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