fs/nfs/flexfilelayout/flexfilelayout.c

   1 /*
   2  * Module for pnfs flexfile layout driver.
   3  *
   4  * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
   5  *
   6  * Tao Peng <bergwolf@primarydata.com>
   7  */
   8
   9 #include <linux/nfs_fs.h>
  10 #include <linux/nfs_page.h>
  11 #include <linux/module.h>
  12
  13 #include <linux/sunrpc/metrics.h>
  14
  15 #include "flexfilelayout.h"
  16 #include "../nfs4session.h"
  17 #include "../nfs4idmap.h"
  18 #include "../internal.h"
  19 #include "../delegation.h"
  20 #include "../nfs4trace.h"
  21 #include "../iostat.h"
  22 #include "../nfs.h"
  23 #include "../nfs42.h"
  24
  25 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  26
  27 #define FF_LAYOUT_POLL_RETRY_MAX     (15*HZ)
  28
  29 static struct group_info        *ff_zero_group;
  30
  31 static struct pnfs_layout_hdr *
  32 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
  33 {
  34         struct nfs4_flexfile_layout *ffl;
  35
  36         ffl = kzalloc(sizeof(*ffl), gfp_flags);
  37         if (ffl) {
  38                 INIT_LIST_HEAD(&ffl->error_list);
  39                 INIT_LIST_HEAD(&ffl->mirrors);
  40                 return &ffl->generic_hdr;
  41         } else
  42                 return NULL;
  43 }
  44
  45 static void
  46 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo)
  47 {
  48         struct nfs4_ff_layout_ds_err *err, *n;
  49
  50         list_for_each_entry_safe(err, n, &FF_LAYOUT_FROM_HDR(lo)->error_list,
  51                                  list) {
  52                 list_del(&err->list);
  53                 kfree(err);
  54         }
  55         kfree(FF_LAYOUT_FROM_HDR(lo));
  56 }
  57
  58 static int decode_pnfs_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid)
  59 {
  60         __be32 *p;
  61
  62         p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE);
  63         if (unlikely(p == NULL))
  64                 return -ENOBUFS;
  65         stateid->type = NFS4_PNFS_DS_STATEID_TYPE;
  66         memcpy(stateid->data, p, NFS4_STATEID_SIZE);
  67         dprintk("%s: stateid id= [%x%x%x%x]\n", __func__,
  68                 p[0], p[1], p[2], p[3]);
  69         return 0;
  70 }
  71
  72 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid)
  73 {
  74         __be32 *p;
  75
  76         p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE);
  77         if (unlikely(!p))
  78                 return -ENOBUFS;
  79         memcpy(devid, p, NFS4_DEVICEID4_SIZE);
  80         nfs4_print_deviceid(devid);
  81         return 0;
  82 }
  83
  84 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh)
  85 {
  86         __be32 *p;
  87
  88         p = xdr_inline_decode(xdr, 4);
  89         if (unlikely(!p))
  90                 return -ENOBUFS;
  91         fh->size = be32_to_cpup(p++);
  92         if (fh->size > sizeof(struct nfs_fh)) {
  93                 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n",
  94                        fh->size);
  95                 return -EOVERFLOW;
  96         }
  97         /* fh.data */
  98         p = xdr_inline_decode(xdr, fh->size);
  99         if (unlikely(!p))
 100                 return -ENOBUFS;
 101         memcpy(&fh->data, p, fh->size);
 102         dprintk("%s: fh len %d\n", __func__, fh->size);
 103
 104         return 0;
 105 }
 106
 107 /*
 108  * Currently only stringified uids and gids are accepted.
 109  * I.e., kerberos is not supported to the DSes, so no pricipals.
 110  *
 111  * That means that one common function will suffice, but when
 112  * principals are added, this should be split to accomodate
 113  * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid().
 114  */
 115 static int
 116 decode_name(struct xdr_stream *xdr, u32 *id)
 117 {
 118         __be32 *p;
 119         int len;
 120
 121         /* opaque_length(4)*/
 122         p = xdr_inline_decode(xdr, 4);
 123         if (unlikely(!p))
 124                 return -ENOBUFS;
 125         len = be32_to_cpup(p++);
 126         if (len < 0)
 127                 return -EINVAL;
 128
 129         dprintk("%s: len %u\n", __func__, len);
 130
 131         /* opaque body */
 132         p = xdr_inline_decode(xdr, len);
 133         if (unlikely(!p))
 134                 return -ENOBUFS;
 135
 136         if (!nfs_map_string_to_numeric((char *)p, len, id))
 137                 return -EINVAL;
 138
 139         return 0;
 140 }
 141
 142 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1,
 143                 const struct nfs4_ff_layout_mirror *m2)
 144 {
 145         int i, j;
 146
 147         if (m1->fh_versions_cnt != m2->fh_versions_cnt)
 148                 return false;
 149         for (i = 0; i < m1->fh_versions_cnt; i++) {
 150                 bool found_fh = false;
 151                 for (j = 0; j < m2->fh_versions_cnt; j++) {
 152                         if (nfs_compare_fh(&m1->fh_versions[i],
 153                                         &m2->fh_versions[j]) == 0) {
 154                                 found_fh = true;
 155                                 break;
 156                         }
 157                 }
 158                 if (!found_fh)
 159                         return false;
 160         }
 161         return true;
 162 }
 163
 164 static struct nfs4_ff_layout_mirror *
 165 ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
 166                 struct nfs4_ff_layout_mirror *mirror)
 167 {
 168         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
 169         struct nfs4_ff_layout_mirror *pos;
 170         struct inode *inode = lo->plh_inode;
 171
 172         spin_lock(&inode->i_lock);
 173         list_for_each_entry(pos, &ff_layout->mirrors, mirrors) {
 174                 if (mirror->mirror_ds != pos->mirror_ds)
 175                         continue;
 176                 if (!ff_mirror_match_fh(mirror, pos))
 177                         continue;
 178                 if (atomic_inc_not_zero(&pos->ref)) {
 179                         spin_unlock(&inode->i_lock);
 180                         return pos;
 181                 }
 182         }
 183         list_add(&mirror->mirrors, &ff_layout->mirrors);
 184         mirror->layout = lo;
 185         spin_unlock(&inode->i_lock);
 186         return mirror;
 187 }
 188
 189 static void
 190 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror)
 191 {
 192         struct inode *inode;
 193         if (mirror->layout == NULL)
 194                 return;
 195         inode = mirror->layout->plh_inode;
 196         spin_lock(&inode->i_lock);
 197         list_del(&mirror->mirrors);
 198         spin_unlock(&inode->i_lock);
 199         mirror->layout = NULL;
 200 }
 201
 202 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(gfp_t gfp_flags)
 203 {
 204         struct nfs4_ff_layout_mirror *mirror;
 205
 206         mirror = kzalloc(sizeof(*mirror), gfp_flags);
 207         if (mirror != NULL) {
 208                 spin_lock_init(&mirror->lock);
 209                 atomic_set(&mirror->ref, 1);
 210                 INIT_LIST_HEAD(&mirror->mirrors);
 211         }
 212         return mirror;
 213 }
 214
 215 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror)
 216 {
 217         struct rpc_cred *cred;
 218
 219         ff_layout_remove_mirror(mirror);
 220         kfree(mirror->fh_versions);
 221         cred = rcu_access_pointer(mirror->ro_cred);
 222         if (cred)
 223                 put_rpccred(cred);
 224         cred = rcu_access_pointer(mirror->rw_cred);
 225         if (cred)
 226                 put_rpccred(cred);
 227         nfs4_ff_layout_put_deviceid(mirror->mirror_ds);
 228         kfree(mirror);
 229 }
 230
 231 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror)
 232 {
 233         if (mirror != NULL && atomic_dec_and_test(&mirror->ref))
 234                 ff_layout_free_mirror(mirror);
 235 }
 236
 237 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls)
 238 {
 239         int i;
 240
 241         if (fls->mirror_array) {
 242                 for (i = 0; i < fls->mirror_array_cnt; i++) {
 243                         /* normally mirror_ds is freed in
 244                          * .free_deviceid_node but we still do it here
 245                          * for .alloc_lseg error path */
 246                         ff_layout_put_mirror(fls->mirror_array[i]);
 247                 }
 248                 kfree(fls->mirror_array);
 249                 fls->mirror_array = NULL;
 250         }
 251 }
 252
 253 static int ff_layout_check_layout(struct nfs4_layoutget_res *lgr)
 254 {
 255         int ret = 0;
 256
 257         dprintk("--> %s\n", __func__);
 258
 259         /* FIXME: remove this check when layout segment support is added */
 260         if (lgr->range.offset != 0 ||
 261             lgr->range.length != NFS4_MAX_UINT64) {
 262                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 263                         __func__);
 264                 ret = -EINVAL;
 265         }
 266
 267         dprintk("--> %s returns %d\n", __func__, ret);
 268         return ret;
 269 }
 270
 271 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls)
 272 {
 273         if (fls) {
 274                 ff_layout_free_mirror_array(fls);
 275                 kfree(fls);
 276         }
 277 }
 278
 279 static bool
 280 ff_lseg_range_is_after(const struct pnfs_layout_range *l1,
 281                 const struct pnfs_layout_range *l2)
 282 {
 283         u64 end1, end2;
 284
 285         if (l1->iomode != l2->iomode)
 286                 return l1->iomode != IOMODE_READ;
 287         end1 = pnfs_calc_offset_end(l1->offset, l1->length);
 288         end2 = pnfs_calc_offset_end(l2->offset, l2->length);
 289         if (end1 < l2->offset)
 290                 return false;
 291         if (end2 < l1->offset)
 292                 return true;
 293         return l2->offset <= l1->offset;
 294 }
 295
 296 static bool
 297 ff_lseg_merge(struct pnfs_layout_segment *new,
 298                 struct pnfs_layout_segment *old)
 299 {
 300         u64 new_end, old_end;
 301
 302         if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags))
 303                 return false;
 304         if (new->pls_range.iomode != old->pls_range.iomode)
 305                 return false;
 306         old_end = pnfs_calc_offset_end(old->pls_range.offset,
 307                         old->pls_range.length);
 308         if (old_end < new->pls_range.offset)
 309                 return false;
 310         new_end = pnfs_calc_offset_end(new->pls_range.offset,
 311                         new->pls_range.length);
 312         if (new_end < old->pls_range.offset)
 313                 return false;
 314
 315         /* Mergeable: copy info from 'old' to 'new' */
 316         if (new_end < old_end)
 317                 new_end = old_end;
 318         if (new->pls_range.offset < old->pls_range.offset)
 319                 new->pls_range.offset = old->pls_range.offset;
 320         new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset,
 321                         new_end);
 322         if (test_bit(NFS_LSEG_ROC, &old->pls_flags))
 323                 set_bit(NFS_LSEG_ROC, &new->pls_flags);
 324         return true;
 325 }
 326
 327 static void
 328 ff_layout_add_lseg(struct pnfs_layout_hdr *lo,
 329                 struct pnfs_layout_segment *lseg,
 330                 struct list_head *free_me)
 331 {
 332         pnfs_generic_layout_insert_lseg(lo, lseg,
 333                         ff_lseg_range_is_after,
 334                         ff_lseg_merge,
 335                         free_me);
 336 }
 337
 338 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls)
 339 {
 340         int i, j;
 341
 342         for (i = 0; i < fls->mirror_array_cnt - 1; i++) {
 343                 for (j = i + 1; j < fls->mirror_array_cnt; j++)
 344                         if (fls->mirror_array[i]->efficiency <
 345                             fls->mirror_array[j]->efficiency)
 346                                 swap(fls->mirror_array[i],
 347                                      fls->mirror_array[j]);
 348         }
 349 }
 350
 351 static void ff_layout_mark_devices_valid(struct nfs4_ff_layout_segment *fls)
 352 {
 353         struct nfs4_deviceid_node *node;
 354         int i;
 355
 356         if (!(fls->flags & FF_FLAGS_NO_IO_THRU_MDS))
 357                 return;
 358         for (i = 0; i < fls->mirror_array_cnt; i++) {
 359                 node = &fls->mirror_array[i]->mirror_ds->id_node;
 360                 clear_bit(NFS_DEVICEID_UNAVAILABLE, &node->flags);
 361         }
 362 }
 363
 364 static struct pnfs_layout_segment *
 365 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh,
 366                      struct nfs4_layoutget_res *lgr,
 367                      gfp_t gfp_flags)
 368 {
 369         struct pnfs_layout_segment *ret;
 370         struct nfs4_ff_layout_segment *fls = NULL;
 371         struct xdr_stream stream;
 372         struct xdr_buf buf;
 373         struct page *scratch;
 374         u64 stripe_unit;
 375         u32 mirror_array_cnt;
 376         __be32 *p;
 377         int i, rc;
 378
 379         dprintk("--> %s\n", __func__);
 380         scratch = alloc_page(gfp_flags);
 381         if (!scratch)
 382                 return ERR_PTR(-ENOMEM);
 383
 384         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages,
 385                               lgr->layoutp->len);
 386         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 387
 388         /* stripe unit and mirror_array_cnt */
 389         rc = -EIO;
 390         p = xdr_inline_decode(&stream, 8 + 4);
 391         if (!p)
 392                 goto out_err_free;
 393
 394         p = xdr_decode_hyper(p, &stripe_unit);
 395         mirror_array_cnt = be32_to_cpup(p++);
 396         dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__,
 397                 stripe_unit, mirror_array_cnt);
 398
 399         if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT ||
 400             mirror_array_cnt == 0)
 401                 goto out_err_free;
 402
 403         rc = -ENOMEM;
 404         fls = kzalloc(sizeof(*fls), gfp_flags);
 405         if (!fls)
 406                 goto out_err_free;
 407
 408         fls->mirror_array_cnt = mirror_array_cnt;
 409         fls->stripe_unit = stripe_unit;
 410         fls->mirror_array = kcalloc(fls->mirror_array_cnt,
 411                                     sizeof(fls->mirror_array[0]), gfp_flags);
 412         if (fls->mirror_array == NULL)
 413                 goto out_err_free;
 414
 415         for (i = 0; i < fls->mirror_array_cnt; i++) {
 416                 struct nfs4_ff_layout_mirror *mirror;
 417                 struct nfs4_deviceid devid;
 418                 struct nfs4_deviceid_node *idnode;
 419                 struct auth_cred acred = { .group_info = ff_zero_group };
 420                 struct rpc_cred __rcu *cred;
 421                 u32 ds_count, fh_count, id;
 422                 int j;
 423
 424                 rc = -EIO;
 425                 p = xdr_inline_decode(&stream, 4);
 426                 if (!p)
 427                         goto out_err_free;
 428                 ds_count = be32_to_cpup(p);
 429
 430                 /* FIXME: allow for striping? */
 431                 if (ds_count != 1)
 432                         goto out_err_free;
 433
 434                 fls->mirror_array[i] = ff_layout_alloc_mirror(gfp_flags);
 435                 if (fls->mirror_array[i] == NULL) {
 436                         rc = -ENOMEM;
 437                         goto out_err_free;
 438                 }
 439
 440                 fls->mirror_array[i]->ds_count = ds_count;
 441
 442                 /* deviceid */
 443                 rc = decode_deviceid(&stream, &devid);
 444                 if (rc)
 445                         goto out_err_free;
 446
 447                 idnode = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
 448                                                 &devid, lh->plh_lc_cred,
 449                                                 gfp_flags);
 450                 /*
 451                  * upon success, mirror_ds is allocated by previous
 452                  * getdeviceinfo, or newly by .alloc_deviceid_node
 453                  * nfs4_find_get_deviceid failure is indeed getdeviceinfo falure
 454                  */
 455                 if (idnode)
 456                         fls->mirror_array[i]->mirror_ds =
 457                                 FF_LAYOUT_MIRROR_DS(idnode);
 458                 else
 459                         goto out_err_free;
 460
 461                 /* efficiency */
 462                 rc = -EIO;
 463                 p = xdr_inline_decode(&stream, 4);
 464                 if (!p)
 465                         goto out_err_free;
 466                 fls->mirror_array[i]->efficiency = be32_to_cpup(p);
 467
 468                 /* stateid */
 469                 rc = decode_pnfs_stateid(&stream, &fls->mirror_array[i]->stateid);
 470                 if (rc)
 471                         goto out_err_free;
 472
 473                 /* fh */
 474                 p = xdr_inline_decode(&stream, 4);
 475                 if (!p)
 476                         goto out_err_free;
 477                 fh_count = be32_to_cpup(p);
 478
 479                 fls->mirror_array[i]->fh_versions =
 480                         kzalloc(fh_count * sizeof(struct nfs_fh),
 481                                 gfp_flags);
 482                 if (fls->mirror_array[i]->fh_versions == NULL) {
 483                         rc = -ENOMEM;
 484                         goto out_err_free;
 485                 }
 486
 487                 for (j = 0; j < fh_count; j++) {
 488                         rc = decode_nfs_fh(&stream,
 489                                            &fls->mirror_array[i]->fh_versions[j]);
 490                         if (rc)
 491                                 goto out_err_free;
 492                 }
 493
 494                 fls->mirror_array[i]->fh_versions_cnt = fh_count;
 495
 496                 /* user */
 497                 rc = decode_name(&stream, &id);
 498                 if (rc)
 499                         goto out_err_free;
 500
 501                 acred.uid = make_kuid(&init_user_ns, id);
 502
 503                 /* group */
 504                 rc = decode_name(&stream, &id);
 505                 if (rc)
 506                         goto out_err_free;
 507
 508                 acred.gid = make_kgid(&init_user_ns, id);
 509
 510                 /* find the cred for it */
 511                 rcu_assign_pointer(cred, rpc_lookup_generic_cred(&acred, 0, gfp_flags));
 512                 if (IS_ERR(cred)) {
 513                         rc = PTR_ERR(cred);
 514                         goto out_err_free;
 515                 }
 516
 517                 if (lgr->range.iomode == IOMODE_READ)
 518                         rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred);
 519                 else
 520                         rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred);
 521
 522                 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]);
 523                 if (mirror != fls->mirror_array[i]) {
 524                         /* swap cred ptrs so free_mirror will clean up old */
 525                         if (lgr->range.iomode == IOMODE_READ) {
 526                                 cred = xchg(&mirror->ro_cred, cred);
 527                                 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred);
 528                         } else {
 529                                 cred = xchg(&mirror->rw_cred, cred);
 530                                 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred);
 531                         }
 532                         ff_layout_free_mirror(fls->mirror_array[i]);
 533                         fls->mirror_array[i] = mirror;
 534                 }
 535
 536                 dprintk("%s: iomode %s uid %u gid %u\n", __func__,
 537                         lgr->range.iomode == IOMODE_READ ? "READ" : "RW",
 538                         from_kuid(&init_user_ns, acred.uid),
 539                         from_kgid(&init_user_ns, acred.gid));
 540         }
 541
 542         p = xdr_inline_decode(&stream, 4);
 543         if (!p)
 544                 goto out_sort_mirrors;
 545         fls->flags = be32_to_cpup(p);
 546
 547         p = xdr_inline_decode(&stream, 4);
 548         if (!p)
 549                 goto out_sort_mirrors;
 550         for (i=0; i < fls->mirror_array_cnt; i++)
 551                 fls->mirror_array[i]->report_interval = be32_to_cpup(p);
 552
 553 out_sort_mirrors:
 554         ff_layout_sort_mirrors(fls);
 555         rc = ff_layout_check_layout(lgr);
 556         if (rc)
 557                 goto out_err_free;
 558         ff_layout_mark_devices_valid(fls);
 559
 560         ret = &fls->generic_hdr;
 561         dprintk("<-- %s (success)\n", __func__);
 562 out_free_page:
 563         __free_page(scratch);
 564         return ret;
 565 out_err_free:
 566         _ff_layout_free_lseg(fls);
 567         ret = ERR_PTR(rc);
 568         dprintk("<-- %s (%d)\n", __func__, rc);
 569         goto out_free_page;
 570 }
 571
 572 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout)
 573 {
 574         struct pnfs_layout_segment *lseg;
 575
 576         list_for_each_entry(lseg, &layout->plh_segs, pls_list)
 577                 if (lseg->pls_range.iomode == IOMODE_RW)
 578                         return true;
 579
 580         return false;
 581 }
 582
 583 static void
 584 ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
 585 {
 586         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 587
 588         dprintk("--> %s\n", __func__);
 589
 590         if (lseg->pls_range.iomode == IOMODE_RW) {
 591                 struct nfs4_flexfile_layout *ffl;
 592                 struct inode *inode;
 593
 594                 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
 595                 inode = ffl->generic_hdr.plh_inode;
 596                 spin_lock(&inode->i_lock);
 597                 if (!ff_layout_has_rw_segments(lseg->pls_layout)) {
 598                         ffl->commit_info.nbuckets = 0;
 599                         kfree(ffl->commit_info.buckets);
 600                         ffl->commit_info.buckets = NULL;
 601                 }
 602                 spin_unlock(&inode->i_lock);
 603         }
 604         _ff_layout_free_lseg(fls);
 605 }
 606
 607 /* Return 1 until we have multiple lsegs support */
 608 static int
 609 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
 610 {
 611         return 1;
 612 }
 613
 614 static void
 615 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 616 {
 617         /* first IO request? */
 618         if (atomic_inc_return(&timer->n_ops) == 1) {
 619                 timer->start_time = now;
 620         }
 621 }
 622
 623 static ktime_t
 624 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 625 {
 626         ktime_t start;
 627
 628         if (atomic_dec_return(&timer->n_ops) < 0)
 629                 WARN_ON_ONCE(1);
 630
 631         start = timer->start_time;
 632         timer->start_time = now;
 633         return ktime_sub(now, start);
 634 }
 635
 636 static bool
 637 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
 638                             struct nfs4_ff_layoutstat *layoutstat,
 639                             ktime_t now)
 640 {
 641         static const ktime_t notime = {0};
 642         s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL;
 643
 644         nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now);
 645         if (ktime_equal(mirror->start_time, notime))
 646                 mirror->start_time = now;
 647         if (ktime_equal(mirror->last_report_time, notime))
 648                 mirror->last_report_time = now;
 649         if (mirror->report_interval != 0)
 650                 report_interval = (s64)mirror->report_interval * 1000LL;
 651         else if (layoutstats_timer != 0)
 652                 report_interval = (s64)layoutstats_timer * 1000LL;
 653         if (ktime_to_ms(ktime_sub(now, mirror->last_report_time)) >=
 654                         report_interval) {
 655                 mirror->last_report_time = now;
 656                 return true;
 657         }
 658
 659         return false;
 660 }
 661
 662 static void
 663 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
 664                 __u64 requested)
 665 {
 666         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 667
 668         iostat->ops_requested++;
 669         iostat->bytes_requested += requested;
 670 }
 671
 672 static void
 673 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
 674                 __u64 requested,
 675                 __u64 completed,
 676                 ktime_t time_completed,
 677                 ktime_t time_started)
 678 {
 679         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 680         ktime_t completion_time = ktime_sub(time_completed, time_started);
 681         ktime_t timer;
 682
 683         iostat->ops_completed++;
 684         iostat->bytes_completed += completed;
 685         iostat->bytes_not_delivered += requested - completed;
 686
 687         timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed);
 688         iostat->total_busy_time =
 689                         ktime_add(iostat->total_busy_time, timer);
 690         iostat->aggregate_completion_time =
 691                         ktime_add(iostat->aggregate_completion_time,
 692                                         completion_time);
 693 }
 694
 695 static void
 696 nfs4_ff_layout_stat_io_start_read(struct inode *inode,
 697                 struct nfs4_ff_layout_mirror *mirror,
 698                 __u64 requested, ktime_t now)
 699 {
 700         bool report;
 701
 702         spin_lock(&mirror->lock);
 703         report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now);
 704         nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
 705         spin_unlock(&mirror->lock);
 706
 707         if (report)
 708                 pnfs_report_layoutstat(inode, GFP_KERNEL);
 709 }
 710
 711 static void
 712 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
 713                 struct nfs4_ff_layout_mirror *mirror,
 714                 __u64 requested,
 715                 __u64 completed)
 716 {
 717         spin_lock(&mirror->lock);
 718         nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
 719                         requested, completed,
 720                         ktime_get(), task->tk_start);
 721         spin_unlock(&mirror->lock);
 722 }
 723
 724 static void
 725 nfs4_ff_layout_stat_io_start_write(struct inode *inode,
 726                 struct nfs4_ff_layout_mirror *mirror,
 727                 __u64 requested, ktime_t now)
 728 {
 729         bool report;
 730
 731         spin_lock(&mirror->lock);
 732         report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now);
 733         nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
 734         spin_unlock(&mirror->lock);
 735
 736         if (report)
 737                 pnfs_report_layoutstat(inode, GFP_NOIO);
 738 }
 739
 740 static void
 741 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
 742                 struct nfs4_ff_layout_mirror *mirror,
 743                 __u64 requested,
 744                 __u64 completed,
 745                 enum nfs3_stable_how committed)
 746 {
 747         if (committed == NFS_UNSTABLE)
 748                 requested = completed = 0;
 749
 750         spin_lock(&mirror->lock);
 751         nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
 752                         requested, completed, ktime_get(), task->tk_start);
 753         spin_unlock(&mirror->lock);
 754 }
 755
 756 static int
 757 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 758                             struct nfs_commit_info *cinfo,
 759                             gfp_t gfp_flags)
 760 {
 761         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 762         struct pnfs_commit_bucket *buckets;
 763         int size;
 764
 765         if (cinfo->ds->nbuckets != 0) {
 766                 /* This assumes there is only one RW lseg per file.
 767                  * To support multiple lseg per file, we need to
 768                  * change struct pnfs_commit_bucket to allow dynamic
 769                  * increasing nbuckets.
 770                  */
 771                 return 0;
 772         }
 773
 774         size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg);
 775
 776         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 777                           gfp_flags);
 778         if (!buckets)
 779                 return -ENOMEM;
 780         else {
 781                 int i;
 782
 783                 spin_lock(&cinfo->inode->i_lock);
 784                 if (cinfo->ds->nbuckets != 0)
 785                         kfree(buckets);
 786                 else {
 787                         cinfo->ds->buckets = buckets;
 788                         cinfo->ds->nbuckets = size;
 789                         for (i = 0; i < size; i++) {
 790                                 INIT_LIST_HEAD(&buckets[i].written);
 791                                 INIT_LIST_HEAD(&buckets[i].committing);
 792                                 /* mark direct verifier as unset */
 793                                 buckets[i].direct_verf.committed =
 794                                         NFS_INVALID_STABLE_HOW;
 795                         }
 796                 }
 797                 spin_unlock(&cinfo->inode->i_lock);
 798                 return 0;
 799         }
 800 }
 801
 802 static struct nfs4_pnfs_ds *
 803 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg,
 804                                   int start_idx,
 805                                   int *best_idx)
 806 {
 807         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 808         struct nfs4_pnfs_ds *ds;
 809         int idx;
 810
 811         /* mirrors are sorted by efficiency */
 812         for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) {
 813                 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
 814                 if (ds) {
 815                         *best_idx = idx;
 816                         return ds;
 817                 }
 818         }
 819
 820         return NULL;
 821 }
 822
 823 static void
 824 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio,
 825                       struct nfs_page *req,
 826                       bool strict_iomode)
 827 {
 828 retry_strict:
 829         pnfs_put_lseg(pgio->pg_lseg);
 830         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 831                                            req->wb_context,
 832                                            0,
 833                                            NFS4_MAX_UINT64,
 834                                            IOMODE_READ,
 835                                            strict_iomode,
 836                                            GFP_KERNEL);
 837         if (IS_ERR(pgio->pg_lseg)) {
 838                 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 839                 pgio->pg_lseg = NULL;
 840         }
 841
 842         /* If we don't have checking, do get a IOMODE_RW
 843          * segment, and the server wants to avoid READs
 844          * there, then retry!
 845          */
 846         if (pgio->pg_lseg && !strict_iomode &&
 847             ff_layout_avoid_read_on_rw(pgio->pg_lseg)) {
 848                 strict_iomode = true;
 849                 goto retry_strict;
 850         }
 851 }
 852
 853 static void
 854 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 855                         struct nfs_page *req)
 856 {
 857         struct nfs_pgio_mirror *pgm;
 858         struct nfs4_ff_layout_mirror *mirror;
 859         struct nfs4_pnfs_ds *ds;
 860         int ds_idx;
 861
 862         /* Use full layout for now */
 863         if (!pgio->pg_lseg)
 864                 ff_layout_pg_get_read(pgio, req, false);
 865         else if (ff_layout_avoid_read_on_rw(pgio->pg_lseg))
 866                 ff_layout_pg_get_read(pgio, req, true);
 867
 868         /* If no lseg, fall back to read through mds */
 869         if (pgio->pg_lseg == NULL)
 870                 goto out_mds;
 871
 872         ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx);
 873         if (!ds) {
 874                 if (ff_layout_no_fallback_to_mds(pgio->pg_lseg))
 875                         goto out_pnfs;
 876                 else
 877                         goto out_mds;
 878         }
 879
 880         mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
 881
 882         pgio->pg_mirror_idx = ds_idx;
 883
 884         /* read always uses only one mirror - idx 0 for pgio layer */
 885         pgm = &pgio->pg_mirrors[0];
 886         pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
 887
 888         return;
 889 out_mds:
 890         pnfs_put_lseg(pgio->pg_lseg);
 891         pgio->pg_lseg = NULL;
 892         nfs_pageio_reset_read_mds(pgio);
 893         return;
 894
 895 out_pnfs:
 896         pnfs_set_lo_fail(pgio->pg_lseg);
 897         pnfs_put_lseg(pgio->pg_lseg);
 898         pgio->pg_lseg = NULL;
 899 }
 900
 901 static void
 902 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 903                         struct nfs_page *req)
 904 {
 905         struct nfs4_ff_layout_mirror *mirror;
 906         struct nfs_pgio_mirror *pgm;
 907         struct nfs_commit_info cinfo;
 908         struct nfs4_pnfs_ds *ds;
 909         int i;
 910         int status;
 911
 912         if (!pgio->pg_lseg) {
 913                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 914                                                    req->wb_context,
 915                                                    0,
 916                                                    NFS4_MAX_UINT64,
 917                                                    IOMODE_RW,
 918                                                    false,
 919                                                    GFP_NOFS);
 920                 if (IS_ERR(pgio->pg_lseg)) {
 921                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 922                         pgio->pg_lseg = NULL;
 923                         return;
 924                 }
 925         }
 926         /* If no lseg, fall back to write through mds */
 927         if (pgio->pg_lseg == NULL)
 928                 goto out_mds;
 929
 930         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 931         status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 932         if (status < 0)
 933                 goto out_mds;
 934
 935         /* Use a direct mapping of ds_idx to pgio mirror_idx */
 936         if (WARN_ON_ONCE(pgio->pg_mirror_count !=
 937             FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)))
 938                 goto out_mds;
 939
 940         for (i = 0; i < pgio->pg_mirror_count; i++) {
 941                 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
 942                 if (!ds) {
 943                         if (ff_layout_no_fallback_to_mds(pgio->pg_lseg))
 944                                 goto out_pnfs;
 945                         else
 946                                 goto out_mds;
 947                 }
 948                 pgm = &pgio->pg_mirrors[i];
 949                 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
 950                 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
 951         }
 952
 953         return;
 954
 955 out_mds:
 956         pnfs_put_lseg(pgio->pg_lseg);
 957         pgio->pg_lseg = NULL;
 958         nfs_pageio_reset_write_mds(pgio);
 959         return;
 960
 961 out_pnfs:
 962         pnfs_set_lo_fail(pgio->pg_lseg);
 963         pnfs_put_lseg(pgio->pg_lseg);
 964         pgio->pg_lseg = NULL;
 965 }
 966
 967 static unsigned int
 968 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
 969                                     struct nfs_page *req)
 970 {
 971         if (!pgio->pg_lseg) {
 972                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 973                                                    req->wb_context,
 974                                                    0,
 975                                                    NFS4_MAX_UINT64,
 976                                                    IOMODE_RW,
 977                                                    false,
 978                                                    GFP_NOFS);
 979                 if (IS_ERR(pgio->pg_lseg)) {
 980                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 981                         pgio->pg_lseg = NULL;
 982                         goto out;
 983                 }
 984         }
 985         if (pgio->pg_lseg)
 986                 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
 987
 988         /* no lseg means that pnfs is not in use, so no mirroring here */
 989         nfs_pageio_reset_write_mds(pgio);
 990 out:
 991         return 1;
 992 }
 993
 994 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
 995         .pg_init = ff_layout_pg_init_read,
 996         .pg_test = pnfs_generic_pg_test,
 997         .pg_doio = pnfs_generic_pg_readpages,
 998         .pg_cleanup = pnfs_generic_pg_cleanup,
 999 };
1000
1001 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
1002         .pg_init = ff_layout_pg_init_write,
1003         .pg_test = pnfs_generic_pg_test,
1004         .pg_doio = pnfs_generic_pg_writepages,
1005         .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
1006         .pg_cleanup = pnfs_generic_pg_cleanup,
1007 };
1008
1009 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
1010 {
1011         struct rpc_task *task = &hdr->task;
1012
1013         pnfs_layoutcommit_inode(hdr->inode, false);
1014
1015         if (retry_pnfs) {
1016                 dprintk("%s Reset task %5u for i/o through pNFS "
1017                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1018                         hdr->task.tk_pid,
1019                         hdr->inode->i_sb->s_id,
1020                         (unsigned long long)NFS_FILEID(hdr->inode),
1021                         hdr->args.count,
1022                         (unsigned long long)hdr->args.offset);
1023
1024                 hdr->completion_ops->reschedule_io(hdr);
1025                 return;
1026         }
1027
1028         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1029                 dprintk("%s Reset task %5u for i/o through MDS "
1030                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1031                         hdr->task.tk_pid,
1032                         hdr->inode->i_sb->s_id,
1033                         (unsigned long long)NFS_FILEID(hdr->inode),
1034                         hdr->args.count,
1035                         (unsigned long long)hdr->args.offset);
1036
1037                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
1038         }
1039 }
1040
1041 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
1042 {
1043         struct rpc_task *task = &hdr->task;
1044
1045         pnfs_layoutcommit_inode(hdr->inode, false);
1046
1047         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1048                 dprintk("%s Reset task %5u for i/o through MDS "
1049                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1050                         hdr->task.tk_pid,
1051                         hdr->inode->i_sb->s_id,
1052                         (unsigned long long)NFS_FILEID(hdr->inode),
1053                         hdr->args.count,
1054                         (unsigned long long)hdr->args.offset);
1055
1056                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
1057         }
1058 }
1059
1060 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
1061                                            struct nfs4_state *state,
1062                                            struct nfs_client *clp,
1063                                            struct pnfs_layout_segment *lseg,
1064                                            int idx)
1065 {
1066         struct pnfs_layout_hdr *lo = lseg->pls_layout;
1067         struct inode *inode = lo->plh_inode;
1068         struct nfs_server *mds_server = NFS_SERVER(inode);
1069
1070         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1071         struct nfs_client *mds_client = mds_server->nfs_client;
1072         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
1073
1074         if (task->tk_status >= 0)
1075                 return 0;
1076
1077         switch (task->tk_status) {
1078         /* MDS state errors */
1079         case -NFS4ERR_DELEG_REVOKED:
1080         case -NFS4ERR_ADMIN_REVOKED:
1081         case -NFS4ERR_BAD_STATEID:
1082                 if (state == NULL)
1083                         break;
1084                 nfs_remove_bad_delegation(state->inode);
1085         case -NFS4ERR_OPENMODE:
1086                 if (state == NULL)
1087                         break;
1088                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1089                         goto out_bad_stateid;
1090                 goto wait_on_recovery;
1091         case -NFS4ERR_EXPIRED:
1092                 if (state != NULL) {
1093                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1094                                 goto out_bad_stateid;
1095                 }
1096                 nfs4_schedule_lease_recovery(mds_client);
1097                 goto wait_on_recovery;
1098         /* DS session errors */
1099         case -NFS4ERR_BADSESSION:
1100         case -NFS4ERR_BADSLOT:
1101         case -NFS4ERR_BAD_HIGH_SLOT:
1102         case -NFS4ERR_DEADSESSION:
1103         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1104         case -NFS4ERR_SEQ_FALSE_RETRY:
1105         case -NFS4ERR_SEQ_MISORDERED:
1106                 dprintk("%s ERROR %d, Reset session. Exchangeid "
1107                         "flags 0x%x\n", __func__, task->tk_status,
1108                         clp->cl_exchange_flags);
1109                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1110                 break;
1111         case -NFS4ERR_DELAY:
1112         case -NFS4ERR_GRACE:
1113                 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1114                 break;
1115         case -NFS4ERR_RETRY_UNCACHED_REP:
1116                 break;
1117         /* Invalidate Layout errors */
1118         case -NFS4ERR_PNFS_NO_LAYOUT:
1119         case -ESTALE:           /* mapped NFS4ERR_STALE */
1120         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
1121         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
1122         case -NFS4ERR_FHEXPIRED:
1123         case -NFS4ERR_WRONG_TYPE:
1124                 dprintk("%s Invalid layout error %d\n", __func__,
1125                         task->tk_status);
1126                 /*
1127                  * Destroy layout so new i/o will get a new layout.
1128                  * Layout will not be destroyed until all current lseg
1129                  * references are put. Mark layout as invalid to resend failed
1130                  * i/o and all i/o waiting on the slot table to the MDS until
1131                  * layout is destroyed and a new valid layout is obtained.
1132                  */
1133                 pnfs_destroy_layout(NFS_I(inode));
1134                 rpc_wake_up(&tbl->slot_tbl_waitq);
1135                 goto reset;
1136         /* RPC connection errors */
1137         case -ECONNREFUSED:
1138         case -EHOSTDOWN:
1139         case -EHOSTUNREACH:
1140         case -ENETUNREACH:
1141         case -EIO:
1142         case -ETIMEDOUT:
1143         case -EPIPE:
1144                 dprintk("%s DS connection error %d\n", __func__,
1145                         task->tk_status);
1146                 nfs4_mark_deviceid_unavailable(devid);
1147                 rpc_wake_up(&tbl->slot_tbl_waitq);
1148                 /* fall through */
1149         default:
1150                 if (ff_layout_avoid_mds_available_ds(lseg))
1151                         return -NFS4ERR_RESET_TO_PNFS;
1152 reset:
1153                 dprintk("%s Retry through MDS. Error %d\n", __func__,
1154                         task->tk_status);
1155                 return -NFS4ERR_RESET_TO_MDS;
1156         }
1157 out:
1158         task->tk_status = 0;
1159         return -EAGAIN;
1160 out_bad_stateid:
1161         task->tk_status = -EIO;
1162         return 0;
1163 wait_on_recovery:
1164         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
1165         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
1166                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
1167         goto out;
1168 }
1169
1170 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
1171 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1172                                            struct pnfs_layout_segment *lseg,
1173                                            int idx)
1174 {
1175         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1176
1177         if (task->tk_status >= 0)
1178                 return 0;
1179
1180         switch (task->tk_status) {
1181         /* File access problems. Don't mark the device as unavailable */
1182         case -EACCES:
1183         case -ESTALE:
1184         case -EISDIR:
1185         case -EBADHANDLE:
1186         case -ELOOP:
1187         case -ENOSPC:
1188                 break;
1189         case -EJUKEBOX:
1190                 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1191                 goto out_retry;
1192         default:
1193                 dprintk("%s DS connection error %d\n", __func__,
1194                         task->tk_status);
1195                 nfs4_mark_deviceid_unavailable(devid);
1196         }
1197         /* FIXME: Need to prevent infinite looping here. */
1198         return -NFS4ERR_RESET_TO_PNFS;
1199 out_retry:
1200         task->tk_status = 0;
1201         rpc_restart_call_prepare(task);
1202         rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
1203         return -EAGAIN;
1204 }
1205
1206 static int ff_layout_async_handle_error(struct rpc_task *task,
1207                                         struct nfs4_state *state,
1208                                         struct nfs_client *clp,
1209                                         struct pnfs_layout_segment *lseg,
1210                                         int idx)
1211 {
1212         int vers = clp->cl_nfs_mod->rpc_vers->number;
1213
1214         switch (vers) {
1215         case 3:
1216                 return ff_layout_async_handle_error_v3(task, lseg, idx);
1217         case 4:
1218                 return ff_layout_async_handle_error_v4(task, state, clp,
1219                                                        lseg, idx);
1220         default:
1221                 /* should never happen */
1222                 WARN_ON_ONCE(1);
1223                 return 0;
1224         }
1225 }
1226
1227 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1228                                         int idx, u64 offset, u64 length,
1229                                         u32 status, int opnum, int error)
1230 {
1231         struct nfs4_ff_layout_mirror *mirror;
1232         int err;
1233
1234         if (status == 0) {
1235                 switch (error) {
1236                 case -ETIMEDOUT:
1237                 case -EPFNOSUPPORT:
1238                 case -EPROTONOSUPPORT:
1239                 case -EOPNOTSUPP:
1240                 case -ECONNREFUSED:
1241                 case -ECONNRESET:
1242                 case -EHOSTDOWN:
1243                 case -EHOSTUNREACH:
1244                 case -ENETUNREACH:
1245                 case -EADDRINUSE:
1246                 case -ENOBUFS:
1247                 case -EPIPE:
1248                 case -EPERM:
1249                         status = NFS4ERR_NXIO;
1250                         break;
1251                 case -EACCES:
1252                         status = NFS4ERR_ACCESS;
1253                         break;
1254                 default:
1255                         return;
1256                 }
1257         }
1258
1259         switch (status) {
1260         case NFS4ERR_DELAY:
1261         case NFS4ERR_GRACE:
1262                 return;
1263         default:
1264                 break;
1265         }
1266
1267         mirror = FF_LAYOUT_COMP(lseg, idx);
1268         err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1269                                        mirror, offset, length, status, opnum,
1270                                        GFP_NOIO);
1271         pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
1272         dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1273 }
1274
1275 /* NFS_PROTO call done callback routines */
1276 static int ff_layout_read_done_cb(struct rpc_task *task,
1277                                 struct nfs_pgio_header *hdr)
1278 {
1279         int err;
1280
1281         trace_nfs4_pnfs_read(hdr, task->tk_status);
1282         if (task->tk_status < 0)
1283                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1284                                             hdr->args.offset, hdr->args.count,
1285                                             hdr->res.op_status, OP_READ,
1286                                             task->tk_status);
1287         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1288                                            hdr->ds_clp, hdr->lseg,
1289                                            hdr->pgio_mirror_idx);
1290
1291         switch (err) {
1292         case -NFS4ERR_RESET_TO_PNFS:
1293                 if (ff_layout_choose_best_ds_for_read(hdr->lseg,
1294                                         hdr->pgio_mirror_idx + 1,
1295                                         &hdr->pgio_mirror_idx))
1296                         goto out_eagain;
1297                 pnfs_read_resend_pnfs(hdr);
1298                 return task->tk_status;
1299         case -NFS4ERR_RESET_TO_MDS:
1300                 ff_layout_reset_read(hdr);
1301                 return task->tk_status;
1302         case -EAGAIN:
1303                 goto out_eagain;
1304         }
1305
1306         return 0;
1307 out_eagain:
1308         rpc_restart_call_prepare(task);
1309         return -EAGAIN;
1310 }
1311
1312 static bool
1313 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1314 {
1315         return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1316 }
1317
1318 /*
1319  * We reference the rpc_cred of the first WRITE that triggers the need for
1320  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1321  * rfc5661 is not clear about which credential should be used.
1322  *
1323  * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1324  * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1325  * we always send layoutcommit after DS writes.
1326  */
1327 static void
1328 ff_layout_set_layoutcommit(struct nfs_pgio_header *hdr)
1329 {
1330         if (!ff_layout_need_layoutcommit(hdr->lseg))
1331                 return;
1332
1333         pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1334                         hdr->mds_offset + hdr->res.count);
1335         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
1336                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
1337 }
1338
1339 static bool
1340 ff_layout_device_unavailable(struct pnfs_layout_segment *lseg, int idx)
1341 {
1342         /* No mirroring for now */
1343         struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1344
1345         return ff_layout_test_devid_unavailable(node);
1346 }
1347
1348 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task,
1349                 struct nfs_pgio_header *hdr)
1350 {
1351         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1352                 return;
1353         nfs4_ff_layout_stat_io_start_read(hdr->inode,
1354                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1355                         hdr->args.count,
1356                         task->tk_start);
1357 }
1358
1359 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
1360                 struct nfs_pgio_header *hdr)
1361 {
1362         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1363                 return;
1364         nfs4_ff_layout_stat_io_end_read(task,
1365                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1366                         hdr->args.count,
1367                         hdr->res.count);
1368 }
1369
1370 static int ff_layout_read_prepare_common(struct rpc_task *task,
1371                                          struct nfs_pgio_header *hdr)
1372 {
1373         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1374                 rpc_exit(task, -EIO);
1375                 return -EIO;
1376         }
1377         if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) {
1378                 rpc_exit(task, -EHOSTDOWN);
1379                 return -EAGAIN;
1380         }
1381
1382         ff_layout_read_record_layoutstats_start(task, hdr);
1383         return 0;
1384 }
1385
1386 /*
1387  * Call ops for the async read/write cases
1388  * In the case of dense layouts, the offset needs to be reset to its
1389  * original value.
1390  */
1391 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1392 {
1393         struct nfs_pgio_header *hdr = data;
1394
1395         if (ff_layout_read_prepare_common(task, hdr))
1396                 return;
1397
1398         rpc_call_start(task);
1399 }
1400
1401 static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1402                                     struct nfs4_sequence_args *args,
1403                                     struct nfs4_sequence_res *res,
1404                                     struct rpc_task *task)
1405 {
1406         if (ds_clp->cl_session)
1407                 return nfs41_setup_sequence(ds_clp->cl_session,
1408                                            args,
1409                                            res,
1410                                            task);
1411         return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1412                                    args,
1413                                    res,
1414                                    task);
1415 }
1416
1417 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1418 {
1419         struct nfs_pgio_header *hdr = data;
1420
1421         if (ff_layout_setup_sequence(hdr->ds_clp,
1422                                      &hdr->args.seq_args,
1423                                      &hdr->res.seq_res,
1424                                      task))
1425                 return;
1426
1427         if (ff_layout_read_prepare_common(task, hdr))
1428                 return;
1429
1430         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1431                         hdr->args.lock_context, FMODE_READ) == -EIO)
1432                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1433 }
1434
1435 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1436 {
1437         struct nfs_pgio_header *hdr = data;
1438
1439         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1440
1441         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1442             task->tk_status == 0) {
1443                 nfs4_sequence_done(task, &hdr->res.seq_res);
1444                 return;
1445         }
1446
1447         /* Note this may cause RPC to be resent */
1448         hdr->mds_ops->rpc_call_done(task, hdr);
1449 }
1450
1451 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1452 {
1453         struct nfs_pgio_header *hdr = data;
1454
1455         ff_layout_read_record_layoutstats_done(task, hdr);
1456         rpc_count_iostats_metrics(task,
1457             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1458 }
1459
1460 static void ff_layout_read_release(void *data)
1461 {
1462         struct nfs_pgio_header *hdr = data;
1463
1464         ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
1465         pnfs_generic_rw_release(data);
1466 }
1467
1468
1469 static int ff_layout_write_done_cb(struct rpc_task *task,
1470                                 struct nfs_pgio_header *hdr)
1471 {
1472         int err;
1473
1474         trace_nfs4_pnfs_write(hdr, task->tk_status);
1475         if (task->tk_status < 0)
1476                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1477                                             hdr->args.offset, hdr->args.count,
1478                                             hdr->res.op_status, OP_WRITE,
1479                                             task->tk_status);
1480         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1481                                            hdr->ds_clp, hdr->lseg,
1482                                            hdr->pgio_mirror_idx);
1483
1484         switch (err) {
1485         case -NFS4ERR_RESET_TO_PNFS:
1486                 ff_layout_reset_write(hdr, true);
1487                 return task->tk_status;
1488         case -NFS4ERR_RESET_TO_MDS:
1489                 ff_layout_reset_write(hdr, false);
1490                 return task->tk_status;
1491         case -EAGAIN:
1492                 return -EAGAIN;
1493         }
1494
1495         if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1496             hdr->res.verf->committed == NFS_DATA_SYNC)
1497                 ff_layout_set_layoutcommit(hdr);
1498
1499         /* zero out fattr since we don't care DS attr at all */
1500         hdr->fattr.valid = 0;
1501         if (task->tk_status >= 0)
1502                 nfs_writeback_update_inode(hdr);
1503
1504         return 0;
1505 }
1506
1507 static int ff_layout_commit_done_cb(struct rpc_task *task,
1508                                      struct nfs_commit_data *data)
1509 {
1510         int err;
1511
1512         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1513         if (task->tk_status < 0)
1514                 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1515                                             data->args.offset, data->args.count,
1516                                             data->res.op_status, OP_COMMIT,
1517                                             task->tk_status);
1518         err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1519                                            data->lseg, data->ds_commit_index);
1520
1521         switch (err) {
1522         case -NFS4ERR_RESET_TO_PNFS:
1523                 pnfs_generic_prepare_to_resend_writes(data);
1524                 return -EAGAIN;
1525         case -NFS4ERR_RESET_TO_MDS:
1526                 pnfs_generic_prepare_to_resend_writes(data);
1527                 return -EAGAIN;
1528         case -EAGAIN:
1529                 rpc_restart_call_prepare(task);
1530                 return -EAGAIN;
1531         }
1532
1533         if (data->verf.committed == NFS_UNSTABLE
1534             && ff_layout_need_layoutcommit(data->lseg))
1535                 pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
1536
1537         return 0;
1538 }
1539
1540 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task,
1541                 struct nfs_pgio_header *hdr)
1542 {
1543         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1544                 return;
1545         nfs4_ff_layout_stat_io_start_write(hdr->inode,
1546                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1547                         hdr->args.count,
1548                         task->tk_start);
1549 }
1550
1551 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task,
1552                 struct nfs_pgio_header *hdr)
1553 {
1554         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1555                 return;
1556         nfs4_ff_layout_stat_io_end_write(task,
1557                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1558                         hdr->args.count, hdr->res.count,
1559                         hdr->res.verf->committed);
1560 }
1561
1562 static int ff_layout_write_prepare_common(struct rpc_task *task,
1563                                           struct nfs_pgio_header *hdr)
1564 {
1565         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1566                 rpc_exit(task, -EIO);
1567                 return -EIO;
1568         }
1569
1570         if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) {
1571                 rpc_exit(task, -EHOSTDOWN);
1572                 return -EAGAIN;
1573         }
1574
1575         ff_layout_write_record_layoutstats_start(task, hdr);
1576         return 0;
1577 }
1578
1579 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1580 {
1581         struct nfs_pgio_header *hdr = data;
1582
1583         if (ff_layout_write_prepare_common(task, hdr))
1584                 return;
1585
1586         rpc_call_start(task);
1587 }
1588
1589 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1590 {
1591         struct nfs_pgio_header *hdr = data;
1592
1593         if (ff_layout_setup_sequence(hdr->ds_clp,
1594                                      &hdr->args.seq_args,
1595                                      &hdr->res.seq_res,
1596                                      task))
1597                 return;
1598
1599         if (ff_layout_write_prepare_common(task, hdr))
1600                 return;
1601
1602         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1603                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
1604                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1605 }
1606
1607 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1608 {
1609         struct nfs_pgio_header *hdr = data;
1610
1611         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1612             task->tk_status == 0) {
1613                 nfs4_sequence_done(task, &hdr->res.seq_res);
1614                 return;
1615         }
1616
1617         /* Note this may cause RPC to be resent */
1618         hdr->mds_ops->rpc_call_done(task, hdr);
1619 }
1620
1621 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1622 {
1623         struct nfs_pgio_header *hdr = data;
1624
1625         ff_layout_write_record_layoutstats_done(task, hdr);
1626         rpc_count_iostats_metrics(task,
1627             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1628 }
1629
1630 static void ff_layout_write_release(void *data)
1631 {
1632         struct nfs_pgio_header *hdr = data;
1633
1634         ff_layout_write_record_layoutstats_done(&hdr->task, hdr);
1635         pnfs_generic_rw_release(data);
1636 }
1637
1638 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task,
1639                 struct nfs_commit_data *cdata)
1640 {
1641         if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags))
1642                 return;
1643         nfs4_ff_layout_stat_io_start_write(cdata->inode,
1644                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1645                         0, task->tk_start);
1646 }
1647
1648 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task,
1649                 struct nfs_commit_data *cdata)
1650 {
1651         struct nfs_page *req;
1652         __u64 count = 0;
1653
1654         if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags))
1655                 return;
1656
1657         if (task->tk_status == 0) {
1658                 list_for_each_entry(req, &cdata->pages, wb_list)
1659                         count += req->wb_bytes;
1660         }
1661         nfs4_ff_layout_stat_io_end_write(task,
1662                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1663                         count, count, NFS_FILE_SYNC);
1664 }
1665
1666 static void ff_layout_commit_prepare_common(struct rpc_task *task,
1667                 struct nfs_commit_data *cdata)
1668 {
1669         ff_layout_commit_record_layoutstats_start(task, cdata);
1670 }
1671
1672 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1673 {
1674         ff_layout_commit_prepare_common(task, data);
1675         rpc_call_start(task);
1676 }
1677
1678 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1679 {
1680         struct nfs_commit_data *wdata = data;
1681
1682         if (ff_layout_setup_sequence(wdata->ds_clp,
1683                                  &wdata->args.seq_args,
1684                                  &wdata->res.seq_res,
1685                                  task))
1686                 return;
1687         ff_layout_commit_prepare_common(task, data);
1688 }
1689
1690 static void ff_layout_commit_done(struct rpc_task *task, void *data)
1691 {
1692         pnfs_generic_write_commit_done(task, data);
1693 }
1694
1695 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1696 {
1697         struct nfs_commit_data *cdata = data;
1698
1699         ff_layout_commit_record_layoutstats_done(task, cdata);
1700         rpc_count_iostats_metrics(task,
1701             &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1702 }
1703
1704 static void ff_layout_commit_release(void *data)
1705 {
1706         struct nfs_commit_data *cdata = data;
1707
1708         ff_layout_commit_record_layoutstats_done(&cdata->task, cdata);
1709         pnfs_generic_commit_release(data);
1710 }
1711
1712 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1713         .rpc_call_prepare = ff_layout_read_prepare_v3,
1714         .rpc_call_done = ff_layout_read_call_done,
1715         .rpc_count_stats = ff_layout_read_count_stats,
1716         .rpc_release = ff_layout_read_release,
1717 };
1718
1719 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1720         .rpc_call_prepare = ff_layout_read_prepare_v4,
1721         .rpc_call_done = ff_layout_read_call_done,
1722         .rpc_count_stats = ff_layout_read_count_stats,
1723         .rpc_release = ff_layout_read_release,
1724 };
1725
1726 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1727         .rpc_call_prepare = ff_layout_write_prepare_v3,
1728         .rpc_call_done = ff_layout_write_call_done,
1729         .rpc_count_stats = ff_layout_write_count_stats,
1730         .rpc_release = ff_layout_write_release,
1731 };
1732
1733 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1734         .rpc_call_prepare = ff_layout_write_prepare_v4,
1735         .rpc_call_done = ff_layout_write_call_done,
1736         .rpc_count_stats = ff_layout_write_count_stats,
1737         .rpc_release = ff_layout_write_release,
1738 };
1739
1740 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1741         .rpc_call_prepare = ff_layout_commit_prepare_v3,
1742         .rpc_call_done = ff_layout_commit_done,
1743         .rpc_count_stats = ff_layout_commit_count_stats,
1744         .rpc_release = ff_layout_commit_release,
1745 };
1746
1747 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1748         .rpc_call_prepare = ff_layout_commit_prepare_v4,
1749         .rpc_call_done = ff_layout_commit_done,
1750         .rpc_count_stats = ff_layout_commit_count_stats,
1751         .rpc_release = ff_layout_commit_release,
1752 };
1753
1754 static enum pnfs_try_status
1755 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1756 {
1757         struct pnfs_layout_segment *lseg = hdr->lseg;
1758         struct nfs4_pnfs_ds *ds;
1759         struct rpc_clnt *ds_clnt;
1760         struct rpc_cred *ds_cred;
1761         loff_t offset = hdr->args.offset;
1762         u32 idx = hdr->pgio_mirror_idx;
1763         int vers;
1764         struct nfs_fh *fh;
1765
1766         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1767                 __func__, hdr->inode->i_ino,
1768                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1769
1770         ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1771         if (!ds)
1772                 goto out_failed;
1773
1774         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1775                                                    hdr->inode);
1776         if (IS_ERR(ds_clnt))
1777                 goto out_failed;
1778
1779         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1780         if (!ds_cred)
1781                 goto out_failed;
1782
1783         vers = nfs4_ff_layout_ds_version(lseg, idx);
1784
1785         dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1786                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1787
1788         hdr->pgio_done_cb = ff_layout_read_done_cb;
1789         atomic_inc(&ds->ds_clp->cl_count);
1790         hdr->ds_clp = ds->ds_clp;
1791         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1792         if (fh)
1793                 hdr->args.fh = fh;
1794         /*
1795          * Note that if we ever decide to split across DSes,
1796          * then we may need to handle dense-like offsets.
1797          */
1798         hdr->args.offset = offset;
1799         hdr->mds_offset = offset;
1800
1801         /* Perform an asynchronous read to ds */
1802         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1803                           vers == 3 ? &ff_layout_read_call_ops_v3 :
1804                                       &ff_layout_read_call_ops_v4,
1805                           0, RPC_TASK_SOFTCONN);
1806         put_rpccred(ds_cred);
1807         return PNFS_ATTEMPTED;
1808
1809 out_failed:
1810         if (ff_layout_avoid_mds_available_ds(lseg))
1811                 return PNFS_TRY_AGAIN;
1812         return PNFS_NOT_ATTEMPTED;
1813 }
1814
1815 /* Perform async writes. */
1816 static enum pnfs_try_status
1817 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1818 {
1819         struct pnfs_layout_segment *lseg = hdr->lseg;
1820         struct nfs4_pnfs_ds *ds;
1821         struct rpc_clnt *ds_clnt;
1822         struct rpc_cred *ds_cred;
1823         loff_t offset = hdr->args.offset;
1824         int vers;
1825         struct nfs_fh *fh;
1826         int idx = hdr->pgio_mirror_idx;
1827
1828         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1829         if (!ds)
1830                 return PNFS_NOT_ATTEMPTED;
1831
1832         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1833                                                    hdr->inode);
1834         if (IS_ERR(ds_clnt))
1835                 return PNFS_NOT_ATTEMPTED;
1836
1837         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1838         if (!ds_cred)
1839                 return PNFS_NOT_ATTEMPTED;
1840
1841         vers = nfs4_ff_layout_ds_version(lseg, idx);
1842
1843         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1844                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1845                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1846                 vers);
1847
1848         hdr->pgio_done_cb = ff_layout_write_done_cb;
1849         atomic_inc(&ds->ds_clp->cl_count);
1850         hdr->ds_clp = ds->ds_clp;
1851         hdr->ds_commit_idx = idx;
1852         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1853         if (fh)
1854                 hdr->args.fh = fh;
1855
1856         /*
1857          * Note that if we ever decide to split across DSes,
1858          * then we may need to handle dense-like offsets.
1859          */
1860         hdr->args.offset = offset;
1861
1862         /* Perform an asynchronous write */
1863         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1864                           vers == 3 ? &ff_layout_write_call_ops_v3 :
1865                                       &ff_layout_write_call_ops_v4,
1866                           sync, RPC_TASK_SOFTCONN);
1867         put_rpccred(ds_cred);
1868         return PNFS_ATTEMPTED;
1869 }
1870
1871 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1872 {
1873         return i;
1874 }
1875
1876 static struct nfs_fh *
1877 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1878 {
1879         struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1880
1881         /* FIXME: Assume that there is only one NFS version available
1882          * for the DS.
1883          */
1884         return &flseg->mirror_array[i]->fh_versions[0];
1885 }
1886
1887 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1888 {
1889         struct pnfs_layout_segment *lseg = data->lseg;
1890         struct nfs4_pnfs_ds *ds;
1891         struct rpc_clnt *ds_clnt;
1892         struct rpc_cred *ds_cred;
1893         u32 idx;
1894         int vers, ret;
1895         struct nfs_fh *fh;
1896
1897         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1898         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1899         if (!ds)
1900                 goto out_err;
1901
1902         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1903                                                    data->inode);
1904         if (IS_ERR(ds_clnt))
1905                 goto out_err;
1906
1907         ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1908         if (!ds_cred)
1909                 goto out_err;
1910
1911         vers = nfs4_ff_layout_ds_version(lseg, idx);
1912
1913         dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1914                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1915                 vers);
1916         data->commit_done_cb = ff_layout_commit_done_cb;
1917         data->cred = ds_cred;
1918         atomic_inc(&ds->ds_clp->cl_count);
1919         data->ds_clp = ds->ds_clp;
1920         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1921         if (fh)
1922                 data->args.fh = fh;
1923
1924         ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1925                                    vers == 3 ? &ff_layout_commit_call_ops_v3 :
1926                                                &ff_layout_commit_call_ops_v4,
1927                                    how, RPC_TASK_SOFTCONN);
1928         put_rpccred(ds_cred);
1929         return ret;
1930 out_err:
1931         pnfs_generic_prepare_to_resend_writes(data);
1932         pnfs_generic_commit_release(data);
1933         return -EAGAIN;
1934 }
1935
1936 static int
1937 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1938                            int how, struct nfs_commit_info *cinfo)
1939 {
1940         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1941                                             ff_layout_initiate_commit);
1942 }
1943
1944 static struct pnfs_ds_commit_info *
1945 ff_layout_get_ds_info(struct inode *inode)
1946 {
1947         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1948
1949         if (layout == NULL)
1950                 return NULL;
1951
1952         return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1953 }
1954
1955 static void
1956 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
1957 {
1958         nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1959                                                   id_node));
1960 }
1961
1962 static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1963                                   struct xdr_stream *xdr,
1964                                   const struct nfs4_layoutreturn_args *args)
1965 {
1966         struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1967         __be32 *start;
1968         int count = 0, ret = 0;
1969
1970         start = xdr_reserve_space(xdr, 4);
1971         if (unlikely(!start))
1972                 return -E2BIG;
1973
1974         /* This assume we always return _ALL_ layouts */
1975         spin_lock(&hdr->plh_inode->i_lock);
1976         ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1977         spin_unlock(&hdr->plh_inode->i_lock);
1978
1979         *start = cpu_to_be32(count);
1980
1981         return ret;
1982 }
1983
1984 /* report nothing for now */
1985 static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1986                                      struct xdr_stream *xdr,
1987                                      const struct nfs4_layoutreturn_args *args)
1988 {
1989         __be32 *p;
1990
1991         p = xdr_reserve_space(xdr, 4);
1992         if (likely(p))
1993                 *p = cpu_to_be32(0);
1994 }
1995
1996 static struct nfs4_deviceid_node *
1997 ff_layout_alloc_deviceid_node(struct nfs_server *server,
1998                               struct pnfs_device *pdev, gfp_t gfp_flags)
1999 {
2000         struct nfs4_ff_layout_ds *dsaddr;
2001
2002         dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
2003         if (!dsaddr)
2004                 return NULL;
2005         return &dsaddr->id_node;
2006 }
2007
2008 static void
2009 ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
2010                               struct xdr_stream *xdr,
2011                               const struct nfs4_layoutreturn_args *args)
2012 {
2013         struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
2014         __be32 *start;
2015
2016         dprintk("%s: Begin\n", __func__);
2017         start = xdr_reserve_space(xdr, 4);
2018         BUG_ON(!start);
2019
2020         ff_layout_encode_ioerr(flo, xdr, args);
2021         ff_layout_encode_iostats(flo, xdr, args);
2022
2023         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2024         dprintk("%s: Return\n", __func__);
2025 }
2026
2027 static int
2028 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
2029 {
2030         const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
2031
2032         return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
2033 }
2034
2035 static size_t
2036 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
2037                           const int buflen)
2038 {
2039         const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
2040         const struct in6_addr *addr = &sin6->sin6_addr;
2041
2042         /*
2043          * RFC 4291, Section 2.2.2
2044          *
2045          * Shorthanded ANY address
2046          */
2047         if (ipv6_addr_any(addr))
2048                 return snprintf(buf, buflen, "::");
2049
2050         /*
2051          * RFC 4291, Section 2.2.2
2052          *
2053          * Shorthanded loopback address
2054          */
2055         if (ipv6_addr_loopback(addr))
2056                 return snprintf(buf, buflen, "::1");
2057
2058         /*
2059          * RFC 4291, Section 2.2.3
2060          *
2061          * Special presentation address format for mapped v4
2062          * addresses.
2063          */
2064         if (ipv6_addr_v4mapped(addr))
2065                 return snprintf(buf, buflen, "::ffff:%pI4",
2066                                         &addr->s6_addr32[3]);
2067
2068         /*
2069          * RFC 4291, Section 2.2.1
2070          */
2071         return snprintf(buf, buflen, "%pI6c", addr);
2072 }
2073
2074 /* Derived from rpc_sockaddr2uaddr */
2075 static void
2076 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
2077 {
2078         struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
2079         char portbuf[RPCBIND_MAXUADDRPLEN];
2080         char addrbuf[RPCBIND_MAXUADDRLEN];
2081         char *netid;
2082         unsigned short port;
2083         int len, netid_len;
2084         __be32 *p;
2085
2086         switch (sap->sa_family) {
2087         case AF_INET:
2088                 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
2089                         return;
2090                 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
2091                 netid = "tcp";
2092                 netid_len = 3;
2093                 break;
2094         case AF_INET6:
2095                 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
2096                         return;
2097                 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
2098                 netid = "tcp6";
2099                 netid_len = 4;
2100                 break;
2101         default:
2102                 /* we only support tcp and tcp6 */
2103                 WARN_ON_ONCE(1);
2104                 return;
2105         }
2106
2107         snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
2108         len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
2109
2110         p = xdr_reserve_space(xdr, 4 + netid_len);
2111         xdr_encode_opaque(p, netid, netid_len);
2112
2113         p = xdr_reserve_space(xdr, 4 + len);
2114         xdr_encode_opaque(p, addrbuf, len);
2115 }
2116
2117 static void
2118 ff_layout_encode_nfstime(struct xdr_stream *xdr,
2119                          ktime_t t)
2120 {
2121         struct timespec64 ts;
2122         __be32 *p;
2123
2124         p = xdr_reserve_space(xdr, 12);
2125         ts = ktime_to_timespec64(t);
2126         p = xdr_encode_hyper(p, ts.tv_sec);
2127         *p++ = cpu_to_be32(ts.tv_nsec);
2128 }
2129
2130 static void
2131 ff_layout_encode_io_latency(struct xdr_stream *xdr,
2132                             struct nfs4_ff_io_stat *stat)
2133 {
2134         __be32 *p;
2135
2136         p = xdr_reserve_space(xdr, 5 * 8);
2137         p = xdr_encode_hyper(p, stat->ops_requested);
2138         p = xdr_encode_hyper(p, stat->bytes_requested);
2139         p = xdr_encode_hyper(p, stat->ops_completed);
2140         p = xdr_encode_hyper(p, stat->bytes_completed);
2141         p = xdr_encode_hyper(p, stat->bytes_not_delivered);
2142         ff_layout_encode_nfstime(xdr, stat->total_busy_time);
2143         ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
2144 }
2145
2146 static void
2147 ff_layout_encode_layoutstats(struct xdr_stream *xdr,
2148                              struct nfs42_layoutstat_args *args,
2149                              struct nfs42_layoutstat_devinfo *devinfo)
2150 {
2151         struct nfs4_ff_layout_mirror *mirror = devinfo->layout_private;
2152         struct nfs4_pnfs_ds_addr *da;
2153         struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds;
2154         struct nfs_fh *fh = &mirror->fh_versions[0];
2155         __be32 *p, *start;
2156
2157         da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2158         dprintk("%s: DS %s: encoding address %s\n",
2159                 __func__, ds->ds_remotestr, da->da_remotestr);
2160         /* layoutupdate length */
2161         start = xdr_reserve_space(xdr, 4);
2162         /* netaddr4 */
2163         ff_layout_encode_netaddr(xdr, da);
2164         /* nfs_fh4 */
2165         p = xdr_reserve_space(xdr, 4 + fh->size);
2166         xdr_encode_opaque(p, fh->data, fh->size);
2167         /* ff_io_latency4 read */
2168         spin_lock(&mirror->lock);
2169         ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat);
2170         /* ff_io_latency4 write */
2171         ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat);
2172         spin_unlock(&mirror->lock);
2173         /* nfstime4 */
2174         ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time));
2175         /* bool */
2176         p = xdr_reserve_space(xdr, 4);
2177         *p = cpu_to_be32(false);
2178
2179         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2180 }
2181
2182 static int
2183 ff_layout_mirror_prepare_stats(struct nfs42_layoutstat_args *args,
2184                                struct pnfs_layout_hdr *lo,
2185                                int dev_limit)
2186 {
2187         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2188         struct nfs4_ff_layout_mirror *mirror;
2189         struct nfs4_deviceid_node *dev;
2190         struct nfs42_layoutstat_devinfo *devinfo;
2191         int i = 0;
2192
2193         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2194                 if (i >= dev_limit)
2195                         break;
2196                 if (!mirror->mirror_ds)
2197                         continue;
2198                 /* mirror refcount put in cleanup_layoutstats */
2199                 if (!atomic_inc_not_zero(&mirror->ref))
2200                         continue;
2201                 dev = &mirror->mirror_ds->id_node;
2202                 devinfo = &args->devinfo[i];
2203                 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE);
2204                 devinfo->offset = 0;
2205                 devinfo->length = NFS4_MAX_UINT64;
2206                 devinfo->read_count = mirror->read_stat.io_stat.ops_completed;
2207                 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed;
2208                 devinfo->write_count = mirror->write_stat.io_stat.ops_completed;
2209                 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed;
2210                 devinfo->layout_type = LAYOUT_FLEX_FILES;
2211                 devinfo->layoutstats_encode = ff_layout_encode_layoutstats;
2212                 devinfo->layout_private = mirror;
2213
2214                 i++;
2215         }
2216         return i;
2217 }
2218
2219 static int
2220 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2221 {
2222         struct nfs4_flexfile_layout *ff_layout;
2223         struct nfs4_ff_layout_mirror *mirror;
2224         int dev_count = 0;
2225
2226         spin_lock(&args->inode->i_lock);
2227         ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout);
2228         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2229                 if (atomic_read(&mirror->ref) != 0)
2230                         dev_count ++;
2231         }
2232         spin_unlock(&args->inode->i_lock);
2233         /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2234         if (dev_count > PNFS_LAYOUTSTATS_MAXDEV) {
2235                 dprintk("%s: truncating devinfo to limit (%d:%d)\n",
2236                         __func__, dev_count, PNFS_LAYOUTSTATS_MAXDEV);
2237                 dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2238         }
2239         args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO);
2240         if (!args->devinfo)
2241                 return -ENOMEM;
2242
2243         spin_lock(&args->inode->i_lock);
2244         args->num_dev = ff_layout_mirror_prepare_stats(args,
2245                         &ff_layout->generic_hdr, dev_count);
2246         spin_unlock(&args->inode->i_lock);
2247
2248         return 0;
2249 }
2250
2251 static void
2252 ff_layout_cleanup_layoutstats(struct nfs42_layoutstat_data *data)
2253 {
2254         struct nfs4_ff_layout_mirror *mirror;
2255         int i;
2256
2257         for (i = 0; i < data->args.num_dev; i++) {
2258                 mirror = data->args.devinfo[i].layout_private;
2259                 data->args.devinfo[i].layout_private = NULL;
2260                 ff_layout_put_mirror(mirror);
2261         }
2262 }
2263
2264 static struct pnfs_layoutdriver_type flexfilelayout_type = {
2265         .id                     = LAYOUT_FLEX_FILES,
2266         .name                   = "LAYOUT_FLEX_FILES",
2267         .owner                  = THIS_MODULE,
2268         .alloc_layout_hdr       = ff_layout_alloc_layout_hdr,
2269         .free_layout_hdr        = ff_layout_free_layout_hdr,
2270         .alloc_lseg             = ff_layout_alloc_lseg,
2271         .free_lseg              = ff_layout_free_lseg,
2272         .add_lseg               = ff_layout_add_lseg,
2273         .pg_read_ops            = &ff_layout_pg_read_ops,
2274         .pg_write_ops           = &ff_layout_pg_write_ops,
2275         .get_ds_info            = ff_layout_get_ds_info,
2276         .free_deviceid_node     = ff_layout_free_deviceid_node,
2277         .mark_request_commit    = pnfs_layout_mark_request_commit,
2278         .clear_request_commit   = pnfs_generic_clear_request_commit,
2279         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
2280         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
2281         .commit_pagelist        = ff_layout_commit_pagelist,
2282         .read_pagelist          = ff_layout_read_pagelist,
2283         .write_pagelist         = ff_layout_write_pagelist,
2284         .alloc_deviceid_node    = ff_layout_alloc_deviceid_node,
2285         .encode_layoutreturn    = ff_layout_encode_layoutreturn,
2286         .sync                   = pnfs_nfs_generic_sync,
2287         .prepare_layoutstats    = ff_layout_prepare_layoutstats,
2288         .cleanup_layoutstats    = ff_layout_cleanup_layoutstats,
2289 };
2290
2291 static int __init nfs4flexfilelayout_init(void)
2292 {
2293         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
2294                __func__);
2295         if (!ff_zero_group) {
2296                 ff_zero_group = groups_alloc(0);
2297                 if (!ff_zero_group)
2298                         return -ENOMEM;
2299         }
2300         return pnfs_register_layoutdriver(&flexfilelayout_type);
2301 }
2302
2303 static void __exit nfs4flexfilelayout_exit(void)
2304 {
2305         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
2306                __func__);
2307         pnfs_unregister_layoutdriver(&flexfilelayout_type);
2308         if (ff_zero_group) {
2309                 put_group_info(ff_zero_group);
2310                 ff_zero_group = NULL;
2311         }
2312 }
2313
2314 MODULE_ALIAS("nfs-layouttype4-4");
2315
2316 MODULE_LICENSE("GPL");
2317 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
2318
2319 module_init(nfs4flexfilelayout_init);
2320 module_exit(nfs4flexfilelayout_exit);