ret;
 936 }
 937
 938 /*
 939  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
 940  * to the MDS or over pNFS
 941  *
 942  * The nfs_inode read_io and write_io fields are cumulative counters reset
 943  * when there are no layout segments. Note that in pnfs_update_layout iomode
 944  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
 945  * WRITE request.
 946  *
 947  * A return of true means use MDS I/O.
 948  *
 949  * From rfc 5661:
 950  * If a file's size is smaller than the file size threshold, data accesses
 951  * SHOULD be sent to the metadata server.  If an I/O request has a length that
 952  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
 953  * server.  If both file size and I/O size are provided, the client SHOULD
 954  * reach or exceed  both thresholds before sending its read or write
 955  * requests to the data server.
 956  */
 957 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
 958                                      struct inode *ino, int iomode)
 959 {
 960         struct nfs4_threshold *t = ctx->mdsthreshold;
 961         struct nfs_inode *nfsi = NFS_I(ino);
 962         loff_t fsize = i_size_read(ino);
 963         bool size = false, size_set = false, io = false, io_set = false, ret = false;
 964
 965         if (t == NULL)
 966                 return ret;
 967
 968         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
 969                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
 970
 971         switch (iomode) {
 972         case IOMODE_READ:
 973                 if (t->bm & THRESHOLD_RD) {
 974                         dprintk("%s fsize %llu\n", __func__, fsize);
 975                         size_set = true;
 976                         if (fsize < t->rd_sz)
 977                                 size = true;
 978                 }
 979                 if (t->bm & THRESHOLD_RD_IO) {
 980                         dprintk("%s nfsi->read_io %llu\n", __func__,
 981                                 nfsi->read_io);
 982                         io_set = true;
 983                         if (nfsi->read_io < t->rd_io_sz)
 984                                 io = true;
 985                 }
 986                 break;
 987         case IOMODE_RW:
 988                 if (t->bm & THRESHOLD_WR) {
 989                         dprintk("%s fsize %llu\n", __func__, fsize);
 990                         size_set = true;
 991                         if (fsize < t->wr_sz)
 992                                 size = true;
 993                 }
 994                 if (t->bm & THRESHOLD_WR_IO) {
 995                         dprintk("%s nfsi->write_io %llu\n", __func__,
 996                                 nfsi->write_io);
 997                         io_set = true;
 998                         if (nfsi->write_io < t->wr_io_sz)
 999                                 io = true;
1000                 }
1001                 break;
1002         }
1003         if (size_set && io_set) {
1004                 if (size && io)
1005                         ret = true;
1006         } else if (size || io)
1007                 ret = true;
1008
1009         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1010         return ret;
1011 }
1012
1013 /*
1014  * Layout segment is retreived from the server if not cached.
1015  * The appropriate layout segment is referenced and returned to the caller.
1016  */
1017 struct pnfs_layout_segment *
1018 pnfs_update_layout(struct inode *ino,
1019                    struct nfs_open_context *ctx,
1020                    loff_t pos,
1021                    u64 count,
1022                    enum pnfs_iomode iomode,
1023                    gfp_t gfp_flags)
1024 {
1025         struct pnfs_layout_range arg = {
1026                 .iomode = iomode,
1027                 .offset = pos,
1028                 .length = count,
1029         };
1030         unsigned pg_offset;
1031         struct nfs_inode *nfsi = NFS_I(ino);
1032         struct nfs_server *server = NFS_SERVER(ino);
1033         struct nfs_client *clp = server->nfs_client;
1034         struct pnfs_layout_hdr *lo;
1035         struct pnfs_layout_segment *lseg = NULL;
1036         bool first = false;
1037
1038         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1039                 return NULL;
1040
1041         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1042                 return NULL;
1043
1044         spin_lock(&ino->i_lock);
1045         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1046         if (lo == NULL) {
1047                 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
1048                 goto out_unlock;
1049         }
1050
1051         /* Do we even need to bother with this? */
1052         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1053                 dprintk("%s matches recall, use MDS\n", __func__);
1054                 goto out_unlock;
1055         }
1056
1057         /* if LAYOUTGET already failed once we don't try again */
1058         if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
1059                 goto out_unlock;
1060
1061         /* Check to see if the layout for the given range already exists */
1062         lseg = pnfs_find_lseg(lo, &arg);
1063         if (lseg)
1064                 goto out_unlock;
1065
1066         if (pnfs_layoutgets_blocked(lo, NULL, 0))
1067                 goto out_unlock;
1068         atomic_inc(&lo->plh_outstanding);
1069
1070         get_layout_hdr(lo);
1071         if (list_empty(&lo->plh_segs))
1072                 first = true;
1073         spin_unlock(&ino->i_lock);
1074         if (first) {
1075                 /* The lo must be on the clp list if there is any
1076                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1077                  */
1078                 spin_lock(&clp->cl_lock);
1079                 BUG_ON(!list_empty(&lo->plh_layouts));
1080                 list_add_tail(&lo->plh_layouts, &server->layouts);
1081                 spin_unlock(&clp->cl_lock);
1082         }
1083
1084         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1085         if (pg_offset) {
1086                 arg.offset -= pg_offset;
1087                 arg.length += pg_offset;
1088         }
1089         if (arg.length != NFS4_MAX_UINT64)
1090                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1091
1092         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1093         if (!lseg && first) {
1094                 spin_lock(&clp->cl_lock);
1095                 list_del_init(&lo->plh_layouts);
1096                 spin_unlock(&clp->cl_lock);
1097         }
1098         atomic_dec(&lo->plh_outstanding);
1099         put_layout_hdr(lo);
1100 out:
1101         dprintk("%s end, state 0x%lx lseg %p\n", __func__,
1102                 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
1103         return lseg;
1104 out_unlock:
1105         spin_unlock(&ino->i_lock);
1106         goto out;
1107 }
1108 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1109
1110 int
1111 pnfs_layout_process(struct nfs4_layoutget *lgp)
1112 {
1113         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1114         struct nfs4_layoutget_res *res = &lgp->res;
1115         struct pnfs_layout_segment *lseg;
1116         struct inode *ino = lo->plh_inode;
1117         int status = 0;
1118
1119         /* Inject layout blob into I/O device driver */
1120         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1121         if (!lseg || IS_ERR(lseg)) {
1122                 if (!lseg)
1123                         status = -ENOMEM;
1124                 else
1125                         status = PTR_ERR(lseg);
1126                 dprintk("%s: Could not allocate layout: error %d\n",
1127                        __func__, status);
1128                 goto out;
1129         }
1130
1131         spin_lock(&ino->i_lock);
1132         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1133                 dprintk("%s forget reply due to recall\n", __func__);
1134                 goto out_forget_reply;
1135         }
1136
1137         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1138                 dprintk("%s forget reply due to state\n", __func__);
1139                 goto out_forget_reply;
1140         }
1141         init_lseg(lo, lseg);
1142         lseg->pls_range = res->range;
1143         *lgp->lsegpp = get_lseg(lseg);
1144         pnfs_insert_layout(lo, lseg);
1145
1146         if (res->return_on_close) {
1147                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1148                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1149         }
1150
1151         /* Done processing layoutget. Set the layout stateid */
1152         pnfs_set_layout_stateid(lo, &res->stateid, false);
1153         spin_unlock(&ino->i_lock);
1154 out:
1155         return status;
1156
1157 out_forget_reply:
1158         spin_unlock(&ino->i_lock);
1159         lseg->pls_layout = lo;
1160         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1161         goto out;
1162 }
1163
1164 void
1165 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1166 {
1167         BUG_ON(pgio->pg_lseg != NULL);
1168
1169         if (req->wb_offset != req->wb_pgbase) {
1170                 nfs_pageio_reset_read_mds(pgio);
1171                 return;
1172         }
1173         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1174                                            req->wb_context,
1175                                            req_offset(req),
1176                                            req->wb_bytes,
1177                                            IOMODE_READ,
1178                                            GFP_KERNEL);
1179         /* If no lseg, fall back to read through mds */
1180         if (pgio->pg_lseg == NULL)
1181                 nfs_pageio_reset_read_mds(pgio);
1182
1183 }
1184 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1185
1186 void
1187 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1188 {
1189         BUG_ON(pgio->pg_lseg != NULL);
1190
1191         if (req->wb_offset != req->wb_pgbase) {
1192                 nfs_pageio_reset_write_mds(pgio);
1193                 return;
1194         }
1195         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1196                                            req->wb_context,
1197                                            req_offset(req),
1198                                            req->wb_bytes,
1199                                            IOMODE_RW,
1200                                            GFP_NOFS);
1201         /* If no lseg, fall back to write through mds */
1202         if (pgio->pg_lseg == NULL)
1203                 nfs_pageio_reset_write_mds(pgio);
1204 }
1205 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1206
1207 bool
1208 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1209                       const struct nfs_pgio_completion_ops *compl_ops)
1210 {
1211         struct nfs_server *server = NFS_SERVER(inode);
1212         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1213
1214         if (ld == NULL)
1215                 return false;
1216         nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops,
1217                         server->rsize, 0);
1218         return true;
1219 }
1220
1221 bool
1222 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1223                        int ioflags,
1224                        const struct nfs_pgio_completion_ops *compl_ops)
1225 {
1226         struct nfs_server *server = NFS_SERVER(inode);
1227         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1228
1229         if (ld == NULL)
1230                 return false;
1231         nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops,
1232                         server->wsize, ioflags);
1233         return true;
1234 }
1235
1236 bool
1237 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1238                      struct nfs_page *req)
1239 {
1240         if (pgio->pg_lseg == NULL)
1241                 return nfs_generic_pg_test(pgio, prev, req);
1242
1243         /*
1244          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1245          * Note that this test makes several assumptions:
1246          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1247          *   is known to lie within the range.
1248          *   - that the nfs_page being tested is known to be contiguous with the
1249          *   previous nfs_page.
1250          *   - Layout ranges are page aligned, so we only have to test the
1251          *   start offset of the request.
1252          *
1253          * Please also note that 'end_offset' is actually the offset of the
1254          * first byte that lies outside the pnfs_layout_range. FIXME?
1255          *
1256          */
1257         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1258                                          pgio->pg_lseg->pls_range.length);
1259 }
1260 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1261
1262 int pnfs_write_done_resend_to_mds(struct inode *inode,
1263                                 struct list_head *head,
1264                                 const struct nfs_pgio_completion_ops *compl_ops)
1265 {
1266         struct nfs_pageio_descriptor pgio;
1267         LIST_HEAD(failed);
1268
1269         /* Resend all requests through the MDS */
1270         nfs_pageio_init_write_mds(&pgio, inode, FLUSH_STABLE, compl_ops);
1271         while (!list_empty(head)) {
1272                 struct nfs_page *req = nfs_list_entry(head->next);
1273
1274                 nfs_list_remove_request(req);
1275                 if (!nfs_pageio_add_request(&pgio, req))
1276                         nfs_list_add_request(req, &failed);
1277         }
1278         nfs_pageio_complete(&pgio);
1279
1280         if (!list_empty(&failed)) {
1281                 /* For some reason our attempt to resend pages. Mark the
1282                  * overall send request as having failed, and let
1283                  * nfs_writeback_release_full deal with the error.
1284                  */
1285                 list_move(&failed, head);
1286                 return -EIO;
1287         }
1288         return 0;
1289 }
1290 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1291
1292 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1293 {
1294         struct nfs_pgio_header *hdr = data->header;
1295
1296         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1297         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1298             PNFS_LAYOUTRET_ON_ERROR) {
1299                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1300                 pnfs_return_layout(hdr->inode);
1301         }
1302         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1303                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1304                                                         &hdr->pages,
1305                                                         hdr->completion_ops);
1306 }
1307
1308 /*
1309  * Called by non rpc-based layout drivers
1310  */
1311 void pnfs_ld_write_done(struct nfs_write_data *data)
1312 {
1313         struct nfs_pgio_header *hdr = data->header;
1314
1315         if (!hdr->pnfs_error) {
1316                 pnfs_set_layoutcommit(data);
1317                 hdr->mds_ops->rpc_call_done(&data->task, data);
1318         } else
1319                 pnfs_ld_handle_write_error(data);
1320         hdr->mds_ops->rpc_release(data);
1321 }
1322 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1323
1324 static void
1325 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1326                 struct nfs_write_data *data)
1327 {
1328         struct nfs_pgio_header *hdr = data->header;
1329
1330         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1331                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1332                 nfs_pageio_reset_write_mds(desc);
1333                 desc->pg_recoalesce = 1;
1334         }
1335         nfs_writedata_release(data);
1336 }
1337
1338 static enum pnfs_try_status
1339 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1340                         const struct rpc_call_ops *call_ops,
1341                         struct pnfs_layout_segment *lseg,
1342                         int how)
1343 {
1344         struct nfs_pgio_header *hdr = wdata->header;
1345         struct inode *inode = hdr->inode;
1346         enum pnfs_try_status trypnfs;
1347         struct nfs_server *nfss = NFS_SERVER(inode);
1348
1349         hdr->mds_ops = call_ops;
1350
1351         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1352                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1353         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1354         if (trypnfs != PNFS_NOT_ATTEMPTED)
1355                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1356         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1357         return trypnfs;
1358 }
1359
1360 static void
1361 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1362 {
1363         struct nfs_write_data *data;
1364         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1365         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1366
1367         desc->pg_lseg = NULL;
1368         while (!list_empty(head)) {
1369                 enum pnfs_try_status trypnfs;
1370
1371                 data = list_first_entry(head, struct nfs_write_data, list);
1372                 list_del_init(&data->list);
1373
1374                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1375                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1376                         pnfs_write_through_mds(desc, data);
1377         }
1378         put_lseg(lseg);
1379 }
1380
1381 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1382 {
1383         put_lseg(hdr->lseg);
1384         nfs_writehdr_free(hdr);
1385 }
1386
1387 int
1388 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1389 {
1390         struct nfs_write_header *whdr;
1391         struct nfs_pgio_header *hdr;
1392         int ret;
1393
1394         whdr = nfs_writehdr_alloc();
1395         if (!whdr) {
1396                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1397                 put_lseg(desc->pg_lseg);
1398                 desc->pg_lseg = NULL;
1399                 return -ENOMEM;
1400         }
1401         hdr = &whdr->header;
1402         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1403         hdr->lseg = get_lseg(desc->pg_lseg);
1404         atomic_inc(&hdr->refcnt);
1405         ret = nfs_generic_flush(desc, hdr);
1406         if (ret != 0) {
1407                 put_lseg(desc->pg_lseg);
1408                 desc->pg_lseg = NULL;
1409         } else
1410                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1411         if (atomic_dec_and_test(&hdr->refcnt))
1412                 hdr->completion_ops->completion(hdr);
1413         return ret;
1414 }
1415 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1416
1417 int pnfs_read_done_resend_to_mds(struct inode *inode,
1418                                 struct list_head *head,
1419                                 const struct nfs_pgio_completion_ops *compl_ops)
1420 {
1421         struct nfs_pageio_descriptor pgio;
1422         LIST_HEAD(failed);
1423
1424         /* Resend all requests through the MDS */
1425         nfs_pageio_init_read_mds(&pgio, inode, compl_ops);
1426         while (!list_empty(head)) {
1427                 struct nfs_page *req = nfs_list_entry(head->next);
1428
1429                 nfs_list_remove_request(req);
1430                 if (!nfs_pageio_add_request(&pgio, req))
1431                         nfs_list_add_request(req, &failed);
1432         }
1433         nfs_pageio_complete(&pgio);
1434
1435         if (!list_empty(&failed)) {
1436                 list_move(&failed, head);
1437                 return -EIO;
1438         }
1439         return 0;
1440 }
1441 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1442
1443 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1444 {
1445         struct nfs_pgio_header *hdr = data->header;
1446
1447         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1448         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1449             PNFS_LAYOUTRET_ON_ERROR) {
1450                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1451                 pnfs_return_layout(hdr->inode);
1452         }
1453         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1454                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1455                                                         &hdr->pages,
1456                                                         hdr->completion_ops);
1457 }
1458
1459 /*
1460  * Called by non rpc-based layout drivers
1461  */
1462 void pnfs_ld_read_done(struct nfs_read_data *data)
1463 {
1464         struct nfs_pgio_header *hdr = data->header;
1465
1466         if (likely(!hdr->pnfs_error)) {
1467                 __nfs4_read_done_cb(data);
1468                 hdr->mds_ops->rpc_call_done(&data->task, data);
1469         } else
1470                 pnfs_ld_handle_read_error(data);
1471         hdr->mds_ops->rpc_release(data);
1472 }
1473 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1474
1475 static void
1476 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1477                 struct nfs_read_data *data)
1478 {
1479         struct nfs_pgio_header *hdr = data->header;
1480
1481         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1482                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1483                 nfs_pageio_reset_read_mds(desc);
1484                 desc->pg_recoalesce = 1;
1485         }
1486         nfs_readdata_release(data);
1487 }
1488
1489 /*
1490  * Call the appropriate parallel I/O subsystem read function.
1491  */
1492 static enum pnfs_try_status
1493 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1494                        const struct rpc_call_ops *call_ops,
1495                        struct pnfs_layout_segment *lseg)
1496 {
1497         struct nfs_pgio_header *hdr = rdata->header;
1498         struct inode *inode = hdr->inode;
1499         struct nfs_server *nfss = NFS_SERVER(inode);
1500         enum pnfs_try_status trypnfs;
1501
1502         hdr->mds_ops = call_ops;
1503
1504         dprintk("%s: Reading ino:%lu %u@%llu\n",
1505                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1506
1507         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1508         if (trypnfs != PNFS_NOT_ATTEMPTED)
1509                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1510         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1511         return trypnfs;
1512 }
1513
1514 static void
1515 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1516 {
1517         struct nfs_read_data *data;
1518         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1519         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1520
1521         desc->pg_lseg = NULL;
1522         while (!list_empty(head)) {
1523                 enum pnfs_try_status trypnfs;
1524
1525                 data = list_first_entry(head, struct nfs_read_data, list);
1526                 list_del_init(&data->list);
1527
1528                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1529                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1530                         pnfs_read_through_mds(desc, data);
1531         }
1532         put_lseg(lseg);
1533 }
1534
1535 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1536 {
1537         put_lseg(hdr->lseg);
1538         nfs_readhdr_free(hdr);
1539 }
1540
1541 int
1542 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1543 {
1544         struct nfs_read_header *rhdr;
1545         struct nfs_pgio_header *hdr;
1546         int ret;
1547
1548         rhdr = nfs_readhdr_alloc();
1549         if (!rhdr) {
1550                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1551                 ret = -ENOMEM;
1552                 put_lseg(desc->pg_lseg);
1553                 desc->pg_lseg = NULL;
1554                 return ret;
1555         }
1556         hdr = &rhdr->header;
1557         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1558         hdr->lseg = get_lseg(desc->pg_lseg);
1559         atomic_inc(&hdr->refcnt);
1560         ret = nfs_generic_pagein(desc, hdr);
1561         if (ret != 0) {
1562                 put_lseg(desc->pg_lseg);
1563                 desc->pg_lseg = NULL;
1564         } else
1565                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1566         if (atomic_dec_and_test(&hdr->refcnt))
1567                 hdr->completion_ops->completion(hdr);
1568         return ret;
1569 }
1570 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1571
1572 /*
1573  * There can be multiple RW segments.
1574  */
1575 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1576 {
1577         struct pnfs_layout_segment *lseg;
1578
1579         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1580                 if (lseg->pls_range.iomode == IOMODE_RW &&
1581                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1582                         list_add(&lseg->pls_lc_list, listp);
1583         }
1584 }
1585
1586 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1587 {
1588         if (lseg->pls_range.iomode == IOMODE_RW) {
1589                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
1590                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
1591         } else {
1592                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
1593                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
1594         }
1595 }
1596 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1597
1598 void
1599 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1600 {
1601         struct nfs_pgio_header *hdr = wdata->header;
1602         struct inode *inode = hdr->inode;
1603         struct nfs_inode *nfsi = NFS_I(inode);
1604         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1605         bool mark_as_dirty = false;
1606
1607         spin_lock(&inode->i_lock);
1608         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1609                 mark_as_dirty = true;
1610                 dprintk("%s: Set layoutcommit for inode %lu ",
1611                         __func__, inode->i_ino);
1612         }
1613         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1614                 /* references matched in nfs4_layoutcommit_release */
1615                 get_lseg(hdr->lseg);
1616         }
1617         if (end_pos > nfsi->layout->plh_lwb)
1618                 nfsi->layout->plh_lwb = end_pos;
1619         spin_unlock(&inode->i_lock);
1620         dprintk("%s: lseg %p end_pos %llu\n",
1621                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1622
1623         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1624          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1625         if (mark_as_dirty)
1626                 mark_inode_dirty_sync(inode);
1627 }
1628 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1629
1630 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1631 {
1632         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1633
1634         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1635                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1636 }
1637
1638 /*
1639  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1640  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1641  * data to disk to allow the server to recover the data if it crashes.
1642  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1643  * is off, and a COMMIT is sent to a data server, or
1644  * if WRITEs to a data server return NFS_DATA_SYNC.
1645  */
1646 int
1647 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1648 {
1649         struct nfs4_layoutcommit_data *data;
1650         struct nfs_inode *nfsi = NFS_I(inode);
1651         loff_t end_pos;
1652         int status = 0;
1653
1654         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1655
1656         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1657                 return 0;
1658
1659         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1660         data = kzalloc(sizeof(*data), GFP_NOFS);
1661         if (!data) {
1662                 status = -ENOMEM;
1663                 goto out;
1664         }
1665
1666         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1667                 goto out_free;
1668
1669         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1670                 if (!sync) {
1671                         status = -EAGAIN;
1672                         goto out_free;
1673                 }
1674                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1675                                         nfs_wait_bit_killable, TASK_KILLABLE);
1676                 if (status)
1677                         goto out_free;
1678         }
1679
1680         INIT_LIST_HEAD(&data->lseg_list);
1681         spin_lock(&inode->i_lock);
1682         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1683                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1684                 spin_unlock(&inode->i_lock);
1685                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1686                 goto out_free;
1687         }
1688
1689         pnfs_list_write_lseg(inode, &data->lseg_list);
1690
1691         end_pos = nfsi->layout->plh_lwb;
1692         nfsi->layout->plh_lwb = 0;
1693
1694         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1695         spin_unlock(&inode->i_lock);
1696
1697         data->args.inode = inode;
1698         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1699         nfs_fattr_init(&data->fattr);
1700         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1701         data->res.fattr = &data->fattr;
1702         data->args.lastbytewritten = end_pos - 1;
1703         data->res.server = NFS_SERVER(inode);
1704
1705         status = nfs4_proc_layoutcommit(data, sync);
1706 out:
1707         if (status)
1708                 mark_inode_dirty_sync(inode);
1709         dprintk("<-- %s status %d\n", __func__, status);
1710         return status;
1711 out_free:
1712         kfree(data);
1713         goto out;
1714 }
1715
1716 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1717 {
1718         struct nfs4_threshold *thp;
1719
1720         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1721         if (!thp) {
1722                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1723                 return NULL;
1724         }
1725         return thp;
1726 }