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