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 (new->pls_range.iomode != old->pls_range.iomode)
 303                 return false;
 304         old_end = pnfs_calc_offset_end(old->pls_range.offset,
 305                         old->pls_range.length);
 306         if (old_end < new->pls_range.offset)
 307                 return false;
 308         new_end = pnfs_calc_offset_end(new->pls_range.offset,
 309                         new->pls_range.length);
 310         if (new_end < old->pls_range.offset)
 311                 return false;
 312
 313         /* Mergeable: copy info from 'old' to 'new' */
 314         if (new_end < old_end)
 315                 new_end = old_end;
 316         if (new->pls_range.offset < old->pls_range.offset)
 317                 new->pls_range.offset = old->pls_range.offset;
 318         new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset,
 319                         new_end);
 320         if (test_bit(NFS_LSEG_ROC, &old->pls_flags))
 321                 set_bit(NFS_LSEG_ROC, &new->pls_flags);
 322         if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags))
 323                 set_bit(NFS_LSEG_LAYOUTRETURN, &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_init_read(struct nfs_pageio_descriptor *pgio,
 825                         struct nfs_page *req)
 826 {
 827         struct nfs_pgio_mirror *pgm;
 828         struct nfs4_ff_layout_mirror *mirror;
 829         struct nfs4_pnfs_ds *ds;
 830         int ds_idx;
 831
 832         /* Use full layout for now */
 833         if (!pgio->pg_lseg) {
 834                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 835                                                    req->wb_context,
 836                                                    0,
 837                                                    NFS4_MAX_UINT64,
 838                                                    IOMODE_READ,
 839                                                    GFP_KERNEL);
 840                 if (IS_ERR(pgio->pg_lseg)) {
 841                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 842                         pgio->pg_lseg = NULL;
 843                         return;
 844                 }
 845         }
 846         /* If no lseg, fall back to read through mds */
 847         if (pgio->pg_lseg == NULL)
 848                 goto out_mds;
 849
 850         ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx);
 851         if (!ds)
 852                 goto out_mds;
 853         mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
 854
 855         pgio->pg_mirror_idx = ds_idx;
 856
 857         /* read always uses only one mirror - idx 0 for pgio layer */
 858         pgm = &pgio->pg_mirrors[0];
 859         pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
 860
 861         return;
 862 out_mds:
 863         pnfs_put_lseg(pgio->pg_lseg);
 864         pgio->pg_lseg = NULL;
 865         nfs_pageio_reset_read_mds(pgio);
 866 }
 867
 868 static void
 869 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 870                         struct nfs_page *req)
 871 {
 872         struct nfs4_ff_layout_mirror *mirror;
 873         struct nfs_pgio_mirror *pgm;
 874         struct nfs_commit_info cinfo;
 875         struct nfs4_pnfs_ds *ds;
 876         int i;
 877         int status;
 878
 879         if (!pgio->pg_lseg) {
 880                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 881                                                    req->wb_context,
 882                                                    0,
 883                                                    NFS4_MAX_UINT64,
 884                                                    IOMODE_RW,
 885                                                    GFP_NOFS);
 886                 if (IS_ERR(pgio->pg_lseg)) {
 887                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 888                         pgio->pg_lseg = NULL;
 889                         return;
 890                 }
 891         }
 892         /* If no lseg, fall back to write through mds */
 893         if (pgio->pg_lseg == NULL)
 894                 goto out_mds;
 895
 896         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 897         status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 898         if (status < 0)
 899                 goto out_mds;
 900
 901         /* Use a direct mapping of ds_idx to pgio mirror_idx */
 902         if (WARN_ON_ONCE(pgio->pg_mirror_count !=
 903             FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)))
 904                 goto out_mds;
 905
 906         for (i = 0; i < pgio->pg_mirror_count; i++) {
 907                 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
 908                 if (!ds)
 909                         goto out_mds;
 910                 pgm = &pgio->pg_mirrors[i];
 911                 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
 912                 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
 913         }
 914
 915         return;
 916
 917 out_mds:
 918         pnfs_put_lseg(pgio->pg_lseg);
 919         pgio->pg_lseg = NULL;
 920         nfs_pageio_reset_write_mds(pgio);
 921 }
 922
 923 static unsigned int
 924 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
 925                                     struct nfs_page *req)
 926 {
 927         if (!pgio->pg_lseg) {
 928                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 929                                                    req->wb_context,
 930                                                    0,
 931                                                    NFS4_MAX_UINT64,
 932                                                    IOMODE_RW,
 933                                                    GFP_NOFS);
 934                 if (IS_ERR(pgio->pg_lseg)) {
 935                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 936                         pgio->pg_lseg = NULL;
 937                         goto out;
 938                 }
 939         }
 940         if (pgio->pg_lseg)
 941                 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
 942
 943         /* no lseg means that pnfs is not in use, so no mirroring here */
 944         nfs_pageio_reset_write_mds(pgio);
 945 out:
 946         return 1;
 947 }
 948
 949 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
 950         .pg_init = ff_layout_pg_init_read,
 951         .pg_test = pnfs_generic_pg_test,
 952         .pg_doio = pnfs_generic_pg_readpages,
 953         .pg_cleanup = pnfs_generic_pg_cleanup,
 954 };
 955
 956 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
 957         .pg_init = ff_layout_pg_init_write,
 958         .pg_test = pnfs_generic_pg_test,
 959         .pg_doio = pnfs_generic_pg_writepages,
 960         .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
 961         .pg_cleanup = pnfs_generic_pg_cleanup,
 962 };
 963
 964 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
 965 {
 966         struct rpc_task *task = &hdr->task;
 967
 968         pnfs_layoutcommit_inode(hdr->inode, false);
 969
 970         if (retry_pnfs) {
 971                 dprintk("%s Reset task %5u for i/o through pNFS "
 972                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 973                         hdr->task.tk_pid,
 974                         hdr->inode->i_sb->s_id,
 975                         (unsigned long long)NFS_FILEID(hdr->inode),
 976                         hdr->args.count,
 977                         (unsigned long long)hdr->args.offset);
 978
 979                 hdr->completion_ops->reschedule_io(hdr);
 980                 return;
 981         }
 982
 983         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 984                 dprintk("%s Reset task %5u for i/o through MDS "
 985                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 986                         hdr->task.tk_pid,
 987                         hdr->inode->i_sb->s_id,
 988                         (unsigned long long)NFS_FILEID(hdr->inode),
 989                         hdr->args.count,
 990                         (unsigned long long)hdr->args.offset);
 991
 992                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
 993         }
 994 }
 995
 996 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
 997 {
 998         struct rpc_task *task = &hdr->task;
 999
1000         pnfs_layoutcommit_inode(hdr->inode, false);
1001
1002         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1003                 dprintk("%s Reset task %5u for i/o through MDS "
1004                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1005                         hdr->task.tk_pid,
1006                         hdr->inode->i_sb->s_id,
1007                         (unsigned long long)NFS_FILEID(hdr->inode),
1008                         hdr->args.count,
1009                         (unsigned long long)hdr->args.offset);
1010
1011                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
1012         }
1013 }
1014
1015 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
1016                                            struct nfs4_state *state,
1017                                            struct nfs_client *clp,
1018                                            struct pnfs_layout_segment *lseg,
1019                                            int idx)
1020 {
1021         struct pnfs_layout_hdr *lo = lseg->pls_layout;
1022         struct inode *inode = lo->plh_inode;
1023         struct nfs_server *mds_server = NFS_SERVER(inode);
1024
1025         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1026         struct nfs_client *mds_client = mds_server->nfs_client;
1027         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
1028
1029         if (task->tk_status >= 0)
1030                 return 0;
1031
1032         switch (task->tk_status) {
1033         /* MDS state errors */
1034         case -NFS4ERR_DELEG_REVOKED:
1035         case -NFS4ERR_ADMIN_REVOKED:
1036         case -NFS4ERR_BAD_STATEID:
1037                 if (state == NULL)
1038                         break;
1039                 nfs_remove_bad_delegation(state->inode);
1040         case -NFS4ERR_OPENMODE:
1041                 if (state == NULL)
1042                         break;
1043                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1044                         goto out_bad_stateid;
1045                 goto wait_on_recovery;
1046         case -NFS4ERR_EXPIRED:
1047                 if (state != NULL) {
1048                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1049                                 goto out_bad_stateid;
1050                 }
1051                 nfs4_schedule_lease_recovery(mds_client);
1052                 goto wait_on_recovery;
1053         /* DS session errors */
1054         case -NFS4ERR_BADSESSION:
1055         case -NFS4ERR_BADSLOT:
1056         case -NFS4ERR_BAD_HIGH_SLOT:
1057         case -NFS4ERR_DEADSESSION:
1058         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1059         case -NFS4ERR_SEQ_FALSE_RETRY:
1060         case -NFS4ERR_SEQ_MISORDERED:
1061                 dprintk("%s ERROR %d, Reset session. Exchangeid "
1062                         "flags 0x%x\n", __func__, task->tk_status,
1063                         clp->cl_exchange_flags);
1064                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1065                 break;
1066         case -NFS4ERR_DELAY:
1067         case -NFS4ERR_GRACE:
1068                 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1069                 break;
1070         case -NFS4ERR_RETRY_UNCACHED_REP:
1071                 break;
1072         /* Invalidate Layout errors */
1073         case -NFS4ERR_PNFS_NO_LAYOUT:
1074         case -ESTALE:           /* mapped NFS4ERR_STALE */
1075         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
1076         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
1077         case -NFS4ERR_FHEXPIRED:
1078         case -NFS4ERR_WRONG_TYPE:
1079                 dprintk("%s Invalid layout error %d\n", __func__,
1080                         task->tk_status);
1081                 /*
1082                  * Destroy layout so new i/o will get a new layout.
1083                  * Layout will not be destroyed until all current lseg
1084                  * references are put. Mark layout as invalid to resend failed
1085                  * i/o and all i/o waiting on the slot table to the MDS until
1086                  * layout is destroyed and a new valid layout is obtained.
1087                  */
1088                 pnfs_destroy_layout(NFS_I(inode));
1089                 rpc_wake_up(&tbl->slot_tbl_waitq);
1090                 goto reset;
1091         /* RPC connection errors */
1092         case -ECONNREFUSED:
1093         case -EHOSTDOWN:
1094         case -EHOSTUNREACH:
1095         case -ENETUNREACH:
1096         case -EIO:
1097         case -ETIMEDOUT:
1098         case -EPIPE:
1099                 dprintk("%s DS connection error %d\n", __func__,
1100                         task->tk_status);
1101                 nfs4_mark_deviceid_unavailable(devid);
1102                 rpc_wake_up(&tbl->slot_tbl_waitq);
1103                 /* fall through */
1104         default:
1105                 if (ff_layout_no_fallback_to_mds(lseg) ||
1106                     ff_layout_has_available_ds(lseg))
1107                         return -NFS4ERR_RESET_TO_PNFS;
1108 reset:
1109                 dprintk("%s Retry through MDS. Error %d\n", __func__,
1110                         task->tk_status);
1111                 return -NFS4ERR_RESET_TO_MDS;
1112         }
1113 out:
1114         task->tk_status = 0;
1115         return -EAGAIN;
1116 out_bad_stateid:
1117         task->tk_status = -EIO;
1118         return 0;
1119 wait_on_recovery:
1120         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
1121         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
1122                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
1123         goto out;
1124 }
1125
1126 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
1127 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1128                                            struct pnfs_layout_segment *lseg,
1129                                            int idx)
1130 {
1131         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1132
1133         if (task->tk_status >= 0)
1134                 return 0;
1135
1136         switch (task->tk_status) {
1137         /* File access problems. Don't mark the device as unavailable */
1138         case -EACCES:
1139         case -ESTALE:
1140         case -EISDIR:
1141         case -EBADHANDLE:
1142         case -ELOOP:
1143         case -ENOSPC:
1144                 break;
1145         case -EJUKEBOX:
1146                 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1147                 goto out_retry;
1148         default:
1149                 dprintk("%s DS connection error %d\n", __func__,
1150                         task->tk_status);
1151                 nfs4_mark_deviceid_unavailable(devid);
1152         }
1153         /* FIXME: Need to prevent infinite looping here. */
1154         return -NFS4ERR_RESET_TO_PNFS;
1155 out_retry:
1156         task->tk_status = 0;
1157         rpc_restart_call_prepare(task);
1158         rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
1159         return -EAGAIN;
1160 }
1161
1162 static int ff_layout_async_handle_error(struct rpc_task *task,
1163                                         struct nfs4_state *state,
1164                                         struct nfs_client *clp,
1165                                         struct pnfs_layout_segment *lseg,
1166                                         int idx)
1167 {
1168         int vers = clp->cl_nfs_mod->rpc_vers->number;
1169
1170         switch (vers) {
1171         case 3:
1172                 return ff_layout_async_handle_error_v3(task, lseg, idx);
1173         case 4:
1174                 return ff_layout_async_handle_error_v4(task, state, clp,
1175                                                        lseg, idx);
1176         default:
1177                 /* should never happen */
1178                 WARN_ON_ONCE(1);
1179                 return 0;
1180         }
1181 }
1182
1183 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1184                                         int idx, u64 offset, u64 length,
1185                                         u32 status, int opnum, int error)
1186 {
1187         struct nfs4_ff_layout_mirror *mirror;
1188         int err;
1189
1190         if (status == 0) {
1191                 switch (error) {
1192                 case -ETIMEDOUT:
1193                 case -EPFNOSUPPORT:
1194                 case -EPROTONOSUPPORT:
1195                 case -EOPNOTSUPP:
1196                 case -ECONNREFUSED:
1197                 case -ECONNRESET:
1198                 case -EHOSTDOWN:
1199                 case -EHOSTUNREACH:
1200                 case -ENETUNREACH:
1201                 case -EADDRINUSE:
1202                 case -ENOBUFS:
1203                 case -EPIPE:
1204                 case -EPERM:
1205                         status = NFS4ERR_NXIO;
1206                         break;
1207                 case -EACCES:
1208                         status = NFS4ERR_ACCESS;
1209                         break;
1210                 default:
1211                         return;
1212                 }
1213         }
1214
1215         switch (status) {
1216         case NFS4ERR_DELAY:
1217         case NFS4ERR_GRACE:
1218                 return;
1219         default:
1220                 break;
1221         }
1222
1223         mirror = FF_LAYOUT_COMP(lseg, idx);
1224         err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1225                                        mirror, offset, length, status, opnum,
1226                                        GFP_NOIO);
1227         pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
1228         dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1229 }
1230
1231 /* NFS_PROTO call done callback routines */
1232 static int ff_layout_read_done_cb(struct rpc_task *task,
1233                                 struct nfs_pgio_header *hdr)
1234 {
1235         int err;
1236
1237         trace_nfs4_pnfs_read(hdr, task->tk_status);
1238         if (task->tk_status < 0)
1239                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1240                                             hdr->args.offset, hdr->args.count,
1241                                             hdr->res.op_status, OP_READ,
1242                                             task->tk_status);
1243         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1244                                            hdr->ds_clp, hdr->lseg,
1245                                            hdr->pgio_mirror_idx);
1246
1247         switch (err) {
1248         case -NFS4ERR_RESET_TO_PNFS:
1249                 if (ff_layout_choose_best_ds_for_read(hdr->lseg,
1250                                         hdr->pgio_mirror_idx + 1,
1251                                         &hdr->pgio_mirror_idx))
1252                         goto out_eagain;
1253                 set_bit(NFS_LAYOUT_RETURN_REQUESTED,
1254                         &hdr->lseg->pls_layout->plh_flags);
1255                 pnfs_read_resend_pnfs(hdr);
1256                 return task->tk_status;
1257         case -NFS4ERR_RESET_TO_MDS:
1258                 ff_layout_reset_read(hdr);
1259                 return task->tk_status;
1260         case -EAGAIN:
1261                 goto out_eagain;
1262         }
1263
1264         return 0;
1265 out_eagain:
1266         rpc_restart_call_prepare(task);
1267         return -EAGAIN;
1268 }
1269
1270 static bool
1271 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1272 {
1273         return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1274 }
1275
1276 /*
1277  * We reference the rpc_cred of the first WRITE that triggers the need for
1278  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1279  * rfc5661 is not clear about which credential should be used.
1280  *
1281  * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1282  * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1283  * we always send layoutcommit after DS writes.
1284  */
1285 static void
1286 ff_layout_set_layoutcommit(struct nfs_pgio_header *hdr)
1287 {
1288         if (!ff_layout_need_layoutcommit(hdr->lseg))
1289                 return;
1290
1291         pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1292                         hdr->mds_offset + hdr->res.count);
1293         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
1294                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
1295 }
1296
1297 static bool
1298 ff_layout_reset_to_mds(struct pnfs_layout_segment *lseg, int idx)
1299 {
1300         /* No mirroring for now */
1301         struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1302
1303         return ff_layout_test_devid_unavailable(node);
1304 }
1305
1306 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task,
1307                 struct nfs_pgio_header *hdr)
1308 {
1309         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1310                 return;
1311         nfs4_ff_layout_stat_io_start_read(hdr->inode,
1312                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1313                         hdr->args.count,
1314                         task->tk_start);
1315 }
1316
1317 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
1318                 struct nfs_pgio_header *hdr)
1319 {
1320         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1321                 return;
1322         nfs4_ff_layout_stat_io_end_read(task,
1323                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1324                         hdr->args.count,
1325                         hdr->res.count);
1326 }
1327
1328 static int ff_layout_read_prepare_common(struct rpc_task *task,
1329                                          struct nfs_pgio_header *hdr)
1330 {
1331         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1332                 rpc_exit(task, -EIO);
1333                 return -EIO;
1334         }
1335         if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1336                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
1337                 if (ff_layout_has_available_ds(hdr->lseg))
1338                         pnfs_read_resend_pnfs(hdr);
1339                 else
1340                         ff_layout_reset_read(hdr);
1341                 rpc_exit(task, 0);
1342                 return -EAGAIN;
1343         }
1344         hdr->pgio_done_cb = ff_layout_read_done_cb;
1345
1346         ff_layout_read_record_layoutstats_start(task, hdr);
1347         return 0;
1348 }
1349
1350 /*
1351  * Call ops for the async read/write cases
1352  * In the case of dense layouts, the offset needs to be reset to its
1353  * original value.
1354  */
1355 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1356 {
1357         struct nfs_pgio_header *hdr = data;
1358
1359         if (ff_layout_read_prepare_common(task, hdr))
1360                 return;
1361
1362         rpc_call_start(task);
1363 }
1364
1365 static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1366                                     struct nfs4_sequence_args *args,
1367                                     struct nfs4_sequence_res *res,
1368                                     struct rpc_task *task)
1369 {
1370         if (ds_clp->cl_session)
1371                 return nfs41_setup_sequence(ds_clp->cl_session,
1372                                            args,
1373                                            res,
1374                                            task);
1375         return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1376                                    args,
1377                                    res,
1378                                    task);
1379 }
1380
1381 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1382 {
1383         struct nfs_pgio_header *hdr = data;
1384
1385         if (ff_layout_setup_sequence(hdr->ds_clp,
1386                                      &hdr->args.seq_args,
1387                                      &hdr->res.seq_res,
1388                                      task))
1389                 return;
1390
1391         if (ff_layout_read_prepare_common(task, hdr))
1392                 return;
1393
1394         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1395                         hdr->args.lock_context, FMODE_READ) == -EIO)
1396                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1397 }
1398
1399 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1400 {
1401         struct nfs_pgio_header *hdr = data;
1402
1403         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1404
1405         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1406             task->tk_status == 0) {
1407                 nfs4_sequence_done(task, &hdr->res.seq_res);
1408                 return;
1409         }
1410
1411         /* Note this may cause RPC to be resent */
1412         hdr->mds_ops->rpc_call_done(task, hdr);
1413 }
1414
1415 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1416 {
1417         struct nfs_pgio_header *hdr = data;
1418
1419         ff_layout_read_record_layoutstats_done(task, hdr);
1420         rpc_count_iostats_metrics(task,
1421             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1422 }
1423
1424 static void ff_layout_read_release(void *data)
1425 {
1426         struct nfs_pgio_header *hdr = data;
1427
1428         ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
1429         pnfs_generic_rw_release(data);
1430 }
1431
1432
1433 static int ff_layout_write_done_cb(struct rpc_task *task,
1434                                 struct nfs_pgio_header *hdr)
1435 {
1436         int err;
1437
1438         trace_nfs4_pnfs_write(hdr, task->tk_status);
1439         if (task->tk_status < 0)
1440                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1441                                             hdr->args.offset, hdr->args.count,
1442                                             hdr->res.op_status, OP_WRITE,
1443                                             task->tk_status);
1444         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1445                                            hdr->ds_clp, hdr->lseg,
1446                                            hdr->pgio_mirror_idx);
1447
1448         switch (err) {
1449         case -NFS4ERR_RESET_TO_PNFS:
1450                 ff_layout_reset_write(hdr, true);
1451                 return task->tk_status;
1452         case -NFS4ERR_RESET_TO_MDS:
1453                 ff_layout_reset_write(hdr, false);
1454                 return task->tk_status;
1455         case -EAGAIN:
1456                 return -EAGAIN;
1457         }
1458
1459         if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1460             hdr->res.verf->committed == NFS_DATA_SYNC)
1461                 ff_layout_set_layoutcommit(hdr);
1462
1463         /* zero out fattr since we don't care DS attr at all */
1464         hdr->fattr.valid = 0;
1465         if (task->tk_status >= 0)
1466                 nfs_writeback_update_inode(hdr);
1467
1468         return 0;
1469 }
1470
1471 static int ff_layout_commit_done_cb(struct rpc_task *task,
1472                                      struct nfs_commit_data *data)
1473 {
1474         int err;
1475
1476         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1477         if (task->tk_status < 0)
1478                 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1479                                             data->args.offset, data->args.count,
1480                                             data->res.op_status, OP_COMMIT,
1481                                             task->tk_status);
1482         err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1483                                            data->lseg, data->ds_commit_index);
1484
1485         switch (err) {
1486         case -NFS4ERR_RESET_TO_PNFS:
1487                 pnfs_generic_prepare_to_resend_writes(data);
1488                 return -EAGAIN;
1489         case -NFS4ERR_RESET_TO_MDS:
1490                 pnfs_generic_prepare_to_resend_writes(data);
1491                 return -EAGAIN;
1492         case -EAGAIN:
1493                 rpc_restart_call_prepare(task);
1494                 return -EAGAIN;
1495         }
1496
1497         if (data->verf.committed == NFS_UNSTABLE
1498             && ff_layout_need_layoutcommit(data->lseg))
1499                 pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
1500
1501         return 0;
1502 }
1503
1504 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task,
1505                 struct nfs_pgio_header *hdr)
1506 {
1507         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1508                 return;
1509         nfs4_ff_layout_stat_io_start_write(hdr->inode,
1510                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1511                         hdr->args.count,
1512                         task->tk_start);
1513 }
1514
1515 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task,
1516                 struct nfs_pgio_header *hdr)
1517 {
1518         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1519                 return;
1520         nfs4_ff_layout_stat_io_end_write(task,
1521                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1522                         hdr->args.count, hdr->res.count,
1523                         hdr->res.verf->committed);
1524 }
1525
1526 static int ff_layout_write_prepare_common(struct rpc_task *task,
1527                                           struct nfs_pgio_header *hdr)
1528 {
1529         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1530                 rpc_exit(task, -EIO);
1531                 return -EIO;
1532         }
1533
1534         if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1535                 bool retry_pnfs;
1536
1537                 retry_pnfs = ff_layout_has_available_ds(hdr->lseg);
1538                 dprintk("%s task %u reset io to %s\n", __func__,
1539                         task->tk_pid, retry_pnfs ? "pNFS" : "MDS");
1540                 ff_layout_reset_write(hdr, retry_pnfs);
1541                 rpc_exit(task, 0);
1542                 return -EAGAIN;
1543         }
1544
1545         ff_layout_write_record_layoutstats_start(task, hdr);
1546         return 0;
1547 }
1548
1549 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1550 {
1551         struct nfs_pgio_header *hdr = data;
1552
1553         if (ff_layout_write_prepare_common(task, hdr))
1554                 return;
1555
1556         rpc_call_start(task);
1557 }
1558
1559 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1560 {
1561         struct nfs_pgio_header *hdr = data;
1562
1563         if (ff_layout_setup_sequence(hdr->ds_clp,
1564                                      &hdr->args.seq_args,
1565                                      &hdr->res.seq_res,
1566                                      task))
1567                 return;
1568
1569         if (ff_layout_write_prepare_common(task, hdr))
1570                 return;
1571
1572         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1573                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
1574                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1575 }
1576
1577 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1578 {
1579         struct nfs_pgio_header *hdr = data;
1580
1581         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1582             task->tk_status == 0) {
1583                 nfs4_sequence_done(task, &hdr->res.seq_res);
1584                 return;
1585         }
1586
1587         /* Note this may cause RPC to be resent */
1588         hdr->mds_ops->rpc_call_done(task, hdr);
1589 }
1590
1591 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1592 {
1593         struct nfs_pgio_header *hdr = data;
1594
1595         ff_layout_write_record_layoutstats_done(task, hdr);
1596         rpc_count_iostats_metrics(task,
1597             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1598 }
1599
1600 static void ff_layout_write_release(void *data)
1601 {
1602         struct nfs_pgio_header *hdr = data;
1603
1604         ff_layout_write_record_layoutstats_done(&hdr->task, hdr);
1605         pnfs_generic_rw_release(data);
1606 }
1607
1608 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task,
1609                 struct nfs_commit_data *cdata)
1610 {
1611         if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags))
1612                 return;
1613         nfs4_ff_layout_stat_io_start_write(cdata->inode,
1614                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1615                         0, task->tk_start);
1616 }
1617
1618 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task,
1619                 struct nfs_commit_data *cdata)
1620 {
1621         struct nfs_page *req;
1622         __u64 count = 0;
1623
1624         if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags))
1625                 return;
1626
1627         if (task->tk_status == 0) {
1628                 list_for_each_entry(req, &cdata->pages, wb_list)
1629                         count += req->wb_bytes;
1630         }
1631         nfs4_ff_layout_stat_io_end_write(task,
1632                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1633                         count, count, NFS_FILE_SYNC);
1634 }
1635
1636 static void ff_layout_commit_prepare_common(struct rpc_task *task,
1637                 struct nfs_commit_data *cdata)
1638 {
1639         ff_layout_commit_record_layoutstats_start(task, cdata);
1640 }
1641
1642 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1643 {
1644         ff_layout_commit_prepare_common(task, data);
1645         rpc_call_start(task);
1646 }
1647
1648 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1649 {
1650         struct nfs_commit_data *wdata = data;
1651
1652         if (ff_layout_setup_sequence(wdata->ds_clp,
1653                                  &wdata->args.seq_args,
1654                                  &wdata->res.seq_res,
1655                                  task))
1656                 return;
1657         ff_layout_commit_prepare_common(task, data);
1658 }
1659
1660 static void ff_layout_commit_done(struct rpc_task *task, void *data)
1661 {
1662         pnfs_generic_write_commit_done(task, data);
1663 }
1664
1665 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1666 {
1667         struct nfs_commit_data *cdata = data;
1668
1669         ff_layout_commit_record_layoutstats_done(task, cdata);
1670         rpc_count_iostats_metrics(task,
1671             &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1672 }
1673
1674 static void ff_layout_commit_release(void *data)
1675 {
1676         struct nfs_commit_data *cdata = data;
1677
1678         ff_layout_commit_record_layoutstats_done(&cdata->task, cdata);
1679         pnfs_generic_commit_release(data);
1680 }
1681
1682 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1683         .rpc_call_prepare = ff_layout_read_prepare_v3,
1684         .rpc_call_done = ff_layout_read_call_done,
1685         .rpc_count_stats = ff_layout_read_count_stats,
1686         .rpc_release = ff_layout_read_release,
1687 };
1688
1689 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1690         .rpc_call_prepare = ff_layout_read_prepare_v4,
1691         .rpc_call_done = ff_layout_read_call_done,
1692         .rpc_count_stats = ff_layout_read_count_stats,
1693         .rpc_release = ff_layout_read_release,
1694 };
1695
1696 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1697         .rpc_call_prepare = ff_layout_write_prepare_v3,
1698         .rpc_call_done = ff_layout_write_call_done,
1699         .rpc_count_stats = ff_layout_write_count_stats,
1700         .rpc_release = ff_layout_write_release,
1701 };
1702
1703 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1704         .rpc_call_prepare = ff_layout_write_prepare_v4,
1705         .rpc_call_done = ff_layout_write_call_done,
1706         .rpc_count_stats = ff_layout_write_count_stats,
1707         .rpc_release = ff_layout_write_release,
1708 };
1709
1710 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1711         .rpc_call_prepare = ff_layout_commit_prepare_v3,
1712         .rpc_call_done = ff_layout_commit_done,
1713         .rpc_count_stats = ff_layout_commit_count_stats,
1714         .rpc_release = ff_layout_commit_release,
1715 };
1716
1717 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1718         .rpc_call_prepare = ff_layout_commit_prepare_v4,
1719         .rpc_call_done = ff_layout_commit_done,
1720         .rpc_count_stats = ff_layout_commit_count_stats,
1721         .rpc_release = ff_layout_commit_release,
1722 };
1723
1724 static enum pnfs_try_status
1725 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1726 {
1727         struct pnfs_layout_segment *lseg = hdr->lseg;
1728         struct nfs4_pnfs_ds *ds;
1729         struct rpc_clnt *ds_clnt;
1730         struct rpc_cred *ds_cred;
1731         loff_t offset = hdr->args.offset;
1732         u32 idx = hdr->pgio_mirror_idx;
1733         int vers;
1734         struct nfs_fh *fh;
1735
1736         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1737                 __func__, hdr->inode->i_ino,
1738                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1739
1740         ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1741         if (!ds)
1742                 goto out_failed;
1743
1744         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1745                                                    hdr->inode);
1746         if (IS_ERR(ds_clnt))
1747                 goto out_failed;
1748
1749         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1750         if (!ds_cred)
1751                 goto out_failed;
1752
1753         vers = nfs4_ff_layout_ds_version(lseg, idx);
1754
1755         dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1756                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1757
1758         atomic_inc(&ds->ds_clp->cl_count);
1759         hdr->ds_clp = ds->ds_clp;
1760         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1761         if (fh)
1762                 hdr->args.fh = fh;
1763         /*
1764          * Note that if we ever decide to split across DSes,
1765          * then we may need to handle dense-like offsets.
1766          */
1767         hdr->args.offset = offset;
1768         hdr->mds_offset = offset;
1769
1770         /* Perform an asynchronous read to ds */
1771         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1772                           vers == 3 ? &ff_layout_read_call_ops_v3 :
1773                                       &ff_layout_read_call_ops_v4,
1774                           0, RPC_TASK_SOFTCONN);
1775         put_rpccred(ds_cred);
1776         return PNFS_ATTEMPTED;
1777
1778 out_failed:
1779         if (ff_layout_has_available_ds(lseg))
1780                 return PNFS_TRY_AGAIN;
1781         return PNFS_NOT_ATTEMPTED;
1782 }
1783
1784 /* Perform async writes. */
1785 static enum pnfs_try_status
1786 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1787 {
1788         struct pnfs_layout_segment *lseg = hdr->lseg;
1789         struct nfs4_pnfs_ds *ds;
1790         struct rpc_clnt *ds_clnt;
1791         struct rpc_cred *ds_cred;
1792         loff_t offset = hdr->args.offset;
1793         int vers;
1794         struct nfs_fh *fh;
1795         int idx = hdr->pgio_mirror_idx;
1796
1797         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1798         if (!ds)
1799                 return PNFS_NOT_ATTEMPTED;
1800
1801         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1802                                                    hdr->inode);
1803         if (IS_ERR(ds_clnt))
1804                 return PNFS_NOT_ATTEMPTED;
1805
1806         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1807         if (!ds_cred)
1808                 return PNFS_NOT_ATTEMPTED;
1809
1810         vers = nfs4_ff_layout_ds_version(lseg, idx);
1811
1812         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1813                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1814                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1815                 vers);
1816
1817         hdr->pgio_done_cb = ff_layout_write_done_cb;
1818         atomic_inc(&ds->ds_clp->cl_count);
1819         hdr->ds_clp = ds->ds_clp;
1820         hdr->ds_commit_idx = idx;
1821         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1822         if (fh)
1823                 hdr->args.fh = fh;
1824
1825         /*
1826          * Note that if we ever decide to split across DSes,
1827          * then we may need to handle dense-like offsets.
1828          */
1829         hdr->args.offset = offset;
1830
1831         /* Perform an asynchronous write */
1832         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1833                           vers == 3 ? &ff_layout_write_call_ops_v3 :
1834                                       &ff_layout_write_call_ops_v4,
1835                           sync, RPC_TASK_SOFTCONN);
1836         put_rpccred(ds_cred);
1837         return PNFS_ATTEMPTED;
1838 }
1839
1840 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1841 {
1842         return i;
1843 }
1844
1845 static struct nfs_fh *
1846 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1847 {
1848         struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1849
1850         /* FIXME: Assume that there is only one NFS version available
1851          * for the DS.
1852          */
1853         return &flseg->mirror_array[i]->fh_versions[0];
1854 }
1855
1856 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1857 {
1858         struct pnfs_layout_segment *lseg = data->lseg;
1859         struct nfs4_pnfs_ds *ds;
1860         struct rpc_clnt *ds_clnt;
1861         struct rpc_cred *ds_cred;
1862         u32 idx;
1863         int vers, ret;
1864         struct nfs_fh *fh;
1865
1866         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1867         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1868         if (!ds)
1869                 goto out_err;
1870
1871         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1872                                                    data->inode);
1873         if (IS_ERR(ds_clnt))
1874                 goto out_err;
1875
1876         ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1877         if (!ds_cred)
1878                 goto out_err;
1879
1880         vers = nfs4_ff_layout_ds_version(lseg, idx);
1881
1882         dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1883                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1884                 vers);
1885         data->commit_done_cb = ff_layout_commit_done_cb;
1886         data->cred = ds_cred;
1887         atomic_inc(&ds->ds_clp->cl_count);
1888         data->ds_clp = ds->ds_clp;
1889         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1890         if (fh)
1891                 data->args.fh = fh;
1892
1893         ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1894                                    vers == 3 ? &ff_layout_commit_call_ops_v3 :
1895                                                &ff_layout_commit_call_ops_v4,
1896                                    how, RPC_TASK_SOFTCONN);
1897         put_rpccred(ds_cred);
1898         return ret;
1899 out_err:
1900         pnfs_generic_prepare_to_resend_writes(data);
1901         pnfs_generic_commit_release(data);
1902         return -EAGAIN;
1903 }
1904
1905 static int
1906 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1907                            int how, struct nfs_commit_info *cinfo)
1908 {
1909         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1910                                             ff_layout_initiate_commit);
1911 }
1912
1913 static struct pnfs_ds_commit_info *
1914 ff_layout_get_ds_info(struct inode *inode)
1915 {
1916         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1917
1918         if (layout == NULL)
1919                 return NULL;
1920
1921         return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1922 }
1923
1924 static void
1925 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
1926 {
1927         nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1928                                                   id_node));
1929 }
1930
1931 static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1932                                   struct xdr_stream *xdr,
1933                                   const struct nfs4_layoutreturn_args *args)
1934 {
1935         struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1936         __be32 *start;
1937         int count = 0, ret = 0;
1938
1939         start = xdr_reserve_space(xdr, 4);
1940         if (unlikely(!start))
1941                 return -E2BIG;
1942
1943         /* This assume we always return _ALL_ layouts */
1944         spin_lock(&hdr->plh_inode->i_lock);
1945         ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1946         spin_unlock(&hdr->plh_inode->i_lock);
1947
1948         *start = cpu_to_be32(count);
1949
1950         return ret;
1951 }
1952
1953 /* report nothing for now */
1954 static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1955                                      struct xdr_stream *xdr,
1956                                      const struct nfs4_layoutreturn_args *args)
1957 {
1958         __be32 *p;
1959
1960         p = xdr_reserve_space(xdr, 4);
1961         if (likely(p))
1962                 *p = cpu_to_be32(0);
1963 }
1964
1965 static struct nfs4_deviceid_node *
1966 ff_layout_alloc_deviceid_node(struct nfs_server *server,
1967                               struct pnfs_device *pdev, gfp_t gfp_flags)
1968 {
1969         struct nfs4_ff_layout_ds *dsaddr;
1970
1971         dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
1972         if (!dsaddr)
1973                 return NULL;
1974         return &dsaddr->id_node;
1975 }
1976
1977 static void
1978 ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
1979                               struct xdr_stream *xdr,
1980                               const struct nfs4_layoutreturn_args *args)
1981 {
1982         struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
1983         __be32 *start;
1984
1985         dprintk("%s: Begin\n", __func__);
1986         start = xdr_reserve_space(xdr, 4);
1987         BUG_ON(!start);
1988
1989         ff_layout_encode_ioerr(flo, xdr, args);
1990         ff_layout_encode_iostats(flo, xdr, args);
1991
1992         *start = cpu_to_be32((xdr->p - start - 1) * 4);
1993         dprintk("%s: Return\n", __func__);
1994 }
1995
1996 static int
1997 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
1998 {
1999         const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
2000
2001         return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
2002 }
2003
2004 static size_t
2005 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
2006                           const int buflen)
2007 {
2008         const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
2009         const struct in6_addr *addr = &sin6->sin6_addr;
2010
2011         /*
2012          * RFC 4291, Section 2.2.2
2013          *
2014          * Shorthanded ANY address
2015          */
2016         if (ipv6_addr_any(addr))
2017                 return snprintf(buf, buflen, "::");
2018
2019         /*
2020          * RFC 4291, Section 2.2.2
2021          *
2022          * Shorthanded loopback address
2023          */
2024         if (ipv6_addr_loopback(addr))
2025                 return snprintf(buf, buflen, "::1");
2026
2027         /*
2028          * RFC 4291, Section 2.2.3
2029          *
2030          * Special presentation address format for mapped v4
2031          * addresses.
2032          */
2033         if (ipv6_addr_v4mapped(addr))
2034                 return snprintf(buf, buflen, "::ffff:%pI4",
2035                                         &addr->s6_addr32[3]);
2036
2037         /*
2038          * RFC 4291, Section 2.2.1
2039          */
2040         return snprintf(buf, buflen, "%pI6c", addr);
2041 }
2042
2043 /* Derived from rpc_sockaddr2uaddr */
2044 static void
2045 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
2046 {
2047         struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
2048         char portbuf[RPCBIND_MAXUADDRPLEN];
2049         char addrbuf[RPCBIND_MAXUADDRLEN];
2050         char *netid;
2051         unsigned short port;
2052         int len, netid_len;
2053         __be32 *p;
2054
2055         switch (sap->sa_family) {
2056         case AF_INET:
2057                 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
2058                         return;
2059                 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
2060                 netid = "tcp";
2061                 netid_len = 3;
2062                 break;
2063         case AF_INET6:
2064                 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
2065                         return;
2066                 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
2067                 netid = "tcp6";
2068                 netid_len = 4;
2069                 break;
2070         default:
2071                 /* we only support tcp and tcp6 */
2072                 WARN_ON_ONCE(1);
2073                 return;
2074         }
2075
2076         snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
2077         len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
2078
2079         p = xdr_reserve_space(xdr, 4 + netid_len);
2080         xdr_encode_opaque(p, netid, netid_len);
2081
2082         p = xdr_reserve_space(xdr, 4 + len);
2083         xdr_encode_opaque(p, addrbuf, len);
2084 }
2085
2086 static void
2087 ff_layout_encode_nfstime(struct xdr_stream *xdr,
2088                          ktime_t t)
2089 {
2090         struct timespec64 ts;
2091         __be32 *p;
2092
2093         p = xdr_reserve_space(xdr, 12);
2094         ts = ktime_to_timespec64(t);
2095         p = xdr_encode_hyper(p, ts.tv_sec);
2096         *p++ = cpu_to_be32(ts.tv_nsec);
2097 }
2098
2099 static void
2100 ff_layout_encode_io_latency(struct xdr_stream *xdr,
2101                             struct nfs4_ff_io_stat *stat)
2102 {
2103         __be32 *p;
2104
2105         p = xdr_reserve_space(xdr, 5 * 8);
2106         p = xdr_encode_hyper(p, stat->ops_requested);
2107         p = xdr_encode_hyper(p, stat->bytes_requested);
2108         p = xdr_encode_hyper(p, stat->ops_completed);
2109         p = xdr_encode_hyper(p, stat->bytes_completed);
2110         p = xdr_encode_hyper(p, stat->bytes_not_delivered);
2111         ff_layout_encode_nfstime(xdr, stat->total_busy_time);
2112         ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
2113 }
2114
2115 static void
2116 ff_layout_encode_layoutstats(struct xdr_stream *xdr,
2117                              struct nfs42_layoutstat_args *args,
2118                              struct nfs42_layoutstat_devinfo *devinfo)
2119 {
2120         struct nfs4_ff_layout_mirror *mirror = devinfo->layout_private;
2121         struct nfs4_pnfs_ds_addr *da;
2122         struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds;
2123         struct nfs_fh *fh = &mirror->fh_versions[0];
2124         __be32 *p, *start;
2125
2126         da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2127         dprintk("%s: DS %s: encoding address %s\n",
2128                 __func__, ds->ds_remotestr, da->da_remotestr);
2129         /* layoutupdate length */
2130         start = xdr_reserve_space(xdr, 4);
2131         /* netaddr4 */
2132         ff_layout_encode_netaddr(xdr, da);
2133         /* nfs_fh4 */
2134         p = xdr_reserve_space(xdr, 4 + fh->size);
2135         xdr_encode_opaque(p, fh->data, fh->size);
2136         /* ff_io_latency4 read */
2137         spin_lock(&mirror->lock);
2138         ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat);
2139         /* ff_io_latency4 write */
2140         ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat);
2141         spin_unlock(&mirror->lock);
2142         /* nfstime4 */
2143         ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time));
2144         /* bool */
2145         p = xdr_reserve_space(xdr, 4);
2146         *p = cpu_to_be32(false);
2147
2148         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2149 }
2150
2151 static int
2152 ff_layout_mirror_prepare_stats(struct nfs42_layoutstat_args *args,
2153                                struct pnfs_layout_hdr *lo,
2154                                int dev_limit)
2155 {
2156         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2157         struct nfs4_ff_layout_mirror *mirror;
2158         struct nfs4_deviceid_node *dev;
2159         struct nfs42_layoutstat_devinfo *devinfo;
2160         int i = 0;
2161
2162         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2163                 if (i >= dev_limit)
2164                         break;
2165                 if (!mirror->mirror_ds)
2166                         continue;
2167                 /* mirror refcount put in cleanup_layoutstats */
2168                 if (!atomic_inc_not_zero(&mirror->ref))
2169                         continue;
2170                 dev = &mirror->mirror_ds->id_node;
2171                 devinfo = &args->devinfo[i];
2172                 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE);
2173                 devinfo->offset = 0;
2174                 devinfo->length = NFS4_MAX_UINT64;
2175                 devinfo->read_count = mirror->read_stat.io_stat.ops_completed;
2176                 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed;
2177                 devinfo->write_count = mirror->write_stat.io_stat.ops_completed;
2178                 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed;
2179                 devinfo->layout_type = LAYOUT_FLEX_FILES;
2180                 devinfo->layoutstats_encode = ff_layout_encode_layoutstats;
2181                 devinfo->layout_private = mirror;
2182
2183                 i++;
2184         }
2185         return i;
2186 }
2187
2188 static int
2189 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2190 {
2191         struct nfs4_flexfile_layout *ff_layout;
2192         struct nfs4_ff_layout_mirror *mirror;
2193         int dev_count = 0;
2194
2195         spin_lock(&args->inode->i_lock);
2196         ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout);
2197         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2198                 if (atomic_read(&mirror->ref) != 0)
2199                         dev_count ++;
2200         }
2201         spin_unlock(&args->inode->i_lock);
2202         /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2203         if (dev_count > PNFS_LAYOUTSTATS_MAXDEV) {
2204                 dprintk("%s: truncating devinfo to limit (%d:%d)\n",
2205                         __func__, dev_count, PNFS_LAYOUTSTATS_MAXDEV);
2206                 dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2207         }
2208         args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO);
2209         if (!args->devinfo)
2210                 return -ENOMEM;
2211
2212         spin_lock(&args->inode->i_lock);
2213         args->num_dev = ff_layout_mirror_prepare_stats(args,
2214                         &ff_layout->generic_hdr, dev_count);
2215         spin_unlock(&args->inode->i_lock);
2216
2217         return 0;
2218 }
2219
2220 static void
2221 ff_layout_cleanup_layoutstats(struct nfs42_layoutstat_data *data)
2222 {
2223         struct nfs4_ff_layout_mirror *mirror;
2224         int i;
2225
2226         for (i = 0; i < data->args.num_dev; i++) {
2227                 mirror = data->args.devinfo[i].layout_private;
2228                 data->args.devinfo[i].layout_private = NULL;
2229                 ff_layout_put_mirror(mirror);
2230         }
2231 }
2232
2233 static struct pnfs_layoutdriver_type flexfilelayout_type = {
2234         .id                     = LAYOUT_FLEX_FILES,
2235         .name                   = "LAYOUT_FLEX_FILES",
2236         .owner                  = THIS_MODULE,
2237         .alloc_layout_hdr       = ff_layout_alloc_layout_hdr,
2238         .free_layout_hdr        = ff_layout_free_layout_hdr,
2239         .alloc_lseg             = ff_layout_alloc_lseg,
2240         .free_lseg              = ff_layout_free_lseg,
2241         .add_lseg               = ff_layout_add_lseg,
2242         .pg_read_ops            = &ff_layout_pg_read_ops,
2243         .pg_write_ops           = &ff_layout_pg_write_ops,
2244         .get_ds_info            = ff_layout_get_ds_info,
2245         .free_deviceid_node     = ff_layout_free_deviceid_node,
2246         .mark_request_commit    = pnfs_layout_mark_request_commit,
2247         .clear_request_commit   = pnfs_generic_clear_request_commit,
2248         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
2249         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
2250         .commit_pagelist        = ff_layout_commit_pagelist,
2251         .read_pagelist          = ff_layout_read_pagelist,
2252         .write_pagelist         = ff_layout_write_pagelist,
2253         .alloc_deviceid_node    = ff_layout_alloc_deviceid_node,
2254         .encode_layoutreturn    = ff_layout_encode_layoutreturn,
2255         .sync                   = pnfs_nfs_generic_sync,
2256         .prepare_layoutstats    = ff_layout_prepare_layoutstats,
2257         .cleanup_layoutstats    = ff_layout_cleanup_layoutstats,
2258 };
2259
2260 static int __init nfs4flexfilelayout_init(void)
2261 {
2262         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
2263                __func__);
2264         if (!ff_zero_group) {
2265                 ff_zero_group = groups_alloc(0);
2266                 if (!ff_zero_group)
2267                         return -ENOMEM;
2268         }
2269         return pnfs_register_layoutdriver(&flexfilelayout_type);
2270 }
2271
2272 static void __exit nfs4flexfilelayout_exit(void)
2273 {
2274         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
2275                __func__);
2276         pnfs_unregister_layoutdriver(&flexfilelayout_type);
2277         if (ff_zero_group) {
2278                 put_group_info(ff_zero_group);
2279                 ff_zero_group = NULL;
2280         }
2281 }
2282
2283 MODULE_ALIAS("nfs-layouttype4-4");
2284
2285 MODULE_LICENSE("GPL");
2286 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
2287
2288 module_init(nfs4flexfilelayout_init);
2289 module_exit(nfs4flexfilelayout_exit);