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