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 <[email protected]>
   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 the inode (/cinfo) 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->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->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->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 i;
 163
 164         lockdep_assert_held(cinfo->lock);
 165 restart:
 166         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
 167                 if (pnfs_generic_transfer_commit_list(&b->written, dst,
 168                                                       cinfo, 0)) {
 169                         freeme = b->wlseg;
 170                         b->wlseg = NULL;
 171                         spin_unlock(cinfo->lock);
 172                         pnfs_put_lseg(freeme);
 173                         spin_lock(cinfo->lock);
 174                         goto restart;
 175                 }
 176         }
 177         cinfo->ds->nwritten = 0;
 178 }
 179 EXPORT_SYMBOL_GPL(pnfs_generic_recover_commit_reqs);
 180
 181 static void pnfs_generic_retry_commit(struct nfs_commit_info *cinfo, int idx)
 182 {
 183         struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
 184         struct pnfs_commit_bucket *bucket;
 185         struct pnfs_layout_segment *freeme;
 186         LIST_HEAD(pages);
 187         int i;
 188
 189         spin_lock(cinfo->lock);
 190         for (i = idx; i < fl_cinfo->nbuckets; i++) {
 191                 bucket = &fl_cinfo->buckets[i];
 192                 if (list_empty(&bucket->committing))
 193                         continue;
 194                 freeme = bucket->clseg;
 195                 bucket->clseg = NULL;
 196                 list_splice_init(&bucket->committing, &pages);
 197                 spin_unlock(cinfo->lock);
 198                 nfs_retry_commit(&pages, freeme, cinfo, i);
 199                 pnfs_put_lseg(freeme);
 200                 spin_lock(cinfo->lock);
 201         }
 202         spin_unlock(cinfo->lock);
 203 }
 204
 205 static unsigned int
 206 pnfs_generic_alloc_ds_commits(struct nfs_commit_info *cinfo,
 207                               struct list_head *list)
 208 {
 209         struct pnfs_ds_commit_info *fl_cinfo;
 210         struct pnfs_commit_bucket *bucket;
 211         struct nfs_commit_data *data;
 212         int i;
 213         unsigned int nreq = 0;
 214
 215         fl_cinfo = cinfo->ds;
 216         bucket = fl_cinfo->buckets;
 217         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
 218                 if (list_empty(&bucket->committing))
 219                         continue;
 220                 data = nfs_commitdata_alloc();
 221                 if (!data)
 222                         break;
 223                 data->ds_commit_index = i;
 224                 list_add(&data->pages, list);
 225                 nreq++;
 226         }
 227
 228         /* Clean up on error */
 229         pnfs_generic_retry_commit(cinfo, i);
 230         return nreq;
 231 }
 232
 233 static inline
 234 void pnfs_fetch_commit_bucket_list(struct list_head *pages,
 235                 struct nfs_commit_data *data,
 236                 struct nfs_commit_info *cinfo)
 237 {
 238         struct pnfs_commit_bucket *bucket;
 239
 240         bucket = &cinfo->ds->buckets[data->ds_commit_index];
 241         spin_lock(cinfo->lock);
 242         list_splice_init(&bucket->committing, pages);
 243         data->lseg = bucket->clseg;
 244         bucket->clseg = NULL;
 245         spin_unlock(cinfo->lock);
 246
 247 }
 248
 249 /* This follows nfs_commit_list pretty closely */
 250 int
 251 pnfs_generic_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 252                              int how, struct nfs_commit_info *cinfo,
 253                              int (*initiate_commit)(struct nfs_commit_data *data,
 254                                                     int how))
 255 {
 256         struct nfs_commit_data *data, *tmp;
 257         LIST_HEAD(list);
 258         unsigned int nreq = 0;
 259
 260         if (!list_empty(mds_pages)) {
 261                 data = nfs_commitdata_alloc();
 262                 if (data != NULL) {
 263                         data->ds_commit_index = -1;
 264                         list_add(&data->pages, &list);
 265                         nreq++;
 266                 } else {
 267                         nfs_retry_commit(mds_pages, NULL, cinfo, 0);
 268                         pnfs_generic_retry_commit(cinfo, 0);
 269                         return -ENOMEM;
 270                 }
 271         }
 272
 273         nreq += pnfs_generic_alloc_ds_commits(cinfo, &list);
 274
 275         if (nreq == 0)
 276                 goto out;
 277
 278         atomic_add(nreq, &cinfo->mds->rpcs_out);
 279
 280         list_for_each_entry_safe(data, tmp, &list, pages) {
 281                 list_del_init(&data->pages);
 282                 if (data->ds_commit_index < 0) {
 283                         nfs_init_commit(data, mds_pages, NULL, cinfo);
 284                         nfs_initiate_commit(NFS_CLIENT(inode), data,
 285                                             NFS_PROTO(data->inode),
 286                                             data->mds_ops, how, 0);
 287                 } else {
 288                         LIST_HEAD(pages);
 289
 290                         pnfs_fetch_commit_bucket_list(&pages, data, cinfo);
 291                         nfs_init_commit(data, &pages, data->lseg, cinfo);
 292                         initiate_commit(data, how);
 293                 }
 294         }
 295 out:
 296         cinfo->ds->ncommitting = 0;
 297         return PNFS_ATTEMPTED;
 298 }
 299 EXPORT_SYMBOL_GPL(pnfs_generic_commit_pagelist);
 300
 301 /*
 302  * Data server cache
 303  *
 304  * Data servers can be mapped to different device ids.
 305  * nfs4_pnfs_ds reference counting
 306  *   - set to 1 on allocation
 307  *   - incremented when a device id maps a data server already in the cache.
 308  *   - decremented when deviceid is removed from the cache.
 309  */
 310 static DEFINE_SPINLOCK(nfs4_ds_cache_lock);
 311 static LIST_HEAD(nfs4_data_server_cache);
 312
 313 /* Debug routines */
 314 static void
 315 print_ds(struct nfs4_pnfs_ds *ds)
 316 {
 317         if (ds == NULL) {
 318                 printk(KERN_WARNING "%s NULL device\n", __func__);
 319                 return;
 320         }
 321         printk(KERN_WARNING "        ds %s\n"
 322                 "        ref count %d\n"
 323                 "        client %p\n"
 324                 "        cl_exchange_flags %x\n",
 325                 ds->ds_remotestr,
 326                 atomic_read(&ds->ds_count), ds->ds_clp,
 327                 ds->ds_clp ? ds->ds_clp->cl_exchange_flags : 0);
 328 }
 329
 330 static bool
 331 same_sockaddr(struct sockaddr *addr1, struct sockaddr *addr2)
 332 {
 333         struct sockaddr_in *a, *b;
 334         struct sockaddr_in6 *a6, *b6;
 335
 336         if (addr1->sa_family != addr2->sa_family)
 337                 return false;
 338
 339         switch (addr1->sa_family) {
 340         case AF_INET:
 341                 a = (struct sockaddr_in *)addr1;
 342                 b = (struct sockaddr_in *)addr2;
 343
 344                 if (a->sin_addr.s_addr == b->sin_addr.s_addr &&
 345                     a->sin_port == b->sin_port)
 346                         return true;
 347                 break;
 348
 349         case AF_INET6:
 350                 a6 = (struct sockaddr_in6 *)addr1;
 351                 b6 = (struct sockaddr_in6 *)addr2;
 352
 353                 /* LINKLOCAL addresses must have matching scope_id */
 354                 if (ipv6_addr_src_scope(&a6->sin6_addr) ==
 355                     IPV6_ADDR_SCOPE_LINKLOCAL &&
 356                     a6->sin6_scope_id != b6->sin6_scope_id)
 357                         return false;
 358
 359                 if (ipv6_addr_equal(&a6->sin6_addr, &b6->sin6_addr) &&
 360                     a6->sin6_port == b6->sin6_port)
 361                         return true;
 362                 break;
 363
 364         default:
 365                 dprintk("%s: unhandled address family: %u\n",
 366                         __func__, addr1->sa_family);
 367                 return false;
 368         }
 369
 370         return false;
 371 }
 372
 373 /*
 374  * Checks if 'dsaddrs1' contains a subset of 'dsaddrs2'. If it does,
 375  * declare a match.
 376  */
 377 static bool
 378 _same_data_server_addrs_locked(const struct list_head *dsaddrs1,
 379                                const struct list_head *dsaddrs2)
 380 {
 381         struct nfs4_pnfs_ds_addr *da1, *da2;
 382         struct sockaddr *sa1, *sa2;
 383         bool match = false;
 384
 385         list_for_each_entry(da1, dsaddrs1, da_node) {
 386                 sa1 = (struct sockaddr *)&da1->da_addr;
 387                 match = false;
 388                 list_for_each_entry(da2, dsaddrs2, da_node) {
 389                         sa2 = (struct sockaddr *)&da2->da_addr;
 390                         match = same_sockaddr(sa1, sa2);
 391                         if (match)
 392                                 break;
 393                 }
 394                 if (!match)
 395                         break;
 396         }
 397         return match;
 398 }
 399
 400 /*
 401  * Lookup DS by addresses.  nfs4_ds_cache_lock is held
 402  */
 403 static struct nfs4_pnfs_ds *
 404 _data_server_lookup_locked(const struct list_head *dsaddrs)
 405 {
 406         struct nfs4_pnfs_ds *ds;
 407
 408         list_for_each_entry(ds, &nfs4_data_server_cache, ds_node)
 409                 if (_same_data_server_addrs_locked(&ds->ds_addrs, dsaddrs))
 410                         return ds;
 411         return NULL;
 412 }
 413
 414 static void destroy_ds(struct nfs4_pnfs_ds *ds)
 415 {
 416         struct nfs4_pnfs_ds_addr *da;
 417
 418         dprintk("--> %s\n", __func__);
 419         ifdebug(FACILITY)
 420                 print_ds(ds);
 421
 422         nfs_put_client(ds->ds_clp);
 423
 424         while (!list_empty(&ds->ds_addrs)) {
 425                 da = list_first_entry(&ds->ds_addrs,
 426                                       struct nfs4_pnfs_ds_addr,
 427                                       da_node);
 428                 list_del_init(&da->da_node);
 429                 kfree(da->da_remotestr);
 430                 kfree(da);
 431         }
 432
 433         kfree(ds->ds_remotestr);
 434         kfree(ds);
 435 }
 436
 437 void nfs4_pnfs_ds_put(struct nfs4_pnfs_ds *ds)
 438 {
 439         if (atomic_dec_and_lock(&ds->ds_count,
 440                                 &nfs4_ds_cache_lock)) {
 441                 list_del_init(&ds->ds_node);
 442                 spin_unlock(&nfs4_ds_cache_lock);
 443                 destroy_ds(ds);
 444         }
 445 }
 446 EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_put);
 447
 448 /*
 449  * Create a string with a human readable address and port to avoid
 450  * complicated setup around many dprinks.
 451  */
 452 static char *
 453 nfs4_pnfs_remotestr(struct list_head *dsaddrs, gfp_t gfp_flags)
 454 {
 455         struct nfs4_pnfs_ds_addr *da;
 456         char *remotestr;
 457         size_t len;
 458         char *p;
 459
 460         len = 3;        /* '{', '}' and eol */
 461         list_for_each_entry(da, dsaddrs, da_node) {
 462                 len += strlen(da->da_remotestr) + 1;    /* string plus comma */
 463         }
 464
 465         remotestr = kzalloc(len, gfp_flags);
 466         if (!remotestr)
 467                 return NULL;
 468
 469         p = remotestr;
 470         *(p++) = '{';
 471         len--;
 472         list_for_each_entry(da, dsaddrs, da_node) {
 473                 size_t ll = strlen(da->da_remotestr);
 474
 475                 if (ll > len)
 476                         goto out_err;
 477
 478                 memcpy(p, da->da_remotestr, ll);
 479                 p += ll;
 480                 len -= ll;
 481
 482                 if (len < 1)
 483                         goto out_err;
 484                 (*p++) = ',';
 485                 len--;
 486         }
 487         if (len < 2)
 488                 goto out_err;
 489         *(p++) = '}';
 490         *p = '\0';
 491         return remotestr;
 492 out_err:
 493         kfree(remotestr);
 494         return NULL;
 495 }
 496
 497 /*
 498  * Given a list of multipath struct nfs4_pnfs_ds_addr, add it to ds cache if
 499  * uncached and return cached struct nfs4_pnfs_ds.
 500  */
 501 struct nfs4_pnfs_ds *
 502 nfs4_pnfs_ds_add(struct list_head *dsaddrs, gfp_t gfp_flags)
 503 {
 504         struct nfs4_pnfs_ds *tmp_ds, *ds = NULL;
 505         char *remotestr;
 506
 507         if (list_empty(dsaddrs)) {
 508                 dprintk("%s: no addresses defined\n", __func__);
 509                 goto out;
 510         }
 511
 512         ds = kzalloc(sizeof(*ds), gfp_flags);
 513         if (!ds)
 514                 goto out;
 515
 516         /* this is only used for debugging, so it's ok if its NULL */
 517         remotestr = nfs4_pnfs_remotestr(dsaddrs, gfp_flags);
 518
 519         spin_lock(&nfs4_ds_cache_lock);
 520         tmp_ds = _data_server_lookup_locked(dsaddrs);
 521         if (tmp_ds == NULL) {
 522                 INIT_LIST_HEAD(&ds->ds_addrs);
 523                 list_splice_init(dsaddrs, &ds->ds_addrs);
 524                 ds->ds_remotestr = remotestr;
 525                 atomic_set(&ds->ds_count, 1);
 526                 INIT_LIST_HEAD(&ds->ds_node);
 527                 ds->ds_clp = NULL;
 528                 list_add(&ds->ds_node, &nfs4_data_server_cache);
 529                 dprintk("%s add new data server %s\n", __func__,
 530                         ds->ds_remotestr);
 531         } else {
 532                 kfree(remotestr);
 533                 kfree(ds);
 534                 atomic_inc(&tmp_ds->ds_count);
 535                 dprintk("%s data server %s found, inc'ed ds_count to %d\n",
 536                         __func__, tmp_ds->ds_remotestr,
 537                         atomic_read(&tmp_ds->ds_count));
 538                 ds = tmp_ds;
 539         }
 540         spin_unlock(&nfs4_ds_cache_lock);
 541 out:
 542         return ds;
 543 }
 544 EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_add);
 545
 546 static void nfs4_wait_ds_connect(struct nfs4_pnfs_ds *ds)
 547 {
 548         might_sleep();
 549         wait_on_bit(&ds->ds_state, NFS4DS_CONNECTING,
 550                         TASK_KILLABLE);
 551 }
 552
 553 static void nfs4_clear_ds_conn_bit(struct nfs4_pnfs_ds *ds)
 554 {
 555         smp_mb__before_atomic();
 556         clear_bit(NFS4DS_CONNECTING, &ds->ds_state);
 557         smp_mb__after_atomic();
 558         wake_up_bit(&ds->ds_state, NFS4DS_CONNECTING);
 559 }
 560
 561 static struct nfs_client *(*get_v3_ds_connect)(
 562                         struct nfs_client *mds_clp,
 563                         const struct sockaddr *ds_addr,
 564                         int ds_addrlen,
 565                         int ds_proto,
 566                         unsigned int ds_timeo,
 567                         unsigned int ds_retrans,
 568                         rpc_authflavor_t au_flavor);
 569
 570 static bool load_v3_ds_connect(void)
 571 {
 572         if (!get_v3_ds_connect) {
 573                 get_v3_ds_connect = symbol_request(nfs3_set_ds_client);
 574                 WARN_ON_ONCE(!get_v3_ds_connect);
 575         }
 576
 577         return(get_v3_ds_connect != NULL);
 578 }
 579
 580 void nfs4_pnfs_v3_ds_connect_unload(void)
 581 {
 582         if (get_v3_ds_connect) {
 583                 symbol_put(nfs3_set_ds_client);
 584                 get_v3_ds_connect = NULL;
 585         }
 586 }
 587 EXPORT_SYMBOL_GPL(nfs4_pnfs_v3_ds_connect_unload);
 588
 589 static int _nfs4_pnfs_v3_ds_connect(struct nfs_server *mds_srv,
 590                                  struct nfs4_pnfs_ds *ds,
 591                                  unsigned int timeo,
 592                                  unsigned int retrans,
 593                                  rpc_authflavor_t au_flavor)
 594 {
 595         struct nfs_client *clp = ERR_PTR(-EIO);
 596         struct nfs4_pnfs_ds_addr *da;
 597         int status = 0;
 598
 599         dprintk("--> %s DS %s au_flavor %d\n", __func__,
 600                 ds->ds_remotestr, au_flavor);
 601
 602         if (!load_v3_ds_connect())
 603                 goto out;
 604
 605         list_for_each_entry(da, &ds->ds_addrs, da_node) {
 606                 dprintk("%s: DS %s: trying address %s\n",
 607                         __func__, ds->ds_remotestr, da->da_remotestr);
 608
 609                 clp = get_v3_ds_connect(mds_srv->nfs_client,
 610                                         (struct sockaddr *)&da->da_addr,
 611                                         da->da_addrlen, IPPROTO_TCP,
 612                                         timeo, retrans, au_flavor);
 613                 if (!IS_ERR(clp))
 614                         break;
 615         }
 616
 617         if (IS_ERR(clp)) {
 618                 status = PTR_ERR(clp);
 619                 goto out;
 620         }
 621
 622         smp_wmb();
 623         ds->ds_clp = clp;
 624         dprintk("%s [new] addr: %s\n", __func__, ds->ds_remotestr);
 625 out:
 626         return status;
 627 }
 628
 629 static int _nfs4_pnfs_v4_ds_connect(struct nfs_server *mds_srv,
 630                                  struct nfs4_pnfs_ds *ds,
 631                                  unsigned int timeo,
 632                                  unsigned int retrans,
 633                                  u32 minor_version,
 634                                  rpc_authflavor_t au_flavor)
 635 {
 636         struct nfs_client *clp = ERR_PTR(-EIO);
 637         struct nfs4_pnfs_ds_addr *da;
 638         int status = 0;
 639
 640         dprintk("--> %s DS %s au_flavor %d\n", __func__, ds->ds_remotestr,
 641                 au_flavor);
 642
 643         list_for_each_entry(da, &ds->ds_addrs, da_node) {
 644                 dprintk("%s: DS %s: trying address %s\n",
 645                         __func__, ds->ds_remotestr, da->da_remotestr);
 646
 647                 clp = nfs4_set_ds_client(mds_srv->nfs_client,
 648                                         (struct sockaddr *)&da->da_addr,
 649                                         da->da_addrlen, IPPROTO_TCP,
 650                                         timeo, retrans, minor_version,
 651                                         au_flavor);
 652                 if (!IS_ERR(clp))
 653                         break;
 654         }
 655
 656         if (IS_ERR(clp)) {
 657                 status = PTR_ERR(clp);
 658                 goto out;
 659         }
 660
 661         status = nfs4_init_ds_session(clp, mds_srv->nfs_client->cl_lease_time);
 662         if (status)
 663                 goto out_put;
 664
 665         smp_wmb();
 666         ds->ds_clp = clp;
 667         dprintk("%s [new] addr: %s\n", __func__, ds->ds_remotestr);
 668 out:
 669         return status;
 670 out_put:
 671         nfs_put_client(clp);
 672         goto out;
 673 }
 674
 675 /*
 676  * Create an rpc connection to the nfs4_pnfs_ds data server.
 677  * Currently only supports IPv4 and IPv6 addresses.
 678  * If connection fails, make devid unavailable.
 679  */
 680 void nfs4_pnfs_ds_connect(struct nfs_server *mds_srv, struct nfs4_pnfs_ds *ds,
 681                           struct nfs4_deviceid_node *devid, unsigned int timeo,
 682                           unsigned int retrans, u32 version,
 683                           u32 minor_version, rpc_authflavor_t au_flavor)
 684 {
 685         if (test_and_set_bit(NFS4DS_CONNECTING, &ds->ds_state) == 0) {
 686                 int err = 0;
 687
 688                 if (version == 3) {
 689                         err = _nfs4_pnfs_v3_ds_connect(mds_srv, ds, timeo,
 690                                                        retrans, au_flavor);
 691                 } else if (version == 4) {
 692                         err = _nfs4_pnfs_v4_ds_connect(mds_srv, ds, timeo,
 693                                                        retrans, minor_version,
 694                                                        au_flavor);
 695                 } else {
 696                         dprintk("%s: unsupported DS version %d\n", __func__,
 697                                 version);
 698                         err = -EPROTONOSUPPORT;
 699                 }
 700
 701                 if (err)
 702                         nfs4_mark_deviceid_unavailable(devid);
 703                 nfs4_clear_ds_conn_bit(ds);
 704         } else {
 705                 nfs4_wait_ds_connect(ds);
 706         }
 707 }
 708 EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_connect);
 709
 710 /*
 711  * Currently only supports ipv4, ipv6 and one multi-path address.
 712  */
 713 struct nfs4_pnfs_ds_addr *
 714 nfs4_decode_mp_ds_addr(struct net *net, struct xdr_stream *xdr, gfp_t gfp_flags)
 715 {
 716         struct nfs4_pnfs_ds_addr *da = NULL;
 717         char *buf, *portstr;
 718         __be16 port;
 719         int nlen, rlen;
 720         int tmp[2];
 721         __be32 *p;
 722         char *netid, *match_netid;
 723         size_t len, match_netid_len;
 724         char *startsep = "";
 725         char *endsep = "";
 726
 727
 728         /* r_netid */
 729         p = xdr_inline_decode(xdr, 4);
 730         if (unlikely(!p))
 731                 goto out_err;
 732         nlen = be32_to_cpup(p++);
 733
 734         p = xdr_inline_decode(xdr, nlen);
 735         if (unlikely(!p))
 736                 goto out_err;
 737
 738         netid = kmalloc(nlen+1, gfp_flags);
 739         if (unlikely(!netid))
 740                 goto out_err;
 741
 742         netid[nlen] = '\0';
 743         memcpy(netid, p, nlen);
 744
 745         /* r_addr: ip/ip6addr with port in dec octets - see RFC 5665 */
 746         p = xdr_inline_decode(xdr, 4);
 747         if (unlikely(!p))
 748                 goto out_free_netid;
 749         rlen = be32_to_cpup(p);
 750
 751         p = xdr_inline_decode(xdr, rlen);
 752         if (unlikely(!p))
 753                 goto out_free_netid;
 754
 755         /* port is ".ABC.DEF", 8 chars max */
 756         if (rlen > INET6_ADDRSTRLEN + IPV6_SCOPE_ID_LEN + 8) {
 757                 dprintk("%s: Invalid address, length %d\n", __func__,
 758                         rlen);
 759                 goto out_free_netid;
 760         }
 761         buf = kmalloc(rlen + 1, gfp_flags);
 762         if (!buf) {
 763                 dprintk("%s: Not enough memory\n", __func__);
 764                 goto out_free_netid;
 765         }
 766         buf[rlen] = '\0';
 767         memcpy(buf, p, rlen);
 768
 769         /* replace port '.' with '-' */
 770         portstr = strrchr(buf, '.');
 771         if (!portstr) {
 772                 dprintk("%s: Failed finding expected dot in port\n",
 773                         __func__);
 774                 goto out_free_buf;
 775         }
 776         *portstr = '-';
 777
 778         /* find '.' between address and port */
 779         portstr = strrchr(buf, '.');
 780         if (!portstr) {
 781                 dprintk("%s: Failed finding expected dot between address and "
 782                         "port\n", __func__);
 783                 goto out_free_buf;
 784         }
 785         *portstr = '\0';
 786
 787         da = kzalloc(sizeof(*da), gfp_flags);
 788         if (unlikely(!da))
 789                 goto out_free_buf;
 790
 791         INIT_LIST_HEAD(&da->da_node);
 792
 793         if (!rpc_pton(net, buf, portstr-buf, (struct sockaddr *)&da->da_addr,
 794                       sizeof(da->da_addr))) {
 795                 dprintk("%s: error parsing address %s\n", __func__, buf);
 796                 goto out_free_da;
 797         }
 798
 799         portstr++;
 800         sscanf(portstr, "%d-%d", &tmp[0], &tmp[1]);
 801         port = htons((tmp[0] << 8) | (tmp[1]));
 802
 803         switch (da->da_addr.ss_family) {
 804         case AF_INET:
 805                 ((struct sockaddr_in *)&da->da_addr)->sin_port = port;
 806                 da->da_addrlen = sizeof(struct sockaddr_in);
 807                 match_netid = "tcp";
 808                 match_netid_len = 3;
 809                 break;
 810
 811         case AF_INET6:
 812                 ((struct sockaddr_in6 *)&da->da_addr)->sin6_port = port;
 813                 da->da_addrlen = sizeof(struct sockaddr_in6);
 814                 match_netid = "tcp6";
 815                 match_netid_len = 4;
 816                 startsep = "[";
 817                 endsep = "]";
 818                 break;
 819
 820         default:
 821                 dprintk("%s: unsupported address family: %u\n",
 822                         __func__, da->da_addr.ss_family);
 823                 goto out_free_da;
 824         }
 825
 826         if (nlen != match_netid_len || strncmp(netid, match_netid, nlen)) {
 827                 dprintk("%s: ERROR: r_netid \"%s\" != \"%s\"\n",
 828                         __func__, netid, match_netid);
 829                 goto out_free_da;
 830         }
 831
 832         /* save human readable address */
 833         len = strlen(startsep) + strlen(buf) + strlen(endsep) + 7;
 834         da->da_remotestr = kzalloc(len, gfp_flags);
 835
 836         /* NULL is ok, only used for dprintk */
 837         if (da->da_remotestr)
 838                 snprintf(da->da_remotestr, len, "%s%s%s:%u", startsep,
 839                          buf, endsep, ntohs(port));
 840
 841         dprintk("%s: Parsed DS addr %s\n", __func__, da->da_remotestr);
 842         kfree(buf);
 843         kfree(netid);
 844         return da;
 845
 846 out_free_da:
 847         kfree(da);
 848 out_free_buf:
 849         dprintk("%s: Error parsing DS addr: %s\n", __func__, buf);
 850         kfree(buf);
 851 out_free_netid:
 852         kfree(netid);
 853 out_err:
 854         return NULL;
 855 }
 856 EXPORT_SYMBOL_GPL(nfs4_decode_mp_ds_addr);
 857
 858 void
 859 pnfs_layout_mark_request_commit(struct nfs_page *req,
 860                                 struct pnfs_layout_segment *lseg,
 861                                 struct nfs_commit_info *cinfo,
 862                                 u32 ds_commit_idx)
 863 {
 864         struct list_head *list;
 865         struct pnfs_commit_bucket *buckets;
 866
 867         spin_lock(cinfo->lock);
 868         buckets = cinfo->ds->buckets;
 869         list = &buckets[ds_commit_idx].written;
 870         if (list_empty(list)) {
 871                 if (!pnfs_is_valid_lseg(lseg)) {
 872                         spin_unlock(cinfo->lock);
 873                         cinfo->completion_ops->resched_write(cinfo, req);
 874                         return;
 875                 }
 876                 /* Non-empty buckets hold a reference on the lseg.  That ref
 877                  * is normally transferred to the COMMIT call and released
 878                  * there.  It could also be released if the last req is pulled
 879                  * off due to a rewrite, in which case it will be done in
 880                  * pnfs_common_clear_request_commit
 881                  */
 882                 WARN_ON_ONCE(buckets[ds_commit_idx].wlseg != NULL);
 883                 buckets[ds_commit_idx].wlseg = pnfs_get_lseg(lseg);
 884         }
 885         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
 886         cinfo->ds->nwritten++;
 887
 888         nfs_request_add_commit_list_locked(req, list, cinfo);
 889         spin_unlock(cinfo->lock);
 890         nfs_mark_page_unstable(req->wb_page, cinfo);
 891 }
 892 EXPORT_SYMBOL_GPL(pnfs_layout_mark_request_commit);
 893
 894 int
 895 pnfs_nfs_generic_sync(struct inode *inode, bool datasync)
 896 {
 897         if (datasync)
 898                 return 0;
 899         return pnfs_layoutcommit_inode(inode, true);
 900 }
 901 EXPORT_SYMBOL_GPL(pnfs_nfs_generic_sync);
 902