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 void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
  32                 struct nfs_pgio_header *hdr);
  33
  34 static struct pnfs_layout_hdr *
  35 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
  36 {
  37         struct nfs4_flexfile_layout *ffl;
  38
  39         ffl = kzalloc(sizeof(*ffl), gfp_flags);
  40         if (ffl) {
  41                 INIT_LIST_HEAD(&ffl->error_list);
  42                 INIT_LIST_HEAD(&ffl->mirrors);
  43                 ffl->last_report_time = ktime_get();
  44                 return &ffl->generic_hdr;
  45         } else
  46                 return NULL;
  47 }
  48
  49 static void
  50 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo)
  51 {
  52         struct nfs4_ff_layout_ds_err *err, *n;
  53
  54         list_for_each_entry_safe(err, n, &FF_LAYOUT_FROM_HDR(lo)->error_list,
  55                                  list) {
  56                 list_del(&err->list);
  57                 kfree(err);
  58         }
  59         kfree(FF_LAYOUT_FROM_HDR(lo));
  60 }
  61
  62 static int decode_pnfs_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid)
  63 {
  64         __be32 *p;
  65
  66         p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE);
  67         if (unlikely(p == NULL))
  68                 return -ENOBUFS;
  69         stateid->type = NFS4_PNFS_DS_STATEID_TYPE;
  70         memcpy(stateid->data, p, NFS4_STATEID_SIZE);
  71         dprintk("%s: stateid id= [%x%x%x%x]\n", __func__,
  72                 p[0], p[1], p[2], p[3]);
  73         return 0;
  74 }
  75
  76 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid)
  77 {
  78         __be32 *p;
  79
  80         p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE);
  81         if (unlikely(!p))
  82                 return -ENOBUFS;
  83         memcpy(devid, p, NFS4_DEVICEID4_SIZE);
  84         nfs4_print_deviceid(devid);
  85         return 0;
  86 }
  87
  88 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh)
  89 {
  90         __be32 *p;
  91
  92         p = xdr_inline_decode(xdr, 4);
  93         if (unlikely(!p))
  94                 return -ENOBUFS;
  95         fh->size = be32_to_cpup(p++);
  96         if (fh->size > NFS_MAXFHSIZE) {
  97                 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n",
  98                        fh->size);
  99                 return -EOVERFLOW;
 100         }
 101         /* fh.data */
 102         p = xdr_inline_decode(xdr, fh->size);
 103         if (unlikely(!p))
 104                 return -ENOBUFS;
 105         memcpy(&fh->data, p, fh->size);
 106         dprintk("%s: fh len %d\n", __func__, fh->size);
 107
 108         return 0;
 109 }
 110
 111 /*
 112  * Currently only stringified uids and gids are accepted.
 113  * I.e., kerberos is not supported to the DSes, so no pricipals.
 114  *
 115  * That means that one common function will suffice, but when
 116  * principals are added, this should be split to accomodate
 117  * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid().
 118  */
 119 static int
 120 decode_name(struct xdr_stream *xdr, u32 *id)
 121 {
 122         __be32 *p;
 123         int len;
 124
 125         /* opaque_length(4)*/
 126         p = xdr_inline_decode(xdr, 4);
 127         if (unlikely(!p))
 128                 return -ENOBUFS;
 129         len = be32_to_cpup(p++);
 130         if (len < 0)
 131                 return -EINVAL;
 132
 133         dprintk("%s: len %u\n", __func__, len);
 134
 135         /* opaque body */
 136         p = xdr_inline_decode(xdr, len);
 137         if (unlikely(!p))
 138                 return -ENOBUFS;
 139
 140         if (!nfs_map_string_to_numeric((char *)p, len, id))
 141                 return -EINVAL;
 142
 143         return 0;
 144 }
 145
 146 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1,
 147                 const struct nfs4_ff_layout_mirror *m2)
 148 {
 149         int i, j;
 150
 151         if (m1->fh_versions_cnt != m2->fh_versions_cnt)
 152                 return false;
 153         for (i = 0; i < m1->fh_versions_cnt; i++) {
 154                 bool found_fh = false;
 155                 for (j = 0; j < m2->fh_versions_cnt; j++) {
 156                         if (nfs_compare_fh(&m1->fh_versions[i],
 157                                         &m2->fh_versions[j]) == 0) {
 158                                 found_fh = true;
 159                                 break;
 160                         }
 161                 }
 162                 if (!found_fh)
 163                         return false;
 164         }
 165         return true;
 166 }
 167
 168 static struct nfs4_ff_layout_mirror *
 169 ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
 170                 struct nfs4_ff_layout_mirror *mirror)
 171 {
 172         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
 173         struct nfs4_ff_layout_mirror *pos;
 174         struct inode *inode = lo->plh_inode;
 175
 176         spin_lock(&inode->i_lock);
 177         list_for_each_entry(pos, &ff_layout->mirrors, mirrors) {
 178                 if (mirror->mirror_ds != pos->mirror_ds)
 179                         continue;
 180                 if (!ff_mirror_match_fh(mirror, pos))
 181                         continue;
 182                 if (atomic_inc_not_zero(&pos->ref)) {
 183                         spin_unlock(&inode->i_lock);
 184                         return pos;
 185                 }
 186         }
 187         list_add(&mirror->mirrors, &ff_layout->mirrors);
 188         mirror->layout = lo;
 189         spin_unlock(&inode->i_lock);
 190         return mirror;
 191 }
 192
 193 static void
 194 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror)
 195 {
 196         struct inode *inode;
 197         if (mirror->layout == NULL)
 198                 return;
 199         inode = mirror->layout->plh_inode;
 200         spin_lock(&inode->i_lock);
 201         list_del(&mirror->mirrors);
 202         spin_unlock(&inode->i_lock);
 203         mirror->layout = NULL;
 204 }
 205
 206 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(gfp_t gfp_flags)
 207 {
 208         struct nfs4_ff_layout_mirror *mirror;
 209
 210         mirror = kzalloc(sizeof(*mirror), gfp_flags);
 211         if (mirror != NULL) {
 212                 spin_lock_init(&mirror->lock);
 213                 atomic_set(&mirror->ref, 1);
 214                 INIT_LIST_HEAD(&mirror->mirrors);
 215         }
 216         return mirror;
 217 }
 218
 219 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror)
 220 {
 221         struct rpc_cred *cred;
 222
 223         ff_layout_remove_mirror(mirror);
 224         kfree(mirror->fh_versions);
 225         cred = rcu_access_pointer(mirror->ro_cred);
 226         if (cred)
 227                 put_rpccred(cred);
 228         cred = rcu_access_pointer(mirror->rw_cred);
 229         if (cred)
 230                 put_rpccred(cred);
 231         nfs4_ff_layout_put_deviceid(mirror->mirror_ds);
 232         kfree(mirror);
 233 }
 234
 235 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror)
 236 {
 237         if (mirror != NULL && atomic_dec_and_test(&mirror->ref))
 238                 ff_layout_free_mirror(mirror);
 239 }
 240
 241 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls)
 242 {
 243         int i;
 244
 245         if (fls->mirror_array) {
 246                 for (i = 0; i < fls->mirror_array_cnt; i++) {
 247                         /* normally mirror_ds is freed in
 248                          * .free_deviceid_node but we still do it here
 249                          * for .alloc_lseg error path */
 250                         ff_layout_put_mirror(fls->mirror_array[i]);
 251                 }
 252                 kfree(fls->mirror_array);
 253                 fls->mirror_array = NULL;
 254         }
 255 }
 256
 257 static int ff_layout_check_layout(struct nfs4_layoutget_res *lgr)
 258 {
 259         int ret = 0;
 260
 261         dprintk("--> %s\n", __func__);
 262
 263         /* FIXME: remove this check when layout segment support is added */
 264         if (lgr->range.offset != 0 ||
 265             lgr->range.length != NFS4_MAX_UINT64) {
 266                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 267                         __func__);
 268                 ret = -EINVAL;
 269         }
 270
 271         dprintk("--> %s returns %d\n", __func__, ret);
 272         return ret;
 273 }
 274
 275 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls)
 276 {
 277         if (fls) {
 278                 ff_layout_free_mirror_array(fls);
 279                 kfree(fls);
 280         }
 281 }
 282
 283 static bool
 284 ff_lseg_range_is_after(const struct pnfs_layout_range *l1,
 285                 const struct pnfs_layout_range *l2)
 286 {
 287         u64 end1, end2;
 288
 289         if (l1->iomode != l2->iomode)
 290                 return l1->iomode != IOMODE_READ;
 291         end1 = pnfs_calc_offset_end(l1->offset, l1->length);
 292         end2 = pnfs_calc_offset_end(l2->offset, l2->length);
 293         if (end1 < l2->offset)
 294                 return false;
 295         if (end2 < l1->offset)
 296                 return true;
 297         return l2->offset <= l1->offset;
 298 }
 299
 300 static bool
 301 ff_lseg_merge(struct pnfs_layout_segment *new,
 302                 struct pnfs_layout_segment *old)
 303 {
 304         u64 new_end, old_end;
 305
 306         if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags))
 307                 return false;
 308         if (new->pls_range.iomode != old->pls_range.iomode)
 309                 return false;
 310         old_end = pnfs_calc_offset_end(old->pls_range.offset,
 311                         old->pls_range.length);
 312         if (old_end < new->pls_range.offset)
 313                 return false;
 314         new_end = pnfs_calc_offset_end(new->pls_range.offset,
 315                         new->pls_range.length);
 316         if (new_end < old->pls_range.offset)
 317                 return false;
 318
 319         /* Mergeable: copy info from 'old' to 'new' */
 320         if (new_end < old_end)
 321                 new_end = old_end;
 322         if (new->pls_range.offset < old->pls_range.offset)
 323                 new->pls_range.offset = old->pls_range.offset;
 324         new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset,
 325                         new_end);
 326         if (test_bit(NFS_LSEG_ROC, &old->pls_flags))
 327                 set_bit(NFS_LSEG_ROC, &new->pls_flags);
 328         return true;
 329 }
 330
 331 static void
 332 ff_layout_add_lseg(struct pnfs_layout_hdr *lo,
 333                 struct pnfs_layout_segment *lseg,
 334                 struct list_head *free_me)
 335 {
 336         pnfs_generic_layout_insert_lseg(lo, lseg,
 337                         ff_lseg_range_is_after,
 338                         ff_lseg_merge,
 339                         free_me);
 340 }
 341
 342 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls)
 343 {
 344         int i, j;
 345
 346         for (i = 0; i < fls->mirror_array_cnt - 1; i++) {
 347                 for (j = i + 1; j < fls->mirror_array_cnt; j++)
 348                         if (fls->mirror_array[i]->efficiency <
 349                             fls->mirror_array[j]->efficiency)
 350                                 swap(fls->mirror_array[i],
 351                                      fls->mirror_array[j]);
 352         }
 353 }
 354
 355 static void ff_layout_mark_devices_valid(struct nfs4_ff_layout_segment *fls)
 356 {
 357         struct nfs4_deviceid_node *node;
 358         int i;
 359
 360         if (!(fls->flags & FF_FLAGS_NO_IO_THRU_MDS))
 361                 return;
 362         for (i = 0; i < fls->mirror_array_cnt; i++) {
 363                 node = &fls->mirror_array[i]->mirror_ds->id_node;
 364                 clear_bit(NFS_DEVICEID_UNAVAILABLE, &node->flags);
 365         }
 366 }
 367
 368 static struct pnfs_layout_segment *
 369 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh,
 370                      struct nfs4_layoutget_res *lgr,
 371                      gfp_t gfp_flags)
 372 {
 373         struct pnfs_layout_segment *ret;
 374         struct nfs4_ff_layout_segment *fls = NULL;
 375         struct xdr_stream stream;
 376         struct xdr_buf buf;
 377         struct page *scratch;
 378         u64 stripe_unit;
 379         u32 mirror_array_cnt;
 380         __be32 *p;
 381         int i, rc;
 382
 383         dprintk("--> %s\n", __func__);
 384         scratch = alloc_page(gfp_flags);
 385         if (!scratch)
 386                 return ERR_PTR(-ENOMEM);
 387
 388         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages,
 389                               lgr->layoutp->len);
 390         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 391
 392         /* stripe unit and mirror_array_cnt */
 393         rc = -EIO;
 394         p = xdr_inline_decode(&stream, 8 + 4);
 395         if (!p)
 396                 goto out_err_free;
 397
 398         p = xdr_decode_hyper(p, &stripe_unit);
 399         mirror_array_cnt = be32_to_cpup(p++);
 400         dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__,
 401                 stripe_unit, mirror_array_cnt);
 402
 403         if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT ||
 404             mirror_array_cnt == 0)
 405                 goto out_err_free;
 406
 407         rc = -ENOMEM;
 408         fls = kzalloc(sizeof(*fls), gfp_flags);
 409         if (!fls)
 410                 goto out_err_free;
 411
 412         fls->mirror_array_cnt = mirror_array_cnt;
 413         fls->stripe_unit = stripe_unit;
 414         fls->mirror_array = kcalloc(fls->mirror_array_cnt,
 415                                     sizeof(fls->mirror_array[0]), gfp_flags);
 416         if (fls->mirror_array == NULL)
 417                 goto out_err_free;
 418
 419         for (i = 0; i < fls->mirror_array_cnt; i++) {
 420                 struct nfs4_ff_layout_mirror *mirror;
 421                 struct nfs4_deviceid devid;
 422                 struct nfs4_deviceid_node *idnode;
 423                 struct auth_cred acred = { .group_info = ff_zero_group };
 424                 struct rpc_cred __rcu *cred;
 425                 u32 ds_count, fh_count, id;
 426                 int j;
 427
 428                 rc = -EIO;
 429                 p = xdr_inline_decode(&stream, 4);
 430                 if (!p)
 431                         goto out_err_free;
 432                 ds_count = be32_to_cpup(p);
 433
 434                 /* FIXME: allow for striping? */
 435                 if (ds_count != 1)
 436                         goto out_err_free;
 437
 438                 fls->mirror_array[i] = ff_layout_alloc_mirror(gfp_flags);
 439                 if (fls->mirror_array[i] == NULL) {
 440                         rc = -ENOMEM;
 441                         goto out_err_free;
 442                 }
 443
 444                 fls->mirror_array[i]->ds_count = ds_count;
 445
 446                 /* deviceid */
 447                 rc = decode_deviceid(&stream, &devid);
 448                 if (rc)
 449                         goto out_err_free;
 450
 451                 idnode = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
 452                                                 &devid, lh->plh_lc_cred,
 453                                                 gfp_flags);
 454                 /*
 455                  * upon success, mirror_ds is allocated by previous
 456                  * getdeviceinfo, or newly by .alloc_deviceid_node
 457                  * nfs4_find_get_deviceid failure is indeed getdeviceinfo falure
 458                  */
 459                 if (idnode)
 460                         fls->mirror_array[i]->mirror_ds =
 461                                 FF_LAYOUT_MIRROR_DS(idnode);
 462                 else
 463                         goto out_err_free;
 464
 465                 /* efficiency */
 466                 rc = -EIO;
 467                 p = xdr_inline_decode(&stream, 4);
 468                 if (!p)
 469                         goto out_err_free;
 470                 fls->mirror_array[i]->efficiency = be32_to_cpup(p);
 471
 472                 /* stateid */
 473                 rc = decode_pnfs_stateid(&stream, &fls->mirror_array[i]->stateid);
 474                 if (rc)
 475                         goto out_err_free;
 476
 477                 /* fh */
 478                 rc = -EIO;
 479                 p = xdr_inline_decode(&stream, 4);
 480                 if (!p)
 481                         goto out_err_free;
 482                 fh_count = be32_to_cpup(p);
 483
 484                 fls->mirror_array[i]->fh_versions =
 485                         kzalloc(fh_count * sizeof(struct nfs_fh),
 486                                 gfp_flags);
 487                 if (fls->mirror_array[i]->fh_versions == NULL) {
 488                         rc = -ENOMEM;
 489                         goto out_err_free;
 490                 }
 491
 492                 for (j = 0; j < fh_count; j++) {
 493                         rc = decode_nfs_fh(&stream,
 494                                            &fls->mirror_array[i]->fh_versions[j]);
 495                         if (rc)
 496                                 goto out_err_free;
 497                 }
 498
 499                 fls->mirror_array[i]->fh_versions_cnt = fh_count;
 500
 501                 /* user */
 502                 rc = decode_name(&stream, &id);
 503                 if (rc)
 504                         goto out_err_free;
 505
 506                 acred.uid = make_kuid(&init_user_ns, id);
 507
 508                 /* group */
 509                 rc = decode_name(&stream, &id);
 510                 if (rc)
 511                         goto out_err_free;
 512
 513                 acred.gid = make_kgid(&init_user_ns, id);
 514
 515                 /* find the cred for it */
 516                 rcu_assign_pointer(cred, rpc_lookup_generic_cred(&acred, 0, gfp_flags));
 517                 if (IS_ERR(cred)) {
 518                         rc = PTR_ERR(cred);
 519                         goto out_err_free;
 520                 }
 521
 522                 if (lgr->range.iomode == IOMODE_READ)
 523                         rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred);
 524                 else
 525                         rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred);
 526
 527                 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]);
 528                 if (mirror != fls->mirror_array[i]) {
 529                         /* swap cred ptrs so free_mirror will clean up old */
 530                         if (lgr->range.iomode == IOMODE_READ) {
 531                                 cred = xchg(&mirror->ro_cred, cred);
 532                                 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred);
 533                         } else {
 534                                 cred = xchg(&mirror->rw_cred, cred);
 535                                 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred);
 536                         }
 537                         ff_layout_free_mirror(fls->mirror_array[i]);
 538                         fls->mirror_array[i] = mirror;
 539                 }
 540
 541                 dprintk("%s: iomode %s uid %u gid %u\n", __func__,
 542                         lgr->range.iomode == IOMODE_READ ? "READ" : "RW",
 543                         from_kuid(&init_user_ns, acred.uid),
 544                         from_kgid(&init_user_ns, acred.gid));
 545         }
 546
 547         p = xdr_inline_decode(&stream, 4);
 548         if (!p)
 549                 goto out_sort_mirrors;
 550         fls->flags = be32_to_cpup(p);
 551
 552         p = xdr_inline_decode(&stream, 4);
 553         if (!p)
 554                 goto out_sort_mirrors;
 555         for (i=0; i < fls->mirror_array_cnt; i++)
 556                 fls->mirror_array[i]->report_interval = be32_to_cpup(p);
 557
 558 out_sort_mirrors:
 559         ff_layout_sort_mirrors(fls);
 560         rc = ff_layout_check_layout(lgr);
 561         if (rc)
 562                 goto out_err_free;
 563         ff_layout_mark_devices_valid(fls);
 564
 565         ret = &fls->generic_hdr;
 566         dprintk("<-- %s (success)\n", __func__);
 567 out_free_page:
 568         __free_page(scratch);
 569         return ret;
 570 out_err_free:
 571         _ff_layout_free_lseg(fls);
 572         ret = ERR_PTR(rc);
 573         dprintk("<-- %s (%d)\n", __func__, rc);
 574         goto out_free_page;
 575 }
 576
 577 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout)
 578 {
 579         struct pnfs_layout_segment *lseg;
 580
 581         list_for_each_entry(lseg, &layout->plh_segs, pls_list)
 582                 if (lseg->pls_range.iomode == IOMODE_RW)
 583                         return true;
 584
 585         return false;
 586 }
 587
 588 static void
 589 ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
 590 {
 591         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 592
 593         dprintk("--> %s\n", __func__);
 594
 595         if (lseg->pls_range.iomode == IOMODE_RW) {
 596                 struct nfs4_flexfile_layout *ffl;
 597                 struct inode *inode;
 598
 599                 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
 600                 inode = ffl->generic_hdr.plh_inode;
 601                 spin_lock(&inode->i_lock);
 602                 if (!ff_layout_has_rw_segments(lseg->pls_layout)) {
 603                         ffl->commit_info.nbuckets = 0;
 604                         kfree(ffl->commit_info.buckets);
 605                         ffl->commit_info.buckets = NULL;
 606                 }
 607                 spin_unlock(&inode->i_lock);
 608         }
 609         _ff_layout_free_lseg(fls);
 610 }
 611
 612 /* Return 1 until we have multiple lsegs support */
 613 static int
 614 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
 615 {
 616         return 1;
 617 }
 618
 619 static void
 620 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 621 {
 622         /* first IO request? */
 623         if (atomic_inc_return(&timer->n_ops) == 1) {
 624                 timer->start_time = now;
 625         }
 626 }
 627
 628 static ktime_t
 629 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 630 {
 631         ktime_t start;
 632
 633         if (atomic_dec_return(&timer->n_ops) < 0)
 634                 WARN_ON_ONCE(1);
 635
 636         start = timer->start_time;
 637         timer->start_time = now;
 638         return ktime_sub(now, start);
 639 }
 640
 641 static bool
 642 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
 643                             struct nfs4_ff_layoutstat *layoutstat,
 644                             ktime_t now)
 645 {
 646         static const ktime_t notime = {0};
 647         s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL;
 648         struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(mirror->layout);
 649
 650         nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now);
 651         if (ktime_equal(mirror->start_time, notime))
 652                 mirror->start_time = now;
 653         if (mirror->report_interval != 0)
 654                 report_interval = (s64)mirror->report_interval * 1000LL;
 655         else if (layoutstats_timer != 0)
 656                 report_interval = (s64)layoutstats_timer * 1000LL;
 657         if (ktime_to_ms(ktime_sub(now, ffl->last_report_time)) >=
 658                         report_interval) {
 659                 ffl->last_report_time = now;
 660                 return true;
 661         }
 662
 663         return false;
 664 }
 665
 666 static void
 667 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
 668                 __u64 requested)
 669 {
 670         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 671
 672         iostat->ops_requested++;
 673         iostat->bytes_requested += requested;
 674 }
 675
 676 static void
 677 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
 678                 __u64 requested,
 679                 __u64 completed,
 680                 ktime_t time_completed,
 681                 ktime_t time_started)
 682 {
 683         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 684         ktime_t completion_time = ktime_sub(time_completed, time_started);
 685         ktime_t timer;
 686
 687         iostat->ops_completed++;
 688         iostat->bytes_completed += completed;
 689         iostat->bytes_not_delivered += requested - completed;
 690
 691         timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed);
 692         iostat->total_busy_time =
 693                         ktime_add(iostat->total_busy_time, timer);
 694         iostat->aggregate_completion_time =
 695                         ktime_add(iostat->aggregate_completion_time,
 696                                         completion_time);
 697 }
 698
 699 static void
 700 nfs4_ff_layout_stat_io_start_read(struct inode *inode,
 701                 struct nfs4_ff_layout_mirror *mirror,
 702                 __u64 requested, ktime_t now)
 703 {
 704         bool report;
 705
 706         spin_lock(&mirror->lock);
 707         report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now);
 708         nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
 709         spin_unlock(&mirror->lock);
 710
 711         if (report)
 712                 pnfs_report_layoutstat(inode, GFP_KERNEL);
 713 }
 714
 715 static void
 716 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
 717                 struct nfs4_ff_layout_mirror *mirror,
 718                 __u64 requested,
 719                 __u64 completed)
 720 {
 721         spin_lock(&mirror->lock);
 722         nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
 723                         requested, completed,
 724                         ktime_get(), task->tk_start);
 725         spin_unlock(&mirror->lock);
 726 }
 727
 728 static void
 729 nfs4_ff_layout_stat_io_start_write(struct inode *inode,
 730                 struct nfs4_ff_layout_mirror *mirror,
 731                 __u64 requested, ktime_t now)
 732 {
 733         bool report;
 734
 735         spin_lock(&mirror->lock);
 736         report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now);
 737         nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
 738         spin_unlock(&mirror->lock);
 739
 740         if (report)
 741                 pnfs_report_layoutstat(inode, GFP_NOIO);
 742 }
 743
 744 static void
 745 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
 746                 struct nfs4_ff_layout_mirror *mirror,
 747                 __u64 requested,
 748                 __u64 completed,
 749                 enum nfs3_stable_how committed)
 750 {
 751         if (committed == NFS_UNSTABLE)
 752                 requested = completed = 0;
 753
 754         spin_lock(&mirror->lock);
 755         nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
 756                         requested, completed, ktime_get(), task->tk_start);
 757         spin_unlock(&mirror->lock);
 758 }
 759
 760 static int
 761 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 762                             struct nfs_commit_info *cinfo,
 763                             gfp_t gfp_flags)
 764 {
 765         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 766         struct pnfs_commit_bucket *buckets;
 767         int size;
 768
 769         if (cinfo->ds->nbuckets != 0) {
 770                 /* This assumes there is only one RW lseg per file.
 771                  * To support multiple lseg per file, we need to
 772                  * change struct pnfs_commit_bucket to allow dynamic
 773                  * increasing nbuckets.
 774                  */
 775                 return 0;
 776         }
 777
 778         size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg);
 779
 780         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 781                           gfp_flags);
 782         if (!buckets)
 783                 return -ENOMEM;
 784         else {
 785                 int i;
 786
 787                 spin_lock(&cinfo->inode->i_lock);
 788                 if (cinfo->ds->nbuckets != 0)
 789                         kfree(buckets);
 790                 else {
 791                         cinfo->ds->buckets = buckets;
 792                         cinfo->ds->nbuckets = size;
 793                         for (i = 0; i < size; i++) {
 794                                 INIT_LIST_HEAD(&buckets[i].written);
 795                                 INIT_LIST_HEAD(&buckets[i].committing);
 796                                 /* mark direct verifier as unset */
 797                                 buckets[i].direct_verf.committed =
 798                                         NFS_INVALID_STABLE_HOW;
 799                         }
 800                 }
 801                 spin_unlock(&cinfo->inode->i_lock);
 802                 return 0;
 803         }
 804 }
 805
 806 static struct nfs4_pnfs_ds *
 807 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg,
 808                                   int start_idx,
 809                                   int *best_idx)
 810 {
 811         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 812         struct nfs4_pnfs_ds *ds;
 813         bool fail_return = false;
 814         int idx;
 815
 816         /* mirrors are sorted by efficiency */
 817         for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) {
 818                 if (idx+1 == fls->mirror_array_cnt)
 819                         fail_return = true;
 820                 ds = nfs4_ff_layout_prepare_ds(lseg, idx, fail_return);
 821                 if (ds) {
 822                         *best_idx = idx;
 823                         return ds;
 824                 }
 825         }
 826
 827         return NULL;
 828 }
 829
 830 static void
 831 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio,
 832                       struct nfs_page *req,
 833                       bool strict_iomode)
 834 {
 835 retry_strict:
 836         pnfs_put_lseg(pgio->pg_lseg);
 837         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 838                                            req->wb_context,
 839                                            0,
 840                                            NFS4_MAX_UINT64,
 841                                            IOMODE_READ,
 842                                            strict_iomode,
 843                                            GFP_KERNEL);
 844         if (IS_ERR(pgio->pg_lseg)) {
 845                 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 846                 pgio->pg_lseg = NULL;
 847         }
 848
 849         /* If we don't have checking, do get a IOMODE_RW
 850          * segment, and the server wants to avoid READs
 851          * there, then retry!
 852          */
 853         if (pgio->pg_lseg && !strict_iomode &&
 854             ff_layout_avoid_read_on_rw(pgio->pg_lseg)) {
 855                 strict_iomode = true;
 856                 goto retry_strict;
 857         }
 858 }
 859
 860 static void
 861 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 862                         struct nfs_page *req)
 863 {
 864         struct nfs_pgio_mirror *pgm;
 865         struct nfs4_ff_layout_mirror *mirror;
 866         struct nfs4_pnfs_ds *ds;
 867         int ds_idx;
 868
 869 retry:
 870         /* Use full layout for now */
 871         if (!pgio->pg_lseg)
 872                 ff_layout_pg_get_read(pgio, req, false);
 873         else if (ff_layout_avoid_read_on_rw(pgio->pg_lseg))
 874                 ff_layout_pg_get_read(pgio, req, true);
 875
 876         /* If no lseg, fall back to read through mds */
 877         if (pgio->pg_lseg == NULL)
 878                 goto out_mds;
 879
 880         ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx);
 881         if (!ds) {
 882                 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg))
 883                         goto out_mds;
 884                 pnfs_put_lseg(pgio->pg_lseg);
 885                 pgio->pg_lseg = NULL;
 886                 /* Sleep for 1 second before retrying */
 887                 ssleep(1);
 888                 goto retry;
 889         }
 890
 891         mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
 892
 893         pgio->pg_mirror_idx = ds_idx;
 894
 895         /* read always uses only one mirror - idx 0 for pgio layer */
 896         pgm = &pgio->pg_mirrors[0];
 897         pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
 898
 899         return;
 900 out_mds:
 901         pnfs_put_lseg(pgio->pg_lseg);
 902         pgio->pg_lseg = NULL;
 903         nfs_pageio_reset_read_mds(pgio);
 904 }
 905
 906 static void
 907 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 908                         struct nfs_page *req)
 909 {
 910         struct nfs4_ff_layout_mirror *mirror;
 911         struct nfs_pgio_mirror *pgm;
 912         struct nfs_commit_info cinfo;
 913         struct nfs4_pnfs_ds *ds;
 914         int i;
 915         int status;
 916
 917 retry:
 918         if (!pgio->pg_lseg) {
 919                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 920                                                    req->wb_context,
 921                                                    0,
 922                                                    NFS4_MAX_UINT64,
 923                                                    IOMODE_RW,
 924                                                    false,
 925                                                    GFP_NOFS);
 926                 if (IS_ERR(pgio->pg_lseg)) {
 927                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 928                         pgio->pg_lseg = NULL;
 929                         return;
 930                 }
 931         }
 932         /* If no lseg, fall back to write through mds */
 933         if (pgio->pg_lseg == NULL)
 934                 goto out_mds;
 935
 936         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 937         status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 938         if (status < 0)
 939                 goto out_mds;
 940
 941         /* Use a direct mapping of ds_idx to pgio mirror_idx */
 942         if (pgio->pg_mirror_count != FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg))
 943                 goto out_eagain;
 944
 945         for (i = 0; i < pgio->pg_mirror_count; i++) {
 946                 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
 947                 if (!ds) {
 948                         if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg))
 949                                 goto out_mds;
 950                         pnfs_put_lseg(pgio->pg_lseg);
 951                         pgio->pg_lseg = NULL;
 952                         /* Sleep for 1 second before retrying */
 953                         ssleep(1);
 954                         goto retry;
 955                 }
 956                 pgm = &pgio->pg_mirrors[i];
 957                 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
 958                 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
 959         }
 960
 961         return;
 962 out_eagain:
 963         pnfs_generic_pg_cleanup(pgio);
 964         pgio->pg_error = -EAGAIN;
 965         return;
 966 out_mds:
 967         pnfs_put_lseg(pgio->pg_lseg);
 968         pgio->pg_lseg = NULL;
 969         nfs_pageio_reset_write_mds(pgio);
 970         pgio->pg_error = -EAGAIN;
 971 }
 972
 973 static unsigned int
 974 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
 975                                     struct nfs_page *req)
 976 {
 977         if (!pgio->pg_lseg) {
 978                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 979                                                    req->wb_context,
 980                                                    0,
 981                                                    NFS4_MAX_UINT64,
 982                                                    IOMODE_RW,
 983                                                    false,
 984                                                    GFP_NOFS);
 985                 if (IS_ERR(pgio->pg_lseg)) {
 986                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 987                         pgio->pg_lseg = NULL;
 988                         goto out;
 989                 }
 990         }
 991         if (pgio->pg_lseg)
 992                 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
 993
 994         /* no lseg means that pnfs is not in use, so no mirroring here */
 995         nfs_pageio_reset_write_mds(pgio);
 996 out:
 997         return 1;
 998 }
 999
1000 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
1001         .pg_init = ff_layout_pg_init_read,
1002         .pg_test = pnfs_generic_pg_test,
1003         .pg_doio = pnfs_generic_pg_readpages,
1004         .pg_cleanup = pnfs_generic_pg_cleanup,
1005 };
1006
1007 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
1008         .pg_init = ff_layout_pg_init_write,
1009         .pg_test = pnfs_generic_pg_test,
1010         .pg_doio = pnfs_generic_pg_writepages,
1011         .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
1012         .pg_cleanup = pnfs_generic_pg_cleanup,
1013 };
1014
1015 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
1016 {
1017         struct rpc_task *task = &hdr->task;
1018
1019         pnfs_layoutcommit_inode(hdr->inode, false);
1020
1021         if (retry_pnfs) {
1022                 dprintk("%s Reset task %5u for i/o through pNFS "
1023                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1024                         hdr->task.tk_pid,
1025                         hdr->inode->i_sb->s_id,
1026                         (unsigned long long)NFS_FILEID(hdr->inode),
1027                         hdr->args.count,
1028                         (unsigned long long)hdr->args.offset);
1029
1030                 hdr->completion_ops->reschedule_io(hdr);
1031                 return;
1032         }
1033
1034         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1035                 dprintk("%s Reset task %5u for i/o through MDS "
1036                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1037                         hdr->task.tk_pid,
1038                         hdr->inode->i_sb->s_id,
1039                         (unsigned long long)NFS_FILEID(hdr->inode),
1040                         hdr->args.count,
1041                         (unsigned long long)hdr->args.offset);
1042
1043                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
1044         }
1045 }
1046
1047 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
1048 {
1049         struct rpc_task *task = &hdr->task;
1050
1051         pnfs_layoutcommit_inode(hdr->inode, false);
1052
1053         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1054                 dprintk("%s Reset task %5u for i/o through MDS "
1055                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1056                         hdr->task.tk_pid,
1057                         hdr->inode->i_sb->s_id,
1058                         (unsigned long long)NFS_FILEID(hdr->inode),
1059                         hdr->args.count,
1060                         (unsigned long long)hdr->args.offset);
1061
1062                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
1063         }
1064 }
1065
1066 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
1067                                            struct nfs4_state *state,
1068                                            struct nfs_client *clp,
1069                                            struct pnfs_layout_segment *lseg,
1070                                            int idx)
1071 {
1072         struct pnfs_layout_hdr *lo = lseg->pls_layout;
1073         struct inode *inode = lo->plh_inode;
1074         struct nfs_server *mds_server = NFS_SERVER(inode);
1075
1076         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1077         struct nfs_client *mds_client = mds_server->nfs_client;
1078         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
1079
1080         switch (task->tk_status) {
1081         /* MDS state errors */
1082         case -NFS4ERR_DELEG_REVOKED:
1083         case -NFS4ERR_ADMIN_REVOKED:
1084         case -NFS4ERR_BAD_STATEID:
1085                 if (state == NULL)
1086                         break;
1087                 nfs_remove_bad_delegation(state->inode, NULL);
1088         case -NFS4ERR_OPENMODE:
1089                 if (state == NULL)
1090                         break;
1091                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1092                         goto out_bad_stateid;
1093                 goto wait_on_recovery;
1094         case -NFS4ERR_EXPIRED:
1095                 if (state != NULL) {
1096                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1097                                 goto out_bad_stateid;
1098                 }
1099                 nfs4_schedule_lease_recovery(mds_client);
1100                 goto wait_on_recovery;
1101         /* DS session errors */
1102         case -NFS4ERR_BADSESSION:
1103         case -NFS4ERR_BADSLOT:
1104         case -NFS4ERR_BAD_HIGH_SLOT:
1105         case -NFS4ERR_DEADSESSION:
1106         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1107         case -NFS4ERR_SEQ_FALSE_RETRY:
1108         case -NFS4ERR_SEQ_MISORDERED:
1109                 dprintk("%s ERROR %d, Reset session. Exchangeid "
1110                         "flags 0x%x\n", __func__, task->tk_status,
1111                         clp->cl_exchange_flags);
1112                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1113                 break;
1114         case -NFS4ERR_DELAY:
1115         case -NFS4ERR_GRACE:
1116                 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1117                 break;
1118         case -NFS4ERR_RETRY_UNCACHED_REP:
1119                 break;
1120         /* Invalidate Layout errors */
1121         case -NFS4ERR_PNFS_NO_LAYOUT:
1122         case -ESTALE:           /* mapped NFS4ERR_STALE */
1123         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
1124         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
1125         case -NFS4ERR_FHEXPIRED:
1126         case -NFS4ERR_WRONG_TYPE:
1127                 dprintk("%s Invalid layout error %d\n", __func__,
1128                         task->tk_status);
1129                 /*
1130                  * Destroy layout so new i/o will get a new layout.
1131                  * Layout will not be destroyed until all current lseg
1132                  * references are put. Mark layout as invalid to resend failed
1133                  * i/o and all i/o waiting on the slot table to the MDS until
1134                  * layout is destroyed and a new valid layout is obtained.
1135                  */
1136                 pnfs_destroy_layout(NFS_I(inode));
1137                 rpc_wake_up(&tbl->slot_tbl_waitq);
1138                 goto reset;
1139         /* RPC connection errors */
1140         case -ECONNREFUSED:
1141         case -EHOSTDOWN:
1142         case -EHOSTUNREACH:
1143         case -ENETUNREACH:
1144         case -EIO:
1145         case -ETIMEDOUT:
1146         case -EPIPE:
1147                 dprintk("%s DS connection error %d\n", __func__,
1148                         task->tk_status);
1149                 nfs4_mark_deviceid_unavailable(devid);
1150                 rpc_wake_up(&tbl->slot_tbl_waitq);
1151                 /* fall through */
1152         default:
1153                 if (ff_layout_avoid_mds_available_ds(lseg))
1154                         return -NFS4ERR_RESET_TO_PNFS;
1155 reset:
1156                 dprintk("%s Retry through MDS. Error %d\n", __func__,
1157                         task->tk_status);
1158                 return -NFS4ERR_RESET_TO_MDS;
1159         }
1160 out:
1161         task->tk_status = 0;
1162         return -EAGAIN;
1163 out_bad_stateid:
1164         task->tk_status = -EIO;
1165         return 0;
1166 wait_on_recovery:
1167         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
1168         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
1169                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
1170         goto out;
1171 }
1172
1173 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
1174 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1175                                            struct pnfs_layout_segment *lseg,
1176                                            int idx)
1177 {
1178         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
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         if (task->tk_status >= 0)
1215                 return 0;
1216
1217         /* Handle the case of an invalid layout segment */
1218         if (!pnfs_is_valid_lseg(lseg))
1219                 return -NFS4ERR_RESET_TO_PNFS;
1220
1221         switch (vers) {
1222         case 3:
1223                 return ff_layout_async_handle_error_v3(task, lseg, idx);
1224         case 4:
1225                 return ff_layout_async_handle_error_v4(task, state, clp,
1226                                                        lseg, idx);
1227         default:
1228                 /* should never happen */
1229                 WARN_ON_ONCE(1);
1230                 return 0;
1231         }
1232 }
1233
1234 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1235                                         int idx, u64 offset, u64 length,
1236                                         u32 status, int opnum, int error)
1237 {
1238         struct nfs4_ff_layout_mirror *mirror;
1239         int err;
1240
1241         if (status == 0) {
1242                 switch (error) {
1243                 case -ETIMEDOUT:
1244                 case -EPFNOSUPPORT:
1245                 case -EPROTONOSUPPORT:
1246                 case -EOPNOTSUPP:
1247                 case -ECONNREFUSED:
1248                 case -ECONNRESET:
1249                 case -EHOSTDOWN:
1250                 case -EHOSTUNREACH:
1251                 case -ENETUNREACH:
1252                 case -EADDRINUSE:
1253                 case -ENOBUFS:
1254                 case -EPIPE:
1255                 case -EPERM:
1256                         status = NFS4ERR_NXIO;
1257                         break;
1258                 case -EACCES:
1259                         status = NFS4ERR_ACCESS;
1260                         break;
1261                 default:
1262                         return;
1263                 }
1264         }
1265
1266         switch (status) {
1267         case NFS4ERR_DELAY:
1268         case NFS4ERR_GRACE:
1269                 return;
1270         default:
1271                 break;
1272         }
1273
1274         mirror = FF_LAYOUT_COMP(lseg, idx);
1275         err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1276                                        mirror, offset, length, status, opnum,
1277                                        GFP_NOIO);
1278         pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
1279         dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1280 }
1281
1282 /* NFS_PROTO call done callback routines */
1283 static int ff_layout_read_done_cb(struct rpc_task *task,
1284                                 struct nfs_pgio_header *hdr)
1285 {
1286         int err;
1287
1288         trace_nfs4_pnfs_read(hdr, task->tk_status);
1289         if (task->tk_status < 0)
1290                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1291                                             hdr->args.offset, hdr->args.count,
1292                                             hdr->res.op_status, OP_READ,
1293                                             task->tk_status);
1294         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1295                                            hdr->ds_clp, hdr->lseg,
1296                                            hdr->pgio_mirror_idx);
1297
1298         switch (err) {
1299         case -NFS4ERR_RESET_TO_PNFS:
1300                 if (ff_layout_choose_best_ds_for_read(hdr->lseg,
1301                                         hdr->pgio_mirror_idx + 1,
1302                                         &hdr->pgio_mirror_idx))
1303                         goto out_eagain;
1304                 ff_layout_read_record_layoutstats_done(task, hdr);
1305                 pnfs_read_resend_pnfs(hdr);
1306                 return task->tk_status;
1307         case -NFS4ERR_RESET_TO_MDS:
1308                 ff_layout_reset_read(hdr);
1309                 return task->tk_status;
1310         case -EAGAIN:
1311                 goto out_eagain;
1312         }
1313
1314         return 0;
1315 out_eagain:
1316         rpc_restart_call_prepare(task);
1317         return -EAGAIN;
1318 }
1319
1320 static bool
1321 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1322 {
1323         return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1324 }
1325
1326 /*
1327  * We reference the rpc_cred of the first WRITE that triggers the need for
1328  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1329  * rfc5661 is not clear about which credential should be used.
1330  *
1331  * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1332  * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1333  * we always send layoutcommit after DS writes.
1334  */
1335 static void
1336 ff_layout_set_layoutcommit(struct inode *inode,
1337                 struct pnfs_layout_segment *lseg,
1338                 loff_t end_offset)
1339 {
1340         if (!ff_layout_need_layoutcommit(lseg))
1341                 return;
1342
1343         pnfs_set_layoutcommit(inode, lseg, end_offset);
1344         dprintk("%s inode %lu pls_end_pos %llu\n", __func__, inode->i_ino,
1345                 (unsigned long long) NFS_I(inode)->layout->plh_lwb);
1346 }
1347
1348 static bool
1349 ff_layout_device_unavailable(struct pnfs_layout_segment *lseg, int idx)
1350 {
1351         /* No mirroring for now */
1352         struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1353
1354         return ff_layout_test_devid_unavailable(node);
1355 }
1356
1357 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task,
1358                 struct nfs_pgio_header *hdr)
1359 {
1360         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1361                 return;
1362         nfs4_ff_layout_stat_io_start_read(hdr->inode,
1363                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1364                         hdr->args.count,
1365                         task->tk_start);
1366 }
1367
1368 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
1369                 struct nfs_pgio_header *hdr)
1370 {
1371         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1372                 return;
1373         nfs4_ff_layout_stat_io_end_read(task,
1374                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1375                         hdr->args.count,
1376                         hdr->res.count);
1377 }
1378
1379 static int ff_layout_read_prepare_common(struct rpc_task *task,
1380                                          struct nfs_pgio_header *hdr)
1381 {
1382         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1383                 rpc_exit(task, -EIO);
1384                 return -EIO;
1385         }
1386         if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) {
1387                 rpc_exit(task, -EHOSTDOWN);
1388                 return -EAGAIN;
1389         }
1390
1391         ff_layout_read_record_layoutstats_start(task, hdr);
1392         return 0;
1393 }
1394
1395 /*
1396  * Call ops for the async read/write cases
1397  * In the case of dense layouts, the offset needs to be reset to its
1398  * original value.
1399  */
1400 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1401 {
1402         struct nfs_pgio_header *hdr = data;
1403
1404         if (ff_layout_read_prepare_common(task, hdr))
1405                 return;
1406
1407         rpc_call_start(task);
1408 }
1409
1410 static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1411                                     struct nfs4_sequence_args *args,
1412                                     struct nfs4_sequence_res *res,
1413                                     struct rpc_task *task)
1414 {
1415         if (ds_clp->cl_session)
1416                 return nfs41_setup_sequence(ds_clp->cl_session,
1417                                            args,
1418                                            res,
1419                                            task);
1420         return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1421                                    args,
1422                                    res,
1423                                    task);
1424 }
1425
1426 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1427 {
1428         struct nfs_pgio_header *hdr = data;
1429
1430         if (ff_layout_setup_sequence(hdr->ds_clp,
1431                                      &hdr->args.seq_args,
1432                                      &hdr->res.seq_res,
1433                                      task))
1434                 return;
1435
1436         if (ff_layout_read_prepare_common(task, hdr))
1437                 return;
1438
1439         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1440                         hdr->args.lock_context, FMODE_READ) == -EIO)
1441                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1442 }
1443
1444 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1445 {
1446         struct nfs_pgio_header *hdr = data;
1447
1448         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1449
1450         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1451             task->tk_status == 0) {
1452                 nfs4_sequence_done(task, &hdr->res.seq_res);
1453                 return;
1454         }
1455
1456         /* Note this may cause RPC to be resent */
1457         hdr->mds_ops->rpc_call_done(task, hdr);
1458 }
1459
1460 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1461 {
1462         struct nfs_pgio_header *hdr = data;
1463
1464         ff_layout_read_record_layoutstats_done(task, hdr);
1465         rpc_count_iostats_metrics(task,
1466             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1467 }
1468
1469 static void ff_layout_read_release(void *data)
1470 {
1471         struct nfs_pgio_header *hdr = data;
1472
1473         ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
1474         pnfs_generic_rw_release(data);
1475 }
1476
1477
1478 static int ff_layout_write_done_cb(struct rpc_task *task,
1479                                 struct nfs_pgio_header *hdr)
1480 {
1481         loff_t end_offs = 0;
1482         int err;
1483
1484         trace_nfs4_pnfs_write(hdr, task->tk_status);
1485         if (task->tk_status < 0)
1486                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1487                                             hdr->args.offset, hdr->args.count,
1488                                             hdr->res.op_status, OP_WRITE,
1489                                             task->tk_status);
1490         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1491                                            hdr->ds_clp, hdr->lseg,
1492                                            hdr->pgio_mirror_idx);
1493
1494         switch (err) {
1495         case -NFS4ERR_RESET_TO_PNFS:
1496                 ff_layout_reset_write(hdr, true);
1497                 return task->tk_status;
1498         case -NFS4ERR_RESET_TO_MDS:
1499                 ff_layout_reset_write(hdr, false);
1500                 return task->tk_status;
1501         case -EAGAIN:
1502                 return -EAGAIN;
1503         }
1504
1505         if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1506             hdr->res.verf->committed == NFS_DATA_SYNC)
1507                 end_offs = hdr->mds_offset + (loff_t)hdr->res.count;
1508
1509         /* Note: if the write is unstable, don't set end_offs until commit */
1510         ff_layout_set_layoutcommit(hdr->inode, hdr->lseg, end_offs);
1511
1512         /* zero out fattr since we don't care DS attr at all */
1513         hdr->fattr.valid = 0;
1514         if (task->tk_status >= 0)
1515                 nfs_writeback_update_inode(hdr);
1516
1517         return 0;
1518 }
1519
1520 static int ff_layout_commit_done_cb(struct rpc_task *task,
1521                                      struct nfs_commit_data *data)
1522 {
1523         int err;
1524
1525         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1526         if (task->tk_status < 0)
1527                 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1528                                             data->args.offset, data->args.count,
1529                                             data->res.op_status, OP_COMMIT,
1530                                             task->tk_status);
1531         err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1532                                            data->lseg, data->ds_commit_index);
1533
1534         switch (err) {
1535         case -NFS4ERR_RESET_TO_PNFS:
1536                 pnfs_generic_prepare_to_resend_writes(data);
1537                 return -EAGAIN;
1538         case -NFS4ERR_RESET_TO_MDS:
1539                 pnfs_generic_prepare_to_resend_writes(data);
1540                 return -EAGAIN;
1541         case -EAGAIN:
1542                 rpc_restart_call_prepare(task);
1543                 return -EAGAIN;
1544         }
1545
1546         ff_layout_set_layoutcommit(data->inode, data->lseg, data->lwb);
1547
1548         return 0;
1549 }
1550
1551 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task,
1552                 struct nfs_pgio_header *hdr)
1553 {
1554         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1555                 return;
1556         nfs4_ff_layout_stat_io_start_write(hdr->inode,
1557                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1558                         hdr->args.count,
1559                         task->tk_start);
1560 }
1561
1562 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task,
1563                 struct nfs_pgio_header *hdr)
1564 {
1565         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1566                 return;
1567         nfs4_ff_layout_stat_io_end_write(task,
1568                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1569                         hdr->args.count, hdr->res.count,
1570                         hdr->res.verf->committed);
1571 }
1572
1573 static int ff_layout_write_prepare_common(struct rpc_task *task,
1574                                           struct nfs_pgio_header *hdr)
1575 {
1576         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1577                 rpc_exit(task, -EIO);
1578                 return -EIO;
1579         }
1580
1581         if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) {
1582                 rpc_exit(task, -EHOSTDOWN);
1583                 return -EAGAIN;
1584         }
1585
1586         ff_layout_write_record_layoutstats_start(task, hdr);
1587         return 0;
1588 }
1589
1590 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1591 {
1592         struct nfs_pgio_header *hdr = data;
1593
1594         if (ff_layout_write_prepare_common(task, hdr))
1595                 return;
1596
1597         rpc_call_start(task);
1598 }
1599
1600 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1601 {
1602         struct nfs_pgio_header *hdr = data;
1603
1604         if (ff_layout_setup_sequence(hdr->ds_clp,
1605                                      &hdr->args.seq_args,
1606                                      &hdr->res.seq_res,
1607                                      task))
1608                 return;
1609
1610         if (ff_layout_write_prepare_common(task, hdr))
1611                 return;
1612
1613         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1614                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
1615                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1616 }
1617
1618 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1619 {
1620         struct nfs_pgio_header *hdr = data;
1621
1622         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1623             task->tk_status == 0) {
1624                 nfs4_sequence_done(task, &hdr->res.seq_res);
1625                 return;
1626         }
1627
1628         /* Note this may cause RPC to be resent */
1629         hdr->mds_ops->rpc_call_done(task, hdr);
1630 }
1631
1632 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1633 {
1634         struct nfs_pgio_header *hdr = data;
1635
1636         ff_layout_write_record_layoutstats_done(task, hdr);
1637         rpc_count_iostats_metrics(task,
1638             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1639 }
1640
1641 static void ff_layout_write_release(void *data)
1642 {
1643         struct nfs_pgio_header *hdr = data;
1644
1645         ff_layout_write_record_layoutstats_done(&hdr->task, hdr);
1646         pnfs_generic_rw_release(data);
1647 }
1648
1649 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task,
1650                 struct nfs_commit_data *cdata)
1651 {
1652         if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags))
1653                 return;
1654         nfs4_ff_layout_stat_io_start_write(cdata->inode,
1655                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1656                         0, task->tk_start);
1657 }
1658
1659 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task,
1660                 struct nfs_commit_data *cdata)
1661 {
1662         struct nfs_page *req;
1663         __u64 count = 0;
1664
1665         if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags))
1666                 return;
1667
1668         if (task->tk_status == 0) {
1669                 list_for_each_entry(req, &cdata->pages, wb_list)
1670                         count += req->wb_bytes;
1671         }
1672         nfs4_ff_layout_stat_io_end_write(task,
1673                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1674                         count, count, NFS_FILE_SYNC);
1675 }
1676
1677 static void ff_layout_commit_prepare_common(struct rpc_task *task,
1678                 struct nfs_commit_data *cdata)
1679 {
1680         ff_layout_commit_record_layoutstats_start(task, cdata);
1681 }
1682
1683 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1684 {
1685         ff_layout_commit_prepare_common(task, data);
1686         rpc_call_start(task);
1687 }
1688
1689 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1690 {
1691         struct nfs_commit_data *wdata = data;
1692
1693         if (ff_layout_setup_sequence(wdata->ds_clp,
1694                                  &wdata->args.seq_args,
1695                                  &wdata->res.seq_res,
1696                                  task))
1697                 return;
1698         ff_layout_commit_prepare_common(task, data);
1699 }
1700
1701 static void ff_layout_commit_done(struct rpc_task *task, void *data)
1702 {
1703         pnfs_generic_write_commit_done(task, data);
1704 }
1705
1706 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1707 {
1708         struct nfs_commit_data *cdata = data;
1709
1710         ff_layout_commit_record_layoutstats_done(task, cdata);
1711         rpc_count_iostats_metrics(task,
1712             &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1713 }
1714
1715 static void ff_layout_commit_release(void *data)
1716 {
1717         struct nfs_commit_data *cdata = data;
1718
1719         ff_layout_commit_record_layoutstats_done(&cdata->task, cdata);
1720         pnfs_generic_commit_release(data);
1721 }
1722
1723 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1724         .rpc_call_prepare = ff_layout_read_prepare_v3,
1725         .rpc_call_done = ff_layout_read_call_done,
1726         .rpc_count_stats = ff_layout_read_count_stats,
1727         .rpc_release = ff_layout_read_release,
1728 };
1729
1730 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1731         .rpc_call_prepare = ff_layout_read_prepare_v4,
1732         .rpc_call_done = ff_layout_read_call_done,
1733         .rpc_count_stats = ff_layout_read_count_stats,
1734         .rpc_release = ff_layout_read_release,
1735 };
1736
1737 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1738         .rpc_call_prepare = ff_layout_write_prepare_v3,
1739         .rpc_call_done = ff_layout_write_call_done,
1740         .rpc_count_stats = ff_layout_write_count_stats,
1741         .rpc_release = ff_layout_write_release,
1742 };
1743
1744 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1745         .rpc_call_prepare = ff_layout_write_prepare_v4,
1746         .rpc_call_done = ff_layout_write_call_done,
1747         .rpc_count_stats = ff_layout_write_count_stats,
1748         .rpc_release = ff_layout_write_release,
1749 };
1750
1751 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1752         .rpc_call_prepare = ff_layout_commit_prepare_v3,
1753         .rpc_call_done = ff_layout_commit_done,
1754         .rpc_count_stats = ff_layout_commit_count_stats,
1755         .rpc_release = ff_layout_commit_release,
1756 };
1757
1758 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1759         .rpc_call_prepare = ff_layout_commit_prepare_v4,
1760         .rpc_call_done = ff_layout_commit_done,
1761         .rpc_count_stats = ff_layout_commit_count_stats,
1762         .rpc_release = ff_layout_commit_release,
1763 };
1764
1765 static enum pnfs_try_status
1766 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1767 {
1768         struct pnfs_layout_segment *lseg = hdr->lseg;
1769         struct nfs4_pnfs_ds *ds;
1770         struct rpc_clnt *ds_clnt;
1771         struct rpc_cred *ds_cred;
1772         loff_t offset = hdr->args.offset;
1773         u32 idx = hdr->pgio_mirror_idx;
1774         int vers;
1775         struct nfs_fh *fh;
1776
1777         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1778                 __func__, hdr->inode->i_ino,
1779                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1780
1781         ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1782         if (!ds)
1783                 goto out_failed;
1784
1785         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1786                                                    hdr->inode);
1787         if (IS_ERR(ds_clnt))
1788                 goto out_failed;
1789
1790         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1791         if (!ds_cred)
1792                 goto out_failed;
1793
1794         vers = nfs4_ff_layout_ds_version(lseg, idx);
1795
1796         dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1797                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1798
1799         hdr->pgio_done_cb = ff_layout_read_done_cb;
1800         atomic_inc(&ds->ds_clp->cl_count);
1801         hdr->ds_clp = ds->ds_clp;
1802         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1803         if (fh)
1804                 hdr->args.fh = fh;
1805         /*
1806          * Note that if we ever decide to split across DSes,
1807          * then we may need to handle dense-like offsets.
1808          */
1809         hdr->args.offset = offset;
1810         hdr->mds_offset = offset;
1811
1812         /* Perform an asynchronous read to ds */
1813         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1814                           vers == 3 ? &ff_layout_read_call_ops_v3 :
1815                                       &ff_layout_read_call_ops_v4,
1816                           0, RPC_TASK_SOFTCONN);
1817         put_rpccred(ds_cred);
1818         return PNFS_ATTEMPTED;
1819
1820 out_failed:
1821         if (ff_layout_avoid_mds_available_ds(lseg))
1822                 return PNFS_TRY_AGAIN;
1823         return PNFS_NOT_ATTEMPTED;
1824 }
1825
1826 /* Perform async writes. */
1827 static enum pnfs_try_status
1828 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1829 {
1830         struct pnfs_layout_segment *lseg = hdr->lseg;
1831         struct nfs4_pnfs_ds *ds;
1832         struct rpc_clnt *ds_clnt;
1833         struct rpc_cred *ds_cred;
1834         loff_t offset = hdr->args.offset;
1835         int vers;
1836         struct nfs_fh *fh;
1837         int idx = hdr->pgio_mirror_idx;
1838
1839         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1840         if (!ds)
1841                 return PNFS_NOT_ATTEMPTED;
1842
1843         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1844                                                    hdr->inode);
1845         if (IS_ERR(ds_clnt))
1846                 return PNFS_NOT_ATTEMPTED;
1847
1848         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1849         if (!ds_cred)
1850                 return PNFS_NOT_ATTEMPTED;
1851
1852         vers = nfs4_ff_layout_ds_version(lseg, idx);
1853
1854         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1855                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1856                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1857                 vers);
1858
1859         hdr->pgio_done_cb = ff_layout_write_done_cb;
1860         atomic_inc(&ds->ds_clp->cl_count);
1861         hdr->ds_clp = ds->ds_clp;
1862         hdr->ds_commit_idx = idx;
1863         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1864         if (fh)
1865                 hdr->args.fh = fh;
1866
1867         /*
1868          * Note that if we ever decide to split across DSes,
1869          * then we may need to handle dense-like offsets.
1870          */
1871         hdr->args.offset = offset;
1872
1873         /* Perform an asynchronous write */
1874         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1875                           vers == 3 ? &ff_layout_write_call_ops_v3 :
1876                                       &ff_layout_write_call_ops_v4,
1877                           sync, RPC_TASK_SOFTCONN);
1878         put_rpccred(ds_cred);
1879         return PNFS_ATTEMPTED;
1880 }
1881
1882 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1883 {
1884         return i;
1885 }
1886
1887 static struct nfs_fh *
1888 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1889 {
1890         struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1891
1892         /* FIXME: Assume that there is only one NFS version available
1893          * for the DS.
1894          */
1895         return &flseg->mirror_array[i]->fh_versions[0];
1896 }
1897
1898 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1899 {
1900         struct pnfs_layout_segment *lseg = data->lseg;
1901         struct nfs4_pnfs_ds *ds;
1902         struct rpc_clnt *ds_clnt;
1903         struct rpc_cred *ds_cred;
1904         u32 idx;
1905         int vers, ret;
1906         struct nfs_fh *fh;
1907
1908         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1909         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1910         if (!ds)
1911                 goto out_err;
1912
1913         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1914                                                    data->inode);
1915         if (IS_ERR(ds_clnt))
1916                 goto out_err;
1917
1918         ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1919         if (!ds_cred)
1920                 goto out_err;
1921
1922         vers = nfs4_ff_layout_ds_version(lseg, idx);
1923
1924         dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1925                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1926                 vers);
1927         data->commit_done_cb = ff_layout_commit_done_cb;
1928         data->cred = ds_cred;
1929         atomic_inc(&ds->ds_clp->cl_count);
1930         data->ds_clp = ds->ds_clp;
1931         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1932         if (fh)
1933                 data->args.fh = fh;
1934
1935         ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1936                                    vers == 3 ? &ff_layout_commit_call_ops_v3 :
1937                                                &ff_layout_commit_call_ops_v4,
1938                                    how, RPC_TASK_SOFTCONN);
1939         put_rpccred(ds_cred);
1940         return ret;
1941 out_err:
1942         pnfs_generic_prepare_to_resend_writes(data);
1943         pnfs_generic_commit_release(data);
1944         return -EAGAIN;
1945 }
1946
1947 static int
1948 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1949                            int how, struct nfs_commit_info *cinfo)
1950 {
1951         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1952                                             ff_layout_initiate_commit);
1953 }
1954
1955 static struct pnfs_ds_commit_info *
1956 ff_layout_get_ds_info(struct inode *inode)
1957 {
1958         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1959
1960         if (layout == NULL)
1961                 return NULL;
1962
1963         return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1964 }
1965
1966 static void
1967 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
1968 {
1969         nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1970                                                   id_node));
1971 }
1972
1973 static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1974                                   struct xdr_stream *xdr,
1975                                   const struct nfs4_layoutreturn_args *args)
1976 {
1977         struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1978         __be32 *start;
1979         int count = 0, ret = 0;
1980
1981         start = xdr_reserve_space(xdr, 4);
1982         if (unlikely(!start))
1983                 return -E2BIG;
1984
1985         /* This assume we always return _ALL_ layouts */
1986         spin_lock(&hdr->plh_inode->i_lock);
1987         ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1988         spin_unlock(&hdr->plh_inode->i_lock);
1989
1990         *start = cpu_to_be32(count);
1991
1992         return ret;
1993 }
1994
1995 /* report nothing for now */
1996 static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1997                                      struct xdr_stream *xdr,
1998                                      const struct nfs4_layoutreturn_args *args)
1999 {
2000         __be32 *p;
2001
2002         p = xdr_reserve_space(xdr, 4);
2003         if (likely(p))
2004                 *p = cpu_to_be32(0);
2005 }
2006
2007 static struct nfs4_deviceid_node *
2008 ff_layout_alloc_deviceid_node(struct nfs_server *server,
2009                               struct pnfs_device *pdev, gfp_t gfp_flags)
2010 {
2011         struct nfs4_ff_layout_ds *dsaddr;
2012
2013         dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
2014         if (!dsaddr)
2015                 return NULL;
2016         return &dsaddr->id_node;
2017 }
2018
2019 static void
2020 ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
2021                               struct xdr_stream *xdr,
2022                               const struct nfs4_layoutreturn_args *args)
2023 {
2024         struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
2025         __be32 *start;
2026
2027         dprintk("%s: Begin\n", __func__);
2028         start = xdr_reserve_space(xdr, 4);
2029         BUG_ON(!start);
2030
2031         ff_layout_encode_ioerr(flo, xdr, args);
2032         ff_layout_encode_iostats(flo, xdr, args);
2033
2034         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2035         dprintk("%s: Return\n", __func__);
2036 }
2037
2038 static int
2039 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
2040 {
2041         const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
2042
2043         return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
2044 }
2045
2046 static size_t
2047 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
2048                           const int buflen)
2049 {
2050         const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
2051         const struct in6_addr *addr = &sin6->sin6_addr;
2052
2053         /*
2054          * RFC 4291, Section 2.2.2
2055          *
2056          * Shorthanded ANY address
2057          */
2058         if (ipv6_addr_any(addr))
2059                 return snprintf(buf, buflen, "::");
2060
2061         /*
2062          * RFC 4291, Section 2.2.2
2063          *
2064          * Shorthanded loopback address
2065          */
2066         if (ipv6_addr_loopback(addr))
2067                 return snprintf(buf, buflen, "::1");
2068
2069         /*
2070          * RFC 4291, Section 2.2.3
2071          *
2072          * Special presentation address format for mapped v4
2073          * addresses.
2074          */
2075         if (ipv6_addr_v4mapped(addr))
2076                 return snprintf(buf, buflen, "::ffff:%pI4",
2077                                         &addr->s6_addr32[3]);
2078
2079         /*
2080          * RFC 4291, Section 2.2.1
2081          */
2082         return snprintf(buf, buflen, "%pI6c", addr);
2083 }
2084
2085 /* Derived from rpc_sockaddr2uaddr */
2086 static void
2087 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
2088 {
2089         struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
2090         char portbuf[RPCBIND_MAXUADDRPLEN];
2091         char addrbuf[RPCBIND_MAXUADDRLEN];
2092         char *netid;
2093         unsigned short port;
2094         int len, netid_len;
2095         __be32 *p;
2096
2097         switch (sap->sa_family) {
2098         case AF_INET:
2099                 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
2100                         return;
2101                 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
2102                 netid = "tcp";
2103                 netid_len = 3;
2104                 break;
2105         case AF_INET6:
2106                 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
2107                         return;
2108                 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
2109                 netid = "tcp6";
2110                 netid_len = 4;
2111                 break;
2112         default:
2113                 /* we only support tcp and tcp6 */
2114                 WARN_ON_ONCE(1);
2115                 return;
2116         }
2117
2118         snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
2119         len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
2120
2121         p = xdr_reserve_space(xdr, 4 + netid_len);
2122         xdr_encode_opaque(p, netid, netid_len);
2123
2124         p = xdr_reserve_space(xdr, 4 + len);
2125         xdr_encode_opaque(p, addrbuf, len);
2126 }
2127
2128 static void
2129 ff_layout_encode_nfstime(struct xdr_stream *xdr,
2130                          ktime_t t)
2131 {
2132         struct timespec64 ts;
2133         __be32 *p;
2134
2135         p = xdr_reserve_space(xdr, 12);
2136         ts = ktime_to_timespec64(t);
2137         p = xdr_encode_hyper(p, ts.tv_sec);
2138         *p++ = cpu_to_be32(ts.tv_nsec);
2139 }
2140
2141 static void
2142 ff_layout_encode_io_latency(struct xdr_stream *xdr,
2143                             struct nfs4_ff_io_stat *stat)
2144 {
2145         __be32 *p;
2146
2147         p = xdr_reserve_space(xdr, 5 * 8);
2148         p = xdr_encode_hyper(p, stat->ops_requested);
2149         p = xdr_encode_hyper(p, stat->bytes_requested);
2150         p = xdr_encode_hyper(p, stat->ops_completed);
2151         p = xdr_encode_hyper(p, stat->bytes_completed);
2152         p = xdr_encode_hyper(p, stat->bytes_not_delivered);
2153         ff_layout_encode_nfstime(xdr, stat->total_busy_time);
2154         ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
2155 }
2156
2157 static void
2158 ff_layout_encode_layoutstats(struct xdr_stream *xdr,
2159                              struct nfs42_layoutstat_args *args,
2160                              struct nfs42_layoutstat_devinfo *devinfo)
2161 {
2162         struct nfs4_ff_layout_mirror *mirror = devinfo->layout_private;
2163         struct nfs4_pnfs_ds_addr *da;
2164         struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds;
2165         struct nfs_fh *fh = &mirror->fh_versions[0];
2166         __be32 *p, *start;
2167
2168         da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2169         dprintk("%s: DS %s: encoding address %s\n",
2170                 __func__, ds->ds_remotestr, da->da_remotestr);
2171         /* layoutupdate length */
2172         start = xdr_reserve_space(xdr, 4);
2173         /* netaddr4 */
2174         ff_layout_encode_netaddr(xdr, da);
2175         /* nfs_fh4 */
2176         p = xdr_reserve_space(xdr, 4 + fh->size);
2177         xdr_encode_opaque(p, fh->data, fh->size);
2178         /* ff_io_latency4 read */
2179         spin_lock(&mirror->lock);
2180         ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat);
2181         /* ff_io_latency4 write */
2182         ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat);
2183         spin_unlock(&mirror->lock);
2184         /* nfstime4 */
2185         ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time));
2186         /* bool */
2187         p = xdr_reserve_space(xdr, 4);
2188         *p = cpu_to_be32(false);
2189
2190         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2191 }
2192
2193 static int
2194 ff_layout_mirror_prepare_stats(struct nfs42_layoutstat_args *args,
2195                                struct pnfs_layout_hdr *lo,
2196                                int dev_limit)
2197 {
2198         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2199         struct nfs4_ff_layout_mirror *mirror;
2200         struct nfs4_deviceid_node *dev;
2201         struct nfs42_layoutstat_devinfo *devinfo;
2202         int i = 0;
2203
2204         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2205                 if (i >= dev_limit)
2206                         break;
2207                 if (!mirror->mirror_ds)
2208                         continue;
2209                 /* mirror refcount put in cleanup_layoutstats */
2210                 if (!atomic_inc_not_zero(&mirror->ref))
2211                         continue;
2212                 dev = &mirror->mirror_ds->id_node;
2213                 devinfo = &args->devinfo[i];
2214                 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE);
2215                 devinfo->offset = 0;
2216                 devinfo->length = NFS4_MAX_UINT64;
2217                 devinfo->read_count = mirror->read_stat.io_stat.ops_completed;
2218                 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed;
2219                 devinfo->write_count = mirror->write_stat.io_stat.ops_completed;
2220                 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed;
2221                 devinfo->layout_type = LAYOUT_FLEX_FILES;
2222                 devinfo->layoutstats_encode = ff_layout_encode_layoutstats;
2223                 devinfo->layout_private = mirror;
2224
2225                 i++;
2226         }
2227         return i;
2228 }
2229
2230 static int
2231 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2232 {
2233         struct nfs4_flexfile_layout *ff_layout;
2234         struct nfs4_ff_layout_mirror *mirror;
2235         int dev_count = 0;
2236
2237         spin_lock(&args->inode->i_lock);
2238         ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout);
2239         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2240                 if (atomic_read(&mirror->ref) != 0)
2241                         dev_count ++;
2242         }
2243         spin_unlock(&args->inode->i_lock);
2244         /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2245         if (dev_count > PNFS_LAYOUTSTATS_MAXDEV) {
2246                 dprintk("%s: truncating devinfo to limit (%d:%d)\n",
2247                         __func__, dev_count, PNFS_LAYOUTSTATS_MAXDEV);
2248                 dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2249         }
2250         args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO);
2251         if (!args->devinfo)
2252                 return -ENOMEM;
2253
2254         spin_lock(&args->inode->i_lock);
2255         args->num_dev = ff_layout_mirror_prepare_stats(args,
2256                         &ff_layout->generic_hdr, dev_count);
2257         spin_unlock(&args->inode->i_lock);
2258
2259         return 0;
2260 }
2261
2262 static void
2263 ff_layout_cleanup_layoutstats(struct nfs42_layoutstat_data *data)
2264 {
2265         struct nfs4_ff_layout_mirror *mirror;
2266         int i;
2267
2268         for (i = 0; i < data->args.num_dev; i++) {
2269                 mirror = data->args.devinfo[i].layout_private;
2270                 data->args.devinfo[i].layout_private = NULL;
2271                 ff_layout_put_mirror(mirror);
2272         }
2273 }
2274
2275 static struct pnfs_layoutdriver_type flexfilelayout_type = {
2276         .id                     = LAYOUT_FLEX_FILES,
2277         .name                   = "LAYOUT_FLEX_FILES",
2278         .owner                  = THIS_MODULE,
2279         .alloc_layout_hdr       = ff_layout_alloc_layout_hdr,
2280         .free_layout_hdr        = ff_layout_free_layout_hdr,
2281         .alloc_lseg             = ff_layout_alloc_lseg,
2282         .free_lseg              = ff_layout_free_lseg,
2283         .add_lseg               = ff_layout_add_lseg,
2284         .pg_read_ops            = &ff_layout_pg_read_ops,
2285         .pg_write_ops           = &ff_layout_pg_write_ops,
2286         .get_ds_info            = ff_layout_get_ds_info,
2287         .free_deviceid_node     = ff_layout_free_deviceid_node,
2288         .mark_request_commit    = pnfs_layout_mark_request_commit,
2289         .clear_request_commit   = pnfs_generic_clear_request_commit,
2290         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
2291         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
2292         .commit_pagelist        = ff_layout_commit_pagelist,
2293         .read_pagelist          = ff_layout_read_pagelist,
2294         .write_pagelist         = ff_layout_write_pagelist,
2295         .alloc_deviceid_node    = ff_layout_alloc_deviceid_node,
2296         .encode_layoutreturn    = ff_layout_encode_layoutreturn,
2297         .sync                   = pnfs_nfs_generic_sync,
2298         .prepare_layoutstats    = ff_layout_prepare_layoutstats,
2299         .cleanup_layoutstats    = ff_layout_cleanup_layoutstats,
2300 };
2301
2302 static int __init nfs4flexfilelayout_init(void)
2303 {
2304         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
2305                __func__);
2306         if (!ff_zero_group) {
2307                 ff_zero_group = groups_alloc(0);
2308                 if (!ff_zero_group)
2309                         return -ENOMEM;
2310         }
2311         return pnfs_register_layoutdriver(&flexfilelayout_type);
2312 }
2313
2314 static void __exit nfs4flexfilelayout_exit(void)
2315 {
2316         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
2317                __func__);
2318         pnfs_unregister_layoutdriver(&flexfilelayout_type);
2319         if (ff_zero_group) {
2320                 put_group_info(ff_zero_group);
2321                 ff_zero_group = NULL;
2322         }
2323 }
2324
2325 MODULE_ALIAS("nfs-layouttype4-4");
2326
2327 MODULE_LICENSE("GPL");
2328 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
2329
2330 module_init(nfs4flexfilelayout_init);
2331 module_exit(nfs4flexfilelayout_exit);