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 <linux/sort.h>
  34 #include "internal.h"
  35 #include "pnfs.h"
  36 #include "iostat.h"
  37 #include "nfs4trace.h"
  38 #include "delegation.h"
  39 #include "nfs42.h"
  40 #include "nfs4_fs.h"
  41
  42 #define NFSDBG_FACILITY         NFSDBG_PNFS
  43 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  44
  45 /* Locking:
  46  *
  47  * pnfs_spinlock:
  48  *      protects pnfs_modules_tbl.
  49  */
  50 static DEFINE_SPINLOCK(pnfs_spinlock);
  51
  52 /*
  53  * pnfs_modules_tbl holds all pnfs modules
  54  */
  55 static LIST_HEAD(pnfs_modules_tbl);
  56
  57 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  58 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  59                 struct list_head *free_me,
  60                 const struct pnfs_layout_range *range,
  61                 u32 seq);
  62 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  63                                 struct list_head *tmp_list);
  64
  65 /* Return the registered pnfs layout driver module matching given id */
  66 static struct pnfs_layoutdriver_type *
  67 find_pnfs_driver_locked(u32 id)
  68 {
  69         struct pnfs_layoutdriver_type *local;
  70
  71         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  72                 if (local->id == id)
  73                         goto out;
  74         local = NULL;
  75 out:
  76         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  77         return local;
  78 }
  79
  80 static struct pnfs_layoutdriver_type *
  81 find_pnfs_driver(u32 id)
  82 {
  83         struct pnfs_layoutdriver_type *local;
  84
  85         spin_lock(&pnfs_spinlock);
  86         local = find_pnfs_driver_locked(id);
  87         if (local != NULL && !try_module_get(local->owner)) {
  88                 dprintk("%s: Could not grab reference on module\n", __func__);
  89                 local = NULL;
  90         }
  91         spin_unlock(&pnfs_spinlock);
  92         return local;
  93 }
  94
  95 const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
  96 {
  97         return find_pnfs_driver(id);
  98 }
  99
 100 void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
 101 {
 102         if (ld)
 103                 module_put(ld->owner);
 104 }
 105
 106 void
 107 unset_pnfs_layoutdriver(struct nfs_server *nfss)
 108 {
 109         if (nfss->pnfs_curr_ld) {
 110                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
 111                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 112                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 113                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 114                         nfs4_deviceid_purge_client(nfss->nfs_client);
 115                 module_put(nfss->pnfs_curr_ld->owner);
 116         }
 117         nfss->pnfs_curr_ld = NULL;
 118 }
 119
 120 /*
 121  * When the server sends a list of layout types, we choose one in the order
 122  * given in the list below.
 123  *
 124  * FIXME: should this list be configurable in some fashion? module param?
 125  *        mount option? something else?
 126  */
 127 static const u32 ld_prefs[] = {
 128         LAYOUT_SCSI,
 129         LAYOUT_BLOCK_VOLUME,
 130         LAYOUT_OSD2_OBJECTS,
 131         LAYOUT_FLEX_FILES,
 132         LAYOUT_NFSV4_1_FILES,
 133         0
 134 };
 135
 136 static int
 137 ld_cmp(const void *e1, const void *e2)
 138 {
 139         u32 ld1 = *((u32 *)e1);
 140         u32 ld2 = *((u32 *)e2);
 141         int i;
 142
 143         for (i = 0; ld_prefs[i] != 0; i++) {
 144                 if (ld1 == ld_prefs[i])
 145                         return -1;
 146
 147                 if (ld2 == ld_prefs[i])
 148                         return 1;
 149         }
 150         return 0;
 151 }
 152
 153 /*
 154  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 155  * Currently only one pNFS layout driver per filesystem is supported.
 156  *
 157  * @ids array of layout types supported by MDS.
 158  */
 159 void
 160 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 161                       struct nfs_fsinfo *fsinfo)
 162 {
 163         struct pnfs_layoutdriver_type *ld_type = NULL;
 164         u32 id;
 165         int i;
 166
 167         if (fsinfo->nlayouttypes == 0)
 168                 goto out_no_driver;
 169         if (!(server->nfs_client->cl_exchange_flags &
 170                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 171                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 172                         __func__, server->nfs_client->cl_exchange_flags);
 173                 goto out_no_driver;
 174         }
 175
 176         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 177                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 178
 179         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 180                 id = fsinfo->layouttype[i];
 181                 ld_type = find_pnfs_driver(id);
 182                 if (!ld_type) {
 183                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 184                                         id);
 185                         ld_type = find_pnfs_driver(id);
 186                 }
 187                 if (ld_type)
 188                         break;
 189         }
 190
 191         if (!ld_type) {
 192                 dprintk("%s: No pNFS module found!\n", __func__);
 193                 goto out_no_driver;
 194         }
 195
 196         server->pnfs_curr_ld = ld_type;
 197         if (ld_type->set_layoutdriver
 198             && ld_type->set_layoutdriver(server, mntfh)) {
 199                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 200                         "driver %u.\n", __func__, id);
 201                 module_put(ld_type->owner);
 202                 goto out_no_driver;
 203         }
 204         /* Bump the MDS count */
 205         atomic_inc(&server->nfs_client->cl_mds_count);
 206
 207         dprintk("%s: pNFS module for %u set\n", __func__, id);
 208         return;
 209
 210 out_no_driver:
 211         dprintk("%s: Using NFSv4 I/O\n", __func__);
 212         server->pnfs_curr_ld = NULL;
 213 }
 214
 215 int
 216 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 217 {
 218         int status = -EINVAL;
 219         struct pnfs_layoutdriver_type *tmp;
 220
 221         if (ld_type->id == 0) {
 222                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 223                 return status;
 224         }
 225         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 226                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 227                        "alloc_lseg and free_lseg.\n", __func__);
 228                 return status;
 229         }
 230
 231         spin_lock(&pnfs_spinlock);
 232         tmp = find_pnfs_driver_locked(ld_type->id);
 233         if (!tmp) {
 234                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 235                 status = 0;
 236                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 237                         ld_type->name);
 238         } else {
 239                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 240                         __func__, ld_type->id);
 241         }
 242         spin_unlock(&pnfs_spinlock);
 243
 244         return status;
 245 }
 246 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 247
 248 void
 249 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 250 {
 251         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 252         spin_lock(&pnfs_spinlock);
 253         list_del(&ld_type->pnfs_tblid);
 254         spin_unlock(&pnfs_spinlock);
 255 }
 256 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 257
 258 /*
 259  * pNFS client layout cache
 260  */
 261
 262 /* Need to hold i_lock if caller does not already hold reference */
 263 void
 264 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 265 {
 266         refcount_inc(&lo->plh_refcount);
 267 }
 268
 269 static struct pnfs_layout_hdr *
 270 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 271 {
 272         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 273         return ld->alloc_layout_hdr(ino, gfp_flags);
 274 }
 275
 276 static void
 277 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 278 {
 279         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 280         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 281
 282         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
 283                 struct nfs_client *clp = server->nfs_client;
 284
 285                 spin_lock(&clp->cl_lock);
 286                 list_del_rcu(&lo->plh_layouts);
 287                 spin_unlock(&clp->cl_lock);
 288         }
 289         put_cred(lo->plh_lc_cred);
 290         return ld->free_layout_hdr(lo);
 291 }
 292
 293 static void
 294 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 295 {
 296         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 297         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 298         nfsi->layout = NULL;
 299         /* Reset MDS Threshold I/O counters */
 300         nfsi->write_io = 0;
 301         nfsi->read_io = 0;
 302 }
 303
 304 void
 305 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 306 {
 307         struct inode *inode;
 308         unsigned long i_state;
 309
 310         if (!lo)
 311                 return;
 312         inode = lo->plh_inode;
 313         pnfs_layoutreturn_before_put_layout_hdr(lo);
 314
 315         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 316                 if (!list_empty(&lo->plh_segs))
 317                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 318                 pnfs_detach_layout_hdr(lo);
 319                 i_state = inode->i_state;
 320                 spin_unlock(&inode->i_lock);
 321                 pnfs_free_layout_hdr(lo);
 322                 /* Notify pnfs_destroy_layout_final() that we're done */
 323                 if (i_state & (I_FREEING | I_CLEAR))
 324                         wake_up_var(lo);
 325         }
 326 }
 327
 328 static struct inode *
 329 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
 330 {
 331         struct inode *inode = igrab(lo->plh_inode);
 332         if (inode)
 333                 return inode;
 334         set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 335         return NULL;
 336 }
 337
 338 /*
 339  * Compare 2 layout stateid sequence ids, to see which is newer,
 340  * taking into account wraparound issues.
 341  */
 342 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 343 {
 344         return (s32)(s1 - s2) > 0;
 345 }
 346
 347 static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
 348 {
 349         if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
 350                 lo->plh_barrier = newseq;
 351 }
 352
 353 static void
 354 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 355                          u32 seq)
 356 {
 357         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 358                 iomode = IOMODE_ANY;
 359         lo->plh_return_iomode = iomode;
 360         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 361         /*
 362          * We must set lo->plh_return_seq to avoid livelocks with
 363          * pnfs_layout_need_return()
 364          */
 365         if (seq == 0)
 366                 seq = be32_to_cpu(lo->plh_stateid.seqid);
 367         if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
 368                 lo->plh_return_seq = seq;
 369         pnfs_barrier_update(lo, seq);
 370 }
 371
 372 static void
 373 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 374 {
 375         struct pnfs_layout_segment *lseg;
 376         lo->plh_return_iomode = 0;
 377         lo->plh_return_seq = 0;
 378         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 379         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 380                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 381                         continue;
 382                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 383         }
 384 }
 385
 386 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 387 {
 388         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 389         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 390         smp_mb__after_atomic();
 391         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 392         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 393 }
 394
 395 static void
 396 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 397                 struct list_head *free_me)
 398 {
 399         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 400         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 401         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 402                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 403         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 404                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 405 }
 406
 407 /*
 408  * Update the seqid of a layout stateid after receiving
 409  * NFS4ERR_OLD_STATEID
 410  */
 411 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
 412                 struct pnfs_layout_range *dst_range,
 413                 struct inode *inode)
 414 {
 415         struct pnfs_layout_hdr *lo;
 416         struct pnfs_layout_range range = {
 417                 .iomode = IOMODE_ANY,
 418                 .offset = 0,
 419                 .length = NFS4_MAX_UINT64,
 420         };
 421         bool ret = false;
 422         LIST_HEAD(head);
 423         int err;
 424
 425         spin_lock(&inode->i_lock);
 426         lo = NFS_I(inode)->layout;
 427         if (lo &&  pnfs_layout_is_valid(lo) &&
 428             nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 429                 /* Is our call using the most recent seqid? If so, bump it */
 430                 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
 431                         nfs4_stateid_seqid_inc(dst);
 432                         ret = true;
 433                         goto out;
 434                 }
 435                 /* Try to update the seqid to the most recent */
 436                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 437                 if (err != -EBUSY) {
 438                         dst->seqid = lo->plh_stateid.seqid;
 439                         *dst_range = range;
 440                         ret = true;
 441                 }
 442         }
 443 out:
 444         spin_unlock(&inode->i_lock);
 445         pnfs_free_lseg_list(&head);
 446         return ret;
 447 }
 448
 449 /*
 450  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 451  *
 452  * In order to continue using the pnfs_layout_hdr, a full recovery
 453  * is required.
 454  * Note that caller must hold inode->i_lock.
 455  */
 456 int
 457 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 458                 struct list_head *lseg_list)
 459 {
 460         struct pnfs_layout_range range = {
 461                 .iomode = IOMODE_ANY,
 462                 .offset = 0,
 463                 .length = NFS4_MAX_UINT64,
 464         };
 465         struct pnfs_layout_segment *lseg, *next;
 466
 467         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 468         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 469                 pnfs_clear_lseg_state(lseg, lseg_list);
 470         pnfs_clear_layoutreturn_info(lo);
 471         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 472         set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
 473         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 474             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 475                 pnfs_clear_layoutreturn_waitbit(lo);
 476         return !list_empty(&lo->plh_segs);
 477 }
 478
 479 static int
 480 pnfs_iomode_to_fail_bit(u32 iomode)
 481 {
 482         return iomode == IOMODE_RW ?
 483                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 484 }
 485
 486 static void
 487 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 488 {
 489         lo->plh_retry_timestamp = jiffies;
 490         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 491                 refcount_inc(&lo->plh_refcount);
 492 }
 493
 494 static void
 495 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 496 {
 497         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 498                 refcount_dec(&lo->plh_refcount);
 499 }
 500
 501 static void
 502 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 503 {
 504         struct inode *inode = lo->plh_inode;
 505         struct pnfs_layout_range range = {
 506                 .iomode = iomode,
 507                 .offset = 0,
 508                 .length = NFS4_MAX_UINT64,
 509         };
 510         LIST_HEAD(head);
 511
 512         spin_lock(&inode->i_lock);
 513         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 514         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 515         spin_unlock(&inode->i_lock);
 516         pnfs_free_lseg_list(&head);
 517         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 518                         iomode == IOMODE_RW ?  "RW" : "READ");
 519 }
 520
 521 static bool
 522 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 523 {
 524         unsigned long start, end;
 525         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 526
 527         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 528                 return false;
 529         end = jiffies;
 530         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 531         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 532                 /* It is time to retry the failed layoutgets */
 533                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 534                 return false;
 535         }
 536         return true;
 537 }
 538
 539 static void
 540 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 541                 const struct pnfs_layout_range *range,
 542                 const nfs4_stateid *stateid)
 543 {
 544         INIT_LIST_HEAD(&lseg->pls_list);
 545         INIT_LIST_HEAD(&lseg->pls_lc_list);
 546         INIT_LIST_HEAD(&lseg->pls_commits);
 547         refcount_set(&lseg->pls_refcount, 1);
 548         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 549         lseg->pls_layout = lo;
 550         lseg->pls_range = *range;
 551         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 552 }
 553
 554 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 555 {
 556         if (lseg != NULL) {
 557                 struct inode *inode = lseg->pls_layout->plh_inode;
 558                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 559         }
 560 }
 561
 562 static void
 563 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 564                 struct pnfs_layout_segment *lseg)
 565 {
 566         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 567         list_del_init(&lseg->pls_list);
 568         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 569         refcount_dec(&lo->plh_refcount);
 570         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 571                 return;
 572         if (list_empty(&lo->plh_segs) &&
 573             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 574             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 575                 if (atomic_read(&lo->plh_outstanding) == 0)
 576                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 577                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 578         }
 579 }
 580
 581 static bool
 582 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 583                 struct pnfs_layout_segment *lseg)
 584 {
 585         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 586             pnfs_layout_is_valid(lo)) {
 587                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 588                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 589                 return true;
 590         }
 591         return false;
 592 }
 593
 594 void
 595 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 596 {
 597         struct pnfs_layout_hdr *lo;
 598         struct inode *inode;
 599
 600         if (!lseg)
 601                 return;
 602
 603         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 604                 refcount_read(&lseg->pls_refcount),
 605                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 606
 607         lo = lseg->pls_layout;
 608         inode = lo->plh_inode;
 609
 610         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 611                 pnfs_get_layout_hdr(lo);
 612                 pnfs_layout_remove_lseg(lo, lseg);
 613                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 614                         lseg = NULL;
 615                 spin_unlock(&inode->i_lock);
 616                 pnfs_free_lseg(lseg);
 617                 pnfs_put_layout_hdr(lo);
 618         }
 619 }
 620 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 621
 622 /*
 623  * is l2 fully contained in l1?
 624  *   start1                             end1
 625  *   [----------------------------------)
 626  *           start2           end2
 627  *           [----------------)
 628  */
 629 static bool
 630 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 631                  const struct pnfs_layout_range *l2)
 632 {
 633         u64 start1 = l1->offset;
 634         u64 end1 = pnfs_end_offset(start1, l1->length);
 635         u64 start2 = l2->offset;
 636         u64 end2 = pnfs_end_offset(start2, l2->length);
 637
 638         return (start1 <= start2) && (end1 >= end2);
 639 }
 640
 641 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 642                 struct list_head *tmp_list)
 643 {
 644         if (!refcount_dec_and_test(&lseg->pls_refcount))
 645                 return false;
 646         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 647         list_add(&lseg->pls_list, tmp_list);
 648         return true;
 649 }
 650
 651 /* Returns 1 if lseg is removed from list, 0 otherwise */
 652 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 653                              struct list_head *tmp_list)
 654 {
 655         int rv = 0;
 656
 657         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 658                 /* Remove the reference keeping the lseg in the
 659                  * list.  It will now be removed when all
 660                  * outstanding io is finished.
 661                  */
 662                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 663                         refcount_read(&lseg->pls_refcount));
 664                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 665                         rv = 1;
 666         }
 667         return rv;
 668 }
 669
 670 static bool
 671 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 672                  const struct pnfs_layout_range *recall_range)
 673 {
 674         return (recall_range->iomode == IOMODE_ANY ||
 675                 lseg_range->iomode == recall_range->iomode) &&
 676                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 677 }
 678
 679 static bool
 680 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 681                 const struct pnfs_layout_range *recall_range,
 682                 u32 seq)
 683 {
 684         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 685                 return false;
 686         if (recall_range == NULL)
 687                 return true;
 688         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 689 }
 690
 691 /**
 692  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 693  * @lo: layout header containing the lsegs
 694  * @tmp_list: list head where doomed lsegs should go
 695  * @recall_range: optional recall range argument to match (may be NULL)
 696  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 697  *
 698  * Walk the list of lsegs in the layout header, and tear down any that should
 699  * be destroyed. If "recall_range" is specified then the segment must match
 700  * that range. If "seq" is non-zero, then only match segments that were handed
 701  * out at or before that sequence.
 702  *
 703  * Returns number of matching invalid lsegs remaining in list after scanning
 704  * it and purging them.
 705  */
 706 int
 707 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 708                             struct list_head *tmp_list,
 709                             const struct pnfs_layout_range *recall_range,
 710                             u32 seq)
 711 {
 712         struct pnfs_layout_segment *lseg, *next;
 713         int remaining = 0;
 714
 715         dprintk("%s:Begin lo %p\n", __func__, lo);
 716
 717         if (list_empty(&lo->plh_segs))
 718                 return 0;
 719         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 720                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 721                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 722                                 "offset %llu length %llu\n", __func__,
 723                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 724                                 lseg->pls_range.offset, lseg->pls_range.length);
 725                         if (!mark_lseg_invalid(lseg, tmp_list))
 726                                 remaining++;
 727                 }
 728         dprintk("%s:Return %i\n", __func__, remaining);
 729         return remaining;
 730 }
 731
 732 static void
 733 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 734                 struct list_head *free_me,
 735                 const struct pnfs_layout_range *range,
 736                 u32 seq)
 737 {
 738         struct pnfs_layout_segment *lseg, *next;
 739
 740         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 741                 if (pnfs_match_lseg_recall(lseg, range, seq))
 742                         list_move_tail(&lseg->pls_list, free_me);
 743         }
 744 }
 745
 746 /* note free_me must contain lsegs from a single layout_hdr */
 747 void
 748 pnfs_free_lseg_list(struct list_head *free_me)
 749 {
 750         struct pnfs_layout_segment *lseg, *tmp;
 751
 752         if (list_empty(free_me))
 753                 return;
 754
 755         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 756                 list_del(&lseg->pls_list);
 757                 pnfs_free_lseg(lseg);
 758         }
 759 }
 760
 761 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 762 {
 763         struct pnfs_layout_hdr *lo;
 764         LIST_HEAD(tmp_list);
 765
 766         spin_lock(&nfsi->vfs_inode.i_lock);
 767         lo = nfsi->layout;
 768         if (lo) {
 769                 pnfs_get_layout_hdr(lo);
 770                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 771                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 772                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 773                 spin_unlock(&nfsi->vfs_inode.i_lock);
 774                 pnfs_free_lseg_list(&tmp_list);
 775                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 776                 pnfs_put_layout_hdr(lo);
 777         } else
 778                 spin_unlock(&nfsi->vfs_inode.i_lock);
 779         return lo;
 780 }
 781
 782 void pnfs_destroy_layout(struct nfs_inode *nfsi)
 783 {
 784         __pnfs_destroy_layout(nfsi);
 785 }
 786 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 787
 788 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
 789                                 struct pnfs_layout_hdr *lo)
 790 {
 791         bool ret;
 792
 793         spin_lock(&nfsi->vfs_inode.i_lock);
 794         ret = nfsi->layout != lo;
 795         spin_unlock(&nfsi->vfs_inode.i_lock);
 796         return ret;
 797 }
 798
 799 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 800 {
 801         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 802
 803         if (lo)
 804                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
 805 }
 806
 807 static bool
 808 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 809                 struct list_head *layout_list)
 810 {
 811         struct pnfs_layout_hdr *lo;
 812         bool ret = false;
 813
 814         spin_lock(&inode->i_lock);
 815         lo = NFS_I(inode)->layout;
 816         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 817                 pnfs_get_layout_hdr(lo);
 818                 list_add(&lo->plh_bulk_destroy, layout_list);
 819                 ret = true;
 820         }
 821         spin_unlock(&inode->i_lock);
 822         return ret;
 823 }
 824
 825 /* Caller must hold rcu_read_lock and clp->cl_lock */
 826 static int
 827 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 828                 struct nfs_server *server,
 829                 struct list_head *layout_list)
 830         __must_hold(&clp->cl_lock)
 831         __must_hold(RCU)
 832 {
 833         struct pnfs_layout_hdr *lo, *next;
 834         struct inode *inode;
 835
 836         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 837                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 838                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 839                     !list_empty(&lo->plh_bulk_destroy))
 840                         continue;
 841                 /* If the sb is being destroyed, just bail */
 842                 if (!nfs_sb_active(server->super))
 843                         break;
 844                 inode = pnfs_grab_inode_layout_hdr(lo);
 845                 if (inode != NULL) {
 846                         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
 847                                 list_del_rcu(&lo->plh_layouts);
 848                         if (pnfs_layout_add_bulk_destroy_list(inode,
 849                                                 layout_list))
 850                                 continue;
 851                         rcu_read_unlock();
 852                         spin_unlock(&clp->cl_lock);
 853                         iput(inode);
 854                 } else {
 855                         rcu_read_unlock();
 856                         spin_unlock(&clp->cl_lock);
 857                 }
 858                 nfs_sb_deactive(server->super);
 859                 spin_lock(&clp->cl_lock);
 860                 rcu_read_lock();
 861                 return -EAGAIN;
 862         }
 863         return 0;
 864 }
 865
 866 static int
 867 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 868                 bool is_bulk_recall)
 869 {
 870         struct pnfs_layout_hdr *lo;
 871         struct inode *inode;
 872         LIST_HEAD(lseg_list);
 873         int ret = 0;
 874
 875         while (!list_empty(layout_list)) {
 876                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 877                                 plh_bulk_destroy);
 878                 dprintk("%s freeing layout for inode %lu\n", __func__,
 879                         lo->plh_inode->i_ino);
 880                 inode = lo->plh_inode;
 881
 882                 pnfs_layoutcommit_inode(inode, false);
 883
 884                 spin_lock(&inode->i_lock);
 885                 list_del_init(&lo->plh_bulk_destroy);
 886                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 887                         if (is_bulk_recall)
 888                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 889                         ret = -EAGAIN;
 890                 }
 891                 spin_unlock(&inode->i_lock);
 892                 pnfs_free_lseg_list(&lseg_list);
 893                 /* Free all lsegs that are attached to commit buckets */
 894                 nfs_commit_inode(inode, 0);
 895                 pnfs_put_layout_hdr(lo);
 896                 nfs_iput_and_deactive(inode);
 897         }
 898         return ret;
 899 }
 900
 901 int
 902 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 903                 struct nfs_fsid *fsid,
 904                 bool is_recall)
 905 {
 906         struct nfs_server *server;
 907         LIST_HEAD(layout_list);
 908
 909         spin_lock(&clp->cl_lock);
 910         rcu_read_lock();
 911 restart:
 912         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 913                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 914                         continue;
 915                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 916                                 server,
 917                                 &layout_list) != 0)
 918                         goto restart;
 919         }
 920         rcu_read_unlock();
 921         spin_unlock(&clp->cl_lock);
 922
 923         if (list_empty(&layout_list))
 924                 return 0;
 925         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 926 }
 927
 928 int
 929 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 930                 bool is_recall)
 931 {
 932         struct nfs_server *server;
 933         LIST_HEAD(layout_list);
 934
 935         spin_lock(&clp->cl_lock);
 936         rcu_read_lock();
 937 restart:
 938         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 939                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 940                                         server,
 941                                         &layout_list) != 0)
 942                         goto restart;
 943         }
 944         rcu_read_unlock();
 945         spin_unlock(&clp->cl_lock);
 946
 947         if (list_empty(&layout_list))
 948                 return 0;
 949         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 950 }
 951
 952 /*
 953  * Called by the state manager to remove all layouts established under an
 954  * expired lease.
 955  */
 956 void
 957 pnfs_destroy_all_layouts(struct nfs_client *clp)
 958 {
 959         nfs4_deviceid_mark_client_invalid(clp);
 960         nfs4_deviceid_purge_client(clp);
 961
 962         pnfs_destroy_layouts_byclid(clp, false);
 963 }
 964
 965 static void
 966 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
 967 {
 968         const struct cred *old;
 969
 970         if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
 971                 old = xchg(&lo->plh_lc_cred, get_cred(cred));
 972                 put_cred(old);
 973         }
 974 }
 975
 976 /* update lo->plh_stateid with new if is more recent */
 977 void
 978 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 979                         const struct cred *cred, bool update_barrier)
 980 {
 981         u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 982         u32 newseq = be32_to_cpu(new->seqid);
 983
 984         if (!pnfs_layout_is_valid(lo)) {
 985                 pnfs_set_layout_cred(lo, cred);
 986                 nfs4_stateid_copy(&lo->plh_stateid, new);
 987                 lo->plh_barrier = newseq;
 988                 pnfs_clear_layoutreturn_info(lo);
 989                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 990                 return;
 991         }
 992
 993         if (pnfs_seqid_is_newer(newseq, oldseq))
 994                 nfs4_stateid_copy(&lo->plh_stateid, new);
 995
 996         if (update_barrier) {
 997                 pnfs_barrier_update(lo, newseq);
 998                 return;
 999         }
1000         /*
1001          * Because of wraparound, we want to keep the barrier
1002          * "close" to the current seqids. We really only want to
1003          * get here from a layoutget call.
1004          */
1005         if (atomic_read(&lo->plh_outstanding) == 1)
1006                  pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1007 }
1008
1009 static bool
1010 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1011                 const nfs4_stateid *stateid)
1012 {
1013         u32 seqid = be32_to_cpu(stateid->seqid);
1014
1015         return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1016 }
1017
1018 /* lget is set to 1 if called from inside send_layoutget call chain */
1019 static bool
1020 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1021 {
1022         return lo->plh_block_lgets ||
1023                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1024 }
1025
1026 static struct nfs_server *
1027 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1028 {
1029         struct nfs_server *server;
1030
1031         if (inode) {
1032                 server = NFS_SERVER(inode);
1033         } else {
1034                 struct dentry *parent_dir = dget_parent(ctx->dentry);
1035                 server = NFS_SERVER(parent_dir->d_inode);
1036                 dput(parent_dir);
1037         }
1038         return server;
1039 }
1040
1041 static void nfs4_free_pages(struct page **pages, size_t size)
1042 {
1043         int i;
1044
1045         if (!pages)
1046                 return;
1047
1048         for (i = 0; i < size; i++) {
1049                 if (!pages[i])
1050                         break;
1051                 __free_page(pages[i]);
1052         }
1053         kfree(pages);
1054 }
1055
1056 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1057 {
1058         struct page **pages;
1059         int i;
1060
1061         pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1062         if (!pages) {
1063                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1064                 return NULL;
1065         }
1066
1067         for (i = 0; i < size; i++) {
1068                 pages[i] = alloc_page(gfp_flags);
1069                 if (!pages[i]) {
1070                         dprintk("%s: failed to allocate page\n", __func__);
1071                         nfs4_free_pages(pages, i);
1072                         return NULL;
1073                 }
1074         }
1075
1076         return pages;
1077 }
1078
1079 static struct nfs4_layoutget *
1080 pnfs_alloc_init_layoutget_args(struct inode *ino,
1081            struct nfs_open_context *ctx,
1082            const nfs4_stateid *stateid,
1083            const struct pnfs_layout_range *range,
1084            gfp_t gfp_flags)
1085 {
1086         struct nfs_server *server = pnfs_find_server(ino, ctx);
1087         size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1088         size_t max_pages = max_response_pages(server);
1089         struct nfs4_layoutget *lgp;
1090
1091         dprintk("--> %s\n", __func__);
1092
1093         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1094         if (lgp == NULL)
1095                 return NULL;
1096
1097         if (max_reply_sz) {
1098                 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1099                 if (npages < max_pages)
1100                         max_pages = npages;
1101         }
1102
1103         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1104         if (!lgp->args.layout.pages) {
1105                 kfree(lgp);
1106                 return NULL;
1107         }
1108         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1109         lgp->res.layoutp = &lgp->args.layout;
1110
1111         /* Don't confuse uninitialised result and success */
1112         lgp->res.status = -NFS4ERR_DELAY;
1113
1114         lgp->args.minlength = PAGE_SIZE;
1115         if (lgp->args.minlength > range->length)
1116                 lgp->args.minlength = range->length;
1117         if (ino) {
1118                 loff_t i_size = i_size_read(ino);
1119
1120                 if (range->iomode == IOMODE_READ) {
1121                         if (range->offset >= i_size)
1122                                 lgp->args.minlength = 0;
1123                         else if (i_size - range->offset < lgp->args.minlength)
1124                                 lgp->args.minlength = i_size - range->offset;
1125                 }
1126         }
1127         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1128         pnfs_copy_range(&lgp->args.range, range);
1129         lgp->args.type = server->pnfs_curr_ld->id;
1130         lgp->args.inode = ino;
1131         lgp->args.ctx = get_nfs_open_context(ctx);
1132         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1133         lgp->gfp_flags = gfp_flags;
1134         lgp->cred = ctx->cred;
1135         return lgp;
1136 }
1137
1138 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1139 {
1140         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1141
1142         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1143         pnfs_put_layout_hdr(lgp->lo);
1144         put_nfs_open_context(lgp->args.ctx);
1145         kfree(lgp);
1146 }
1147
1148 static void pnfs_clear_layoutcommit(struct inode *inode,
1149                 struct list_head *head)
1150 {
1151         struct nfs_inode *nfsi = NFS_I(inode);
1152         struct pnfs_layout_segment *lseg, *tmp;
1153
1154         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1155                 return;
1156         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1157                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1158                         continue;
1159                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1160         }
1161 }
1162
1163 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1164                 const nfs4_stateid *arg_stateid,
1165                 const struct pnfs_layout_range *range,
1166                 const nfs4_stateid *stateid)
1167 {
1168         struct inode *inode = lo->plh_inode;
1169         LIST_HEAD(freeme);
1170
1171         spin_lock(&inode->i_lock);
1172         if (!pnfs_layout_is_valid(lo) ||
1173             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1174                 goto out_unlock;
1175         if (stateid) {
1176                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1177
1178                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1179                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1180                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1181         } else
1182                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1183 out_unlock:
1184         pnfs_clear_layoutreturn_waitbit(lo);
1185         spin_unlock(&inode->i_lock);
1186         pnfs_free_lseg_list(&freeme);
1187
1188 }
1189
1190 static bool
1191 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1192                 nfs4_stateid *stateid,
1193                 const struct cred **cred,
1194                 enum pnfs_iomode *iomode)
1195 {
1196         /* Serialise LAYOUTGET/LAYOUTRETURN */
1197         if (atomic_read(&lo->plh_outstanding) != 0)
1198                 return false;
1199         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1200                 return false;
1201         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1202         pnfs_get_layout_hdr(lo);
1203         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1204         *cred = get_cred(lo->plh_lc_cred);
1205         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1206                 if (lo->plh_return_seq != 0)
1207                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1208                 if (iomode != NULL)
1209                         *iomode = lo->plh_return_iomode;
1210                 pnfs_clear_layoutreturn_info(lo);
1211         } else if (iomode != NULL)
1212                 *iomode = IOMODE_ANY;
1213         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1214         return true;
1215 }
1216
1217 static void
1218 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1219                 struct pnfs_layout_hdr *lo,
1220                 const nfs4_stateid *stateid,
1221                 enum pnfs_iomode iomode)
1222 {
1223         struct inode *inode = lo->plh_inode;
1224
1225         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1226         args->inode = inode;
1227         args->range.iomode = iomode;
1228         args->range.offset = 0;
1229         args->range.length = NFS4_MAX_UINT64;
1230         args->layout = lo;
1231         nfs4_stateid_copy(&args->stateid, stateid);
1232 }
1233
1234 static int
1235 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1236                        const nfs4_stateid *stateid,
1237                        const struct cred **pcred,
1238                        enum pnfs_iomode iomode,
1239                        bool sync)
1240 {
1241         struct inode *ino = lo->plh_inode;
1242         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1243         struct nfs4_layoutreturn *lrp;
1244         const struct cred *cred = *pcred;
1245         int status = 0;
1246
1247         *pcred = NULL;
1248         lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1249         if (unlikely(lrp == NULL)) {
1250                 status = -ENOMEM;
1251                 spin_lock(&ino->i_lock);
1252                 pnfs_clear_layoutreturn_waitbit(lo);
1253                 spin_unlock(&ino->i_lock);
1254                 put_cred(cred);
1255                 pnfs_put_layout_hdr(lo);
1256                 goto out;
1257         }
1258
1259         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1260         lrp->args.ld_private = &lrp->ld_private;
1261         lrp->clp = NFS_SERVER(ino)->nfs_client;
1262         lrp->cred = cred;
1263         if (ld->prepare_layoutreturn)
1264                 ld->prepare_layoutreturn(&lrp->args);
1265
1266         status = nfs4_proc_layoutreturn(lrp, sync);
1267 out:
1268         dprintk("<-- %s status: %d\n", __func__, status);
1269         return status;
1270 }
1271
1272 static bool
1273 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1274                                 enum pnfs_iomode iomode,
1275                                 u32 seq)
1276 {
1277         struct pnfs_layout_range recall_range = {
1278                 .length = NFS4_MAX_UINT64,
1279                 .iomode = iomode,
1280         };
1281         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1282                                                &recall_range, seq) != -EBUSY;
1283 }
1284
1285 /* Return true if layoutreturn is needed */
1286 static bool
1287 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1288 {
1289         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1290                 return false;
1291         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1292                                                lo->plh_return_seq);
1293 }
1294
1295 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1296 {
1297         struct inode *inode= lo->plh_inode;
1298
1299         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1300                 return;
1301         spin_lock(&inode->i_lock);
1302         if (pnfs_layout_need_return(lo)) {
1303                 const struct cred *cred;
1304                 nfs4_stateid stateid;
1305                 enum pnfs_iomode iomode;
1306                 bool send;
1307
1308                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1309                 spin_unlock(&inode->i_lock);
1310                 if (send) {
1311                         /* Send an async layoutreturn so we dont deadlock */
1312                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1313                 }
1314         } else
1315                 spin_unlock(&inode->i_lock);
1316 }
1317
1318 /*
1319  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1320  * when the layout segment list is empty.
1321  *
1322  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1323  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1324  * deviceid is marked invalid.
1325  */
1326 int
1327 _pnfs_return_layout(struct inode *ino)
1328 {
1329         struct pnfs_layout_hdr *lo = NULL;
1330         struct nfs_inode *nfsi = NFS_I(ino);
1331         struct pnfs_layout_range range = {
1332                 .iomode         = IOMODE_ANY,
1333                 .offset         = 0,
1334                 .length         = NFS4_MAX_UINT64,
1335         };
1336         LIST_HEAD(tmp_list);
1337         const struct cred *cred;
1338         nfs4_stateid stateid;
1339         int status = 0;
1340         bool send, valid_layout;
1341
1342         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1343
1344         spin_lock(&ino->i_lock);
1345         lo = nfsi->layout;
1346         if (!lo) {
1347                 spin_unlock(&ino->i_lock);
1348                 dprintk("NFS: %s no layout to return\n", __func__);
1349                 goto out;
1350         }
1351         /* Reference matched in nfs4_layoutreturn_release */
1352         pnfs_get_layout_hdr(lo);
1353         /* Is there an outstanding layoutreturn ? */
1354         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1355                 spin_unlock(&ino->i_lock);
1356                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1357                                         TASK_UNINTERRUPTIBLE))
1358                         goto out_put_layout_hdr;
1359                 spin_lock(&ino->i_lock);
1360         }
1361         valid_layout = pnfs_layout_is_valid(lo);
1362         pnfs_clear_layoutcommit(ino, &tmp_list);
1363         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1364
1365         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1366                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1367
1368         /* Don't send a LAYOUTRETURN if list was initially empty */
1369         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1370                         !valid_layout) {
1371                 spin_unlock(&ino->i_lock);
1372                 dprintk("NFS: %s no layout segments to return\n", __func__);
1373                 goto out_wait_layoutreturn;
1374         }
1375
1376         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1377         spin_unlock(&ino->i_lock);
1378         if (send)
1379                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1380 out_wait_layoutreturn:
1381         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1382 out_put_layout_hdr:
1383         pnfs_free_lseg_list(&tmp_list);
1384         pnfs_put_layout_hdr(lo);
1385 out:
1386         dprintk("<-- %s status: %d\n", __func__, status);
1387         return status;
1388 }
1389
1390 int
1391 pnfs_commit_and_return_layout(struct inode *inode)
1392 {
1393         struct pnfs_layout_hdr *lo;
1394         int ret;
1395
1396         spin_lock(&inode->i_lock);
1397         lo = NFS_I(inode)->layout;
1398         if (lo == NULL) {
1399                 spin_unlock(&inode->i_lock);
1400                 return 0;
1401         }
1402         pnfs_get_layout_hdr(lo);
1403         /* Block new layoutgets and read/write to ds */
1404         lo->plh_block_lgets++;
1405         spin_unlock(&inode->i_lock);
1406         filemap_fdatawait(inode->i_mapping);
1407         ret = pnfs_layoutcommit_inode(inode, true);
1408         if (ret == 0)
1409                 ret = _pnfs_return_layout(inode);
1410         spin_lock(&inode->i_lock);
1411         lo->plh_block_lgets--;
1412         spin_unlock(&inode->i_lock);
1413         pnfs_put_layout_hdr(lo);
1414         return ret;
1415 }
1416
1417 bool pnfs_roc(struct inode *ino,
1418                 struct nfs4_layoutreturn_args *args,
1419                 struct nfs4_layoutreturn_res *res,
1420                 const struct cred *cred)
1421 {
1422         struct nfs_inode *nfsi = NFS_I(ino);
1423         struct nfs_open_context *ctx;
1424         struct nfs4_state *state;
1425         struct pnfs_layout_hdr *lo;
1426         struct pnfs_layout_segment *lseg, *next;
1427         const struct cred *lc_cred;
1428         nfs4_stateid stateid;
1429         enum pnfs_iomode iomode = 0;
1430         bool layoutreturn = false, roc = false;
1431         bool skip_read = false;
1432
1433         if (!nfs_have_layout(ino))
1434                 return false;
1435 retry:
1436         rcu_read_lock();
1437         spin_lock(&ino->i_lock);
1438         lo = nfsi->layout;
1439         if (!lo || !pnfs_layout_is_valid(lo) ||
1440             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1441                 lo = NULL;
1442                 goto out_noroc;
1443         }
1444         pnfs_get_layout_hdr(lo);
1445         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1446                 spin_unlock(&ino->i_lock);
1447                 rcu_read_unlock();
1448                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1449                                 TASK_UNINTERRUPTIBLE);
1450                 pnfs_put_layout_hdr(lo);
1451                 goto retry;
1452         }
1453
1454         /* no roc if we hold a delegation */
1455         if (nfs4_check_delegation(ino, FMODE_READ)) {
1456                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1457                         goto out_noroc;
1458                 skip_read = true;
1459         }
1460
1461         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1462                 state = ctx->state;
1463                 if (state == NULL)
1464                         continue;
1465                 /* Don't return layout if there is open file state */
1466                 if (state->state & FMODE_WRITE)
1467                         goto out_noroc;
1468                 if (state->state & FMODE_READ)
1469                         skip_read = true;
1470         }
1471
1472
1473         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1474                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1475                         continue;
1476                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1477                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1478                         continue;
1479                 /*
1480                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1481                  * doesn't invalidate the layout for us.
1482                  */
1483                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1484                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1485                         continue;
1486                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1487         }
1488
1489         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1490                 goto out_noroc;
1491
1492         /* ROC in two conditions:
1493          * 1. there are ROC lsegs
1494          * 2. we don't send layoutreturn
1495          */
1496         /* lo ref dropped in pnfs_roc_release() */
1497         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1498         /* If the creds don't match, we can't compound the layoutreturn */
1499         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1500                 goto out_noroc;
1501
1502         roc = layoutreturn;
1503         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1504         res->lrs_present = 0;
1505         layoutreturn = false;
1506         put_cred(lc_cred);
1507
1508 out_noroc:
1509         spin_unlock(&ino->i_lock);
1510         rcu_read_unlock();
1511         pnfs_layoutcommit_inode(ino, true);
1512         if (roc) {
1513                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1514                 if (ld->prepare_layoutreturn)
1515                         ld->prepare_layoutreturn(args);
1516                 pnfs_put_layout_hdr(lo);
1517                 return true;
1518         }
1519         if (layoutreturn)
1520                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1521         pnfs_put_layout_hdr(lo);
1522         return false;
1523 }
1524
1525 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1526                   struct nfs4_layoutreturn_res **respp, int *ret)
1527 {
1528         struct nfs4_layoutreturn_args *arg = *argpp;
1529         int retval = -EAGAIN;
1530
1531         if (!arg)
1532                 return 0;
1533         /* Handle Layoutreturn errors */
1534         switch (*ret) {
1535         case 0:
1536                 retval = 0;
1537                 break;
1538         case -NFS4ERR_NOMATCHING_LAYOUT:
1539                 /* Was there an RPC level error? If not, retry */
1540                 if (task->tk_rpc_status == 0)
1541                         break;
1542                 /* If the call was not sent, let caller handle it */
1543                 if (!RPC_WAS_SENT(task))
1544                         return 0;
1545                 /*
1546                  * Otherwise, assume the call succeeded and
1547                  * that we need to release the layout
1548                  */
1549                 *ret = 0;
1550                 (*respp)->lrs_present = 0;
1551                 retval = 0;
1552                 break;
1553         case -NFS4ERR_DELAY:
1554                 /* Let the caller handle the retry */
1555                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1556                 return 0;
1557         case -NFS4ERR_OLD_STATEID:
1558                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1559                                                      &arg->range, arg->inode))
1560                         break;
1561                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1562                 return -EAGAIN;
1563         }
1564         *argpp = NULL;
1565         *respp = NULL;
1566         return retval;
1567 }
1568
1569 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1570                 struct nfs4_layoutreturn_res *res,
1571                 int ret)
1572 {
1573         struct pnfs_layout_hdr *lo = args->layout;
1574         struct inode *inode = args->inode;
1575         const nfs4_stateid *res_stateid = NULL;
1576         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1577
1578         switch (ret) {
1579         case -NFS4ERR_NOMATCHING_LAYOUT:
1580                 spin_lock(&inode->i_lock);
1581                 if (pnfs_layout_is_valid(lo) &&
1582                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1583                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1584                 pnfs_clear_layoutreturn_waitbit(lo);
1585                 spin_unlock(&inode->i_lock);
1586                 break;
1587         case 0:
1588                 if (res->lrs_present)
1589                         res_stateid = &res->stateid;
1590                 fallthrough;
1591         default:
1592                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1593                                              res_stateid);
1594         }
1595         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1596         if (ld_private && ld_private->ops && ld_private->ops->free)
1597                 ld_private->ops->free(ld_private);
1598         pnfs_put_layout_hdr(lo);
1599 }
1600
1601 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1602 {
1603         struct nfs_inode *nfsi = NFS_I(ino);
1604         struct pnfs_layout_hdr *lo;
1605         bool sleep = false;
1606
1607         /* we might not have grabbed lo reference. so need to check under
1608          * i_lock */
1609         spin_lock(&ino->i_lock);
1610         lo = nfsi->layout;
1611         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1612                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1613                 sleep = true;
1614         }
1615         spin_unlock(&ino->i_lock);
1616         return sleep;
1617 }
1618
1619 /*
1620  * Compare two layout segments for sorting into layout cache.
1621  * We want to preferentially return RW over RO layouts, so ensure those
1622  * are seen first.
1623  */
1624 static s64
1625 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1626            const struct pnfs_layout_range *l2)
1627 {
1628         s64 d;
1629
1630         /* high offset > low offset */
1631         d = l1->offset - l2->offset;
1632         if (d)
1633                 return d;
1634
1635         /* short length > long length */
1636         d = l2->length - l1->length;
1637         if (d)
1638                 return d;
1639
1640         /* read > read/write */
1641         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1642 }
1643
1644 static bool
1645 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1646                 const struct pnfs_layout_range *l2)
1647 {
1648         return pnfs_lseg_range_cmp(l1, l2) > 0;
1649 }
1650
1651 static bool
1652 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1653                 struct pnfs_layout_segment *old)
1654 {
1655         return false;
1656 }
1657
1658 void
1659 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1660                    struct pnfs_layout_segment *lseg,
1661                    bool (*is_after)(const struct pnfs_layout_range *,
1662                            const struct pnfs_layout_range *),
1663                    bool (*do_merge)(struct pnfs_layout_segment *,
1664                            struct pnfs_layout_segment *),
1665                    struct list_head *free_me)
1666 {
1667         struct pnfs_layout_segment *lp, *tmp;
1668
1669         dprintk("%s:Begin\n", __func__);
1670
1671         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1672                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1673                         continue;
1674                 if (do_merge(lseg, lp)) {
1675                         mark_lseg_invalid(lp, free_me);
1676                         continue;
1677                 }
1678                 if (is_after(&lseg->pls_range, &lp->pls_range))
1679                         continue;
1680                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1681                 dprintk("%s: inserted lseg %p "
1682                         "iomode %d offset %llu length %llu before "
1683                         "lp %p iomode %d offset %llu length %llu\n",
1684                         __func__, lseg, lseg->pls_range.iomode,
1685                         lseg->pls_range.offset, lseg->pls_range.length,
1686                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1687                         lp->pls_range.length);
1688                 goto out;
1689         }
1690         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1691         dprintk("%s: inserted lseg %p "
1692                 "iomode %d offset %llu length %llu at tail\n",
1693                 __func__, lseg, lseg->pls_range.iomode,
1694                 lseg->pls_range.offset, lseg->pls_range.length);
1695 out:
1696         pnfs_get_layout_hdr(lo);
1697
1698         dprintk("%s:Return\n", __func__);
1699 }
1700 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1701
1702 static void
1703 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1704                    struct pnfs_layout_segment *lseg,
1705                    struct list_head *free_me)
1706 {
1707         struct inode *inode = lo->plh_inode;
1708         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1709
1710         if (ld->add_lseg != NULL)
1711                 ld->add_lseg(lo, lseg, free_me);
1712         else
1713                 pnfs_generic_layout_insert_lseg(lo, lseg,
1714                                 pnfs_lseg_range_is_after,
1715                                 pnfs_lseg_no_merge,
1716                                 free_me);
1717 }
1718
1719 static struct pnfs_layout_hdr *
1720 alloc_init_layout_hdr(struct inode *ino,
1721                       struct nfs_open_context *ctx,
1722                       gfp_t gfp_flags)
1723 {
1724         struct pnfs_layout_hdr *lo;
1725
1726         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1727         if (!lo)
1728                 return NULL;
1729         refcount_set(&lo->plh_refcount, 1);
1730         INIT_LIST_HEAD(&lo->plh_layouts);
1731         INIT_LIST_HEAD(&lo->plh_segs);
1732         INIT_LIST_HEAD(&lo->plh_return_segs);
1733         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1734         lo->plh_inode = ino;
1735         lo->plh_lc_cred = get_cred(ctx->cred);
1736         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1737         return lo;
1738 }
1739
1740 static struct pnfs_layout_hdr *
1741 pnfs_find_alloc_layout(struct inode *ino,
1742                        struct nfs_open_context *ctx,
1743                        gfp_t gfp_flags)
1744         __releases(&ino->i_lock)
1745         __acquires(&ino->i_lock)
1746 {
1747         struct nfs_inode *nfsi = NFS_I(ino);
1748         struct pnfs_layout_hdr *new = NULL;
1749
1750         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1751
1752         if (nfsi->layout != NULL)
1753                 goto out_existing;
1754         spin_unlock(&ino->i_lock);
1755         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1756         spin_lock(&ino->i_lock);
1757
1758         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1759                 nfsi->layout = new;
1760                 return new;
1761         } else if (new != NULL)
1762                 pnfs_free_layout_hdr(new);
1763 out_existing:
1764         pnfs_get_layout_hdr(nfsi->layout);
1765         return nfsi->layout;
1766 }
1767
1768 /*
1769  * iomode matching rules:
1770  * iomode       lseg    strict match
1771  *                      iomode
1772  * -----        -----   ------ -----
1773  * ANY          READ    N/A    true
1774  * ANY          RW      N/A    true
1775  * RW           READ    N/A    false
1776  * RW           RW      N/A    true
1777  * READ         READ    N/A    true
1778  * READ         RW      true   false
1779  * READ         RW      false  true
1780  */
1781 static bool
1782 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1783                  const struct pnfs_layout_range *range,
1784                  bool strict_iomode)
1785 {
1786         struct pnfs_layout_range range1;
1787
1788         if ((range->iomode == IOMODE_RW &&
1789              ls_range->iomode != IOMODE_RW) ||
1790             (range->iomode != ls_range->iomode &&
1791              strict_iomode) ||
1792             !pnfs_lseg_range_intersecting(ls_range, range))
1793                 return false;
1794
1795         /* range1 covers only the first byte in the range */
1796         range1 = *range;
1797         range1.length = 1;
1798         return pnfs_lseg_range_contained(ls_range, &range1);
1799 }
1800
1801 /*
1802  * lookup range in layout
1803  */
1804 static struct pnfs_layout_segment *
1805 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1806                 struct pnfs_layout_range *range,
1807                 bool strict_iomode)
1808 {
1809         struct pnfs_layout_segment *lseg, *ret = NULL;
1810
1811         dprintk("%s:Begin\n", __func__);
1812
1813         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1814                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1815                     pnfs_lseg_range_match(&lseg->pls_range, range,
1816                                           strict_iomode)) {
1817                         ret = pnfs_get_lseg(lseg);
1818                         break;
1819                 }
1820         }
1821
1822         dprintk("%s:Return lseg %p ref %d\n",
1823                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1824         return ret;
1825 }
1826
1827 /*
1828  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1829  * to the MDS or over pNFS
1830  *
1831  * The nfs_inode read_io and write_io fields are cumulative counters reset
1832  * when there are no layout segments. Note that in pnfs_update_layout iomode
1833  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1834  * WRITE request.
1835  *
1836  * A return of true means use MDS I/O.
1837  *
1838  * From rfc 5661:
1839  * If a file's size is smaller than the file size threshold, data accesses
1840  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1841  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1842  * server.  If both file size and I/O size are provided, the client SHOULD
1843  * reach or exceed  both thresholds before sending its read or write
1844  * requests to the data server.
1845  */
1846 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1847                                      struct inode *ino, int iomode)
1848 {
1849         struct nfs4_threshold *t = ctx->mdsthreshold;
1850         struct nfs_inode *nfsi = NFS_I(ino);
1851         loff_t fsize = i_size_read(ino);
1852         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1853
1854         if (t == NULL)
1855                 return ret;
1856
1857         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1858                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1859
1860         switch (iomode) {
1861         case IOMODE_READ:
1862                 if (t->bm & THRESHOLD_RD) {
1863                         dprintk("%s fsize %llu\n", __func__, fsize);
1864                         size_set = true;
1865                         if (fsize < t->rd_sz)
1866                                 size = true;
1867                 }
1868                 if (t->bm & THRESHOLD_RD_IO) {
1869                         dprintk("%s nfsi->read_io %llu\n", __func__,
1870                                 nfsi->read_io);
1871                         io_set = true;
1872                         if (nfsi->read_io < t->rd_io_sz)
1873                                 io = true;
1874                 }
1875                 break;
1876         case IOMODE_RW:
1877                 if (t->bm & THRESHOLD_WR) {
1878                         dprintk("%s fsize %llu\n", __func__, fsize);
1879                         size_set = true;
1880                         if (fsize < t->wr_sz)
1881                                 size = true;
1882                 }
1883                 if (t->bm & THRESHOLD_WR_IO) {
1884                         dprintk("%s nfsi->write_io %llu\n", __func__,
1885                                 nfsi->write_io);
1886                         io_set = true;
1887                         if (nfsi->write_io < t->wr_io_sz)
1888                                 io = true;
1889                 }
1890                 break;
1891         }
1892         if (size_set && io_set) {
1893                 if (size && io)
1894                         ret = true;
1895         } else if (size || io)
1896                 ret = true;
1897
1898         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1899         return ret;
1900 }
1901
1902 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1903 {
1904         /*
1905          * send layoutcommit as it can hold up layoutreturn due to lseg
1906          * reference
1907          */
1908         pnfs_layoutcommit_inode(lo->plh_inode, false);
1909         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1910                                    nfs_wait_bit_killable,
1911                                    TASK_KILLABLE);
1912 }
1913
1914 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1915 {
1916         atomic_inc(&lo->plh_outstanding);
1917 }
1918
1919 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1920 {
1921         if (atomic_dec_and_test(&lo->plh_outstanding) &&
1922             test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
1923                 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
1924 }
1925
1926 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1927 {
1928         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1929 }
1930
1931 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1932 {
1933         unsigned long *bitlock = &lo->plh_flags;
1934
1935         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1936         smp_mb__after_atomic();
1937         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1938 }
1939
1940 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1941                                 struct nfs_server *server)
1942 {
1943         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1944                 struct nfs_client *clp = server->nfs_client;
1945
1946                 /* The lo must be on the clp list if there is any
1947                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1948                  */
1949                 spin_lock(&clp->cl_lock);
1950                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1951                 spin_unlock(&clp->cl_lock);
1952         }
1953 }
1954
1955 /*
1956  * Layout segment is retreived from the server if not cached.
1957  * The appropriate layout segment is referenced and returned to the caller.
1958  */
1959 struct pnfs_layout_segment *
1960 pnfs_update_layout(struct inode *ino,
1961                    struct nfs_open_context *ctx,
1962                    loff_t pos,
1963                    u64 count,
1964                    enum pnfs_iomode iomode,
1965                    bool strict_iomode,
1966                    gfp_t gfp_flags)
1967 {
1968         struct pnfs_layout_range arg = {
1969                 .iomode = iomode,
1970                 .offset = pos,
1971                 .length = count,
1972         };
1973         unsigned pg_offset;
1974         struct nfs_server *server = NFS_SERVER(ino);
1975         struct nfs_client *clp = server->nfs_client;
1976         struct pnfs_layout_hdr *lo = NULL;
1977         struct pnfs_layout_segment *lseg = NULL;
1978         struct nfs4_layoutget *lgp;
1979         nfs4_stateid stateid;
1980         long timeout = 0;
1981         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1982         bool first;
1983
1984         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1985                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1986                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1987                 goto out;
1988         }
1989
1990         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1991                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1992                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1993                 goto out;
1994         }
1995
1996 lookup_again:
1997         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1998         if (IS_ERR(lseg))
1999                 goto out;
2000         first = false;
2001         spin_lock(&ino->i_lock);
2002         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2003         if (lo == NULL) {
2004                 spin_unlock(&ino->i_lock);
2005                 lseg = ERR_PTR(-ENOMEM);
2006                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2007                                  PNFS_UPDATE_LAYOUT_NOMEM);
2008                 goto out;
2009         }
2010
2011         /* Do we even need to bother with this? */
2012         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2013                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2014                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2015                 dprintk("%s matches recall, use MDS\n", __func__);
2016                 goto out_unlock;
2017         }
2018
2019         /* if LAYOUTGET already failed once we don't try again */
2020         if (pnfs_layout_io_test_failed(lo, iomode)) {
2021                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2022                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2023                 goto out_unlock;
2024         }
2025
2026         /*
2027          * If the layout segment list is empty, but there are outstanding
2028          * layoutget calls, then they might be subject to a layoutrecall.
2029          */
2030         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2031             atomic_read(&lo->plh_outstanding) != 0) {
2032                 spin_unlock(&ino->i_lock);
2033                 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2034                                            TASK_KILLABLE));
2035                 if (IS_ERR(lseg))
2036                         goto out_put_layout_hdr;
2037                 pnfs_put_layout_hdr(lo);
2038                 goto lookup_again;
2039         }
2040
2041         /*
2042          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2043          * for LAYOUTRETURN.
2044          */
2045         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2046                 spin_unlock(&ino->i_lock);
2047                 dprintk("%s wait for layoutreturn\n", __func__);
2048                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2049                 if (!IS_ERR(lseg)) {
2050                         pnfs_put_layout_hdr(lo);
2051                         dprintk("%s retrying\n", __func__);
2052                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2053                                                  lseg,
2054                                                  PNFS_UPDATE_LAYOUT_RETRY);
2055                         goto lookup_again;
2056                 }
2057                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2058                                          PNFS_UPDATE_LAYOUT_RETURN);
2059                 goto out_put_layout_hdr;
2060         }
2061
2062         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2063         if (lseg) {
2064                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2065                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2066                 goto out_unlock;
2067         }
2068
2069         /*
2070          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2071          * stateid, or it has been invalidated, then we must use the open
2072          * stateid.
2073          */
2074         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2075                 int status;
2076
2077                 /*
2078                  * The first layoutget for the file. Need to serialize per
2079                  * RFC 5661 Errata 3208.
2080                  */
2081                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2082                                      &lo->plh_flags)) {
2083                         spin_unlock(&ino->i_lock);
2084                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2085                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2086                                                 TASK_KILLABLE));
2087                         if (IS_ERR(lseg))
2088                                 goto out_put_layout_hdr;
2089                         pnfs_put_layout_hdr(lo);
2090                         dprintk("%s retrying\n", __func__);
2091                         goto lookup_again;
2092                 }
2093
2094                 spin_unlock(&ino->i_lock);
2095                 first = true;
2096                 status = nfs4_select_rw_stateid(ctx->state,
2097                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2098                                         NULL, &stateid, NULL);
2099                 if (status != 0) {
2100                         lseg = ERR_PTR(status);
2101                         trace_pnfs_update_layout(ino, pos, count,
2102                                         iomode, lo, lseg,
2103                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2104                         nfs4_schedule_stateid_recovery(server, ctx->state);
2105                         pnfs_clear_first_layoutget(lo);
2106                         pnfs_put_layout_hdr(lo);
2107                         goto lookup_again;
2108                 }
2109                 spin_lock(&ino->i_lock);
2110         } else {
2111                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2112         }
2113
2114         if (pnfs_layoutgets_blocked(lo)) {
2115                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2116                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2117                 goto out_unlock;
2118         }
2119         nfs_layoutget_begin(lo);
2120         spin_unlock(&ino->i_lock);
2121
2122         _add_to_server_list(lo, server);
2123
2124         pg_offset = arg.offset & ~PAGE_MASK;
2125         if (pg_offset) {
2126                 arg.offset -= pg_offset;
2127                 arg.length += pg_offset;
2128         }
2129         if (arg.length != NFS4_MAX_UINT64)
2130                 arg.length = PAGE_ALIGN(arg.length);
2131
2132         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2133         if (!lgp) {
2134                 lseg = ERR_PTR(-ENOMEM);
2135                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2136                                          PNFS_UPDATE_LAYOUT_NOMEM);
2137                 nfs_layoutget_end(lo);
2138                 goto out_put_layout_hdr;
2139         }
2140
2141         lgp->lo = lo;
2142         pnfs_get_layout_hdr(lo);
2143
2144         lseg = nfs4_proc_layoutget(lgp, &timeout);
2145         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2146                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2147         nfs_layoutget_end(lo);
2148         if (IS_ERR(lseg)) {
2149                 switch(PTR_ERR(lseg)) {
2150                 case -EBUSY:
2151                         if (time_after(jiffies, giveup))
2152                                 lseg = NULL;
2153                         break;
2154                 case -ERECALLCONFLICT:
2155                 case -EAGAIN:
2156                         break;
2157                 case -ENODATA:
2158                         /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2159                         pnfs_layout_set_fail_bit(
2160                                 lo, pnfs_iomode_to_fail_bit(iomode));
2161                         lseg = NULL;
2162                         goto out_put_layout_hdr;
2163                 default:
2164                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2165                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2166                                 lseg = NULL;
2167                         }
2168                         goto out_put_layout_hdr;
2169                 }
2170                 if (lseg) {
2171                         if (first)
2172                                 pnfs_clear_first_layoutget(lo);
2173                         trace_pnfs_update_layout(ino, pos, count,
2174                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2175                         pnfs_put_layout_hdr(lo);
2176                         goto lookup_again;
2177                 }
2178         } else {
2179                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2180         }
2181
2182 out_put_layout_hdr:
2183         if (first)
2184                 pnfs_clear_first_layoutget(lo);
2185         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2186                                  PNFS_UPDATE_LAYOUT_EXIT);
2187         pnfs_put_layout_hdr(lo);
2188 out:
2189         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2190                         "(%s, offset: %llu, length: %llu)\n",
2191                         __func__, ino->i_sb->s_id,
2192                         (unsigned long long)NFS_FILEID(ino),
2193                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2194                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2195                         (unsigned long long)pos,
2196                         (unsigned long long)count);
2197         return lseg;
2198 out_unlock:
2199         spin_unlock(&ino->i_lock);
2200         goto out_put_layout_hdr;
2201 }
2202 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2203
2204 static bool
2205 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2206 {
2207         switch (range->iomode) {
2208         case IOMODE_READ:
2209         case IOMODE_RW:
2210                 break;
2211         default:
2212                 return false;
2213         }
2214         if (range->offset == NFS4_MAX_UINT64)
2215                 return false;
2216         if (range->length == 0)
2217                 return false;
2218         if (range->length != NFS4_MAX_UINT64 &&
2219             range->length > NFS4_MAX_UINT64 - range->offset)
2220                 return false;
2221         return true;
2222 }
2223
2224 static struct pnfs_layout_hdr *
2225 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2226 {
2227         struct pnfs_layout_hdr *lo;
2228
2229         spin_lock(&ino->i_lock);
2230         lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2231         if (!lo)
2232                 goto out_unlock;
2233         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2234                 goto out_unlock;
2235         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2236                 goto out_unlock;
2237         if (pnfs_layoutgets_blocked(lo))
2238                 goto out_unlock;
2239         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2240                 goto out_unlock;
2241         nfs_layoutget_begin(lo);
2242         spin_unlock(&ino->i_lock);
2243         _add_to_server_list(lo, NFS_SERVER(ino));
2244         return lo;
2245
2246 out_unlock:
2247         spin_unlock(&ino->i_lock);
2248         pnfs_put_layout_hdr(lo);
2249         return NULL;
2250 }
2251
2252 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2253                                      struct nfs_open_context *ctx)
2254 {
2255         struct inode *ino = data->dentry->d_inode;
2256         struct pnfs_layout_range rng = {
2257                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2258                           IOMODE_RW: IOMODE_READ,
2259                 .offset = 0,
2260                 .length = NFS4_MAX_UINT64,
2261         };
2262         struct nfs4_layoutget *lgp;
2263         struct pnfs_layout_hdr *lo;
2264
2265         /* Heuristic: don't send layoutget if we have cached data */
2266         if (rng.iomode == IOMODE_READ &&
2267            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2268                 return;
2269
2270         lo = _pnfs_grab_empty_layout(ino, ctx);
2271         if (!lo)
2272                 return;
2273         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2274                                              nfs_io_gfp_mask());
2275         if (!lgp) {
2276                 pnfs_clear_first_layoutget(lo);
2277                 nfs_layoutget_end(lo);
2278                 pnfs_put_layout_hdr(lo);
2279                 return;
2280         }
2281         lgp->lo = lo;
2282         data->lgp = lgp;
2283         data->o_arg.lg_args = &lgp->args;
2284         data->o_res.lg_res = &lgp->res;
2285 }
2286
2287 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2288                                      struct nfs_open_context *ctx)
2289 {
2290         struct inode *ino = data->dentry->d_inode;
2291         struct pnfs_layout_range rng = {
2292                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2293                           IOMODE_RW: IOMODE_READ,
2294                 .offset = 0,
2295                 .length = NFS4_MAX_UINT64,
2296         };
2297         struct nfs4_layoutget *lgp;
2298
2299         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2300                                              nfs_io_gfp_mask());
2301         if (!lgp)
2302                 return;
2303         data->lgp = lgp;
2304         data->o_arg.lg_args = &lgp->args;
2305         data->o_res.lg_res = &lgp->res;
2306 }
2307
2308 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2309                          struct nfs_open_context *ctx)
2310 {
2311         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2312
2313         if (!(pnfs_enabled_sb(server) &&
2314               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2315                 return;
2316         /* Could check on max_ops, but currently hardcoded high enough */
2317         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2318                 return;
2319         if (data->lgp)
2320                 return;
2321         if (data->state)
2322                 _lgopen_prepare_attached(data, ctx);
2323         else
2324                 _lgopen_prepare_floating(data, ctx);
2325 }
2326
2327 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2328                        struct nfs_open_context *ctx)
2329 {
2330         struct pnfs_layout_hdr *lo;
2331         struct pnfs_layout_segment *lseg;
2332         struct nfs_server *srv = NFS_SERVER(ino);
2333         u32 iomode;
2334
2335         if (!lgp)
2336                 return;
2337         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2338         if (lgp->res.status) {
2339                 switch (lgp->res.status) {
2340                 default:
2341                         break;
2342                 /*
2343                  * Halt lgopen attempts if the server doesn't recognise
2344                  * the "current stateid" value, the layout type, or the
2345                  * layoutget operation as being valid.
2346                  * Also if it complains about too many ops in the compound
2347                  * or of the request/reply being too big.
2348                  */
2349                 case -NFS4ERR_BAD_STATEID:
2350                 case -NFS4ERR_NOTSUPP:
2351                 case -NFS4ERR_REP_TOO_BIG:
2352                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2353                 case -NFS4ERR_REQ_TOO_BIG:
2354                 case -NFS4ERR_TOO_MANY_OPS:
2355                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2356                         srv->caps &= ~NFS_CAP_LGOPEN;
2357                 }
2358                 return;
2359         }
2360         if (!lgp->lo) {
2361                 lo = _pnfs_grab_empty_layout(ino, ctx);
2362                 if (!lo)
2363                         return;
2364                 lgp->lo = lo;
2365         } else
2366                 lo = lgp->lo;
2367
2368         lseg = pnfs_layout_process(lgp);
2369         if (!IS_ERR(lseg)) {
2370                 iomode = lgp->args.range.iomode;
2371                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2372                 pnfs_put_lseg(lseg);
2373         }
2374 }
2375
2376 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2377 {
2378         if (lgp != NULL) {
2379                 if (lgp->lo) {
2380                         pnfs_clear_first_layoutget(lgp->lo);
2381                         nfs_layoutget_end(lgp->lo);
2382                 }
2383                 pnfs_layoutget_free(lgp);
2384         }
2385 }
2386
2387 struct pnfs_layout_segment *
2388 pnfs_layout_process(struct nfs4_layoutget *lgp)
2389 {
2390         struct pnfs_layout_hdr *lo = lgp->lo;
2391         struct nfs4_layoutget_res *res = &lgp->res;
2392         struct pnfs_layout_segment *lseg;
2393         struct inode *ino = lo->plh_inode;
2394         LIST_HEAD(free_me);
2395
2396         if (!pnfs_sanity_check_layout_range(&res->range))
2397                 return ERR_PTR(-EINVAL);
2398
2399         /* Inject layout blob into I/O device driver */
2400         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2401         if (IS_ERR_OR_NULL(lseg)) {
2402                 if (!lseg)
2403                         lseg = ERR_PTR(-ENOMEM);
2404
2405                 dprintk("%s: Could not allocate layout: error %ld\n",
2406                        __func__, PTR_ERR(lseg));
2407                 return lseg;
2408         }
2409
2410         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2411
2412         spin_lock(&ino->i_lock);
2413         if (pnfs_layoutgets_blocked(lo)) {
2414                 dprintk("%s forget reply due to state\n", __func__);
2415                 goto out_forget;
2416         }
2417
2418         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2419             !pnfs_is_first_layoutget(lo))
2420                 goto out_forget;
2421
2422         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2423                 /* existing state ID, make sure the sequence number matches. */
2424                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2425                         if (!pnfs_layout_is_valid(lo))
2426                                 lo->plh_barrier = 0;
2427                         dprintk("%s forget reply due to sequence\n", __func__);
2428                         goto out_forget;
2429                 }
2430                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2431         } else if (pnfs_layout_is_valid(lo)) {
2432                 /*
2433                  * We got an entirely new state ID.  Mark all segments for the
2434                  * inode invalid, and retry the layoutget
2435                  */
2436                 struct pnfs_layout_range range = {
2437                         .iomode = IOMODE_ANY,
2438                         .length = NFS4_MAX_UINT64,
2439                 };
2440                 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2441                 goto out_forget;
2442         } else {
2443                 /* We have a completely new layout */
2444                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2445         }
2446
2447         pnfs_get_lseg(lseg);
2448         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2449
2450
2451         if (res->return_on_close)
2452                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2453
2454         spin_unlock(&ino->i_lock);
2455         pnfs_free_lseg_list(&free_me);
2456         return lseg;
2457
2458 out_forget:
2459         spin_unlock(&ino->i_lock);
2460         lseg->pls_layout = lo;
2461         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2462         return ERR_PTR(-EAGAIN);
2463 }
2464
2465 /**
2466  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2467  * @lo: pointer to layout header
2468  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2469  * @return_range: describe layout segment ranges to be returned
2470  * @seq: stateid seqid to match
2471  *
2472  * This function is mainly intended for use by layoutrecall. It attempts
2473  * to free the layout segment immediately, or else to mark it for return
2474  * as soon as its reference count drops to zero.
2475  *
2476  * Returns
2477  * - 0: a layoutreturn needs to be scheduled.
2478  * - EBUSY: there are layout segment that are still in use.
2479  * - ENOENT: there are no layout segments that need to be returned.
2480  */
2481 int
2482 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2483                                 struct list_head *tmp_list,
2484                                 const struct pnfs_layout_range *return_range,
2485                                 u32 seq)
2486 {
2487         struct pnfs_layout_segment *lseg, *next;
2488         int remaining = 0;
2489
2490         dprintk("%s:Begin lo %p\n", __func__, lo);
2491
2492         assert_spin_locked(&lo->plh_inode->i_lock);
2493
2494         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2495                 tmp_list = &lo->plh_return_segs;
2496
2497         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2498                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2499                         dprintk("%s: marking lseg %p iomode %d "
2500                                 "offset %llu length %llu\n", __func__,
2501                                 lseg, lseg->pls_range.iomode,
2502                                 lseg->pls_range.offset,
2503                                 lseg->pls_range.length);
2504                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2505                                 tmp_list = &lo->plh_return_segs;
2506                         if (mark_lseg_invalid(lseg, tmp_list))
2507                                 continue;
2508                         remaining++;
2509                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2510                 }
2511
2512         if (remaining) {
2513                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2514                 return -EBUSY;
2515         }
2516
2517         if (!list_empty(&lo->plh_return_segs)) {
2518                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2519                 return 0;
2520         }
2521
2522         return -ENOENT;
2523 }
2524
2525 static void
2526 pnfs_mark_layout_for_return(struct inode *inode,
2527                             const struct pnfs_layout_range *range)
2528 {
2529         struct pnfs_layout_hdr *lo;
2530         bool return_now = false;
2531
2532         spin_lock(&inode->i_lock);
2533         lo = NFS_I(inode)->layout;
2534         if (!pnfs_layout_is_valid(lo)) {
2535                 spin_unlock(&inode->i_lock);
2536                 return;
2537         }
2538         pnfs_set_plh_return_info(lo, range->iomode, 0);
2539         /*
2540          * mark all matching lsegs so that we are sure to have no live
2541          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2542          * for how it works.
2543          */
2544         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2545                 const struct cred *cred;
2546                 nfs4_stateid stateid;
2547                 enum pnfs_iomode iomode;
2548
2549                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2550                 spin_unlock(&inode->i_lock);
2551                 if (return_now)
2552                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2553         } else {
2554                 spin_unlock(&inode->i_lock);
2555                 nfs_commit_inode(inode, 0);
2556         }
2557 }
2558
2559 void pnfs_error_mark_layout_for_return(struct inode *inode,
2560                                        struct pnfs_layout_segment *lseg)
2561 {
2562         struct pnfs_layout_range range = {
2563                 .iomode = lseg->pls_range.iomode,
2564                 .offset = 0,
2565                 .length = NFS4_MAX_UINT64,
2566         };
2567
2568         pnfs_mark_layout_for_return(inode, &range);
2569 }
2570 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2571
2572 static bool
2573 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2574 {
2575         return pnfs_layout_is_valid(lo) &&
2576                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2577                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2578 }
2579
2580 static struct pnfs_layout_segment *
2581 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2582                      const struct pnfs_layout_range *range,
2583                      enum pnfs_iomode iomode)
2584 {
2585         struct pnfs_layout_segment *lseg;
2586
2587         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2588                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2589                         continue;
2590                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2591                         continue;
2592                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2593                         continue;
2594                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2595                         return lseg;
2596         }
2597         return NULL;
2598 }
2599
2600 /* Find open file states whose mode matches that of the range */
2601 static bool
2602 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2603                                  const struct pnfs_layout_range *range)
2604 {
2605         struct list_head *head;
2606         struct nfs_open_context *ctx;
2607         fmode_t mode = 0;
2608
2609         if (!pnfs_layout_can_be_returned(lo) ||
2610             !pnfs_find_first_lseg(lo, range, range->iomode))
2611                 return false;
2612
2613         head = &NFS_I(lo->plh_inode)->open_files;
2614         list_for_each_entry_rcu(ctx, head, list) {
2615                 if (ctx->state)
2616                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2617         }
2618
2619         switch (range->iomode) {
2620         default:
2621                 break;
2622         case IOMODE_READ:
2623                 mode &= ~FMODE_WRITE;
2624                 break;
2625         case IOMODE_RW:
2626                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2627                         mode &= ~FMODE_READ;
2628         }
2629         return mode == 0;
2630 }
2631
2632 static int
2633 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2634 {
2635         const struct pnfs_layout_range *range = data;
2636         struct pnfs_layout_hdr *lo;
2637         struct inode *inode;
2638 restart:
2639         rcu_read_lock();
2640         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2641                 if (!pnfs_layout_can_be_returned(lo) ||
2642                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2643                         continue;
2644                 inode = lo->plh_inode;
2645                 spin_lock(&inode->i_lock);
2646                 if (!pnfs_should_return_unused_layout(lo, range)) {
2647                         spin_unlock(&inode->i_lock);
2648                         continue;
2649                 }
2650                 spin_unlock(&inode->i_lock);
2651                 inode = pnfs_grab_inode_layout_hdr(lo);
2652                 if (!inode)
2653                         continue;
2654                 rcu_read_unlock();
2655                 pnfs_mark_layout_for_return(inode, range);
2656                 iput(inode);
2657                 cond_resched();
2658                 goto restart;
2659         }
2660         rcu_read_unlock();
2661         return 0;
2662 }
2663
2664 void
2665 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2666                                  enum pnfs_iomode iomode)
2667 {
2668         struct pnfs_layout_range range = {
2669                 .iomode = iomode,
2670                 .offset = 0,
2671                 .length = NFS4_MAX_UINT64,
2672         };
2673
2674         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2675                         &range);
2676 }
2677
2678 void
2679 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2680 {
2681         if (pgio->pg_lseg == NULL ||
2682             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2683                 return;
2684         pnfs_put_lseg(pgio->pg_lseg);
2685         pgio->pg_lseg = NULL;
2686 }
2687 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2688
2689 /*
2690  * Check for any intersection between the request and the pgio->pg_lseg,
2691  * and if none, put this pgio->pg_lseg away.
2692  */
2693 void
2694 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2695 {
2696         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2697                 pnfs_put_lseg(pgio->pg_lseg);
2698                 pgio->pg_lseg = NULL;
2699         }
2700 }
2701 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2702
2703 void
2704 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2705 {
2706         u64 rd_size;
2707
2708         pnfs_generic_pg_check_layout(pgio);
2709         pnfs_generic_pg_check_range(pgio, req);
2710         if (pgio->pg_lseg == NULL) {
2711                 if (pgio->pg_dreq == NULL)
2712                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2713                 else
2714                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2715
2716                 pgio->pg_lseg =
2717                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2718                                            req_offset(req), rd_size,
2719                                            IOMODE_READ, false,
2720                                            nfs_io_gfp_mask());
2721                 if (IS_ERR(pgio->pg_lseg)) {
2722                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2723                         pgio->pg_lseg = NULL;
2724                         return;
2725                 }
2726         }
2727         /* If no lseg, fall back to read through mds */
2728         if (pgio->pg_lseg == NULL)
2729                 nfs_pageio_reset_read_mds(pgio);
2730
2731 }
2732 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2733
2734 void
2735 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2736                            struct nfs_page *req, u64 wb_size)
2737 {
2738         pnfs_generic_pg_check_layout(pgio);
2739         pnfs_generic_pg_check_range(pgio, req);
2740         if (pgio->pg_lseg == NULL) {
2741                 pgio->pg_lseg =
2742                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2743                                            req_offset(req), wb_size, IOMODE_RW,
2744                                            false, nfs_io_gfp_mask());
2745                 if (IS_ERR(pgio->pg_lseg)) {
2746                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2747                         pgio->pg_lseg = NULL;
2748                         return;
2749                 }
2750         }
2751         /* If no lseg, fall back to write through mds */
2752         if (pgio->pg_lseg == NULL)
2753                 nfs_pageio_reset_write_mds(pgio);
2754 }
2755 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2756
2757 void
2758 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2759 {
2760         if (desc->pg_lseg) {
2761                 pnfs_put_lseg(desc->pg_lseg);
2762                 desc->pg_lseg = NULL;
2763         }
2764 }
2765 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2766
2767 /*
2768  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2769  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2770  */
2771 size_t
2772 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2773                      struct nfs_page *prev, struct nfs_page *req)
2774 {
2775         unsigned int size;
2776         u64 seg_end, req_start, seg_left;
2777
2778         size = nfs_generic_pg_test(pgio, prev, req);
2779         if (!size)
2780                 return 0;
2781
2782         /*
2783          * 'size' contains the number of bytes left in the current page (up
2784          * to the original size asked for in @req->wb_bytes).
2785          *
2786          * Calculate how many bytes are left in the layout segment
2787          * and if there are less bytes than 'size', return that instead.
2788          *
2789          * Please also note that 'end_offset' is actually the offset of the
2790          * first byte that lies outside the pnfs_layout_range. FIXME?
2791          *
2792          */
2793         if (pgio->pg_lseg) {
2794                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2795                                      pgio->pg_lseg->pls_range.length);
2796                 req_start = req_offset(req);
2797
2798                 /* start of request is past the last byte of this segment */
2799                 if (req_start >= seg_end)
2800                         return 0;
2801
2802                 /* adjust 'size' iff there are fewer bytes left in the
2803                  * segment than what nfs_generic_pg_test returned */
2804                 seg_left = seg_end - req_start;
2805                 if (seg_left < size)
2806                         size = (unsigned int)seg_left;
2807         }
2808
2809         return size;
2810 }
2811 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2812
2813 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2814 {
2815         struct nfs_pageio_descriptor pgio;
2816
2817         /* Resend all requests through the MDS */
2818         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2819                               hdr->completion_ops);
2820         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2821         return nfs_pageio_resend(&pgio, hdr);
2822 }
2823 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2824
2825 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2826 {
2827
2828         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2829         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2830             PNFS_LAYOUTRET_ON_ERROR) {
2831                 pnfs_return_layout(hdr->inode);
2832         }
2833         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2834                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2835 }
2836
2837 /*
2838  * Called by non rpc-based layout drivers
2839  */
2840 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2841 {
2842         if (likely(!hdr->pnfs_error)) {
2843                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2844                                 hdr->mds_offset + hdr->res.count);
2845                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2846         }
2847         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2848         if (unlikely(hdr->pnfs_error))
2849                 pnfs_ld_handle_write_error(hdr);
2850         hdr->mds_ops->rpc_release(hdr);
2851 }
2852 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2853
2854 static void
2855 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2856                 struct nfs_pgio_header *hdr)
2857 {
2858         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2859
2860         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2861                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2862                 nfs_pageio_reset_write_mds(desc);
2863                 mirror->pg_recoalesce = 1;
2864         }
2865         hdr->completion_ops->completion(hdr);
2866 }
2867
2868 static enum pnfs_try_status
2869 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2870                         const struct rpc_call_ops *call_ops,
2871                         struct pnfs_layout_segment *lseg,
2872                         int how)
2873 {
2874         struct inode *inode = hdr->inode;
2875         enum pnfs_try_status trypnfs;
2876         struct nfs_server *nfss = NFS_SERVER(inode);
2877
2878         hdr->mds_ops = call_ops;
2879
2880         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2881                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2882         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2883         if (trypnfs != PNFS_NOT_ATTEMPTED)
2884                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2885         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2886         return trypnfs;
2887 }
2888
2889 static void
2890 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2891               struct nfs_pgio_header *hdr, int how)
2892 {
2893         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2894         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2895         enum pnfs_try_status trypnfs;
2896
2897         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2898         switch (trypnfs) {
2899         case PNFS_NOT_ATTEMPTED:
2900                 pnfs_write_through_mds(desc, hdr);
2901                 break;
2902         case PNFS_ATTEMPTED:
2903                 break;
2904         case PNFS_TRY_AGAIN:
2905                 /* cleanup hdr and prepare to redo pnfs */
2906                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2907                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2908                         list_splice_init(&hdr->pages, &mirror->pg_list);
2909                         mirror->pg_recoalesce = 1;
2910                 }
2911                 hdr->mds_ops->rpc_release(hdr);
2912         }
2913 }
2914
2915 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2916 {
2917         pnfs_put_lseg(hdr->lseg);
2918         nfs_pgio_header_free(hdr);
2919 }
2920
2921 int
2922 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2923 {
2924         struct nfs_pgio_header *hdr;
2925         int ret;
2926
2927         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2928         if (!hdr) {
2929                 desc->pg_error = -ENOMEM;
2930                 return desc->pg_error;
2931         }
2932         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2933
2934         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2935         ret = nfs_generic_pgio(desc, hdr);
2936         if (!ret)
2937                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2938
2939         return ret;
2940 }
2941 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2942
2943 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2944 {
2945         struct nfs_pageio_descriptor pgio;
2946
2947         /* Resend all requests through the MDS */
2948         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2949         return nfs_pageio_resend(&pgio, hdr);
2950 }
2951 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2952
2953 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2954 {
2955         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2956         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2957             PNFS_LAYOUTRET_ON_ERROR) {
2958                 pnfs_return_layout(hdr->inode);
2959         }
2960         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2961                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2962 }
2963
2964 /*
2965  * Called by non rpc-based layout drivers
2966  */
2967 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2968 {
2969         if (likely(!hdr->pnfs_error))
2970                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2971         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2972         if (unlikely(hdr->pnfs_error))
2973                 pnfs_ld_handle_read_error(hdr);
2974         hdr->mds_ops->rpc_release(hdr);
2975 }
2976 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2977
2978 static void
2979 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2980                 struct nfs_pgio_header *hdr)
2981 {
2982         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2983
2984         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2985                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2986                 nfs_pageio_reset_read_mds(desc);
2987                 mirror->pg_recoalesce = 1;
2988         }
2989         hdr->completion_ops->completion(hdr);
2990 }
2991
2992 /*
2993  * Call the appropriate parallel I/O subsystem read function.
2994  */
2995 static enum pnfs_try_status
2996 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2997                        const struct rpc_call_ops *call_ops,
2998                        struct pnfs_layout_segment *lseg)
2999 {
3000         struct inode *inode = hdr->inode;
3001         struct nfs_server *nfss = NFS_SERVER(inode);
3002         enum pnfs_try_status trypnfs;
3003
3004         hdr->mds_ops = call_ops;
3005
3006         dprintk("%s: Reading ino:%lu %u@%llu\n",
3007                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3008
3009         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3010         if (trypnfs != PNFS_NOT_ATTEMPTED)
3011                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3012         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3013         return trypnfs;
3014 }
3015
3016 /* Resend all requests through pnfs. */
3017 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3018                            unsigned int mirror_idx)
3019 {
3020         struct nfs_pageio_descriptor pgio;
3021
3022         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3023                 /* Prevent deadlocks with layoutreturn! */
3024                 pnfs_put_lseg(hdr->lseg);
3025                 hdr->lseg = NULL;
3026
3027                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3028                                         hdr->completion_ops);
3029                 pgio.pg_mirror_idx = mirror_idx;
3030                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3031         }
3032 }
3033 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3034
3035 static void
3036 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3037 {
3038         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3039         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3040         enum pnfs_try_status trypnfs;
3041
3042         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3043         switch (trypnfs) {
3044         case PNFS_NOT_ATTEMPTED:
3045                 pnfs_read_through_mds(desc, hdr);
3046                 break;
3047         case PNFS_ATTEMPTED:
3048                 break;
3049         case PNFS_TRY_AGAIN:
3050                 /* cleanup hdr and prepare to redo pnfs */
3051                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3052                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3053                         list_splice_init(&hdr->pages, &mirror->pg_list);
3054                         mirror->pg_recoalesce = 1;
3055                 }
3056                 hdr->mds_ops->rpc_release(hdr);
3057         }
3058 }
3059
3060 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3061 {
3062         pnfs_put_lseg(hdr->lseg);
3063         nfs_pgio_header_free(hdr);
3064 }
3065
3066 int
3067 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3068 {
3069         struct nfs_pgio_header *hdr;
3070         int ret;
3071
3072         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3073         if (!hdr) {
3074                 desc->pg_error = -ENOMEM;
3075                 return desc->pg_error;
3076         }
3077         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3078         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3079         ret = nfs_generic_pgio(desc, hdr);
3080         if (!ret)
3081                 pnfs_do_read(desc, hdr);
3082         return ret;
3083 }
3084 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3085
3086 static void pnfs_clear_layoutcommitting(struct inode *inode)
3087 {
3088         unsigned long *bitlock = &NFS_I(inode)->flags;
3089
3090         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3091         smp_mb__after_atomic();
3092         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3093 }
3094
3095 /*
3096  * There can be multiple RW segments.
3097  */
3098 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3099 {
3100         struct pnfs_layout_segment *lseg;
3101
3102         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3103                 if (lseg->pls_range.iomode == IOMODE_RW &&
3104                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3105                         list_add(&lseg->pls_lc_list, listp);
3106         }
3107 }
3108
3109 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3110 {
3111         struct pnfs_layout_segment *lseg, *tmp;
3112
3113         /* Matched by references in pnfs_set_layoutcommit */
3114         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3115                 list_del_init(&lseg->pls_lc_list);
3116                 pnfs_put_lseg(lseg);
3117         }
3118
3119         pnfs_clear_layoutcommitting(inode);
3120 }
3121
3122 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3123 {
3124         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3125 }
3126 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3127
3128 void
3129 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3130                 loff_t end_pos)
3131 {
3132         struct nfs_inode *nfsi = NFS_I(inode);
3133         bool mark_as_dirty = false;
3134
3135         spin_lock(&inode->i_lock);
3136         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3137                 nfsi->layout->plh_lwb = end_pos;
3138                 mark_as_dirty = true;
3139                 dprintk("%s: Set layoutcommit for inode %lu ",
3140                         __func__, inode->i_ino);
3141         } else if (end_pos > nfsi->layout->plh_lwb)
3142                 nfsi->layout->plh_lwb = end_pos;
3143         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3144                 /* references matched in nfs4_layoutcommit_release */
3145                 pnfs_get_lseg(lseg);
3146         }
3147         spin_unlock(&inode->i_lock);
3148         dprintk("%s: lseg %p end_pos %llu\n",
3149                 __func__, lseg, nfsi->layout->plh_lwb);
3150
3151         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3152          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3153         if (mark_as_dirty)
3154                 mark_inode_dirty_sync(inode);
3155 }
3156 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3157
3158 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3159 {
3160         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3161
3162         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3163                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3164         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3165 }
3166
3167 /*
3168  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3169  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3170  * data to disk to allow the server to recover the data if it crashes.
3171  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3172  * is off, and a COMMIT is sent to a data server, or
3173  * if WRITEs to a data server return NFS_DATA_SYNC.
3174  */
3175 int
3176 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3177 {
3178         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3179         struct nfs4_layoutcommit_data *data;
3180         struct nfs_inode *nfsi = NFS_I(inode);
3181         loff_t end_pos;
3182         int status;
3183
3184         if (!pnfs_layoutcommit_outstanding(inode))
3185                 return 0;
3186
3187         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3188
3189         status = -EAGAIN;
3190         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3191                 if (!sync)
3192                         goto out;
3193                 status = wait_on_bit_lock_action(&nfsi->flags,
3194                                 NFS_INO_LAYOUTCOMMITTING,
3195                                 nfs_wait_bit_killable,
3196                                 TASK_KILLABLE);
3197                 if (status)
3198                         goto out;
3199         }
3200
3201         status = -ENOMEM;
3202         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3203         data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3204         if (!data)
3205                 goto clear_layoutcommitting;
3206
3207         status = 0;
3208         spin_lock(&inode->i_lock);
3209         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3210                 goto out_unlock;
3211
3212         INIT_LIST_HEAD(&data->lseg_list);
3213         pnfs_list_write_lseg(inode, &data->lseg_list);
3214
3215         end_pos = nfsi->layout->plh_lwb;
3216
3217         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3218         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3219         spin_unlock(&inode->i_lock);
3220
3221         data->args.inode = inode;
3222         nfs_fattr_init(&data->fattr);
3223         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3224         data->res.fattr = &data->fattr;
3225         if (end_pos != 0)
3226                 data->args.lastbytewritten = end_pos - 1;
3227         else
3228                 data->args.lastbytewritten = U64_MAX;
3229         data->res.server = NFS_SERVER(inode);
3230
3231         if (ld->prepare_layoutcommit) {
3232                 status = ld->prepare_layoutcommit(&data->args);
3233                 if (status) {
3234                         put_cred(data->cred);
3235                         spin_lock(&inode->i_lock);
3236                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3237                         if (end_pos > nfsi->layout->plh_lwb)
3238                                 nfsi->layout->plh_lwb = end_pos;
3239                         goto out_unlock;
3240                 }
3241         }
3242
3243
3244         status = nfs4_proc_layoutcommit(data, sync);
3245 out:
3246         if (status)
3247                 mark_inode_dirty_sync(inode);
3248         dprintk("<-- %s status %d\n", __func__, status);
3249         return status;
3250 out_unlock:
3251         spin_unlock(&inode->i_lock);
3252         kfree(data);
3253 clear_layoutcommitting:
3254         pnfs_clear_layoutcommitting(inode);
3255         goto out;
3256 }
3257 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3258
3259 int
3260 pnfs_generic_sync(struct inode *inode, bool datasync)
3261 {
3262         return pnfs_layoutcommit_inode(inode, true);
3263 }
3264 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3265
3266 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3267 {
3268         struct nfs4_threshold *thp;
3269
3270         thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3271         if (!thp) {
3272                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3273                 return NULL;
3274         }
3275         return thp;
3276 }
3277
3278 #if IS_ENABLED(CONFIG_NFS_V4_2)
3279 int
3280 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3281 {
3282         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3283         struct nfs_server *server = NFS_SERVER(inode);
3284         struct nfs_inode *nfsi = NFS_I(inode);
3285         struct nfs42_layoutstat_data *data;
3286         struct pnfs_layout_hdr *hdr;
3287         int status = 0;
3288
3289         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3290                 goto out;
3291
3292         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3293                 goto out;
3294
3295         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3296                 goto out;
3297
3298         spin_lock(&inode->i_lock);
3299         if (!NFS_I(inode)->layout) {
3300                 spin_unlock(&inode->i_lock);
3301                 goto out_clear_layoutstats;
3302         }
3303         hdr = NFS_I(inode)->layout;
3304         pnfs_get_layout_hdr(hdr);
3305         spin_unlock(&inode->i_lock);
3306
3307         data = kzalloc(sizeof(*data), gfp_flags);
3308         if (!data) {
3309                 status = -ENOMEM;
3310                 goto out_put;
3311         }
3312
3313         data->args.fh = NFS_FH(inode);
3314         data->args.inode = inode;
3315         status = ld->prepare_layoutstats(&data->args);
3316         if (status)
3317                 goto out_free;
3318
3319         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3320
3321 out:
3322         dprintk("%s returns %d\n", __func__, status);
3323         return status;
3324
3325 out_free:
3326         kfree(data);
3327 out_put:
3328         pnfs_put_layout_hdr(hdr);
3329 out_clear_layoutstats:
3330         smp_mb__before_atomic();
3331         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3332         smp_mb__after_atomic();
3333         goto out;
3334 }
3335 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3336 #endif
3337
3338 unsigned int layoutstats_timer;
3339 module_param(layoutstats_timer, uint, 0644);
3340 EXPORT_SYMBOL_GPL(layoutstats_timer);