fs/nfs/pnfs_nfs.c

   1 /*
   2  * Common NFS I/O  operations for the pnfs file based
   3  * layout drivers.
   4  *
   5  * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
   6  *
   7  * Tom Haynes <loghyr@primarydata.com>
   8  */
   9
  10 #include <linux/nfs_fs.h>
  11 #include <linux/nfs_page.h>
  12 #include <linux/sunrpc/addr.h>
  13 #include <linux/module.h>
  14
  15 #include "nfs4session.h"
  16 #include "internal.h"
  17 #include "pnfs.h"
  18
  19 #define NFSDBG_FACILITY         NFSDBG_PNFS
  20
  21 void pnfs_generic_rw_release(void *data)
  22 {
  23         struct nfs_pgio_header *hdr = data;
  24
  25         nfs_put_client(hdr->ds_clp);
  26         hdr->mds_ops->rpc_release(data);
  27 }
  28 EXPORT_SYMBOL_GPL(pnfs_generic_rw_release);
  29
  30 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
  31 void pnfs_generic_prepare_to_resend_writes(struct nfs_commit_data *data)
  32 {
  33         struct nfs_page *first = nfs_list_entry(data->pages.next);
  34
  35         data->task.tk_status = 0;
  36         memcpy(&data->verf.verifier, &first->wb_verf,
  37                sizeof(data->verf.verifier));
  38         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
  39 }
  40 EXPORT_SYMBOL_GPL(pnfs_generic_prepare_to_resend_writes);
  41
  42 void pnfs_generic_write_commit_done(struct rpc_task *task, void *data)
  43 {
  44         struct nfs_commit_data *wdata = data;
  45
  46         /* Note this may cause RPC to be resent */
  47         wdata->mds_ops->rpc_call_done(task, data);
  48 }
  49 EXPORT_SYMBOL_GPL(pnfs_generic_write_commit_done);
  50
  51 void pnfs_generic_commit_release(void *calldata)
  52 {
  53         struct nfs_commit_data *data = calldata;
  54
  55         data->completion_ops->completion(data);
  56         pnfs_put_lseg(data->lseg);
  57         nfs_put_client(data->ds_clp);
  58         nfs_commitdata_release(data);
  59 }
  60 EXPORT_SYMBOL_GPL(pnfs_generic_commit_release);
  61
  62 /* The generic layer is about to remove the req from the commit list.
  63  * If this will make the bucket empty, it will need to put the lseg reference.
  64  * Note this must be called holding i_lock
  65  */
  66 void
  67 pnfs_generic_clear_request_commit(struct nfs_page *req,
  68                                   struct nfs_commit_info *cinfo)
  69 {
  70         struct pnfs_layout_segment *freeme = NULL;
  71
  72         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
  73                 goto out;
  74         cinfo->ds->nwritten--;
  75         if (list_is_singular(&req->wb_list)) {
  76                 struct pnfs_commit_bucket *bucket;
  77
  78                 bucket = list_first_entry(&req->wb_list,
  79                                           struct pnfs_commit_bucket,
  80                                           written);
  81                 freeme = bucket->wlseg;
  82                 bucket->wlseg = NULL;
  83         }
  84 out:
  85         nfs_request_remove_commit_list(req, cinfo);
  86         pnfs_put_lseg_locked(freeme);
  87 }
  88 EXPORT_SYMBOL_GPL(pnfs_generic_clear_request_commit);
  89
  90 static int
  91 pnfs_generic_transfer_commit_list(struct list_head *src, struct list_head *dst,
  92                                   struct nfs_commit_info *cinfo, int max)
  93 {
  94         struct nfs_page *req, *tmp;
  95         int ret = 0;
  96
  97         list_for_each_entry_safe(req, tmp, src, wb_list) {
  98                 if (!nfs_lock_request(req))
  99                         continue;
 100                 kref_get(&req->wb_kref);
 101                 if (cond_resched_lock(&cinfo->inode->i_lock))
 102                         list_safe_reset_next(req, tmp, wb_list);
 103                 nfs_request_remove_commit_list(req, cinfo);
 104                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
 105                 nfs_list_add_request(req, dst);
 106                 ret++;
 107                 if ((ret == max) && !cinfo->dreq)
 108                         break;
 109         }
 110         return ret;
 111 }
 112
 113 static int
 114 pnfs_generic_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
 115                                  struct nfs_commit_info *cinfo,
 116                                  int max)
 117 {
 118         struct list_head *src = &bucket->written;
 119         struct list_head *dst = &bucket->committing;
 120         int ret;
 121
 122         lockdep_assert_held(&cinfo->inode->i_lock);
 123         ret = pnfs_generic_transfer_commit_list(src, dst, cinfo, max);
 124         if (ret) {
 125                 cinfo->ds->nwritten -= ret;
 126                 cinfo->ds->ncommitting += ret;
 127                 if (bucket->clseg == NULL)
 128                         bucket->clseg = pnfs_get_lseg(bucket->wlseg);
 129                 if (list_empty(src)) {
 130                         pnfs_put_lseg_locked(bucket->wlseg);
 131                         bucket->wlseg = NULL;
 132                 }
 133         }
 134         return ret;
 135 }
 136
 137 /* Move reqs from written to committing lists, returning count
 138  * of number moved.
 139  */
 140 int pnfs_generic_scan_commit_lists(struct nfs_commit_info *cinfo,
 141                                    int max)
 142 {
 143         int i, rv = 0, cnt;
 144
 145         lockdep_assert_held(&cinfo->inode->i_lock);
 146         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
 147                 cnt = pnfs_generic_scan_ds_commit_list(&cinfo->ds->buckets[i],
 148                                                        cinfo, max);
 149                 max -= cnt;
 150                 rv += cnt;
 151         }
 152         return rv;
 153 }
 154 EXPORT_SYMBOL_GPL(pnfs_generic_scan_commit_lists);
 155
 156 /* Pull everything off the committing lists and dump into @dst.  */
 157 void pnfs_generic_recover_commit_reqs(struct list_head *dst,
 158                                       struct nfs_commit_info *cinfo)
 159 {
 160         struct pnfs_commit_bucket *b;
 161         struct pnfs_layout_segment *freeme;
 162         int nwritten;
 163         int i;
 164
 165         lockdep_assert_held(&cinfo->inode->i_lock);
 166 restart:
 167         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
 168                 nwritten = pnfs_generic_transfer_commit_list(&b->written,
 169                                 dst, cinfo, 0);
 170                 if (!nwritten)
 171                         continue;
 172                 cinfo->ds->nwritten -= nwritten;
 173                 if (list_empty(&b->written)) {
 174                         freeme = b->wlseg;
 175                         b->wlseg = NULL;
 176                         spin_unlock(&cinfo->inode->i_lock);
 177                         pnfs_put_lseg(freeme);
 178                         spin_lock(&cinfo->inode->i_lock);
 179                         goto restart;
 180                 }
 181         }
 182 }
 183 EXPORT_SYMBOL_GPL(pnfs_generic_recover_commit_reqs);
 184
 185 static void pnfs_generic_retry_commit(struct nfs_commit_info *cinfo, int idx)
 186 {
 187         struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
 188         struct pnfs_commit_bucket *bucket;
 189         struct pnfs_layout_segment *freeme;
 190         struct list_head *pos;
 191         LIST_HEAD(pages);
 192         int i;
 193
 194         spin_lock(&cinfo->inode->i_lock);
 195         for (i = idx; i < fl_cinfo->nbuckets; i++) {
 196                 bucket = &fl_cinfo->buckets[i];
 197                 if (list_empty(&bucket->committing))
 198                         continue;
 199                 freeme = bucket->clseg;
 200                 bucket->clseg = NULL;
 201                 list_for_each(pos, &bucket->committing)
 202                         cinfo->ds->ncommitting--;
 203                 list_splice_init(&bucket->committing, &pages);
 204                 spin_unlock(&cinfo->inode->i_lock);
 205                 nfs_retry_commit(&pages, freeme, cinfo, i);
 206                 pnfs_put_lseg(freeme);
 207                 spin_lock(&cinfo->inode->i_lock);
 208         }
 209         spin_unlock(&cinfo->inode->i_lock);
 210 }
 211
 212 static unsigned int
 213 pnfs_generic_alloc_ds_commits(struct nfs_commit_info *cinfo,
 214                               struct list_head *list)
 215 {
 216         struct pnfs_ds_commit_info *fl_cinfo;
 217         struct pnfs_commit_bucket *bucket;
 218         struct nfs_commit_data *data;
 219         int i;
 220         unsigned int nreq = 0;
 221
 222         fl_cinfo = cinfo->ds;
 223         bucket = fl_cinfo->buckets;
 224         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
 225                 if (list_empty(&bucket->committing))
 226                         continue;
 227                 data = nfs_commitdata_alloc(false);
 228                 if (!data)
 229                         break;
 230                 data->ds_commit_index = i;
 231                 list_add(&data->pages, list);
 232                 nreq++;
 233         }
 234
 235         /* Clean up on error */
 236         pnfs_generic_retry_commit(cinfo, i);
 237         return nreq;
 238 }
 239
 240 static inline
 241 void pnfs_fetch_commit_bucket_list(struct list_head *pages,
 242                 struct nfs_commit_data *data,
 243                 struct nfs_commit_info *cinfo)
 244 {
 245         struct pnfs_commit_bucket *bucket;
 246         struct list_head *pos;
 247
 248         bucket = &cinfo->ds->buckets[data->ds_commit_index];
 249         spin_lock(&cinfo->inode->i_lock);
 250         list_for_each(pos, &bucket->committing)
 251                 cinfo->ds->ncommitting--;
 252         list_splice_init(&bucket->committing, pages);
 253         data->lseg = bucket->clseg;
 254         bucket->clseg = NULL;
 255         spin_unlock(&cinfo->inode->i_lock);
 256
 257 }
 258
 259 /* Helper function for pnfs_generic_commit_pagelist to catch an empty
 260  * page list. This can happen when two commits race.
 261  *
 262  * This must be called instead of nfs_init_commit - call one or the other, but
 263  * not both!
 264  */
 265 static bool
 266 pnfs_generic_commit_cancel_empty_pagelist(struct list_head *pages,
 267                                           struct nfs_commit_data *data,
 268                                           struct nfs_commit_info *cinfo)
 269 {
 270         if (list_empty(pages)) {
 271                 if (atomic_dec_and_test(&cinfo->mds->rpcs_out))
 272                         wake_up_atomic_t(&cinfo->mds->rpcs_out);
 273                 /* don't call nfs_commitdata_release - it tries to put
 274                  * the open_context which is not acquired until nfs_init_commit
 275                  * which has not been called on @data */
 276                 WARN_ON_ONCE(data->context);
 277                 nfs_commit_free(data);
 278                 return true;
 279         }
 280
 281         return false;
 282 }
 283
 284 /* This follows nfs_commit_list pretty closely */
 285 int
 286 pnfs_generic_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 287                              int how, struct nfs_commit_info *cinfo,
 288                              int (*initiate_commit)(struct nfs_commit_data *data,
 289                                                     int how))
 290 {
 291         struct nfs_commit_data *data, *tmp;
 292         LIST_HEAD(list);
 293         unsigned int nreq = 0;
 294
 295         if (!list_empty(mds_pages)) {
 296                 data = nfs_commitdata_alloc(true);
 297                 data->ds_commit_index = -1;
 298                 list_add(&data->pages, &list);
 299                 nreq++;
 300         }
 301
 302         nreq += pnfs_generic_alloc_ds_commits(cinfo, &list);
 303
 304         if (nreq == 0)
 305                 goto out;
 306
 307         atomic_add(nreq, &cinfo->mds->rpcs_out);
 308
 309         list_for_each_entry_safe(data, tmp, &list, pages) {
 310                 list_del_init(&data->pages);
 311                 if (data->ds_commit_index < 0) {
 312                         /* another commit raced with us */
 313                         if (pnfs_generic_commit_cancel_empty_pagelist(mds_pages,
 314                                 data, cinfo))
 315                                 continue;
 316
 317                         nfs_init_commit(data, mds_pages, NULL, cinfo);
 318                         nfs_initiate_commit(NFS_CLIENT(inode), data,
 319                                             NFS_PROTO(data->inode),
 320                                             data->mds_ops, how, 0);
 321                 } else {
 322                         LIST_HEAD(pages);
 323
 324                         pnfs_fetch_commit_bucket_list(&pages, data, cinfo);
 325
 326                         /* another commit raced with us */
 327                         if (pnfs_generic_commit_cancel_empty_pagelist(&pages,
 328                                 data, cinfo))
 329                                 continue;
 330
 331                         nfs_init_commit(data, &pages, data->lseg, cinfo);
 332                         initiate_commit(data, how);
 333                 }
 334         }
 335 out:
 336         return PNFS_ATTEMPTED;
 337 }
 338 EXPORT_SYMBOL_GPL(pnfs_generic_commit_pagelist);
 339
 340 /*
 341  * Data server cache
 342  *
 343  * Data servers can be mapped to different device ids.
 344  * nfs4_pnfs_ds reference counting
 345  *   - set to 1 on allocation
 346  *   - incremented when a device id maps a data server already in the cache.
 347  *   - decremented when deviceid is removed from the cache.
 348  */
 349 static DEFINE_SPINLOCK(nfs4_ds_cache_lock);
 350 static LIST_HEAD(nfs4_data_server_cache);
 351
 352 /* Debug routines */
 353 static void
 354 print_ds(struct nfs4_pnfs_ds *ds)
 355 {
 356         if (ds == NULL) {
 357                 printk(KERN_WARNING "%s NULL device\n", __func__);
 358                 return;
 359         }
 360         printk(KERN_WARNING "        ds %s\n"
 361                 "        ref count %d\n"
 362                 "        client %p\n"
 363                 "        cl_exchange_flags %x\n",
 364                 ds->ds_remotestr,
 365                 atomic_read(&ds->ds_count), ds->ds_clp,
 366                 ds->ds_clp ? ds->ds_clp->cl_exchange_flags : 0);
 367 }
 368
 369 static bool
 370 same_sockaddr(struct sockaddr *addr1, struct sockaddr *addr2)
 371 {
 372         struct sockaddr_in *a, *b;
 373         struct sockaddr_in6 *a6, *b6;
 374
 375         if (addr1->sa_family != addr2->sa_family)
 376                 return false;
 377
 378         switch (addr1->sa_family) {
 379         case AF_INET:
 380                 a = (struct sockaddr_in *)addr1;
 381                 b = (struct sockaddr_in *)addr2;
 382
 383                 if (a->sin_addr.s_addr == b->sin_addr.s_addr &&
 384                     a->sin_port == b->sin_port)
 385                         return true;
 386                 break;
 387
 388         case AF_INET6:
 389                 a6 = (struct sockaddr_in6 *)addr1;
 390                 b6 = (struct sockaddr_in6 *)addr2;
 391
 392                 /* LINKLOCAL addresses must have matching scope_id */
 393                 if (ipv6_addr_src_scope(&a6->sin6_addr) ==
 394                     IPV6_ADDR_SCOPE_LINKLOCAL &&
 395                     a6->sin6_scope_id != b6->sin6_scope_id)
 396                         return false;
 397
 398                 if (ipv6_addr_equal(&a6->sin6_addr, &b6->sin6_addr) &&
 399                     a6->sin6_port == b6->sin6_port)
 400                         return true;
 401                 break;
 402
 403         default:
 404                 dprintk("%s: unhandled address family: %u\n",
 405                         __func__, addr1->sa_family);
 406                 return false;
 407         }
 408
 409         return false;
 410 }
 411
 412 /*
 413  * Checks if 'dsaddrs1' contains a subset of 'dsaddrs2'. If it does,
 414  * declare a match.
 415  */
 416 static bool
 417 _same_data_server_addrs_locked(const struct list_head *dsaddrs1,
 418                                const struct list_head *dsaddrs2)
 419 {
 420         struct nfs4_pnfs_ds_addr *da1, *da2;
 421         struct sockaddr *sa1, *sa2;
 422         bool match = false;
 423
 424         list_for_each_entry(da1, dsaddrs1, da_node) {
 425                 sa1 = (struct sockaddr *)&da1->da_addr;
 426                 match = false;
 427                 list_for_each_entry(da2, dsaddrs2, da_node) {
 428                         sa2 = (struct sockaddr *)&da2->da_addr;
 429                         match = same_sockaddr(sa1, sa2);
 430                         if (match)
 431                                 break;
 432                 }
 433                 if (!match)
 434                         break;
 435         }
 436         return match;
 437 }
 438
 439 /*
 440  * Lookup DS by addresses.  nfs4_ds_cache_lock is held
 441  */
 442 static struct nfs4_pnfs_ds *
 443 _data_server_lookup_locked(const struct list_head *dsaddrs)
 444 {
 445         struct nfs4_pnfs_ds *ds;
 446
 447         list_for_each_entry(ds, &nfs4_data_server_cache, ds_node)
 448                 if (_same_data_server_addrs_locked(&ds->ds_addrs, dsaddrs))
 449                         return ds;
 450         return NULL;
 451 }
 452
 453 static void destroy_ds(struct nfs4_pnfs_ds *ds)
 454 {
 455         struct nfs4_pnfs_ds_addr *da;
 456
 457         dprintk("--> %s\n", __func__);
 458         ifdebug(FACILITY)
 459                 print_ds(ds);
 460
 461         nfs_put_client(ds->ds_clp);
 462
 463         while (!list_empty(&ds->ds_addrs)) {
 464                 da = list_first_entry(&ds->ds_addrs,
 465                                       struct nfs4_pnfs_ds_addr,
 466                                       da_node);
 467                 list_del_init(&da->da_node);
 468                 kfree(da->da_remotestr);
 469                 kfree(da);
 470         }
 471
 472         kfree(ds->ds_remotestr);
 473         kfree(ds);
 474 }
 475
 476 void nfs4_pnfs_ds_put(struct nfs4_pnfs_ds *ds)
 477 {
 478         if (atomic_dec_and_lock(&ds->ds_count,
 479                                 &nfs4_ds_cache_lock)) {
 480                 list_del_init(&ds->ds_node);
 481                 spin_unlock(&nfs4_ds_cache_lock);
 482                 destroy_ds(ds);
 483         }
 484 }
 485 EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_put);
 486
 487 /*
 488  * Create a string with a human readable address and port to avoid
 489  * complicated setup around many dprinks.
 490  */
 491 static char *
 492 nfs4_pnfs_remotestr(struct list_head *dsaddrs, gfp_t gfp_flags)
 493 {
 494         struct nfs4_pnfs_ds_addr *da;
 495         char *remotestr;
 496         size_t len;
 497         char *p;
 498
 499         len = 3;        /* '{', '}' and eol */
 500         list_for_each_entry(da, dsaddrs, da_node) {
 501                 len += strlen(da->da_remotestr) + 1;    /* string plus comma */
 502         }
 503
 504         remotestr = kzalloc(len, gfp_flags);
 505         if (!remotestr)
 506                 return NULL;
 507
 508         p = remotestr;
 509         *(p++) = '{';
 510         len--;
 511         list_for_each_entry(da, dsaddrs, da_node) {
 512                 size_t ll = strlen(da->da_remotestr);
 513
 514                 if (ll > len)
 515                         goto out_err;
 516
 517                 memcpy(p, da->da_remotestr, ll);
 518                 p += ll;
 519                 len -= ll;
 520
 521                 if (len < 1)
 522                         goto out_err;
 523                 (*p++) = ',';
 524                 len--;
 525         }
 526         if (len < 2)
 527                 goto out_err;
 528         *(p++) = '}';
 529         *p = '\0';
 530         return remotestr;
 531 out_err:
 532         kfree(remotestr);
 533         return NULL;
 534 }
 535
 536 /*
 537  * Given a list of multipath struct nfs4_pnfs_ds_addr, add it to ds cache if
 538  * uncached and return cached struct nfs4_pnfs_ds.
 539  */
 540 struct nfs4_pnfs_ds *
 541 nfs4_pnfs_ds_add(struct list_head *dsaddrs, gfp_t gfp_flags)
 542 {
 543         struct nfs4_pnfs_ds *tmp_ds, *ds = NULL;
 544         char *remotestr;
 545
 546         if (list_empty(dsaddrs)) {
 547                 dprintk("%s: no addresses defined\n", __func__);
 548                 goto out;
 549         }
 550
 551         ds = kzalloc(sizeof(*ds), gfp_flags);
 552         if (!ds)
 553                 goto out;
 554
 555         /* this is only used for debugging, so it's ok if its NULL */
 556         remotestr = nfs4_pnfs_remotestr(dsaddrs, gfp_flags);
 557
 558         spin_lock(&nfs4_ds_cache_lock);
 559         tmp_ds = _data_server_lookup_locked(dsaddrs);
 560         if (tmp_ds == NULL) {
 561                 INIT_LIST_HEAD(&ds->ds_addrs);
 562                 list_splice_init(dsaddrs, &ds->ds_addrs);
 563                 ds->ds_remotestr = remotestr;
 564                 atomic_set(&ds->ds_count, 1);
 565                 INIT_LIST_HEAD(&ds->ds_node);
 566                 ds->ds_clp = NULL;
 567                 list_add(&ds->ds_node, &nfs4_data_server_cache);
 568                 dprintk("%s add new data server %s\n", __func__,
 569                         ds->ds_remotestr);
 570         } else {
 571                 kfree(remotestr);
 572                 kfree(ds);
 573                 atomic_inc(&tmp_ds->ds_count);
 574                 dprintk("%s data server %s found, inc'ed ds_count to %d\n",
 575                         __func__, tmp_ds->ds_remotestr,
 576                         atomic_read(&tmp_ds->ds_count));
 577                 ds = tmp_ds;
 578         }
 579         spin_unlock(&nfs4_ds_cache_lock);
 580 out:
 581         return ds;
 582 }
 583 EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_add);
 584
 585 static void nfs4_wait_ds_connect(struct nfs4_pnfs_ds *ds)
 586 {
 587         might_sleep();
 588         wait_on_bit(&ds->ds_state, NFS4DS_CONNECTING,
 589                         TASK_KILLABLE);
 590 }
 591
 592 static void nfs4_clear_ds_conn_bit(struct nfs4_pnfs_ds *ds)
 593 {
 594         smp_mb__before_atomic();
 595         clear_bit(NFS4DS_CONNECTING, &ds->ds_state);
 596         smp_mb__after_atomic();
 597         wake_up_bit(&ds->ds_state, NFS4DS_CONNECTING);
 598 }
 599
 600 static struct nfs_client *(*get_v3_ds_connect)(
 601                         struct nfs_server *mds_srv,
 602                         const struct sockaddr *ds_addr,
 603                         int ds_addrlen,
 604                         int ds_proto,
 605                         unsigned int ds_timeo,
 606                         unsigned int ds_retrans);
 607
 608 static bool load_v3_ds_connect(void)
 609 {
 610         if (!get_v3_ds_connect) {
 611                 get_v3_ds_connect = symbol_request(nfs3_set_ds_client);
 612                 WARN_ON_ONCE(!get_v3_ds_connect);
 613         }
 614
 615         return(get_v3_ds_connect != NULL);
 616 }
 617
 618 void nfs4_pnfs_v3_ds_connect_unload(void)
 619 {
 620         if (get_v3_ds_connect) {
 621                 symbol_put(nfs3_set_ds_client);
 622                 get_v3_ds_connect = NULL;
 623         }
 624 }
 625
 626 static int _nfs4_pnfs_v3_ds_connect(struct nfs_server *mds_srv,
 627                                  struct nfs4_pnfs_ds *ds,
 628                                  unsigned int timeo,
 629                                  unsigned int retrans)
 630 {
 631         struct nfs_client *clp = ERR_PTR(-EIO);
 632         struct nfs4_pnfs_ds_addr *da;
 633         int status = 0;
 634
 635         dprintk("--> %s DS %s\n", __func__, ds->ds_remotestr);
 636
 637         if (!load_v3_ds_connect())
 638                 goto out;
 639
 640         list_for_each_entry(da, &ds->ds_addrs, da_node) {
 641                 dprintk("%s: DS %s: trying address %s\n",
 642                         __func__, ds->ds_remotestr, da->da_remotestr);
 643
 644                 if (!IS_ERR(clp)) {
 645                         struct xprt_create xprt_args = {
 646                                 .ident = XPRT_TRANSPORT_TCP,
 647                                 .net = clp->cl_net,
 648                                 .dstaddr = (struct sockaddr *)&da->da_addr,
 649                                 .addrlen = da->da_addrlen,
 650                                 .servername = clp->cl_hostname,
 651                         };
 652                         /* Add this address as an alias */
 653                         rpc_clnt_add_xprt(clp->cl_rpcclient, &xprt_args,
 654                                         rpc_clnt_test_and_add_xprt, NULL);
 655                 } else
 656                         clp = get_v3_ds_connect(mds_srv,
 657                                         (struct sockaddr *)&da->da_addr,
 658                                         da->da_addrlen, IPPROTO_TCP,
 659                                         timeo, retrans);
 660         }
 661
 662         if (IS_ERR(clp)) {
 663                 status = PTR_ERR(clp);
 664                 goto out;
 665         }
 666
 667         smp_wmb();
 668         ds->ds_clp = clp;
 669         dprintk("%s [new] addr: %s\n", __func__, ds->ds_remotestr);
 670 out:
 671         return status;
 672 }
 673
 674 static int _nfs4_pnfs_v4_ds_connect(struct nfs_server *mds_srv,
 675                                  struct nfs4_pnfs_ds *ds,
 676                                  unsigned int timeo,
 677                                  unsigned int retrans,
 678                                  u32 minor_version)
 679 {
 680         struct nfs_client *clp = ERR_PTR(-EIO);
 681         struct nfs4_pnfs_ds_addr *da;
 682         int status = 0;
 683
 684         dprintk("--> %s DS %s\n", __func__, ds->ds_remotestr);
 685
 686         list_for_each_entry(da, &ds->ds_addrs, da_node) {
 687                 dprintk("%s: DS %s: trying address %s\n",
 688                         __func__, ds->ds_remotestr, da->da_remotestr);
 689
 690                 if (!IS_ERR(clp) && clp->cl_mvops->session_trunk) {
 691                         struct xprt_create xprt_args = {
 692                                 .ident = XPRT_TRANSPORT_TCP,
 693                                 .net = clp->cl_net,
 694                                 .dstaddr = (struct sockaddr *)&da->da_addr,
 695                                 .addrlen = da->da_addrlen,
 696                                 .servername = clp->cl_hostname,
 697                         };
 698                         struct nfs4_add_xprt_data xprtdata = {
 699                                 .clp = clp,
 700                                 .cred = nfs4_get_clid_cred(clp),
 701                         };
 702                         struct rpc_add_xprt_test rpcdata = {
 703                                 .add_xprt_test = clp->cl_mvops->session_trunk,
 704                                 .data = &xprtdata,
 705                         };
 706
 707                         /**
 708                         * Test this address for session trunking and
 709                         * add as an alias
 710                         */
 711                         rpc_clnt_add_xprt(clp->cl_rpcclient, &xprt_args,
 712                                           rpc_clnt_setup_test_and_add_xprt,
 713                                           &rpcdata);
 714                         if (xprtdata.cred)
 715                                 put_rpccred(xprtdata.cred);
 716                 } else {
 717                         clp = nfs4_set_ds_client(mds_srv,
 718                                                 (struct sockaddr *)&da->da_addr,
 719                                                 da->da_addrlen, IPPROTO_TCP,
 720                                                 timeo, retrans, minor_version);
 721                         if (IS_ERR(clp))
 722                                 continue;
 723
 724                         status = nfs4_init_ds_session(clp,
 725                                         mds_srv->nfs_client->cl_lease_time);
 726                         if (status) {
 727                                 nfs_put_client(clp);
 728                                 clp = ERR_PTR(-EIO);
 729                                 continue;
 730                         }
 731
 732                 }
 733         }
 734
 735         if (IS_ERR(clp)) {
 736                 status = PTR_ERR(clp);
 737                 goto out;
 738         }
 739
 740         smp_wmb();
 741         ds->ds_clp = clp;
 742         dprintk("%s [new] addr: %s\n", __func__, ds->ds_remotestr);
 743 out:
 744         return status;
 745 }
 746
 747 /*
 748  * Create an rpc connection to the nfs4_pnfs_ds data server.
 749  * Currently only supports IPv4 and IPv6 addresses.
 750  * If connection fails, make devid unavailable and return a -errno.
 751  */
 752 int nfs4_pnfs_ds_connect(struct nfs_server *mds_srv, struct nfs4_pnfs_ds *ds,
 753                           struct nfs4_deviceid_node *devid, unsigned int timeo,
 754                           unsigned int retrans, u32 version, u32 minor_version)
 755 {
 756         int err;
 757
 758 again:
 759         err = 0;
 760         if (test_and_set_bit(NFS4DS_CONNECTING, &ds->ds_state) == 0) {
 761                 if (version == 3) {
 762                         err = _nfs4_pnfs_v3_ds_connect(mds_srv, ds, timeo,
 763                                                        retrans);
 764                 } else if (version == 4) {
 765                         err = _nfs4_pnfs_v4_ds_connect(mds_srv, ds, timeo,
 766                                                        retrans, minor_version);
 767                 } else {
 768                         dprintk("%s: unsupported DS version %d\n", __func__,
 769                                 version);
 770                         err = -EPROTONOSUPPORT;
 771                 }
 772
 773                 nfs4_clear_ds_conn_bit(ds);
 774         } else {
 775                 nfs4_wait_ds_connect(ds);
 776
 777                 /* what was waited on didn't connect AND didn't mark unavail */
 778                 if (!ds->ds_clp && !nfs4_test_deviceid_unavailable(devid))
 779                         goto again;
 780         }
 781
 782         /*
 783          * At this point the ds->ds_clp should be ready, but it might have
 784          * hit an error.
 785          */
 786         if (!err) {
 787                 if (!ds->ds_clp || !nfs_client_init_is_complete(ds->ds_clp)) {
 788                         WARN_ON_ONCE(ds->ds_clp ||
 789                                 !nfs4_test_deviceid_unavailable(devid));
 790                         return -EINVAL;
 791                 }
 792                 err = nfs_client_init_status(ds->ds_clp);
 793         }
 794
 795         return err;
 796 }
 797 EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_connect);
 798
 799 /*
 800  * Currently only supports ipv4, ipv6 and one multi-path address.
 801  */
 802 struct nfs4_pnfs_ds_addr *
 803 nfs4_decode_mp_ds_addr(struct net *net, struct xdr_stream *xdr, gfp_t gfp_flags)
 804 {
 805         struct nfs4_pnfs_ds_addr *da = NULL;
 806         char *buf, *portstr;
 807         __be16 port;
 808         int nlen, rlen;
 809         int tmp[2];
 810         __be32 *p;
 811         char *netid, *match_netid;
 812         size_t len, match_netid_len;
 813         char *startsep = "";
 814         char *endsep = "";
 815
 816
 817         /* r_netid */
 818         p = xdr_inline_decode(xdr, 4);
 819         if (unlikely(!p))
 820                 goto out_err;
 821         nlen = be32_to_cpup(p++);
 822
 823         p = xdr_inline_decode(xdr, nlen);
 824         if (unlikely(!p))
 825                 goto out_err;
 826
 827         netid = kmalloc(nlen+1, gfp_flags);
 828         if (unlikely(!netid))
 829                 goto out_err;
 830
 831         netid[nlen] = '\0';
 832         memcpy(netid, p, nlen);
 833
 834         /* r_addr: ip/ip6addr with port in dec octets - see RFC 5665 */
 835         p = xdr_inline_decode(xdr, 4);
 836         if (unlikely(!p))
 837                 goto out_free_netid;
 838         rlen = be32_to_cpup(p);
 839
 840         p = xdr_inline_decode(xdr, rlen);
 841         if (unlikely(!p))
 842                 goto out_free_netid;
 843
 844         /* port is ".ABC.DEF", 8 chars max */
 845         if (rlen > INET6_ADDRSTRLEN + IPV6_SCOPE_ID_LEN + 8) {
 846                 dprintk("%s: Invalid address, length %d\n", __func__,
 847                         rlen);
 848                 goto out_free_netid;
 849         }
 850         buf = kmalloc(rlen + 1, gfp_flags);
 851         if (!buf) {
 852                 dprintk("%s: Not enough memory\n", __func__);
 853                 goto out_free_netid;
 854         }
 855         buf[rlen] = '\0';
 856         memcpy(buf, p, rlen);
 857
 858         /* replace port '.' with '-' */
 859         portstr = strrchr(buf, '.');
 860         if (!portstr) {
 861                 dprintk("%s: Failed finding expected dot in port\n",
 862                         __func__);
 863                 goto out_free_buf;
 864         }
 865         *portstr = '-';
 866
 867         /* find '.' between address and port */
 868         portstr = strrchr(buf, '.');
 869         if (!portstr) {
 870                 dprintk("%s: Failed finding expected dot between address and "
 871                         "port\n", __func__);
 872                 goto out_free_buf;
 873         }
 874         *portstr = '\0';
 875
 876         da = kzalloc(sizeof(*da), gfp_flags);
 877         if (unlikely(!da))
 878                 goto out_free_buf;
 879
 880         INIT_LIST_HEAD(&da->da_node);
 881
 882         if (!rpc_pton(net, buf, portstr-buf, (struct sockaddr *)&da->da_addr,
 883                       sizeof(da->da_addr))) {
 884                 dprintk("%s: error parsing address %s\n", __func__, buf);
 885                 goto out_free_da;
 886         }
 887
 888         portstr++;
 889         sscanf(portstr, "%d-%d", &tmp[0], &tmp[1]);
 890         port = htons((tmp[0] << 8) | (tmp[1]));
 891
 892         switch (da->da_addr.ss_family) {
 893         case AF_INET:
 894                 ((struct sockaddr_in *)&da->da_addr)->sin_port = port;
 895                 da->da_addrlen = sizeof(struct sockaddr_in);
 896                 match_netid = "tcp";
 897                 match_netid_len = 3;
 898                 break;
 899
 900         case AF_INET6:
 901                 ((struct sockaddr_in6 *)&da->da_addr)->sin6_port = port;
 902                 da->da_addrlen = sizeof(struct sockaddr_in6);
 903                 match_netid = "tcp6";
 904                 match_netid_len = 4;
 905                 startsep = "[";
 906                 endsep = "]";
 907                 break;
 908
 909         default:
 910                 dprintk("%s: unsupported address family: %u\n",
 911                         __func__, da->da_addr.ss_family);
 912                 goto out_free_da;
 913         }
 914
 915         if (nlen != match_netid_len || strncmp(netid, match_netid, nlen)) {
 916                 dprintk("%s: ERROR: r_netid \"%s\" != \"%s\"\n",
 917                         __func__, netid, match_netid);
 918                 goto out_free_da;
 919         }
 920
 921         /* save human readable address */
 922         len = strlen(startsep) + strlen(buf) + strlen(endsep) + 7;
 923         da->da_remotestr = kzalloc(len, gfp_flags);
 924
 925         /* NULL is ok, only used for dprintk */
 926         if (da->da_remotestr)
 927                 snprintf(da->da_remotestr, len, "%s%s%s:%u", startsep,
 928                          buf, endsep, ntohs(port));
 929
 930         dprintk("%s: Parsed DS addr %s\n", __func__, da->da_remotestr);
 931         kfree(buf);
 932         kfree(netid);
 933         return da;
 934
 935 out_free_da:
 936         kfree(da);
 937 out_free_buf:
 938         dprintk("%s: Error parsing DS addr: %s\n", __func__, buf);
 939         kfree(buf);
 940 out_free_netid:
 941         kfree(netid);
 942 out_err:
 943         return NULL;
 944 }
 945 EXPORT_SYMBOL_GPL(nfs4_decode_mp_ds_addr);
 946
 947 void
 948 pnfs_layout_mark_request_commit(struct nfs_page *req,
 949                                 struct pnfs_layout_segment *lseg,
 950                                 struct nfs_commit_info *cinfo,
 951                                 u32 ds_commit_idx)
 952 {
 953         struct list_head *list;
 954         struct pnfs_commit_bucket *buckets;
 955
 956         spin_lock(&cinfo->inode->i_lock);
 957         buckets = cinfo->ds->buckets;
 958         list = &buckets[ds_commit_idx].written;
 959         if (list_empty(list)) {
 960                 if (!pnfs_is_valid_lseg(lseg)) {
 961                         spin_unlock(&cinfo->inode->i_lock);
 962                         cinfo->completion_ops->resched_write(cinfo, req);
 963                         return;
 964                 }
 965                 /* Non-empty buckets hold a reference on the lseg.  That ref
 966                  * is normally transferred to the COMMIT call and released
 967                  * there.  It could also be released if the last req is pulled
 968                  * off due to a rewrite, in which case it will be done in
 969                  * pnfs_common_clear_request_commit
 970                  */
 971                 WARN_ON_ONCE(buckets[ds_commit_idx].wlseg != NULL);
 972                 buckets[ds_commit_idx].wlseg = pnfs_get_lseg(lseg);
 973         }
 974         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
 975         cinfo->ds->nwritten++;
 976
 977         nfs_request_add_commit_list_locked(req, list, cinfo);
 978         spin_unlock(&cinfo->inode->i_lock);
 979         nfs_mark_page_unstable(req->wb_page, cinfo);
 980 }
 981 EXPORT_SYMBOL_GPL(pnfs_layout_mark_request_commit);
 982
 983 int
 984 pnfs_nfs_generic_sync(struct inode *inode, bool datasync)
 985 {
 986         int ret;
 987
 988         if (!pnfs_layoutcommit_outstanding(inode))
 989                 return 0;
 990         ret = nfs_commit_inode(inode, FLUSH_SYNC);
 991         if (ret < 0)
 992                 return ret;
 993         if (datasync)
 994                 return 0;
 995         return pnfs_layoutcommit_inode(inode, true);
 996 }
 997 EXPORT_SYMBOL_GPL(pnfs_nfs_generic_sync);
 998