fs/nfs/pnfs_nfs.c

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