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                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 612                         spin_unlock(&inode->i_lock);
 613                         return;
 614                 }
 615                 pnfs_get_layout_hdr(lo);
 616                 pnfs_layout_remove_lseg(lo, lseg);
 617                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 618                         lseg = NULL;
 619                 spin_unlock(&inode->i_lock);
 620                 pnfs_free_lseg(lseg);
 621                 pnfs_put_layout_hdr(lo);
 622         }
 623 }
 624 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 625
 626 /*
 627  * is l2 fully contained in l1?
 628  *   start1                             end1
 629  *   [----------------------------------)
 630  *           start2           end2
 631  *           [----------------)
 632  */
 633 static bool
 634 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 635                  const struct pnfs_layout_range *l2)
 636 {
 637         u64 start1 = l1->offset;
 638         u64 end1 = pnfs_end_offset(start1, l1->length);
 639         u64 start2 = l2->offset;
 640         u64 end2 = pnfs_end_offset(start2, l2->length);
 641
 642         return (start1 <= start2) && (end1 >= end2);
 643 }
 644
 645 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 646                 struct list_head *tmp_list)
 647 {
 648         if (!refcount_dec_and_test(&lseg->pls_refcount))
 649                 return false;
 650         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 651         list_add(&lseg->pls_list, tmp_list);
 652         return true;
 653 }
 654
 655 /* Returns 1 if lseg is removed from list, 0 otherwise */
 656 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 657                              struct list_head *tmp_list)
 658 {
 659         int rv = 0;
 660
 661         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 662                 /* Remove the reference keeping the lseg in the
 663                  * list.  It will now be removed when all
 664                  * outstanding io is finished.
 665                  */
 666                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 667                         refcount_read(&lseg->pls_refcount));
 668                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 669                         rv = 1;
 670         }
 671         return rv;
 672 }
 673
 674 static bool
 675 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 676                  const struct pnfs_layout_range *recall_range)
 677 {
 678         return (recall_range->iomode == IOMODE_ANY ||
 679                 lseg_range->iomode == recall_range->iomode) &&
 680                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 681 }
 682
 683 static bool
 684 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 685                 const struct pnfs_layout_range *recall_range,
 686                 u32 seq)
 687 {
 688         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 689                 return false;
 690         if (recall_range == NULL)
 691                 return true;
 692         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 693 }
 694
 695 /**
 696  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 697  * @lo: layout header containing the lsegs
 698  * @tmp_list: list head where doomed lsegs should go
 699  * @recall_range: optional recall range argument to match (may be NULL)
 700  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 701  *
 702  * Walk the list of lsegs in the layout header, and tear down any that should
 703  * be destroyed. If "recall_range" is specified then the segment must match
 704  * that range. If "seq" is non-zero, then only match segments that were handed
 705  * out at or before that sequence.
 706  *
 707  * Returns number of matching invalid lsegs remaining in list after scanning
 708  * it and purging them.
 709  */
 710 int
 711 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 712                             struct list_head *tmp_list,
 713                             const struct pnfs_layout_range *recall_range,
 714                             u32 seq)
 715 {
 716         struct pnfs_layout_segment *lseg, *next;
 717         int remaining = 0;
 718
 719         dprintk("%s:Begin lo %p\n", __func__, lo);
 720
 721         if (list_empty(&lo->plh_segs))
 722                 return 0;
 723         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 724                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 725                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 726                                 "offset %llu length %llu\n", __func__,
 727                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 728                                 lseg->pls_range.offset, lseg->pls_range.length);
 729                         if (!mark_lseg_invalid(lseg, tmp_list))
 730                                 remaining++;
 731                 }
 732         dprintk("%s:Return %i\n", __func__, remaining);
 733         return remaining;
 734 }
 735
 736 static void
 737 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 738                 struct list_head *free_me,
 739                 const struct pnfs_layout_range *range,
 740                 u32 seq)
 741 {
 742         struct pnfs_layout_segment *lseg, *next;
 743
 744         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 745                 if (pnfs_match_lseg_recall(lseg, range, seq))
 746                         list_move_tail(&lseg->pls_list, free_me);
 747         }
 748 }
 749
 750 /* note free_me must contain lsegs from a single layout_hdr */
 751 void
 752 pnfs_free_lseg_list(struct list_head *free_me)
 753 {
 754         struct pnfs_layout_segment *lseg, *tmp;
 755
 756         if (list_empty(free_me))
 757                 return;
 758
 759         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 760                 list_del(&lseg->pls_list);
 761                 pnfs_free_lseg(lseg);
 762         }
 763 }
 764
 765 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 766 {
 767         struct pnfs_layout_hdr *lo;
 768         LIST_HEAD(tmp_list);
 769
 770         spin_lock(&nfsi->vfs_inode.i_lock);
 771         lo = nfsi->layout;
 772         if (lo) {
 773                 pnfs_get_layout_hdr(lo);
 774                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 775                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 776                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 777                 spin_unlock(&nfsi->vfs_inode.i_lock);
 778                 pnfs_free_lseg_list(&tmp_list);
 779                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 780                 pnfs_put_layout_hdr(lo);
 781         } else
 782                 spin_unlock(&nfsi->vfs_inode.i_lock);
 783         return lo;
 784 }
 785
 786 void pnfs_destroy_layout(struct nfs_inode *nfsi)
 787 {
 788         __pnfs_destroy_layout(nfsi);
 789 }
 790 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 791
 792 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
 793                                 struct pnfs_layout_hdr *lo)
 794 {
 795         bool ret;
 796
 797         spin_lock(&nfsi->vfs_inode.i_lock);
 798         ret = nfsi->layout != lo;
 799         spin_unlock(&nfsi->vfs_inode.i_lock);
 800         return ret;
 801 }
 802
 803 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 804 {
 805         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 806
 807         if (lo)
 808                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
 809 }
 810
 811 static bool
 812 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 813                 struct list_head *layout_list)
 814 {
 815         struct pnfs_layout_hdr *lo;
 816         bool ret = false;
 817
 818         spin_lock(&inode->i_lock);
 819         lo = NFS_I(inode)->layout;
 820         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 821                 pnfs_get_layout_hdr(lo);
 822                 list_add(&lo->plh_bulk_destroy, layout_list);
 823                 ret = true;
 824         }
 825         spin_unlock(&inode->i_lock);
 826         return ret;
 827 }
 828
 829 /* Caller must hold rcu_read_lock and clp->cl_lock */
 830 static int
 831 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 832                 struct nfs_server *server,
 833                 struct list_head *layout_list)
 834         __must_hold(&clp->cl_lock)
 835         __must_hold(RCU)
 836 {
 837         struct pnfs_layout_hdr *lo, *next;
 838         struct inode *inode;
 839
 840         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 841                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 842                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 843                     !list_empty(&lo->plh_bulk_destroy))
 844                         continue;
 845                 /* If the sb is being destroyed, just bail */
 846                 if (!nfs_sb_active(server->super))
 847                         break;
 848                 inode = pnfs_grab_inode_layout_hdr(lo);
 849                 if (inode != NULL) {
 850                         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
 851                                 list_del_rcu(&lo->plh_layouts);
 852                         if (pnfs_layout_add_bulk_destroy_list(inode,
 853                                                 layout_list))
 854                                 continue;
 855                         rcu_read_unlock();
 856                         spin_unlock(&clp->cl_lock);
 857                         iput(inode);
 858                 } else {
 859                         rcu_read_unlock();
 860                         spin_unlock(&clp->cl_lock);
 861                 }
 862                 nfs_sb_deactive(server->super);
 863                 spin_lock(&clp->cl_lock);
 864                 rcu_read_lock();
 865                 return -EAGAIN;
 866         }
 867         return 0;
 868 }
 869
 870 static int
 871 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 872                 bool is_bulk_recall)
 873 {
 874         struct pnfs_layout_hdr *lo;
 875         struct inode *inode;
 876         LIST_HEAD(lseg_list);
 877         int ret = 0;
 878
 879         while (!list_empty(layout_list)) {
 880                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 881                                 plh_bulk_destroy);
 882                 dprintk("%s freeing layout for inode %lu\n", __func__,
 883                         lo->plh_inode->i_ino);
 884                 inode = lo->plh_inode;
 885
 886                 pnfs_layoutcommit_inode(inode, false);
 887
 888                 spin_lock(&inode->i_lock);
 889                 list_del_init(&lo->plh_bulk_destroy);
 890                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 891                         if (is_bulk_recall)
 892                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 893                         ret = -EAGAIN;
 894                 }
 895                 spin_unlock(&inode->i_lock);
 896                 pnfs_free_lseg_list(&lseg_list);
 897                 /* Free all lsegs that are attached to commit buckets */
 898                 nfs_commit_inode(inode, 0);
 899                 pnfs_put_layout_hdr(lo);
 900                 nfs_iput_and_deactive(inode);
 901         }
 902         return ret;
 903 }
 904
 905 int
 906 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 907                 struct nfs_fsid *fsid,
 908                 bool is_recall)
 909 {
 910         struct nfs_server *server;
 911         LIST_HEAD(layout_list);
 912
 913         spin_lock(&clp->cl_lock);
 914         rcu_read_lock();
 915 restart:
 916         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 917                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 918                         continue;
 919                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 920                                 server,
 921                                 &layout_list) != 0)
 922                         goto restart;
 923         }
 924         rcu_read_unlock();
 925         spin_unlock(&clp->cl_lock);
 926
 927         if (list_empty(&layout_list))
 928                 return 0;
 929         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 930 }
 931
 932 int
 933 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 934                 bool is_recall)
 935 {
 936         struct nfs_server *server;
 937         LIST_HEAD(layout_list);
 938
 939         spin_lock(&clp->cl_lock);
 940         rcu_read_lock();
 941 restart:
 942         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 943                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 944                                         server,
 945                                         &layout_list) != 0)
 946                         goto restart;
 947         }
 948         rcu_read_unlock();
 949         spin_unlock(&clp->cl_lock);
 950
 951         if (list_empty(&layout_list))
 952                 return 0;
 953         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 954 }
 955
 956 /*
 957  * Called by the state manager to remove all layouts established under an
 958  * expired lease.
 959  */
 960 void
 961 pnfs_destroy_all_layouts(struct nfs_client *clp)
 962 {
 963         nfs4_deviceid_mark_client_invalid(clp);
 964         nfs4_deviceid_purge_client(clp);
 965
 966         pnfs_destroy_layouts_byclid(clp, false);
 967 }
 968
 969 static void
 970 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
 971 {
 972         const struct cred *old;
 973
 974         if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
 975                 old = xchg(&lo->plh_lc_cred, get_cred(cred));
 976                 put_cred(old);
 977         }
 978 }
 979
 980 /* update lo->plh_stateid with new if is more recent */
 981 void
 982 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 983                         const struct cred *cred, bool update_barrier)
 984 {
 985         u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 986         u32 newseq = be32_to_cpu(new->seqid);
 987
 988         if (!pnfs_layout_is_valid(lo)) {
 989                 pnfs_set_layout_cred(lo, cred);
 990                 nfs4_stateid_copy(&lo->plh_stateid, new);
 991                 lo->plh_barrier = newseq;
 992                 pnfs_clear_layoutreturn_info(lo);
 993                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 994                 return;
 995         }
 996
 997         if (pnfs_seqid_is_newer(newseq, oldseq))
 998                 nfs4_stateid_copy(&lo->plh_stateid, new);
 999
1000         if (update_barrier) {
1001                 pnfs_barrier_update(lo, newseq);
1002                 return;
1003         }
1004         /*
1005          * Because of wraparound, we want to keep the barrier
1006          * "close" to the current seqids. We really only want to
1007          * get here from a layoutget call.
1008          */
1009         if (atomic_read(&lo->plh_outstanding) == 1)
1010                  pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1011 }
1012
1013 static bool
1014 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1015                 const nfs4_stateid *stateid)
1016 {
1017         u32 seqid = be32_to_cpu(stateid->seqid);
1018
1019         return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1020 }
1021
1022 /* lget is set to 1 if called from inside send_layoutget call chain */
1023 static bool
1024 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1025 {
1026         return lo->plh_block_lgets ||
1027                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1028 }
1029
1030 static struct nfs_server *
1031 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1032 {
1033         struct nfs_server *server;
1034
1035         if (inode) {
1036                 server = NFS_SERVER(inode);
1037         } else {
1038                 struct dentry *parent_dir = dget_parent(ctx->dentry);
1039                 server = NFS_SERVER(parent_dir->d_inode);
1040                 dput(parent_dir);
1041         }
1042         return server;
1043 }
1044
1045 static void nfs4_free_pages(struct page **pages, size_t size)
1046 {
1047         int i;
1048
1049         if (!pages)
1050                 return;
1051
1052         for (i = 0; i < size; i++) {
1053                 if (!pages[i])
1054                         break;
1055                 __free_page(pages[i]);
1056         }
1057         kfree(pages);
1058 }
1059
1060 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1061 {
1062         struct page **pages;
1063         int i;
1064
1065         pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1066         if (!pages) {
1067                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1068                 return NULL;
1069         }
1070
1071         for (i = 0; i < size; i++) {
1072                 pages[i] = alloc_page(gfp_flags);
1073                 if (!pages[i]) {
1074                         dprintk("%s: failed to allocate page\n", __func__);
1075                         nfs4_free_pages(pages, i);
1076                         return NULL;
1077                 }
1078         }
1079
1080         return pages;
1081 }
1082
1083 static struct nfs4_layoutget *
1084 pnfs_alloc_init_layoutget_args(struct inode *ino,
1085            struct nfs_open_context *ctx,
1086            const nfs4_stateid *stateid,
1087            const struct pnfs_layout_range *range,
1088            gfp_t gfp_flags)
1089 {
1090         struct nfs_server *server = pnfs_find_server(ino, ctx);
1091         size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1092         size_t max_pages = max_response_pages(server);
1093         struct nfs4_layoutget *lgp;
1094
1095         dprintk("--> %s\n", __func__);
1096
1097         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1098         if (lgp == NULL)
1099                 return NULL;
1100
1101         if (max_reply_sz) {
1102                 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1103                 if (npages < max_pages)
1104                         max_pages = npages;
1105         }
1106
1107         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1108         if (!lgp->args.layout.pages) {
1109                 kfree(lgp);
1110                 return NULL;
1111         }
1112         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1113         lgp->res.layoutp = &lgp->args.layout;
1114
1115         /* Don't confuse uninitialised result and success */
1116         lgp->res.status = -NFS4ERR_DELAY;
1117
1118         lgp->args.minlength = PAGE_SIZE;
1119         if (lgp->args.minlength > range->length)
1120                 lgp->args.minlength = range->length;
1121         if (ino) {
1122                 loff_t i_size = i_size_read(ino);
1123
1124                 if (range->iomode == IOMODE_READ) {
1125                         if (range->offset >= i_size)
1126                                 lgp->args.minlength = 0;
1127                         else if (i_size - range->offset < lgp->args.minlength)
1128                                 lgp->args.minlength = i_size - range->offset;
1129                 }
1130         }
1131         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1132         pnfs_copy_range(&lgp->args.range, range);
1133         lgp->args.type = server->pnfs_curr_ld->id;
1134         lgp->args.inode = ino;
1135         lgp->args.ctx = get_nfs_open_context(ctx);
1136         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1137         lgp->gfp_flags = gfp_flags;
1138         lgp->cred = ctx->cred;
1139         return lgp;
1140 }
1141
1142 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1143 {
1144         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1145
1146         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1147         if (lgp->args.inode)
1148                 pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1149         put_nfs_open_context(lgp->args.ctx);
1150         kfree(lgp);
1151 }
1152
1153 static void pnfs_clear_layoutcommit(struct inode *inode,
1154                 struct list_head *head)
1155 {
1156         struct nfs_inode *nfsi = NFS_I(inode);
1157         struct pnfs_layout_segment *lseg, *tmp;
1158
1159         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1160                 return;
1161         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1162                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1163                         continue;
1164                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1165         }
1166 }
1167
1168 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1169                 const nfs4_stateid *arg_stateid,
1170                 const struct pnfs_layout_range *range,
1171                 const nfs4_stateid *stateid)
1172 {
1173         struct inode *inode = lo->plh_inode;
1174         LIST_HEAD(freeme);
1175
1176         spin_lock(&inode->i_lock);
1177         if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1178             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1179                 goto out_unlock;
1180         if (stateid) {
1181                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1182
1183                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1184                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1185                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1186         } else
1187                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1188 out_unlock:
1189         pnfs_clear_layoutreturn_waitbit(lo);
1190         spin_unlock(&inode->i_lock);
1191         pnfs_free_lseg_list(&freeme);
1192
1193 }
1194
1195 static bool
1196 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1197                 nfs4_stateid *stateid,
1198                 const struct cred **cred,
1199                 enum pnfs_iomode *iomode)
1200 {
1201         /* Serialise LAYOUTGET/LAYOUTRETURN */
1202         if (atomic_read(&lo->plh_outstanding) != 0)
1203                 return false;
1204         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1205                 return false;
1206         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1207         pnfs_get_layout_hdr(lo);
1208         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1209         *cred = get_cred(lo->plh_lc_cred);
1210         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1211                 if (lo->plh_return_seq != 0)
1212                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1213                 if (iomode != NULL)
1214                         *iomode = lo->plh_return_iomode;
1215                 pnfs_clear_layoutreturn_info(lo);
1216         } else if (iomode != NULL)
1217                 *iomode = IOMODE_ANY;
1218         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1219         return true;
1220 }
1221
1222 static void
1223 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1224                 struct pnfs_layout_hdr *lo,
1225                 const nfs4_stateid *stateid,
1226                 enum pnfs_iomode iomode)
1227 {
1228         struct inode *inode = lo->plh_inode;
1229
1230         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1231         args->inode = inode;
1232         args->range.iomode = iomode;
1233         args->range.offset = 0;
1234         args->range.length = NFS4_MAX_UINT64;
1235         args->layout = lo;
1236         nfs4_stateid_copy(&args->stateid, stateid);
1237 }
1238
1239 static int
1240 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1241                        const nfs4_stateid *stateid,
1242                        const struct cred **pcred,
1243                        enum pnfs_iomode iomode,
1244                        bool sync)
1245 {
1246         struct inode *ino = lo->plh_inode;
1247         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1248         struct nfs4_layoutreturn *lrp;
1249         const struct cred *cred = *pcred;
1250         int status = 0;
1251
1252         *pcred = NULL;
1253         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1254         if (unlikely(lrp == NULL)) {
1255                 status = -ENOMEM;
1256                 spin_lock(&ino->i_lock);
1257                 pnfs_clear_layoutreturn_waitbit(lo);
1258                 spin_unlock(&ino->i_lock);
1259                 put_cred(cred);
1260                 pnfs_put_layout_hdr(lo);
1261                 goto out;
1262         }
1263
1264         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1265         lrp->args.ld_private = &lrp->ld_private;
1266         lrp->clp = NFS_SERVER(ino)->nfs_client;
1267         lrp->cred = cred;
1268         if (ld->prepare_layoutreturn)
1269                 ld->prepare_layoutreturn(&lrp->args);
1270
1271         status = nfs4_proc_layoutreturn(lrp, sync);
1272 out:
1273         dprintk("<-- %s status: %d\n", __func__, status);
1274         return status;
1275 }
1276
1277 static bool
1278 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1279                                 enum pnfs_iomode iomode,
1280                                 u32 seq)
1281 {
1282         struct pnfs_layout_range recall_range = {
1283                 .length = NFS4_MAX_UINT64,
1284                 .iomode = iomode,
1285         };
1286         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1287                                                &recall_range, seq) != -EBUSY;
1288 }
1289
1290 /* Return true if layoutreturn is needed */
1291 static bool
1292 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1293 {
1294         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1295                 return false;
1296         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1297                                                lo->plh_return_seq);
1298 }
1299
1300 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1301 {
1302         struct inode *inode= lo->plh_inode;
1303
1304         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1305                 return;
1306         spin_lock(&inode->i_lock);
1307         if (pnfs_layout_need_return(lo)) {
1308                 const struct cred *cred;
1309                 nfs4_stateid stateid;
1310                 enum pnfs_iomode iomode;
1311                 bool send;
1312
1313                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1314                 spin_unlock(&inode->i_lock);
1315                 if (send) {
1316                         /* Send an async layoutreturn so we dont deadlock */
1317                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1318                 }
1319         } else
1320                 spin_unlock(&inode->i_lock);
1321 }
1322
1323 /*
1324  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1325  * when the layout segment list is empty.
1326  *
1327  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1328  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1329  * deviceid is marked invalid.
1330  */
1331 int
1332 _pnfs_return_layout(struct inode *ino)
1333 {
1334         struct pnfs_layout_hdr *lo = NULL;
1335         struct nfs_inode *nfsi = NFS_I(ino);
1336         struct pnfs_layout_range range = {
1337                 .iomode         = IOMODE_ANY,
1338                 .offset         = 0,
1339                 .length         = NFS4_MAX_UINT64,
1340         };
1341         LIST_HEAD(tmp_list);
1342         const struct cred *cred;
1343         nfs4_stateid stateid;
1344         int status = 0;
1345         bool send, valid_layout;
1346
1347         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1348
1349         spin_lock(&ino->i_lock);
1350         lo = nfsi->layout;
1351         if (!lo) {
1352                 spin_unlock(&ino->i_lock);
1353                 dprintk("NFS: %s no layout to return\n", __func__);
1354                 goto out;
1355         }
1356         /* Reference matched in nfs4_layoutreturn_release */
1357         pnfs_get_layout_hdr(lo);
1358         /* Is there an outstanding layoutreturn ? */
1359         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1360                 spin_unlock(&ino->i_lock);
1361                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1362                                         TASK_UNINTERRUPTIBLE))
1363                         goto out_put_layout_hdr;
1364                 spin_lock(&ino->i_lock);
1365         }
1366         valid_layout = pnfs_layout_is_valid(lo);
1367         pnfs_clear_layoutcommit(ino, &tmp_list);
1368         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1369
1370         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1371                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1372
1373         /* Don't send a LAYOUTRETURN if list was initially empty */
1374         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1375                         !valid_layout) {
1376                 spin_unlock(&ino->i_lock);
1377                 dprintk("NFS: %s no layout segments to return\n", __func__);
1378                 goto out_wait_layoutreturn;
1379         }
1380
1381         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1382         spin_unlock(&ino->i_lock);
1383         if (send)
1384                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1385 out_wait_layoutreturn:
1386         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1387 out_put_layout_hdr:
1388         pnfs_free_lseg_list(&tmp_list);
1389         pnfs_put_layout_hdr(lo);
1390 out:
1391         dprintk("<-- %s status: %d\n", __func__, status);
1392         return status;
1393 }
1394
1395 int
1396 pnfs_commit_and_return_layout(struct inode *inode)
1397 {
1398         struct pnfs_layout_hdr *lo;
1399         int ret;
1400
1401         spin_lock(&inode->i_lock);
1402         lo = NFS_I(inode)->layout;
1403         if (lo == NULL) {
1404                 spin_unlock(&inode->i_lock);
1405                 return 0;
1406         }
1407         pnfs_get_layout_hdr(lo);
1408         /* Block new layoutgets and read/write to ds */
1409         lo->plh_block_lgets++;
1410         spin_unlock(&inode->i_lock);
1411         filemap_fdatawait(inode->i_mapping);
1412         ret = pnfs_layoutcommit_inode(inode, true);
1413         if (ret == 0)
1414                 ret = _pnfs_return_layout(inode);
1415         spin_lock(&inode->i_lock);
1416         lo->plh_block_lgets--;
1417         spin_unlock(&inode->i_lock);
1418         pnfs_put_layout_hdr(lo);
1419         return ret;
1420 }
1421
1422 bool pnfs_roc(struct inode *ino,
1423                 struct nfs4_layoutreturn_args *args,
1424                 struct nfs4_layoutreturn_res *res,
1425                 const struct cred *cred)
1426 {
1427         struct nfs_inode *nfsi = NFS_I(ino);
1428         struct nfs_open_context *ctx;
1429         struct nfs4_state *state;
1430         struct pnfs_layout_hdr *lo;
1431         struct pnfs_layout_segment *lseg, *next;
1432         const struct cred *lc_cred;
1433         nfs4_stateid stateid;
1434         enum pnfs_iomode iomode = 0;
1435         bool layoutreturn = false, roc = false;
1436         bool skip_read = false;
1437
1438         if (!nfs_have_layout(ino))
1439                 return false;
1440 retry:
1441         rcu_read_lock();
1442         spin_lock(&ino->i_lock);
1443         lo = nfsi->layout;
1444         if (!lo || !pnfs_layout_is_valid(lo) ||
1445             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1446                 lo = NULL;
1447                 goto out_noroc;
1448         }
1449         pnfs_get_layout_hdr(lo);
1450         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1451                 spin_unlock(&ino->i_lock);
1452                 rcu_read_unlock();
1453                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1454                                 TASK_UNINTERRUPTIBLE);
1455                 pnfs_put_layout_hdr(lo);
1456                 goto retry;
1457         }
1458
1459         /* no roc if we hold a delegation */
1460         if (nfs4_check_delegation(ino, FMODE_READ)) {
1461                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1462                         goto out_noroc;
1463                 skip_read = true;
1464         }
1465
1466         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1467                 state = ctx->state;
1468                 if (state == NULL)
1469                         continue;
1470                 /* Don't return layout if there is open file state */
1471                 if (state->state & FMODE_WRITE)
1472                         goto out_noroc;
1473                 if (state->state & FMODE_READ)
1474                         skip_read = true;
1475         }
1476
1477
1478         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1479                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1480                         continue;
1481                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1482                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1483                         continue;
1484                 /*
1485                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1486                  * doesn't invalidate the layout for us.
1487                  */
1488                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1489                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1490                         continue;
1491                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1492         }
1493
1494         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1495                 goto out_noroc;
1496
1497         /* ROC in two conditions:
1498          * 1. there are ROC lsegs
1499          * 2. we don't send layoutreturn
1500          */
1501         /* lo ref dropped in pnfs_roc_release() */
1502         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1503         /* If the creds don't match, we can't compound the layoutreturn */
1504         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1505                 goto out_noroc;
1506
1507         roc = layoutreturn;
1508         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1509         res->lrs_present = 0;
1510         layoutreturn = false;
1511         put_cred(lc_cred);
1512
1513 out_noroc:
1514         spin_unlock(&ino->i_lock);
1515         rcu_read_unlock();
1516         pnfs_layoutcommit_inode(ino, true);
1517         if (roc) {
1518                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1519                 if (ld->prepare_layoutreturn)
1520                         ld->prepare_layoutreturn(args);
1521                 pnfs_put_layout_hdr(lo);
1522                 return true;
1523         }
1524         if (layoutreturn)
1525                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1526         pnfs_put_layout_hdr(lo);
1527         return false;
1528 }
1529
1530 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1531                   struct nfs4_layoutreturn_res **respp, int *ret)
1532 {
1533         struct nfs4_layoutreturn_args *arg = *argpp;
1534         int retval = -EAGAIN;
1535
1536         if (!arg)
1537                 return 0;
1538         /* Handle Layoutreturn errors */
1539         switch (*ret) {
1540         case 0:
1541                 retval = 0;
1542                 break;
1543         case -NFS4ERR_NOMATCHING_LAYOUT:
1544                 /* Was there an RPC level error? If not, retry */
1545                 if (task->tk_rpc_status == 0)
1546                         break;
1547                 /* If the call was not sent, let caller handle it */
1548                 if (!RPC_WAS_SENT(task))
1549                         return 0;
1550                 /*
1551                  * Otherwise, assume the call succeeded and
1552                  * that we need to release the layout
1553                  */
1554                 *ret = 0;
1555                 (*respp)->lrs_present = 0;
1556                 retval = 0;
1557                 break;
1558         case -NFS4ERR_DELAY:
1559                 /* Let the caller handle the retry */
1560                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1561                 return 0;
1562         case -NFS4ERR_OLD_STATEID:
1563                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1564                                                      &arg->range, arg->inode))
1565                         break;
1566                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1567                 return -EAGAIN;
1568         }
1569         *argpp = NULL;
1570         *respp = NULL;
1571         return retval;
1572 }
1573
1574 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1575                 struct nfs4_layoutreturn_res *res,
1576                 int ret)
1577 {
1578         struct pnfs_layout_hdr *lo = args->layout;
1579         struct inode *inode = args->inode;
1580         const nfs4_stateid *arg_stateid = NULL;
1581         const nfs4_stateid *res_stateid = NULL;
1582         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1583
1584         switch (ret) {
1585         case -NFS4ERR_NOMATCHING_LAYOUT:
1586                 spin_lock(&inode->i_lock);
1587                 if (pnfs_layout_is_valid(lo) &&
1588                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1589                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1590                 spin_unlock(&inode->i_lock);
1591                 break;
1592         case 0:
1593                 if (res->lrs_present)
1594                         res_stateid = &res->stateid;
1595                 fallthrough;
1596         default:
1597                 arg_stateid = &args->stateid;
1598         }
1599         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1600         pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1601                         res_stateid);
1602         if (ld_private && ld_private->ops && ld_private->ops->free)
1603                 ld_private->ops->free(ld_private);
1604         pnfs_put_layout_hdr(lo);
1605 }
1606
1607 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1608 {
1609         struct nfs_inode *nfsi = NFS_I(ino);
1610         struct pnfs_layout_hdr *lo;
1611         bool sleep = false;
1612
1613         /* we might not have grabbed lo reference. so need to check under
1614          * i_lock */
1615         spin_lock(&ino->i_lock);
1616         lo = nfsi->layout;
1617         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1618                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1619                 sleep = true;
1620         }
1621         spin_unlock(&ino->i_lock);
1622         return sleep;
1623 }
1624
1625 /*
1626  * Compare two layout segments for sorting into layout cache.
1627  * We want to preferentially return RW over RO layouts, so ensure those
1628  * are seen first.
1629  */
1630 static s64
1631 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1632            const struct pnfs_layout_range *l2)
1633 {
1634         s64 d;
1635
1636         /* high offset > low offset */
1637         d = l1->offset - l2->offset;
1638         if (d)
1639                 return d;
1640
1641         /* short length > long length */
1642         d = l2->length - l1->length;
1643         if (d)
1644                 return d;
1645
1646         /* read > read/write */
1647         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1648 }
1649
1650 static bool
1651 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1652                 const struct pnfs_layout_range *l2)
1653 {
1654         return pnfs_lseg_range_cmp(l1, l2) > 0;
1655 }
1656
1657 static bool
1658 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1659                 struct pnfs_layout_segment *old)
1660 {
1661         return false;
1662 }
1663
1664 void
1665 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1666                    struct pnfs_layout_segment *lseg,
1667                    bool (*is_after)(const struct pnfs_layout_range *,
1668                            const struct pnfs_layout_range *),
1669                    bool (*do_merge)(struct pnfs_layout_segment *,
1670                            struct pnfs_layout_segment *),
1671                    struct list_head *free_me)
1672 {
1673         struct pnfs_layout_segment *lp, *tmp;
1674
1675         dprintk("%s:Begin\n", __func__);
1676
1677         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1678                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1679                         continue;
1680                 if (do_merge(lseg, lp)) {
1681                         mark_lseg_invalid(lp, free_me);
1682                         continue;
1683                 }
1684                 if (is_after(&lseg->pls_range, &lp->pls_range))
1685                         continue;
1686                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1687                 dprintk("%s: inserted lseg %p "
1688                         "iomode %d offset %llu length %llu before "
1689                         "lp %p iomode %d offset %llu length %llu\n",
1690                         __func__, lseg, lseg->pls_range.iomode,
1691                         lseg->pls_range.offset, lseg->pls_range.length,
1692                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1693                         lp->pls_range.length);
1694                 goto out;
1695         }
1696         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1697         dprintk("%s: inserted lseg %p "
1698                 "iomode %d offset %llu length %llu at tail\n",
1699                 __func__, lseg, lseg->pls_range.iomode,
1700                 lseg->pls_range.offset, lseg->pls_range.length);
1701 out:
1702         pnfs_get_layout_hdr(lo);
1703
1704         dprintk("%s:Return\n", __func__);
1705 }
1706 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1707
1708 static void
1709 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1710                    struct pnfs_layout_segment *lseg,
1711                    struct list_head *free_me)
1712 {
1713         struct inode *inode = lo->plh_inode;
1714         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1715
1716         if (ld->add_lseg != NULL)
1717                 ld->add_lseg(lo, lseg, free_me);
1718         else
1719                 pnfs_generic_layout_insert_lseg(lo, lseg,
1720                                 pnfs_lseg_range_is_after,
1721                                 pnfs_lseg_no_merge,
1722                                 free_me);
1723 }
1724
1725 static struct pnfs_layout_hdr *
1726 alloc_init_layout_hdr(struct inode *ino,
1727                       struct nfs_open_context *ctx,
1728                       gfp_t gfp_flags)
1729 {
1730         struct pnfs_layout_hdr *lo;
1731
1732         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1733         if (!lo)
1734                 return NULL;
1735         refcount_set(&lo->plh_refcount, 1);
1736         INIT_LIST_HEAD(&lo->plh_layouts);
1737         INIT_LIST_HEAD(&lo->plh_segs);
1738         INIT_LIST_HEAD(&lo->plh_return_segs);
1739         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1740         lo->plh_inode = ino;
1741         lo->plh_lc_cred = get_cred(ctx->cred);
1742         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1743         return lo;
1744 }
1745
1746 static struct pnfs_layout_hdr *
1747 pnfs_find_alloc_layout(struct inode *ino,
1748                        struct nfs_open_context *ctx,
1749                        gfp_t gfp_flags)
1750         __releases(&ino->i_lock)
1751         __acquires(&ino->i_lock)
1752 {
1753         struct nfs_inode *nfsi = NFS_I(ino);
1754         struct pnfs_layout_hdr *new = NULL;
1755
1756         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1757
1758         if (nfsi->layout != NULL)
1759                 goto out_existing;
1760         spin_unlock(&ino->i_lock);
1761         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1762         spin_lock(&ino->i_lock);
1763
1764         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1765                 nfsi->layout = new;
1766                 return new;
1767         } else if (new != NULL)
1768                 pnfs_free_layout_hdr(new);
1769 out_existing:
1770         pnfs_get_layout_hdr(nfsi->layout);
1771         return nfsi->layout;
1772 }
1773
1774 /*
1775  * iomode matching rules:
1776  * iomode       lseg    strict match
1777  *                      iomode
1778  * -----        -----   ------ -----
1779  * ANY          READ    N/A    true
1780  * ANY          RW      N/A    true
1781  * RW           READ    N/A    false
1782  * RW           RW      N/A    true
1783  * READ         READ    N/A    true
1784  * READ         RW      true   false
1785  * READ         RW      false  true
1786  */
1787 static bool
1788 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1789                  const struct pnfs_layout_range *range,
1790                  bool strict_iomode)
1791 {
1792         struct pnfs_layout_range range1;
1793
1794         if ((range->iomode == IOMODE_RW &&
1795              ls_range->iomode != IOMODE_RW) ||
1796             (range->iomode != ls_range->iomode &&
1797              strict_iomode) ||
1798             !pnfs_lseg_range_intersecting(ls_range, range))
1799                 return false;
1800
1801         /* range1 covers only the first byte in the range */
1802         range1 = *range;
1803         range1.length = 1;
1804         return pnfs_lseg_range_contained(ls_range, &range1);
1805 }
1806
1807 /*
1808  * lookup range in layout
1809  */
1810 static struct pnfs_layout_segment *
1811 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1812                 struct pnfs_layout_range *range,
1813                 bool strict_iomode)
1814 {
1815         struct pnfs_layout_segment *lseg, *ret = NULL;
1816
1817         dprintk("%s:Begin\n", __func__);
1818
1819         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1820                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1821                     pnfs_lseg_range_match(&lseg->pls_range, range,
1822                                           strict_iomode)) {
1823                         ret = pnfs_get_lseg(lseg);
1824                         break;
1825                 }
1826         }
1827
1828         dprintk("%s:Return lseg %p ref %d\n",
1829                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1830         return ret;
1831 }
1832
1833 /*
1834  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1835  * to the MDS or over pNFS
1836  *
1837  * The nfs_inode read_io and write_io fields are cumulative counters reset
1838  * when there are no layout segments. Note that in pnfs_update_layout iomode
1839  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1840  * WRITE request.
1841  *
1842  * A return of true means use MDS I/O.
1843  *
1844  * From rfc 5661:
1845  * If a file's size is smaller than the file size threshold, data accesses
1846  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1847  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1848  * server.  If both file size and I/O size are provided, the client SHOULD
1849  * reach or exceed  both thresholds before sending its read or write
1850  * requests to the data server.
1851  */
1852 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1853                                      struct inode *ino, int iomode)
1854 {
1855         struct nfs4_threshold *t = ctx->mdsthreshold;
1856         struct nfs_inode *nfsi = NFS_I(ino);
1857         loff_t fsize = i_size_read(ino);
1858         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1859
1860         if (t == NULL)
1861                 return ret;
1862
1863         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1864                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1865
1866         switch (iomode) {
1867         case IOMODE_READ:
1868                 if (t->bm & THRESHOLD_RD) {
1869                         dprintk("%s fsize %llu\n", __func__, fsize);
1870                         size_set = true;
1871                         if (fsize < t->rd_sz)
1872                                 size = true;
1873                 }
1874                 if (t->bm & THRESHOLD_RD_IO) {
1875                         dprintk("%s nfsi->read_io %llu\n", __func__,
1876                                 nfsi->read_io);
1877                         io_set = true;
1878                         if (nfsi->read_io < t->rd_io_sz)
1879                                 io = true;
1880                 }
1881                 break;
1882         case IOMODE_RW:
1883                 if (t->bm & THRESHOLD_WR) {
1884                         dprintk("%s fsize %llu\n", __func__, fsize);
1885                         size_set = true;
1886                         if (fsize < t->wr_sz)
1887                                 size = true;
1888                 }
1889                 if (t->bm & THRESHOLD_WR_IO) {
1890                         dprintk("%s nfsi->write_io %llu\n", __func__,
1891                                 nfsi->write_io);
1892                         io_set = true;
1893                         if (nfsi->write_io < t->wr_io_sz)
1894                                 io = true;
1895                 }
1896                 break;
1897         }
1898         if (size_set && io_set) {
1899                 if (size && io)
1900                         ret = true;
1901         } else if (size || io)
1902                 ret = true;
1903
1904         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1905         return ret;
1906 }
1907
1908 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1909 {
1910         /*
1911          * send layoutcommit as it can hold up layoutreturn due to lseg
1912          * reference
1913          */
1914         pnfs_layoutcommit_inode(lo->plh_inode, false);
1915         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1916                                    nfs_wait_bit_killable,
1917                                    TASK_KILLABLE);
1918 }
1919
1920 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1921 {
1922         atomic_inc(&lo->plh_outstanding);
1923 }
1924
1925 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1926 {
1927         if (atomic_dec_and_test(&lo->plh_outstanding) &&
1928             test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
1929                 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
1930 }
1931
1932 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1933 {
1934         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1935 }
1936
1937 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1938 {
1939         unsigned long *bitlock = &lo->plh_flags;
1940
1941         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1942         smp_mb__after_atomic();
1943         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1944 }
1945
1946 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1947                                 struct nfs_server *server)
1948 {
1949         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1950                 struct nfs_client *clp = server->nfs_client;
1951
1952                 /* The lo must be on the clp list if there is any
1953                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1954                  */
1955                 spin_lock(&clp->cl_lock);
1956                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1957                 spin_unlock(&clp->cl_lock);
1958         }
1959 }
1960
1961 /*
1962  * Layout segment is retreived from the server if not cached.
1963  * The appropriate layout segment is referenced and returned to the caller.
1964  */
1965 struct pnfs_layout_segment *
1966 pnfs_update_layout(struct inode *ino,
1967                    struct nfs_open_context *ctx,
1968                    loff_t pos,
1969                    u64 count,
1970                    enum pnfs_iomode iomode,
1971                    bool strict_iomode,
1972                    gfp_t gfp_flags)
1973 {
1974         struct pnfs_layout_range arg = {
1975                 .iomode = iomode,
1976                 .offset = pos,
1977                 .length = count,
1978         };
1979         unsigned pg_offset;
1980         struct nfs_server *server = NFS_SERVER(ino);
1981         struct nfs_client *clp = server->nfs_client;
1982         struct pnfs_layout_hdr *lo = NULL;
1983         struct pnfs_layout_segment *lseg = NULL;
1984         struct nfs4_layoutget *lgp;
1985         nfs4_stateid stateid;
1986         long timeout = 0;
1987         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1988         bool first;
1989
1990         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1991                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1992                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1993                 goto out;
1994         }
1995
1996         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1997                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1998                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1999                 goto out;
2000         }
2001
2002 lookup_again:
2003         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2004         if (IS_ERR(lseg))
2005                 goto out;
2006         first = false;
2007         spin_lock(&ino->i_lock);
2008         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2009         if (lo == NULL) {
2010                 spin_unlock(&ino->i_lock);
2011                 lseg = ERR_PTR(-ENOMEM);
2012                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2013                                  PNFS_UPDATE_LAYOUT_NOMEM);
2014                 goto out;
2015         }
2016
2017         /* Do we even need to bother with this? */
2018         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2019                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2020                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2021                 dprintk("%s matches recall, use MDS\n", __func__);
2022                 goto out_unlock;
2023         }
2024
2025         /* if LAYOUTGET already failed once we don't try again */
2026         if (pnfs_layout_io_test_failed(lo, iomode)) {
2027                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2028                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2029                 goto out_unlock;
2030         }
2031
2032         /*
2033          * If the layout segment list is empty, but there are outstanding
2034          * layoutget calls, then they might be subject to a layoutrecall.
2035          */
2036         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2037             atomic_read(&lo->plh_outstanding) != 0) {
2038                 spin_unlock(&ino->i_lock);
2039                 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2040                                            TASK_KILLABLE));
2041                 if (IS_ERR(lseg))
2042                         goto out_put_layout_hdr;
2043                 pnfs_put_layout_hdr(lo);
2044                 goto lookup_again;
2045         }
2046
2047         /*
2048          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2049          * for LAYOUTRETURN.
2050          */
2051         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2052                 spin_unlock(&ino->i_lock);
2053                 dprintk("%s wait for layoutreturn\n", __func__);
2054                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2055                 if (!IS_ERR(lseg)) {
2056                         pnfs_put_layout_hdr(lo);
2057                         dprintk("%s retrying\n", __func__);
2058                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2059                                                  lseg,
2060                                                  PNFS_UPDATE_LAYOUT_RETRY);
2061                         goto lookup_again;
2062                 }
2063                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2064                                          PNFS_UPDATE_LAYOUT_RETURN);
2065                 goto out_put_layout_hdr;
2066         }
2067
2068         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2069         if (lseg) {
2070                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2071                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2072                 goto out_unlock;
2073         }
2074
2075         /*
2076          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2077          * stateid, or it has been invalidated, then we must use the open
2078          * stateid.
2079          */
2080         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2081                 int status;
2082
2083                 /*
2084                  * The first layoutget for the file. Need to serialize per
2085                  * RFC 5661 Errata 3208.
2086                  */
2087                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2088                                      &lo->plh_flags)) {
2089                         spin_unlock(&ino->i_lock);
2090                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2091                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2092                                                 TASK_KILLABLE));
2093                         if (IS_ERR(lseg))
2094                                 goto out_put_layout_hdr;
2095                         pnfs_put_layout_hdr(lo);
2096                         dprintk("%s retrying\n", __func__);
2097                         goto lookup_again;
2098                 }
2099
2100                 spin_unlock(&ino->i_lock);
2101                 first = true;
2102                 status = nfs4_select_rw_stateid(ctx->state,
2103                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2104                                         NULL, &stateid, NULL);
2105                 if (status != 0) {
2106                         lseg = ERR_PTR(status);
2107                         trace_pnfs_update_layout(ino, pos, count,
2108                                         iomode, lo, lseg,
2109                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2110                         nfs4_schedule_stateid_recovery(server, ctx->state);
2111                         pnfs_clear_first_layoutget(lo);
2112                         pnfs_put_layout_hdr(lo);
2113                         goto lookup_again;
2114                 }
2115                 spin_lock(&ino->i_lock);
2116         } else {
2117                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2118         }
2119
2120         if (pnfs_layoutgets_blocked(lo)) {
2121                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2122                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2123                 goto out_unlock;
2124         }
2125         nfs_layoutget_begin(lo);
2126         spin_unlock(&ino->i_lock);
2127
2128         _add_to_server_list(lo, server);
2129
2130         pg_offset = arg.offset & ~PAGE_MASK;
2131         if (pg_offset) {
2132                 arg.offset -= pg_offset;
2133                 arg.length += pg_offset;
2134         }
2135         if (arg.length != NFS4_MAX_UINT64)
2136                 arg.length = PAGE_ALIGN(arg.length);
2137
2138         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2139         if (!lgp) {
2140                 lseg = ERR_PTR(-ENOMEM);
2141                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2142                                          PNFS_UPDATE_LAYOUT_NOMEM);
2143                 nfs_layoutget_end(lo);
2144                 goto out_put_layout_hdr;
2145         }
2146
2147         lseg = nfs4_proc_layoutget(lgp, &timeout);
2148         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2149                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2150         nfs_layoutget_end(lo);
2151         if (IS_ERR(lseg)) {
2152                 switch(PTR_ERR(lseg)) {
2153                 case -EBUSY:
2154                         if (time_after(jiffies, giveup))
2155                                 lseg = NULL;
2156                         break;
2157                 case -ERECALLCONFLICT:
2158                 case -EAGAIN:
2159                         break;
2160                 case -ENODATA:
2161                         /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2162                         pnfs_layout_set_fail_bit(
2163                                 lo, pnfs_iomode_to_fail_bit(iomode));
2164                         lseg = NULL;
2165                         goto out_put_layout_hdr;
2166                 default:
2167                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2168                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2169                                 lseg = NULL;
2170                         }
2171                         goto out_put_layout_hdr;
2172                 }
2173                 if (lseg) {
2174                         if (first)
2175                                 pnfs_clear_first_layoutget(lo);
2176                         trace_pnfs_update_layout(ino, pos, count,
2177                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2178                         pnfs_put_layout_hdr(lo);
2179                         goto lookup_again;
2180                 }
2181         } else {
2182                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2183         }
2184
2185 out_put_layout_hdr:
2186         if (first)
2187                 pnfs_clear_first_layoutget(lo);
2188         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2189                                  PNFS_UPDATE_LAYOUT_EXIT);
2190         pnfs_put_layout_hdr(lo);
2191 out:
2192         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2193                         "(%s, offset: %llu, length: %llu)\n",
2194                         __func__, ino->i_sb->s_id,
2195                         (unsigned long long)NFS_FILEID(ino),
2196                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2197                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2198                         (unsigned long long)pos,
2199                         (unsigned long long)count);
2200         return lseg;
2201 out_unlock:
2202         spin_unlock(&ino->i_lock);
2203         goto out_put_layout_hdr;
2204 }
2205 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2206
2207 static bool
2208 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2209 {
2210         switch (range->iomode) {
2211         case IOMODE_READ:
2212         case IOMODE_RW:
2213                 break;
2214         default:
2215                 return false;
2216         }
2217         if (range->offset == NFS4_MAX_UINT64)
2218                 return false;
2219         if (range->length == 0)
2220                 return false;
2221         if (range->length != NFS4_MAX_UINT64 &&
2222             range->length > NFS4_MAX_UINT64 - range->offset)
2223                 return false;
2224         return true;
2225 }
2226
2227 static struct pnfs_layout_hdr *
2228 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2229 {
2230         struct pnfs_layout_hdr *lo;
2231
2232         spin_lock(&ino->i_lock);
2233         lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2234         if (!lo)
2235                 goto out_unlock;
2236         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2237                 goto out_unlock;
2238         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2239                 goto out_unlock;
2240         if (pnfs_layoutgets_blocked(lo))
2241                 goto out_unlock;
2242         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2243                 goto out_unlock;
2244         nfs_layoutget_begin(lo);
2245         spin_unlock(&ino->i_lock);
2246         _add_to_server_list(lo, NFS_SERVER(ino));
2247         return lo;
2248
2249 out_unlock:
2250         spin_unlock(&ino->i_lock);
2251         pnfs_put_layout_hdr(lo);
2252         return NULL;
2253 }
2254
2255 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2256                                      struct nfs_open_context *ctx)
2257 {
2258         struct inode *ino = data->dentry->d_inode;
2259         struct pnfs_layout_range rng = {
2260                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2261                           IOMODE_RW: IOMODE_READ,
2262                 .offset = 0,
2263                 .length = NFS4_MAX_UINT64,
2264         };
2265         struct nfs4_layoutget *lgp;
2266         struct pnfs_layout_hdr *lo;
2267
2268         /* Heuristic: don't send layoutget if we have cached data */
2269         if (rng.iomode == IOMODE_READ &&
2270            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2271                 return;
2272
2273         lo = _pnfs_grab_empty_layout(ino, ctx);
2274         if (!lo)
2275                 return;
2276         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2277                                              &rng, GFP_KERNEL);
2278         if (!lgp) {
2279                 pnfs_clear_first_layoutget(lo);
2280                 nfs_layoutget_end(lo);
2281                 pnfs_put_layout_hdr(lo);
2282                 return;
2283         }
2284         data->lgp = lgp;
2285         data->o_arg.lg_args = &lgp->args;
2286         data->o_res.lg_res = &lgp->res;
2287 }
2288
2289 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2290                                      struct nfs_open_context *ctx)
2291 {
2292         struct pnfs_layout_range rng = {
2293                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2294                           IOMODE_RW: IOMODE_READ,
2295                 .offset = 0,
2296                 .length = NFS4_MAX_UINT64,
2297         };
2298         struct nfs4_layoutget *lgp;
2299
2300         lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2301                                              &rng, GFP_KERNEL);
2302         if (!lgp)
2303                 return;
2304         data->lgp = lgp;
2305         data->o_arg.lg_args = &lgp->args;
2306         data->o_res.lg_res = &lgp->res;
2307 }
2308
2309 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2310                          struct nfs_open_context *ctx)
2311 {
2312         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2313
2314         if (!(pnfs_enabled_sb(server) &&
2315               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2316                 return;
2317         /* Could check on max_ops, but currently hardcoded high enough */
2318         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2319                 return;
2320         if (data->state)
2321                 _lgopen_prepare_attached(data, ctx);
2322         else
2323                 _lgopen_prepare_floating(data, ctx);
2324 }
2325
2326 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2327                        struct nfs_open_context *ctx)
2328 {
2329         struct pnfs_layout_hdr *lo;
2330         struct pnfs_layout_segment *lseg;
2331         struct nfs_server *srv = NFS_SERVER(ino);
2332         u32 iomode;
2333
2334         if (!lgp)
2335                 return;
2336         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2337         if (lgp->res.status) {
2338                 switch (lgp->res.status) {
2339                 default:
2340                         break;
2341                 /*
2342                  * Halt lgopen attempts if the server doesn't recognise
2343                  * the "current stateid" value, the layout type, or the
2344                  * layoutget operation as being valid.
2345                  * Also if it complains about too many ops in the compound
2346                  * or of the request/reply being too big.
2347                  */
2348                 case -NFS4ERR_BAD_STATEID:
2349                 case -NFS4ERR_NOTSUPP:
2350                 case -NFS4ERR_REP_TOO_BIG:
2351                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2352                 case -NFS4ERR_REQ_TOO_BIG:
2353                 case -NFS4ERR_TOO_MANY_OPS:
2354                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2355                         srv->caps &= ~NFS_CAP_LGOPEN;
2356                 }
2357                 return;
2358         }
2359         if (!lgp->args.inode) {
2360                 lo = _pnfs_grab_empty_layout(ino, ctx);
2361                 if (!lo)
2362                         return;
2363                 lgp->args.inode = ino;
2364         } else
2365                 lo = NFS_I(lgp->args.inode)->layout;
2366
2367         lseg = pnfs_layout_process(lgp);
2368         if (!IS_ERR(lseg)) {
2369                 iomode = lgp->args.range.iomode;
2370                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2371                 pnfs_put_lseg(lseg);
2372         }
2373 }
2374
2375 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2376 {
2377         if (lgp != NULL) {
2378                 struct inode *inode = lgp->args.inode;
2379                 if (inode) {
2380                         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2381                         pnfs_clear_first_layoutget(lo);
2382                         nfs_layoutget_end(lo);
2383                 }
2384                 pnfs_layoutget_free(lgp);
2385         }
2386 }
2387
2388 struct pnfs_layout_segment *
2389 pnfs_layout_process(struct nfs4_layoutget *lgp)
2390 {
2391         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2392         struct nfs4_layoutget_res *res = &lgp->res;
2393         struct pnfs_layout_segment *lseg;
2394         struct inode *ino = lo->plh_inode;
2395         LIST_HEAD(free_me);
2396
2397         if (!pnfs_sanity_check_layout_range(&res->range))
2398                 return ERR_PTR(-EINVAL);
2399
2400         /* Inject layout blob into I/O device driver */
2401         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2402         if (IS_ERR_OR_NULL(lseg)) {
2403                 if (!lseg)
2404                         lseg = ERR_PTR(-ENOMEM);
2405
2406                 dprintk("%s: Could not allocate layout: error %ld\n",
2407                        __func__, PTR_ERR(lseg));
2408                 return lseg;
2409         }
2410
2411         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2412
2413         spin_lock(&ino->i_lock);
2414         if (pnfs_layoutgets_blocked(lo)) {
2415                 dprintk("%s forget reply due to state\n", __func__);
2416                 goto out_forget;
2417         }
2418
2419         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2420             !pnfs_is_first_layoutget(lo))
2421                 goto out_forget;
2422
2423         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2424                 /* existing state ID, make sure the sequence number matches. */
2425                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2426                         if (!pnfs_layout_is_valid(lo))
2427                                 lo->plh_barrier = 0;
2428                         dprintk("%s forget reply due to sequence\n", __func__);
2429                         goto out_forget;
2430                 }
2431                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2432         } else if (pnfs_layout_is_valid(lo)) {
2433                 /*
2434                  * We got an entirely new state ID.  Mark all segments for the
2435                  * inode invalid, and retry the layoutget
2436                  */
2437                 struct pnfs_layout_range range = {
2438                         .iomode = IOMODE_ANY,
2439                         .length = NFS4_MAX_UINT64,
2440                 };
2441                 pnfs_set_plh_return_info(lo, IOMODE_ANY, 0);
2442                 pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2443                                                 &range, 0);
2444                 goto out_forget;
2445         } else {
2446                 /* We have a completely new layout */
2447                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2448         }
2449
2450         pnfs_get_lseg(lseg);
2451         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2452
2453
2454         if (res->return_on_close)
2455                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2456
2457         spin_unlock(&ino->i_lock);
2458         pnfs_free_lseg_list(&free_me);
2459         return lseg;
2460
2461 out_forget:
2462         spin_unlock(&ino->i_lock);
2463         lseg->pls_layout = lo;
2464         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2465         pnfs_free_lseg_list(&free_me);
2466         return ERR_PTR(-EAGAIN);
2467 }
2468
2469 /**
2470  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2471  * @lo: pointer to layout header
2472  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2473  * @return_range: describe layout segment ranges to be returned
2474  * @seq: stateid seqid to match
2475  *
2476  * This function is mainly intended for use by layoutrecall. It attempts
2477  * to free the layout segment immediately, or else to mark it for return
2478  * as soon as its reference count drops to zero.
2479  *
2480  * Returns
2481  * - 0: a layoutreturn needs to be scheduled.
2482  * - EBUSY: there are layout segment that are still in use.
2483  * - ENOENT: there are no layout segments that need to be returned.
2484  */
2485 int
2486 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2487                                 struct list_head *tmp_list,
2488                                 const struct pnfs_layout_range *return_range,
2489                                 u32 seq)
2490 {
2491         struct pnfs_layout_segment *lseg, *next;
2492         int remaining = 0;
2493
2494         dprintk("%s:Begin lo %p\n", __func__, lo);
2495
2496         assert_spin_locked(&lo->plh_inode->i_lock);
2497
2498         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2499                 tmp_list = &lo->plh_return_segs;
2500
2501         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2502                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2503                         dprintk("%s: marking lseg %p iomode %d "
2504                                 "offset %llu length %llu\n", __func__,
2505                                 lseg, lseg->pls_range.iomode,
2506                                 lseg->pls_range.offset,
2507                                 lseg->pls_range.length);
2508                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2509                                 tmp_list = &lo->plh_return_segs;
2510                         if (mark_lseg_invalid(lseg, tmp_list))
2511                                 continue;
2512                         remaining++;
2513                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2514                 }
2515
2516         if (remaining) {
2517                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2518                 return -EBUSY;
2519         }
2520
2521         if (!list_empty(&lo->plh_return_segs)) {
2522                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2523                 return 0;
2524         }
2525
2526         return -ENOENT;
2527 }
2528
2529 static void
2530 pnfs_mark_layout_for_return(struct inode *inode,
2531                             const struct pnfs_layout_range *range)
2532 {
2533         struct pnfs_layout_hdr *lo;
2534         bool return_now = false;
2535
2536         spin_lock(&inode->i_lock);
2537         lo = NFS_I(inode)->layout;
2538         if (!pnfs_layout_is_valid(lo)) {
2539                 spin_unlock(&inode->i_lock);
2540                 return;
2541         }
2542         pnfs_set_plh_return_info(lo, range->iomode, 0);
2543         /*
2544          * mark all matching lsegs so that we are sure to have no live
2545          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2546          * for how it works.
2547          */
2548         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2549                 const struct cred *cred;
2550                 nfs4_stateid stateid;
2551                 enum pnfs_iomode iomode;
2552
2553                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2554                 spin_unlock(&inode->i_lock);
2555                 if (return_now)
2556                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2557         } else {
2558                 spin_unlock(&inode->i_lock);
2559                 nfs_commit_inode(inode, 0);
2560         }
2561 }
2562
2563 void pnfs_error_mark_layout_for_return(struct inode *inode,
2564                                        struct pnfs_layout_segment *lseg)
2565 {
2566         struct pnfs_layout_range range = {
2567                 .iomode = lseg->pls_range.iomode,
2568                 .offset = 0,
2569                 .length = NFS4_MAX_UINT64,
2570         };
2571
2572         pnfs_mark_layout_for_return(inode, &range);
2573 }
2574 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2575
2576 static bool
2577 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2578 {
2579         return pnfs_layout_is_valid(lo) &&
2580                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2581                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2582 }
2583
2584 static struct pnfs_layout_segment *
2585 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2586                      const struct pnfs_layout_range *range,
2587                      enum pnfs_iomode iomode)
2588 {
2589         struct pnfs_layout_segment *lseg;
2590
2591         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2592                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2593                         continue;
2594                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2595                         continue;
2596                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2597                         continue;
2598                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2599                         return lseg;
2600         }
2601         return NULL;
2602 }
2603
2604 /* Find open file states whose mode matches that of the range */
2605 static bool
2606 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2607                                  const struct pnfs_layout_range *range)
2608 {
2609         struct list_head *head;
2610         struct nfs_open_context *ctx;
2611         fmode_t mode = 0;
2612
2613         if (!pnfs_layout_can_be_returned(lo) ||
2614             !pnfs_find_first_lseg(lo, range, range->iomode))
2615                 return false;
2616
2617         head = &NFS_I(lo->plh_inode)->open_files;
2618         list_for_each_entry_rcu(ctx, head, list) {
2619                 if (ctx->state)
2620                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2621         }
2622
2623         switch (range->iomode) {
2624         default:
2625                 break;
2626         case IOMODE_READ:
2627                 mode &= ~FMODE_WRITE;
2628                 break;
2629         case IOMODE_RW:
2630                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2631                         mode &= ~FMODE_READ;
2632         }
2633         return mode == 0;
2634 }
2635
2636 static int
2637 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2638 {
2639         const struct pnfs_layout_range *range = data;
2640         struct pnfs_layout_hdr *lo;
2641         struct inode *inode;
2642 restart:
2643         rcu_read_lock();
2644         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2645                 if (!pnfs_layout_can_be_returned(lo) ||
2646                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2647                         continue;
2648                 inode = lo->plh_inode;
2649                 spin_lock(&inode->i_lock);
2650                 if (!pnfs_should_return_unused_layout(lo, range)) {
2651                         spin_unlock(&inode->i_lock);
2652                         continue;
2653                 }
2654                 spin_unlock(&inode->i_lock);
2655                 inode = pnfs_grab_inode_layout_hdr(lo);
2656                 if (!inode)
2657                         continue;
2658                 rcu_read_unlock();
2659                 pnfs_mark_layout_for_return(inode, range);
2660                 iput(inode);
2661                 cond_resched();
2662                 goto restart;
2663         }
2664         rcu_read_unlock();
2665         return 0;
2666 }
2667
2668 void
2669 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2670                                  enum pnfs_iomode iomode)
2671 {
2672         struct pnfs_layout_range range = {
2673                 .iomode = iomode,
2674                 .offset = 0,
2675                 .length = NFS4_MAX_UINT64,
2676         };
2677
2678         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2679                         &range);
2680 }
2681
2682 void
2683 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2684 {
2685         if (pgio->pg_lseg == NULL ||
2686             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2687                 return;
2688         pnfs_put_lseg(pgio->pg_lseg);
2689         pgio->pg_lseg = NULL;
2690 }
2691 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2692
2693 /*
2694  * Check for any intersection between the request and the pgio->pg_lseg,
2695  * and if none, put this pgio->pg_lseg away.
2696  */
2697 void
2698 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2699 {
2700         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2701                 pnfs_put_lseg(pgio->pg_lseg);
2702                 pgio->pg_lseg = NULL;
2703         }
2704 }
2705 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2706
2707 void
2708 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2709 {
2710         u64 rd_size = req->wb_bytes;
2711
2712         pnfs_generic_pg_check_layout(pgio);
2713         pnfs_generic_pg_check_range(pgio, req);
2714         if (pgio->pg_lseg == NULL) {
2715                 if (pgio->pg_dreq == NULL)
2716                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2717                 else
2718                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2719
2720                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2721                                                    nfs_req_openctx(req),
2722                                                    req_offset(req),
2723                                                    rd_size,
2724                                                    IOMODE_READ,
2725                                                    false,
2726                                                    GFP_KERNEL);
2727                 if (IS_ERR(pgio->pg_lseg)) {
2728                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2729                         pgio->pg_lseg = NULL;
2730                         return;
2731                 }
2732         }
2733         /* If no lseg, fall back to read through mds */
2734         if (pgio->pg_lseg == NULL)
2735                 nfs_pageio_reset_read_mds(pgio);
2736
2737 }
2738 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2739
2740 void
2741 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2742                            struct nfs_page *req, u64 wb_size)
2743 {
2744         pnfs_generic_pg_check_layout(pgio);
2745         pnfs_generic_pg_check_range(pgio, req);
2746         if (pgio->pg_lseg == NULL) {
2747                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2748                                                    nfs_req_openctx(req),
2749                                                    req_offset(req),
2750                                                    wb_size,
2751                                                    IOMODE_RW,
2752                                                    false,
2753                                                    GFP_KERNEL);
2754                 if (IS_ERR(pgio->pg_lseg)) {
2755                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2756                         pgio->pg_lseg = NULL;
2757                         return;
2758                 }
2759         }
2760         /* If no lseg, fall back to write through mds */
2761         if (pgio->pg_lseg == NULL)
2762                 nfs_pageio_reset_write_mds(pgio);
2763 }
2764 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2765
2766 void
2767 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2768 {
2769         if (desc->pg_lseg) {
2770                 pnfs_put_lseg(desc->pg_lseg);
2771                 desc->pg_lseg = NULL;
2772         }
2773 }
2774 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2775
2776 /*
2777  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2778  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2779  */
2780 size_t
2781 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2782                      struct nfs_page *prev, struct nfs_page *req)
2783 {
2784         unsigned int size;
2785         u64 seg_end, req_start, seg_left;
2786
2787         size = nfs_generic_pg_test(pgio, prev, req);
2788         if (!size)
2789                 return 0;
2790
2791         /*
2792          * 'size' contains the number of bytes left in the current page (up
2793          * to the original size asked for in @req->wb_bytes).
2794          *
2795          * Calculate how many bytes are left in the layout segment
2796          * and if there are less bytes than 'size', return that instead.
2797          *
2798          * Please also note that 'end_offset' is actually the offset of the
2799          * first byte that lies outside the pnfs_layout_range. FIXME?
2800          *
2801          */
2802         if (pgio->pg_lseg) {
2803                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2804                                      pgio->pg_lseg->pls_range.length);
2805                 req_start = req_offset(req);
2806
2807                 /* start of request is past the last byte of this segment */
2808                 if (req_start >= seg_end)
2809                         return 0;
2810
2811                 /* adjust 'size' iff there are fewer bytes left in the
2812                  * segment than what nfs_generic_pg_test returned */
2813                 seg_left = seg_end - req_start;
2814                 if (seg_left < size)
2815                         size = (unsigned int)seg_left;
2816         }
2817
2818         return size;
2819 }
2820 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2821
2822 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2823 {
2824         struct nfs_pageio_descriptor pgio;
2825
2826         /* Resend all requests through the MDS */
2827         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2828                               hdr->completion_ops);
2829         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2830         return nfs_pageio_resend(&pgio, hdr);
2831 }
2832 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2833
2834 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2835 {
2836
2837         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2838         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2839             PNFS_LAYOUTRET_ON_ERROR) {
2840                 pnfs_return_layout(hdr->inode);
2841         }
2842         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2843                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2844 }
2845
2846 /*
2847  * Called by non rpc-based layout drivers
2848  */
2849 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2850 {
2851         if (likely(!hdr->pnfs_error)) {
2852                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2853                                 hdr->mds_offset + hdr->res.count);
2854                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2855         }
2856         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2857         if (unlikely(hdr->pnfs_error))
2858                 pnfs_ld_handle_write_error(hdr);
2859         hdr->mds_ops->rpc_release(hdr);
2860 }
2861 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2862
2863 static void
2864 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2865                 struct nfs_pgio_header *hdr)
2866 {
2867         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2868
2869         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2870                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2871                 nfs_pageio_reset_write_mds(desc);
2872                 mirror->pg_recoalesce = 1;
2873         }
2874         hdr->completion_ops->completion(hdr);
2875 }
2876
2877 static enum pnfs_try_status
2878 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2879                         const struct rpc_call_ops *call_ops,
2880                         struct pnfs_layout_segment *lseg,
2881                         int how)
2882 {
2883         struct inode *inode = hdr->inode;
2884         enum pnfs_try_status trypnfs;
2885         struct nfs_server *nfss = NFS_SERVER(inode);
2886
2887         hdr->mds_ops = call_ops;
2888
2889         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2890                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2891         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2892         if (trypnfs != PNFS_NOT_ATTEMPTED)
2893                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2894         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2895         return trypnfs;
2896 }
2897
2898 static void
2899 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2900               struct nfs_pgio_header *hdr, int how)
2901 {
2902         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2903         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2904         enum pnfs_try_status trypnfs;
2905
2906         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2907         switch (trypnfs) {
2908         case PNFS_NOT_ATTEMPTED:
2909                 pnfs_write_through_mds(desc, hdr);
2910         case PNFS_ATTEMPTED:
2911                 break;
2912         case PNFS_TRY_AGAIN:
2913                 /* cleanup hdr and prepare to redo pnfs */
2914                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2915                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2916                         list_splice_init(&hdr->pages, &mirror->pg_list);
2917                         mirror->pg_recoalesce = 1;
2918                 }
2919                 hdr->mds_ops->rpc_release(hdr);
2920         }
2921 }
2922
2923 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2924 {
2925         pnfs_put_lseg(hdr->lseg);
2926         nfs_pgio_header_free(hdr);
2927 }
2928
2929 int
2930 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2931 {
2932         struct nfs_pgio_header *hdr;
2933         int ret;
2934
2935         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2936         if (!hdr) {
2937                 desc->pg_error = -ENOMEM;
2938                 return desc->pg_error;
2939         }
2940         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2941
2942         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2943         ret = nfs_generic_pgio(desc, hdr);
2944         if (!ret)
2945                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2946
2947         return ret;
2948 }
2949 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2950
2951 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2952 {
2953         struct nfs_pageio_descriptor pgio;
2954
2955         /* Resend all requests through the MDS */
2956         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2957         return nfs_pageio_resend(&pgio, hdr);
2958 }
2959 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2960
2961 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2962 {
2963         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2964         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2965             PNFS_LAYOUTRET_ON_ERROR) {
2966                 pnfs_return_layout(hdr->inode);
2967         }
2968         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2969                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2970 }
2971
2972 /*
2973  * Called by non rpc-based layout drivers
2974  */
2975 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2976 {
2977         if (likely(!hdr->pnfs_error))
2978                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2979         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2980         if (unlikely(hdr->pnfs_error))
2981                 pnfs_ld_handle_read_error(hdr);
2982         hdr->mds_ops->rpc_release(hdr);
2983 }
2984 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2985
2986 static void
2987 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2988                 struct nfs_pgio_header *hdr)
2989 {
2990         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2991
2992         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2993                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2994                 nfs_pageio_reset_read_mds(desc);
2995                 mirror->pg_recoalesce = 1;
2996         }
2997         hdr->completion_ops->completion(hdr);
2998 }
2999
3000 /*
3001  * Call the appropriate parallel I/O subsystem read function.
3002  */
3003 static enum pnfs_try_status
3004 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3005                        const struct rpc_call_ops *call_ops,
3006                        struct pnfs_layout_segment *lseg)
3007 {
3008         struct inode *inode = hdr->inode;
3009         struct nfs_server *nfss = NFS_SERVER(inode);
3010         enum pnfs_try_status trypnfs;
3011
3012         hdr->mds_ops = call_ops;
3013
3014         dprintk("%s: Reading ino:%lu %u@%llu\n",
3015                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3016
3017         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3018         if (trypnfs != PNFS_NOT_ATTEMPTED)
3019                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3020         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3021         return trypnfs;
3022 }
3023
3024 /* Resend all requests through pnfs. */
3025 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3026                            unsigned int mirror_idx)
3027 {
3028         struct nfs_pageio_descriptor pgio;
3029
3030         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3031                 /* Prevent deadlocks with layoutreturn! */
3032                 pnfs_put_lseg(hdr->lseg);
3033                 hdr->lseg = NULL;
3034
3035                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3036                                         hdr->completion_ops);
3037                 pgio.pg_mirror_idx = mirror_idx;
3038                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3039         }
3040 }
3041 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3042
3043 static void
3044 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3045 {
3046         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3047         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3048         enum pnfs_try_status trypnfs;
3049
3050         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3051         switch (trypnfs) {
3052         case PNFS_NOT_ATTEMPTED:
3053                 pnfs_read_through_mds(desc, hdr);
3054         case PNFS_ATTEMPTED:
3055                 break;
3056         case PNFS_TRY_AGAIN:
3057                 /* cleanup hdr and prepare to redo pnfs */
3058                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3059                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3060                         list_splice_init(&hdr->pages, &mirror->pg_list);
3061                         mirror->pg_recoalesce = 1;
3062                 }
3063                 hdr->mds_ops->rpc_release(hdr);
3064         }
3065 }
3066
3067 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3068 {
3069         pnfs_put_lseg(hdr->lseg);
3070         nfs_pgio_header_free(hdr);
3071 }
3072
3073 int
3074 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3075 {
3076         struct nfs_pgio_header *hdr;
3077         int ret;
3078
3079         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3080         if (!hdr) {
3081                 desc->pg_error = -ENOMEM;
3082                 return desc->pg_error;
3083         }
3084         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3085         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3086         ret = nfs_generic_pgio(desc, hdr);
3087         if (!ret)
3088                 pnfs_do_read(desc, hdr);
3089         return ret;
3090 }
3091 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3092
3093 static void pnfs_clear_layoutcommitting(struct inode *inode)
3094 {
3095         unsigned long *bitlock = &NFS_I(inode)->flags;
3096
3097         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3098         smp_mb__after_atomic();
3099         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3100 }
3101
3102 /*
3103  * There can be multiple RW segments.
3104  */
3105 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3106 {
3107         struct pnfs_layout_segment *lseg;
3108
3109         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3110                 if (lseg->pls_range.iomode == IOMODE_RW &&
3111                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3112                         list_add(&lseg->pls_lc_list, listp);
3113         }
3114 }
3115
3116 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3117 {
3118         struct pnfs_layout_segment *lseg, *tmp;
3119
3120         /* Matched by references in pnfs_set_layoutcommit */
3121         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3122                 list_del_init(&lseg->pls_lc_list);
3123                 pnfs_put_lseg(lseg);
3124         }
3125
3126         pnfs_clear_layoutcommitting(inode);
3127 }
3128
3129 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3130 {
3131         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3132 }
3133 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3134
3135 void
3136 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3137                 loff_t end_pos)
3138 {
3139         struct nfs_inode *nfsi = NFS_I(inode);
3140         bool mark_as_dirty = false;
3141
3142         spin_lock(&inode->i_lock);
3143         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3144                 nfsi->layout->plh_lwb = end_pos;
3145                 mark_as_dirty = true;
3146                 dprintk("%s: Set layoutcommit for inode %lu ",
3147                         __func__, inode->i_ino);
3148         } else if (end_pos > nfsi->layout->plh_lwb)
3149                 nfsi->layout->plh_lwb = end_pos;
3150         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3151                 /* references matched in nfs4_layoutcommit_release */
3152                 pnfs_get_lseg(lseg);
3153         }
3154         spin_unlock(&inode->i_lock);
3155         dprintk("%s: lseg %p end_pos %llu\n",
3156                 __func__, lseg, nfsi->layout->plh_lwb);
3157
3158         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3159          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3160         if (mark_as_dirty)
3161                 mark_inode_dirty_sync(inode);
3162 }
3163 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3164
3165 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3166 {
3167         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3168
3169         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3170                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3171         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3172 }
3173
3174 /*
3175  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3176  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3177  * data to disk to allow the server to recover the data if it crashes.
3178  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3179  * is off, and a COMMIT is sent to a data server, or
3180  * if WRITEs to a data server return NFS_DATA_SYNC.
3181  */
3182 int
3183 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3184 {
3185         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3186         struct nfs4_layoutcommit_data *data;
3187         struct nfs_inode *nfsi = NFS_I(inode);
3188         loff_t end_pos;
3189         int status;
3190
3191         if (!pnfs_layoutcommit_outstanding(inode))
3192                 return 0;
3193
3194         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3195
3196         status = -EAGAIN;
3197         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3198                 if (!sync)
3199                         goto out;
3200                 status = wait_on_bit_lock_action(&nfsi->flags,
3201                                 NFS_INO_LAYOUTCOMMITTING,
3202                                 nfs_wait_bit_killable,
3203                                 TASK_KILLABLE);
3204                 if (status)
3205                         goto out;
3206         }
3207
3208         status = -ENOMEM;
3209         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3210         data = kzalloc(sizeof(*data), GFP_NOFS);
3211         if (!data)
3212                 goto clear_layoutcommitting;
3213
3214         status = 0;
3215         spin_lock(&inode->i_lock);
3216         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3217                 goto out_unlock;
3218
3219         INIT_LIST_HEAD(&data->lseg_list);
3220         pnfs_list_write_lseg(inode, &data->lseg_list);
3221
3222         end_pos = nfsi->layout->plh_lwb;
3223
3224         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3225         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3226         spin_unlock(&inode->i_lock);
3227
3228         data->args.inode = inode;
3229         nfs_fattr_init(&data->fattr);
3230         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3231         data->res.fattr = &data->fattr;
3232         if (end_pos != 0)
3233                 data->args.lastbytewritten = end_pos - 1;
3234         else
3235                 data->args.lastbytewritten = U64_MAX;
3236         data->res.server = NFS_SERVER(inode);
3237
3238         if (ld->prepare_layoutcommit) {
3239                 status = ld->prepare_layoutcommit(&data->args);
3240                 if (status) {
3241                         put_cred(data->cred);
3242                         spin_lock(&inode->i_lock);
3243                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3244                         if (end_pos > nfsi->layout->plh_lwb)
3245                                 nfsi->layout->plh_lwb = end_pos;
3246                         goto out_unlock;
3247                 }
3248         }
3249
3250
3251         status = nfs4_proc_layoutcommit(data, sync);
3252 out:
3253         if (status)
3254                 mark_inode_dirty_sync(inode);
3255         dprintk("<-- %s status %d\n", __func__, status);
3256         return status;
3257 out_unlock:
3258         spin_unlock(&inode->i_lock);
3259         kfree(data);
3260 clear_layoutcommitting:
3261         pnfs_clear_layoutcommitting(inode);
3262         goto out;
3263 }
3264 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3265
3266 int
3267 pnfs_generic_sync(struct inode *inode, bool datasync)
3268 {
3269         return pnfs_layoutcommit_inode(inode, true);
3270 }
3271 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3272
3273 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3274 {
3275         struct nfs4_threshold *thp;
3276
3277         thp = kzalloc(sizeof(*thp), GFP_NOFS);
3278         if (!thp) {
3279                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3280                 return NULL;
3281         }
3282         return thp;
3283 }
3284
3285 #if IS_ENABLED(CONFIG_NFS_V4_2)
3286 int
3287 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3288 {
3289         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3290         struct nfs_server *server = NFS_SERVER(inode);
3291         struct nfs_inode *nfsi = NFS_I(inode);
3292         struct nfs42_layoutstat_data *data;
3293         struct pnfs_layout_hdr *hdr;
3294         int status = 0;
3295
3296         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3297                 goto out;
3298
3299         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3300                 goto out;
3301
3302         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3303                 goto out;
3304
3305         spin_lock(&inode->i_lock);
3306         if (!NFS_I(inode)->layout) {
3307                 spin_unlock(&inode->i_lock);
3308                 goto out_clear_layoutstats;
3309         }
3310         hdr = NFS_I(inode)->layout;
3311         pnfs_get_layout_hdr(hdr);
3312         spin_unlock(&inode->i_lock);
3313
3314         data = kzalloc(sizeof(*data), gfp_flags);
3315         if (!data) {
3316                 status = -ENOMEM;
3317                 goto out_put;
3318         }
3319
3320         data->args.fh = NFS_FH(inode);
3321         data->args.inode = inode;
3322         status = ld->prepare_layoutstats(&data->args);
3323         if (status)
3324                 goto out_free;
3325
3326         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3327
3328 out:
3329         dprintk("%s returns %d\n", __func__, status);
3330         return status;
3331
3332 out_free:
3333         kfree(data);
3334 out_put:
3335         pnfs_put_layout_hdr(hdr);
3336 out_clear_layoutstats:
3337         smp_mb__before_atomic();
3338         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3339         smp_mb__after_atomic();
3340         goto out;
3341 }
3342 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3343 #endif
3344
3345 unsigned int layoutstats_timer;
3346 module_param(layoutstats_timer, uint, 0644);
3347 EXPORT_SYMBOL_GPL(layoutstats_timer);