fs/nfs/flexfilelayout/flexfilelayout.c

   1 /*
   2  * Module for pnfs flexfile layout driver.
   3  *
   4  * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
   5  *
   6  * Tao Peng <bergwolf@primarydata.com>
   7  */
   8
   9 #include <linux/nfs_fs.h>
  10 #include <linux/nfs_page.h>
  11 #include <linux/module.h>
  12
  13 #include <linux/sunrpc/metrics.h>
  14
  15 #include "flexfilelayout.h"
  16 #include "../nfs4session.h"
  17 #include "../nfs4idmap.h"
  18 #include "../internal.h"
  19 #include "../delegation.h"
  20 #include "../nfs4trace.h"
  21 #include "../iostat.h"
  22 #include "../nfs.h"
  23 #include "../nfs42.h"
  24
  25 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  26
  27 #define FF_LAYOUT_POLL_RETRY_MAX     (15*HZ)
  28
  29 static struct group_info        *ff_zero_group;
  30
  31 static struct pnfs_layout_hdr *
  32 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
  33 {
  34         struct nfs4_flexfile_layout *ffl;
  35
  36         ffl = kzalloc(sizeof(*ffl), gfp_flags);
  37         if (ffl) {
  38                 INIT_LIST_HEAD(&ffl->error_list);
  39                 INIT_LIST_HEAD(&ffl->mirrors);
  40                 ffl->last_report_time = ktime_get();
  41                 return &ffl->generic_hdr;
  42         } else
  43                 return NULL;
  44 }
  45
  46 static void
  47 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo)
  48 {
  49         struct nfs4_ff_layout_ds_err *err, *n;
  50
  51         list_for_each_entry_safe(err, n, &FF_LAYOUT_FROM_HDR(lo)->error_list,
  52                                  list) {
  53                 list_del(&err->list);
  54                 kfree(err);
  55         }
  56         kfree(FF_LAYOUT_FROM_HDR(lo));
  57 }
  58
  59 static int decode_pnfs_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid)
  60 {
  61         __be32 *p;
  62
  63         p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE);
  64         if (unlikely(p == NULL))
  65                 return -ENOBUFS;
  66         stateid->type = NFS4_PNFS_DS_STATEID_TYPE;
  67         memcpy(stateid->data, p, NFS4_STATEID_SIZE);
  68         dprintk("%s: stateid id= [%x%x%x%x]\n", __func__,
  69                 p[0], p[1], p[2], p[3]);
  70         return 0;
  71 }
  72
  73 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid)
  74 {
  75         __be32 *p;
  76
  77         p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE);
  78         if (unlikely(!p))
  79                 return -ENOBUFS;
  80         memcpy(devid, p, NFS4_DEVICEID4_SIZE);
  81         nfs4_print_deviceid(devid);
  82         return 0;
  83 }
  84
  85 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh)
  86 {
  87         __be32 *p;
  88
  89         p = xdr_inline_decode(xdr, 4);
  90         if (unlikely(!p))
  91                 return -ENOBUFS;
  92         fh->size = be32_to_cpup(p++);
  93         if (fh->size > sizeof(struct nfs_fh)) {
  94                 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n",
  95                        fh->size);
  96                 return -EOVERFLOW;
  97         }
  98         /* fh.data */
  99         p = xdr_inline_decode(xdr, fh->size);
 100         if (unlikely(!p))
 101                 return -ENOBUFS;
 102         memcpy(&fh->data, p, fh->size);
 103         dprintk("%s: fh len %d\n", __func__, fh->size);
 104
 105         return 0;
 106 }
 107
 108 /*
 109  * Currently only stringified uids and gids are accepted.
 110  * I.e., kerberos is not supported to the DSes, so no pricipals.
 111  *
 112  * That means that one common function will suffice, but when
 113  * principals are added, this should be split to accomodate
 114  * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid().
 115  */
 116 static int
 117 decode_name(struct xdr_stream *xdr, u32 *id)
 118 {
 119         __be32 *p;
 120         int len;
 121
 122         /* opaque_length(4)*/
 123         p = xdr_inline_decode(xdr, 4);
 124         if (unlikely(!p))
 125                 return -ENOBUFS;
 126         len = be32_to_cpup(p++);
 127         if (len < 0)
 128                 return -EINVAL;
 129
 130         dprintk("%s: len %u\n", __func__, len);
 131
 132         /* opaque body */
 133         p = xdr_inline_decode(xdr, len);
 134         if (unlikely(!p))
 135                 return -ENOBUFS;
 136
 137         if (!nfs_map_string_to_numeric((char *)p, len, id))
 138                 return -EINVAL;
 139
 140         return 0;
 141 }
 142
 143 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1,
 144                 const struct nfs4_ff_layout_mirror *m2)
 145 {
 146         int i, j;
 147
 148         if (m1->fh_versions_cnt != m2->fh_versions_cnt)
 149                 return false;
 150         for (i = 0; i < m1->fh_versions_cnt; i++) {
 151                 bool found_fh = false;
 152                 for (j = 0; j < m2->fh_versions_cnt; j++) {
 153                         if (nfs_compare_fh(&m1->fh_versions[i],
 154                                         &m2->fh_versions[j]) == 0) {
 155                                 found_fh = true;
 156                                 break;
 157                         }
 158                 }
 159                 if (!found_fh)
 160                         return false;
 161         }
 162         return true;
 163 }
 164
 165 static struct nfs4_ff_layout_mirror *
 166 ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
 167                 struct nfs4_ff_layout_mirror *mirror)
 168 {
 169         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
 170         struct nfs4_ff_layout_mirror *pos;
 171         struct inode *inode = lo->plh_inode;
 172
 173         spin_lock(&inode->i_lock);
 174         list_for_each_entry(pos, &ff_layout->mirrors, mirrors) {
 175                 if (mirror->mirror_ds != pos->mirror_ds)
 176                         continue;
 177                 if (!ff_mirror_match_fh(mirror, pos))
 178                         continue;
 179                 if (atomic_inc_not_zero(&pos->ref)) {
 180                         spin_unlock(&inode->i_lock);
 181                         return pos;
 182                 }
 183         }
 184         list_add(&mirror->mirrors, &ff_layout->mirrors);
 185         mirror->layout = lo;
 186         spin_unlock(&inode->i_lock);
 187         return mirror;
 188 }
 189
 190 static void
 191 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror)
 192 {
 193         struct inode *inode;
 194         if (mirror->layout == NULL)
 195                 return;
 196         inode = mirror->layout->plh_inode;
 197         spin_lock(&inode->i_lock);
 198         list_del(&mirror->mirrors);
 199         spin_unlock(&inode->i_lock);
 200         mirror->layout = NULL;
 201 }
 202
 203 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(gfp_t gfp_flags)
 204 {
 205         struct nfs4_ff_layout_mirror *mirror;
 206
 207         mirror = kzalloc(sizeof(*mirror), gfp_flags);
 208         if (mirror != NULL) {
 209                 spin_lock_init(&mirror->lock);
 210                 atomic_set(&mirror->ref, 1);
 211                 INIT_LIST_HEAD(&mirror->mirrors);
 212         }
 213         return mirror;
 214 }
 215
 216 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror)
 217 {
 218         struct rpc_cred *cred;
 219
 220         ff_layout_remove_mirror(mirror);
 221         kfree(mirror->fh_versions);
 222         cred = rcu_access_pointer(mirror->ro_cred);
 223         if (cred)
 224                 put_rpccred(cred);
 225         cred = rcu_access_pointer(mirror->rw_cred);
 226         if (cred)
 227                 put_rpccred(cred);
 228         nfs4_ff_layout_put_deviceid(mirror->mirror_ds);
 229         kfree(mirror);
 230 }
 231
 232 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror)
 233 {
 234         if (mirror != NULL && atomic_dec_and_test(&mirror->ref))
 235                 ff_layout_free_mirror(mirror);
 236 }
 237
 238 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls)
 239 {
 240         int i;
 241
 242         if (fls->mirror_array) {
 243                 for (i = 0; i < fls->mirror_array_cnt; i++) {
 244                         /* normally mirror_ds is freed in
 245                          * .free_deviceid_node but we still do it here
 246                          * for .alloc_lseg error path */
 247                         ff_layout_put_mirror(fls->mirror_array[i]);
 248                 }
 249                 kfree(fls->mirror_array);
 250                 fls->mirror_array = NULL;
 251         }
 252 }
 253
 254 static int ff_layout_check_layout(struct nfs4_layoutget_res *lgr)
 255 {
 256         int ret = 0;
 257
 258         dprintk("--> %s\n", __func__);
 259
 260         /* FIXME: remove this check when layout segment support is added */
 261         if (lgr->range.offset != 0 ||
 262             lgr->range.length != NFS4_MAX_UINT64) {
 263                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 264                         __func__);
 265                 ret = -EINVAL;
 266         }
 267
 268         dprintk("--> %s returns %d\n", __func__, ret);
 269         return ret;
 270 }
 271
 272 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls)
 273 {
 274         if (fls) {
 275                 ff_layout_free_mirror_array(fls);
 276                 kfree(fls);
 277         }
 278 }
 279
 280 static bool
 281 ff_lseg_range_is_after(const struct pnfs_layout_range *l1,
 282                 const struct pnfs_layout_range *l2)
 283 {
 284         u64 end1, end2;
 285
 286         if (l1->iomode != l2->iomode)
 287                 return l1->iomode != IOMODE_READ;
 288         end1 = pnfs_calc_offset_end(l1->offset, l1->length);
 289         end2 = pnfs_calc_offset_end(l2->offset, l2->length);
 290         if (end1 < l2->offset)
 291                 return false;
 292         if (end2 < l1->offset)
 293                 return true;
 294         return l2->offset <= l1->offset;
 295 }
 296
 297 static bool
 298 ff_lseg_merge(struct pnfs_layout_segment *new,
 299                 struct pnfs_layout_segment *old)
 300 {
 301         u64 new_end, old_end;
 302
 303         if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags))
 304                 return false;
 305         if (new->pls_range.iomode != old->pls_range.iomode)
 306                 return false;
 307         old_end = pnfs_calc_offset_end(old->pls_range.offset,
 308                         old->pls_range.length);
 309         if (old_end < new->pls_range.offset)
 310                 return false;
 311         new_end = pnfs_calc_offset_end(new->pls_range.offset,
 312                         new->pls_range.length);
 313         if (new_end < old->pls_range.offset)
 314                 return false;
 315
 316         /* Mergeable: copy info from 'old' to 'new' */
 317         if (new_end < old_end)
 318                 new_end = old_end;
 319         if (new->pls_range.offset < old->pls_range.offset)
 320                 new->pls_range.offset = old->pls_range.offset;
 321         new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset,
 322                         new_end);
 323         if (test_bit(NFS_LSEG_ROC, &old->pls_flags))
 324                 set_bit(NFS_LSEG_ROC, &new->pls_flags);
 325         return true;
 326 }
 327
 328 static void
 329 ff_layout_add_lseg(struct pnfs_layout_hdr *lo,
 330                 struct pnfs_layout_segment *lseg,
 331                 struct list_head *free_me)
 332 {
 333         pnfs_generic_layout_insert_lseg(lo, lseg,
 334                         ff_lseg_range_is_after,
 335                         ff_lseg_merge,
 336                         free_me);
 337 }
 338
 339 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls)
 340 {
 341         int i, j;
 342
 343         for (i = 0; i < fls->mirror_array_cnt - 1; i++) {
 344                 for (j = i + 1; j < fls->mirror_array_cnt; j++)
 345                         if (fls->mirror_array[i]->efficiency <
 346                             fls->mirror_array[j]->efficiency)
 347                                 swap(fls->mirror_array[i],
 348                                      fls->mirror_array[j]);
 349         }
 350 }
 351
 352 static void ff_layout_mark_devices_valid(struct nfs4_ff_layout_segment *fls)
 353 {
 354         struct nfs4_deviceid_node *node;
 355         int i;
 356
 357         if (!(fls->flags & FF_FLAGS_NO_IO_THRU_MDS))
 358                 return;
 359         for (i = 0; i < fls->mirror_array_cnt; i++) {
 360                 node = &fls->mirror_array[i]->mirror_ds->id_node;
 361                 clear_bit(NFS_DEVICEID_UNAVAILABLE, &node->flags);
 362         }
 363 }
 364
 365 static struct pnfs_layout_segment *
 366 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh,
 367                      struct nfs4_layoutget_res *lgr,
 368                      gfp_t gfp_flags)
 369 {
 370         struct pnfs_layout_segment *ret;
 371         struct nfs4_ff_layout_segment *fls = NULL;
 372         struct xdr_stream stream;
 373         struct xdr_buf buf;
 374         struct page *scratch;
 375         u64 stripe_unit;
 376         u32 mirror_array_cnt;
 377         __be32 *p;
 378         int i, rc;
 379
 380         dprintk("--> %s\n", __func__);
 381         scratch = alloc_page(gfp_flags);
 382         if (!scratch)
 383                 return ERR_PTR(-ENOMEM);
 384
 385         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages,
 386                               lgr->layoutp->len);
 387         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 388
 389         /* stripe unit and mirror_array_cnt */
 390         rc = -EIO;
 391         p = xdr_inline_decode(&stream, 8 + 4);
 392         if (!p)
 393                 goto out_err_free;
 394
 395         p = xdr_decode_hyper(p, &stripe_unit);
 396         mirror_array_cnt = be32_to_cpup(p++);
 397         dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__,
 398                 stripe_unit, mirror_array_cnt);
 399
 400         if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT ||
 401             mirror_array_cnt == 0)
 402                 goto out_err_free;
 403
 404         rc = -ENOMEM;
 405         fls = kzalloc(sizeof(*fls), gfp_flags);
 406         if (!fls)
 407                 goto out_err_free;
 408
 409         fls->mirror_array_cnt = mirror_array_cnt;
 410         fls->stripe_unit = stripe_unit;
 411         fls->mirror_array = kcalloc(fls->mirror_array_cnt,
 412                                     sizeof(fls->mirror_array[0]), gfp_flags);
 413         if (fls->mirror_array == NULL)
 414                 goto out_err_free;
 415
 416         for (i = 0; i < fls->mirror_array_cnt; i++) {
 417                 struct nfs4_ff_layout_mirror *mirror;
 418                 struct nfs4_deviceid devid;
 419                 struct nfs4_deviceid_node *idnode;
 420                 struct auth_cred acred = { .group_info = ff_zero_group };
 421                 struct rpc_cred __rcu *cred;
 422                 u32 ds_count, fh_count, id;
 423                 int j;
 424
 425                 rc = -EIO;
 426                 p = xdr_inline_decode(&stream, 4);
 427                 if (!p)
 428                         goto out_err_free;
 429                 ds_count = be32_to_cpup(p);
 430
 431                 /* FIXME: allow for striping? */
 432                 if (ds_count != 1)
 433                         goto out_err_free;
 434
 435                 fls->mirror_array[i] = ff_layout_alloc_mirror(gfp_flags);
 436                 if (fls->mirror_array[i] == NULL) {
 437                         rc = -ENOMEM;
 438                         goto out_err_free;
 439                 }
 440
 441                 fls->mirror_array[i]->ds_count = ds_count;
 442
 443                 /* deviceid */
 444                 rc = decode_deviceid(&stream, &devid);
 445                 if (rc)
 446                         goto out_err_free;
 447
 448                 idnode = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
 449                                                 &devid, lh->plh_lc_cred,
 450                                                 gfp_flags);
 451                 /*
 452                  * upon success, mirror_ds is allocated by previous
 453                  * getdeviceinfo, or newly by .alloc_deviceid_node
 454                  * nfs4_find_get_deviceid failure is indeed getdeviceinfo falure
 455                  */
 456                 if (idnode)
 457                         fls->mirror_array[i]->mirror_ds =
 458                                 FF_LAYOUT_MIRROR_DS(idnode);
 459                 else
 460                         goto out_err_free;
 461
 462                 /* efficiency */
 463                 rc = -EIO;
 464                 p = xdr_inline_decode(&stream, 4);
 465                 if (!p)
 466                         goto out_err_free;
 467                 fls->mirror_array[i]->efficiency = be32_to_cpup(p);
 468
 469                 /* stateid */
 470                 rc = decode_pnfs_stateid(&stream, &fls->mirror_array[i]->stateid);
 471                 if (rc)
 472                         goto out_err_free;
 473
 474                 /* fh */
 475                 p = xdr_inline_decode(&stream, 4);
 476                 if (!p)
 477                         goto out_err_free;
 478                 fh_count = be32_to_cpup(p);
 479
 480                 fls->mirror_array[i]->fh_versions =
 481                         kzalloc(fh_count * sizeof(struct nfs_fh),
 482                                 gfp_flags);
 483                 if (fls->mirror_array[i]->fh_versions == NULL) {
 484                         rc = -ENOMEM;
 485                         goto out_err_free;
 486                 }
 487
 488                 for (j = 0; j < fh_count; j++) {
 489                         rc = decode_nfs_fh(&stream,
 490                                            &fls->mirror_array[i]->fh_versions[j]);
 491                         if (rc)
 492                                 goto out_err_free;
 493                 }
 494
 495                 fls->mirror_array[i]->fh_versions_cnt = fh_count;
 496
 497                 /* user */
 498                 rc = decode_name(&stream, &id);
 499                 if (rc)
 500                         goto out_err_free;
 501
 502                 acred.uid = make_kuid(&init_user_ns, id);
 503
 504                 /* group */
 505                 rc = decode_name(&stream, &id);
 506                 if (rc)
 507                         goto out_err_free;
 508
 509                 acred.gid = make_kgid(&init_user_ns, id);
 510
 511                 /* find the cred for it */
 512                 rcu_assign_pointer(cred, rpc_lookup_generic_cred(&acred, 0, gfp_flags));
 513                 if (IS_ERR(cred)) {
 514                         rc = PTR_ERR(cred);
 515                         goto out_err_free;
 516                 }
 517
 518                 if (lgr->range.iomode == IOMODE_READ)
 519                         rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred);
 520                 else
 521                         rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred);
 522
 523                 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]);
 524                 if (mirror != fls->mirror_array[i]) {
 525                         /* swap cred ptrs so free_mirror will clean up old */
 526                         if (lgr->range.iomode == IOMODE_READ) {
 527                                 cred = xchg(&mirror->ro_cred, cred);
 528                                 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred);
 529                         } else {
 530                                 cred = xchg(&mirror->rw_cred, cred);
 531                                 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred);
 532                         }
 533                         ff_layout_free_mirror(fls->mirror_array[i]);
 534                         fls->mirror_array[i] = mirror;
 535                 }
 536
 537                 dprintk("%s: iomode %s uid %u gid %u\n", __func__,
 538                         lgr->range.iomode == IOMODE_READ ? "READ" : "RW",
 539                         from_kuid(&init_user_ns, acred.uid),
 540                         from_kgid(&init_user_ns, acred.gid));
 541         }
 542
 543         p = xdr_inline_decode(&stream, 4);
 544         if (!p)
 545                 goto out_sort_mirrors;
 546         fls->flags = be32_to_cpup(p);
 547
 548         p = xdr_inline_decode(&stream, 4);
 549         if (!p)
 550                 goto out_sort_mirrors;
 551         for (i=0; i < fls->mirror_array_cnt; i++)
 552                 fls->mirror_array[i]->report_interval = be32_to_cpup(p);
 553
 554 out_sort_mirrors:
 555         ff_layout_sort_mirrors(fls);
 556         rc = ff_layout_check_layout(lgr);
 557         if (rc)
 558                 goto out_err_free;
 559         ff_layout_mark_devices_valid(fls);
 560
 561         ret = &fls->generic_hdr;
 562         dprintk("<-- %s (success)\n", __func__);
 563 out_free_page:
 564         __free_page(scratch);
 565         return ret;
 566 out_err_free:
 567         _ff_layout_free_lseg(fls);
 568         ret = ERR_PTR(rc);
 569         dprintk("<-- %s (%d)\n", __func__, rc);
 570         goto out_free_page;
 571 }
 572
 573 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout)
 574 {
 575         struct pnfs_layout_segment *lseg;
 576
 577         list_for_each_entry(lseg, &layout->plh_segs, pls_list)
 578                 if (lseg->pls_range.iomode == IOMODE_RW)
 579                         return true;
 580
 581         return false;
 582 }
 583
 584 static void
 585 ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
 586 {
 587         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 588
 589         dprintk("--> %s\n", __func__);
 590
 591         if (lseg->pls_range.iomode == IOMODE_RW) {
 592                 struct nfs4_flexfile_layout *ffl;
 593                 struct inode *inode;
 594
 595                 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
 596                 inode = ffl->generic_hdr.plh_inode;
 597                 spin_lock(&inode->i_lock);
 598                 if (!ff_layout_has_rw_segments(lseg->pls_layout)) {
 599                         ffl->commit_info.nbuckets = 0;
 600                         kfree(ffl->commit_info.buckets);
 601                         ffl->commit_info.buckets = NULL;
 602                 }
 603                 spin_unlock(&inode->i_lock);
 604         }
 605         _ff_layout_free_lseg(fls);
 606 }
 607
 608 /* Return 1 until we have multiple lsegs support */
 609 static int
 610 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
 611 {
 612         return 1;
 613 }
 614
 615 static void
 616 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 617 {
 618         /* first IO request? */
 619         if (atomic_inc_return(&timer->n_ops) == 1) {
 620                 timer->start_time = now;
 621         }
 622 }
 623
 624 static ktime_t
 625 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 626 {
 627         ktime_t start;
 628
 629         if (atomic_dec_return(&timer->n_ops) < 0)
 630                 WARN_ON_ONCE(1);
 631
 632         start = timer->start_time;
 633         timer->start_time = now;
 634         return ktime_sub(now, start);
 635 }
 636
 637 static bool
 638 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
 639                             struct nfs4_ff_layoutstat *layoutstat,
 640                             ktime_t now)
 641 {
 642         static const ktime_t notime = {0};
 643         s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL;
 644         struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(mirror->layout);
 645
 646         nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now);
 647         if (ktime_equal(mirror->start_time, notime))
 648                 mirror->start_time = now;
 649         if (mirror->report_interval != 0)
 650                 report_interval = (s64)mirror->report_interval * 1000LL;
 651         else if (layoutstats_timer != 0)
 652                 report_interval = (s64)layoutstats_timer * 1000LL;
 653         if (ktime_to_ms(ktime_sub(now, ffl->last_report_time)) >=
 654                         report_interval) {
 655                 ffl->last_report_time = now;
 656                 return true;
 657         }
 658
 659         return false;
 660 }
 661
 662 static void
 663 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
 664                 __u64 requested)
 665 {
 666         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 667
 668         iostat->ops_requested++;
 669         iostat->bytes_requested += requested;
 670 }
 671
 672 static void
 673 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
 674                 __u64 requested,
 675                 __u64 completed,
 676                 ktime_t time_completed,
 677                 ktime_t time_started)
 678 {
 679         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 680         ktime_t completion_time = ktime_sub(time_completed, time_started);
 681         ktime_t timer;
 682
 683         iostat->ops_completed++;
 684         iostat->bytes_completed += completed;
 685         iostat->bytes_not_delivered += requested - completed;
 686
 687         timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed);
 688         iostat->total_busy_time =
 689                         ktime_add(iostat->total_busy_time, timer);
 690         iostat->aggregate_completion_time =
 691                         ktime_add(iostat->aggregate_completion_time,
 692                                         completion_time);
 693 }
 694
 695 static void
 696 nfs4_ff_layout_stat_io_start_read(struct inode *inode,
 697                 struct nfs4_ff_layout_mirror *mirror,
 698                 __u64 requested, ktime_t now)
 699 {
 700         bool report;
 701
 702         spin_lock(&mirror->lock);
 703         report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now);
 704         nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
 705         spin_unlock(&mirror->lock);
 706
 707         if (report)
 708                 pnfs_report_layoutstat(inode, GFP_KERNEL);
 709 }
 710
 711 static void
 712 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
 713                 struct nfs4_ff_layout_mirror *mirror,
 714                 __u64 requested,
 715                 __u64 completed)
 716 {
 717         spin_lock(&mirror->lock);
 718         nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
 719                         requested, completed,
 720                         ktime_get(), task->tk_start);
 721         spin_unlock(&mirror->lock);
 722 }
 723
 724 static void
 725 nfs4_ff_layout_stat_io_start_write(struct inode *inode,
 726                 struct nfs4_ff_layout_mirror *mirror,
 727                 __u64 requested, ktime_t now)
 728 {
 729         bool report;
 730
 731         spin_lock(&mirror->lock);
 732         report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now);
 733         nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
 734         spin_unlock(&mirror->lock);
 735
 736         if (report)
 737                 pnfs_report_layoutstat(inode, GFP_NOIO);
 738 }
 739
 740 static void
 741 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
 742                 struct nfs4_ff_layout_mirror *mirror,
 743                 __u64 requested,
 744                 __u64 completed,
 745                 enum nfs3_stable_how committed)
 746 {
 747         if (committed == NFS_UNSTABLE)
 748                 requested = completed = 0;
 749
 750         spin_lock(&mirror->lock);
 751         nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
 752                         requested, completed, ktime_get(), task->tk_start);
 753         spin_unlock(&mirror->lock);
 754 }
 755
 756 static int
 757 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 758                             struct nfs_commit_info *cinfo,
 759                             gfp_t gfp_flags)
 760 {
 761         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 762         struct pnfs_commit_bucket *buckets;
 763         int size;
 764
 765         if (cinfo->ds->nbuckets != 0) {
 766                 /* This assumes there is only one RW lseg per file.
 767                  * To support multiple lseg per file, we need to
 768                  * change struct pnfs_commit_bucket to allow dynamic
 769                  * increasing nbuckets.
 770                  */
 771                 return 0;
 772         }
 773
 774         size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg);
 775
 776         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 777                           gfp_flags);
 778         if (!buckets)
 779                 return -ENOMEM;
 780         else {
 781                 int i;
 782
 783                 spin_lock(&cinfo->inode->i_lock);
 784                 if (cinfo->ds->nbuckets != 0)
 785                         kfree(buckets);
 786                 else {
 787                         cinfo->ds->buckets = buckets;
 788                         cinfo->ds->nbuckets = size;
 789                         for (i = 0; i < size; i++) {
 790                                 INIT_LIST_HEAD(&buckets[i].written);
 791                                 INIT_LIST_HEAD(&buckets[i].committing);
 792                                 /* mark direct verifier as unset */
 793                                 buckets[i].direct_verf.committed =
 794                                         NFS_INVALID_STABLE_HOW;
 795                         }
 796                 }
 797                 spin_unlock(&cinfo->inode->i_lock);
 798                 return 0;
 799         }
 800 }
 801
 802 static struct nfs4_pnfs_ds *
 803 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg,
 804                                   int start_idx,
 805                                   int *best_idx)
 806 {
 807         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 808         struct nfs4_pnfs_ds *ds;
 809         bool fail_return = false;
 810         int idx;
 811
 812         /* mirrors are sorted by efficiency */
 813         for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) {
 814                 if (idx+1 == fls->mirror_array_cnt)
 815                         fail_return = true;
 816                 ds = nfs4_ff_layout_prepare_ds(lseg, idx, fail_return);
 817                 if (ds) {
 818                         *best_idx = idx;
 819                         return ds;
 820                 }
 821         }
 822
 823         return NULL;
 824 }
 825
 826 static void
 827 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio,
 828                       struct nfs_page *req,
 829                       bool strict_iomode)
 830 {
 831 retry_strict:
 832         pnfs_put_lseg(pgio->pg_lseg);
 833         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 834                                            req->wb_context,
 835                                            0,
 836                                            NFS4_MAX_UINT64,
 837                                            IOMODE_READ,
 838                                            strict_iomode,
 839                                            GFP_KERNEL);
 840         if (IS_ERR(pgio->pg_lseg)) {
 841                 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 842                 pgio->pg_lseg = NULL;
 843         }
 844
 845         /* If we don't have checking, do get a IOMODE_RW
 846          * segment, and the server wants to avoid READs
 847          * there, then retry!
 848          */
 849         if (pgio->pg_lseg && !strict_iomode &&
 850             ff_layout_avoid_read_on_rw(pgio->pg_lseg)) {
 851                 strict_iomode = true;
 852                 goto retry_strict;
 853         }
 854 }
 855
 856 static void
 857 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 858                         struct nfs_page *req)
 859 {
 860         struct nfs_pgio_mirror *pgm;
 861         struct nfs4_ff_layout_mirror *mirror;
 862         struct nfs4_pnfs_ds *ds;
 863         int ds_idx;
 864
 865 retry:
 866         /* Use full layout for now */
 867         if (!pgio->pg_lseg)
 868                 ff_layout_pg_get_read(pgio, req, false);
 869         else if (ff_layout_avoid_read_on_rw(pgio->pg_lseg))
 870                 ff_layout_pg_get_read(pgio, req, true);
 871
 872         /* If no lseg, fall back to read through mds */
 873         if (pgio->pg_lseg == NULL)
 874                 goto out_mds;
 875
 876         ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx);
 877         if (!ds) {
 878                 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg))
 879                         goto out_mds;
 880                 pnfs_put_lseg(pgio->pg_lseg);
 881                 pgio->pg_lseg = NULL;
 882                 /* Sleep for 1 second before retrying */
 883                 ssleep(1);
 884                 goto retry;
 885         }
 886
 887         mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
 888
 889         pgio->pg_mirror_idx = ds_idx;
 890
 891         /* read always uses only one mirror - idx 0 for pgio layer */
 892         pgm = &pgio->pg_mirrors[0];
 893         pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
 894
 895         return;
 896 out_mds:
 897         pnfs_put_lseg(pgio->pg_lseg);
 898         pgio->pg_lseg = NULL;
 899         nfs_pageio_reset_read_mds(pgio);
 900 }
 901
 902 static void
 903 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 904                         struct nfs_page *req)
 905 {
 906         struct nfs4_ff_layout_mirror *mirror;
 907         struct nfs_pgio_mirror *pgm;
 908         struct nfs_commit_info cinfo;
 909         struct nfs4_pnfs_ds *ds;
 910         int i;
 911         int status;
 912
 913 retry:
 914         if (!pgio->pg_lseg) {
 915                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 916                                                    req->wb_context,
 917                                                    0,
 918                                                    NFS4_MAX_UINT64,
 919                                                    IOMODE_RW,
 920                                                    false,
 921                                                    GFP_NOFS);
 922                 if (IS_ERR(pgio->pg_lseg)) {
 923                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 924                         pgio->pg_lseg = NULL;
 925                         return;
 926                 }
 927         }
 928         /* If no lseg, fall back to write through mds */
 929         if (pgio->pg_lseg == NULL)
 930                 goto out_mds;
 931
 932         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 933         status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 934         if (status < 0)
 935                 goto out_mds;
 936
 937         /* Use a direct mapping of ds_idx to pgio mirror_idx */
 938         if (WARN_ON_ONCE(pgio->pg_mirror_count !=
 939             FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)))
 940                 goto out_mds;
 941
 942         for (i = 0; i < pgio->pg_mirror_count; i++) {
 943                 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
 944                 if (!ds) {
 945                         if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg))
 946                                 goto out_mds;
 947                         pnfs_put_lseg(pgio->pg_lseg);
 948                         pgio->pg_lseg = NULL;
 949                         /* Sleep for 1 second before retrying */
 950                         ssleep(1);
 951                         goto retry;
 952                 }
 953                 pgm = &pgio->pg_mirrors[i];
 954                 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
 955                 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
 956         }
 957
 958         return;
 959
 960 out_mds:
 961         pnfs_put_lseg(pgio->pg_lseg);
 962         pgio->pg_lseg = NULL;
 963         nfs_pageio_reset_write_mds(pgio);
 964 }
 965
 966 static unsigned int
 967 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
 968                                     struct nfs_page *req)
 969 {
 970         if (!pgio->pg_lseg) {
 971                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 972                                                    req->wb_context,
 973                                                    0,
 974                                                    NFS4_MAX_UINT64,
 975                                                    IOMODE_RW,
 976                                                    false,
 977                                                    GFP_NOFS);
 978                 if (IS_ERR(pgio->pg_lseg)) {
 979                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 980                         pgio->pg_lseg = NULL;
 981                         goto out;
 982                 }
 983         }
 984         if (pgio->pg_lseg)
 985                 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
 986
 987         /* no lseg means that pnfs is not in use, so no mirroring here */
 988         nfs_pageio_reset_write_mds(pgio);
 989 out:
 990         return 1;
 991 }
 992
 993 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
 994         .pg_init = ff_layout_pg_init_read,
 995         .pg_test = pnfs_generic_pg_test,
 996         .pg_doio = pnfs_generic_pg_readpages,
 997         .pg_cleanup = pnfs_generic_pg_cleanup,
 998 };
 999
1000 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
1001         .pg_init = ff_layout_pg_init_write,
1002         .pg_test = pnfs_generic_pg_test,
1003         .pg_doio = pnfs_generic_pg_writepages,
1004         .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
1005         .pg_cleanup = pnfs_generic_pg_cleanup,
1006 };
1007
1008 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
1009 {
1010         struct rpc_task *task = &hdr->task;
1011
1012         pnfs_layoutcommit_inode(hdr->inode, false);
1013
1014         if (retry_pnfs) {
1015                 dprintk("%s Reset task %5u for i/o through pNFS "
1016                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1017                         hdr->task.tk_pid,
1018                         hdr->inode->i_sb->s_id,
1019                         (unsigned long long)NFS_FILEID(hdr->inode),
1020                         hdr->args.count,
1021                         (unsigned long long)hdr->args.offset);
1022
1023                 hdr->completion_ops->reschedule_io(hdr);
1024                 return;
1025         }
1026
1027         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1028                 dprintk("%s Reset task %5u for i/o through MDS "
1029                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1030                         hdr->task.tk_pid,
1031                         hdr->inode->i_sb->s_id,
1032                         (unsigned long long)NFS_FILEID(hdr->inode),
1033                         hdr->args.count,
1034                         (unsigned long long)hdr->args.offset);
1035
1036                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
1037         }
1038 }
1039
1040 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
1041 {
1042         struct rpc_task *task = &hdr->task;
1043
1044         pnfs_layoutcommit_inode(hdr->inode, false);
1045
1046         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1047                 dprintk("%s Reset task %5u for i/o through MDS "
1048                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1049                         hdr->task.tk_pid,
1050                         hdr->inode->i_sb->s_id,
1051                         (unsigned long long)NFS_FILEID(hdr->inode),
1052                         hdr->args.count,
1053                         (unsigned long long)hdr->args.offset);
1054
1055                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
1056         }
1057 }
1058
1059 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
1060                                            struct nfs4_state *state,
1061                                            struct nfs_client *clp,
1062                                            struct pnfs_layout_segment *lseg,
1063                                            int idx)
1064 {
1065         struct pnfs_layout_hdr *lo = lseg->pls_layout;
1066         struct inode *inode = lo->plh_inode;
1067         struct nfs_server *mds_server = NFS_SERVER(inode);
1068
1069         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1070         struct nfs_client *mds_client = mds_server->nfs_client;
1071         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
1072
1073         if (task->tk_status >= 0)
1074                 return 0;
1075
1076         switch (task->tk_status) {
1077         /* MDS state errors */
1078         case -NFS4ERR_DELEG_REVOKED:
1079         case -NFS4ERR_ADMIN_REVOKED:
1080         case -NFS4ERR_BAD_STATEID:
1081                 if (state == NULL)
1082                         break;
1083                 nfs_remove_bad_delegation(state->inode);
1084         case -NFS4ERR_OPENMODE:
1085                 if (state == NULL)
1086                         break;
1087                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1088                         goto out_bad_stateid;
1089                 goto wait_on_recovery;
1090         case -NFS4ERR_EXPIRED:
1091                 if (state != NULL) {
1092                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
1093                                 goto out_bad_stateid;
1094                 }
1095                 nfs4_schedule_lease_recovery(mds_client);
1096                 goto wait_on_recovery;
1097         /* DS session errors */
1098         case -NFS4ERR_BADSESSION:
1099         case -NFS4ERR_BADSLOT:
1100         case -NFS4ERR_BAD_HIGH_SLOT:
1101         case -NFS4ERR_DEADSESSION:
1102         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1103         case -NFS4ERR_SEQ_FALSE_RETRY:
1104         case -NFS4ERR_SEQ_MISORDERED:
1105                 dprintk("%s ERROR %d, Reset session. Exchangeid "
1106                         "flags 0x%x\n", __func__, task->tk_status,
1107                         clp->cl_exchange_flags);
1108                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1109                 break;
1110         case -NFS4ERR_DELAY:
1111         case -NFS4ERR_GRACE:
1112                 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1113                 break;
1114         case -NFS4ERR_RETRY_UNCACHED_REP:
1115                 break;
1116         /* Invalidate Layout errors */
1117         case -NFS4ERR_PNFS_NO_LAYOUT:
1118         case -ESTALE:           /* mapped NFS4ERR_STALE */
1119         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
1120         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
1121         case -NFS4ERR_FHEXPIRED:
1122         case -NFS4ERR_WRONG_TYPE:
1123                 dprintk("%s Invalid layout error %d\n", __func__,
1124                         task->tk_status);
1125                 /*
1126                  * Destroy layout so new i/o will get a new layout.
1127                  * Layout will not be destroyed until all current lseg
1128                  * references are put. Mark layout as invalid to resend failed
1129                  * i/o and all i/o waiting on the slot table to the MDS until
1130                  * layout is destroyed and a new valid layout is obtained.
1131                  */
1132                 pnfs_destroy_layout(NFS_I(inode));
1133                 rpc_wake_up(&tbl->slot_tbl_waitq);
1134                 goto reset;
1135         /* RPC connection errors */
1136         case -ECONNREFUSED:
1137         case -EHOSTDOWN:
1138         case -EHOSTUNREACH:
1139         case -ENETUNREACH:
1140         case -EIO:
1141         case -ETIMEDOUT:
1142         case -EPIPE:
1143                 dprintk("%s DS connection error %d\n", __func__,
1144                         task->tk_status);
1145                 nfs4_mark_deviceid_unavailable(devid);
1146                 rpc_wake_up(&tbl->slot_tbl_waitq);
1147                 /* fall through */
1148         default:
1149                 if (ff_layout_avoid_mds_available_ds(lseg))
1150                         return -NFS4ERR_RESET_TO_PNFS;
1151 reset:
1152                 dprintk("%s Retry through MDS. Error %d\n", __func__,
1153                         task->tk_status);
1154                 return -NFS4ERR_RESET_TO_MDS;
1155         }
1156 out:
1157         task->tk_status = 0;
1158         return -EAGAIN;
1159 out_bad_stateid:
1160         task->tk_status = -EIO;
1161         return 0;
1162 wait_on_recovery:
1163         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
1164         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
1165                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
1166         goto out;
1167 }
1168
1169 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
1170 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1171                                            struct pnfs_layout_segment *lseg,
1172                                            int idx)
1173 {
1174         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1175
1176         if (task->tk_status >= 0)
1177                 return 0;
1178
1179         switch (task->tk_status) {
1180         /* File access problems. Don't mark the device as unavailable */
1181         case -EACCES:
1182         case -ESTALE:
1183         case -EISDIR:
1184         case -EBADHANDLE:
1185         case -ELOOP:
1186         case -ENOSPC:
1187                 break;
1188         case -EJUKEBOX:
1189                 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1190                 goto out_retry;
1191         default:
1192                 dprintk("%s DS connection error %d\n", __func__,
1193                         task->tk_status);
1194                 nfs4_mark_deviceid_unavailable(devid);
1195         }
1196         /* FIXME: Need to prevent infinite looping here. */
1197         return -NFS4ERR_RESET_TO_PNFS;
1198 out_retry:
1199         task->tk_status = 0;
1200         rpc_restart_call_prepare(task);
1201         rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
1202         return -EAGAIN;
1203 }
1204
1205 static int ff_layout_async_handle_error(struct rpc_task *task,
1206                                         struct nfs4_state *state,
1207                                         struct nfs_client *clp,
1208                                         struct pnfs_layout_segment *lseg,
1209                                         int idx)
1210 {
1211         int vers = clp->cl_nfs_mod->rpc_vers->number;
1212
1213         switch (vers) {
1214         case 3:
1215                 return ff_layout_async_handle_error_v3(task, lseg, idx);
1216         case 4:
1217                 return ff_layout_async_handle_error_v4(task, state, clp,
1218                                                        lseg, idx);
1219         default:
1220                 /* should never happen */
1221                 WARN_ON_ONCE(1);
1222                 return 0;
1223         }
1224 }
1225
1226 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1227                                         int idx, u64 offset, u64 length,
1228                                         u32 status, int opnum, int error)
1229 {
1230         struct nfs4_ff_layout_mirror *mirror;
1231         int err;
1232
1233         if (status == 0) {
1234                 switch (error) {
1235                 case -ETIMEDOUT:
1236                 case -EPFNOSUPPORT:
1237                 case -EPROTONOSUPPORT:
1238                 case -EOPNOTSUPP:
1239                 case -ECONNREFUSED:
1240                 case -ECONNRESET:
1241                 case -EHOSTDOWN:
1242                 case -EHOSTUNREACH:
1243                 case -ENETUNREACH:
1244                 case -EADDRINUSE:
1245                 case -ENOBUFS:
1246                 case -EPIPE:
1247                 case -EPERM:
1248                         status = NFS4ERR_NXIO;
1249                         break;
1250                 case -EACCES:
1251                         status = NFS4ERR_ACCESS;
1252                         break;
1253                 default:
1254                         return;
1255                 }
1256         }
1257
1258         switch (status) {
1259         case NFS4ERR_DELAY:
1260         case NFS4ERR_GRACE:
1261                 return;
1262         default:
1263                 break;
1264         }
1265
1266         mirror = FF_LAYOUT_COMP(lseg, idx);
1267         err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1268                                        mirror, offset, length, status, opnum,
1269                                        GFP_NOIO);
1270         pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
1271         dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1272 }
1273
1274 /* NFS_PROTO call done callback routines */
1275 static int ff_layout_read_done_cb(struct rpc_task *task,
1276                                 struct nfs_pgio_header *hdr)
1277 {
1278         int err;
1279
1280         trace_nfs4_pnfs_read(hdr, task->tk_status);
1281         if (task->tk_status < 0)
1282                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1283                                             hdr->args.offset, hdr->args.count,
1284                                             hdr->res.op_status, OP_READ,
1285                                             task->tk_status);
1286         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1287                                            hdr->ds_clp, hdr->lseg,
1288                                            hdr->pgio_mirror_idx);
1289
1290         switch (err) {
1291         case -NFS4ERR_RESET_TO_PNFS:
1292                 if (ff_layout_choose_best_ds_for_read(hdr->lseg,
1293                                         hdr->pgio_mirror_idx + 1,
1294                                         &hdr->pgio_mirror_idx))
1295                         goto out_eagain;
1296                 pnfs_read_resend_pnfs(hdr);
1297                 return task->tk_status;
1298         case -NFS4ERR_RESET_TO_MDS:
1299                 ff_layout_reset_read(hdr);
1300                 return task->tk_status;
1301         case -EAGAIN:
1302                 goto out_eagain;
1303         }
1304
1305         return 0;
1306 out_eagain:
1307         rpc_restart_call_prepare(task);
1308         return -EAGAIN;
1309 }
1310
1311 static bool
1312 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1313 {
1314         return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1315 }
1316
1317 /*
1318  * We reference the rpc_cred of the first WRITE that triggers the need for
1319  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1320  * rfc5661 is not clear about which credential should be used.
1321  *
1322  * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1323  * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1324  * we always send layoutcommit after DS writes.
1325  */
1326 static void
1327 ff_layout_set_layoutcommit(struct inode *inode,
1328                 struct pnfs_layout_segment *lseg,
1329                 loff_t end_offset)
1330 {
1331         if (!ff_layout_need_layoutcommit(lseg))
1332                 return;
1333
1334         pnfs_set_layoutcommit(inode, lseg, end_offset);
1335         dprintk("%s inode %lu pls_end_pos %llu\n", __func__, inode->i_ino,
1336                 (unsigned long long) NFS_I(inode)->layout->plh_lwb);
1337 }
1338
1339 static bool
1340 ff_layout_device_unavailable(struct pnfs_layout_segment *lseg, int idx)
1341 {
1342         /* No mirroring for now */
1343         struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1344
1345         return ff_layout_test_devid_unavailable(node);
1346 }
1347
1348 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task,
1349                 struct nfs_pgio_header *hdr)
1350 {
1351         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1352                 return;
1353         nfs4_ff_layout_stat_io_start_read(hdr->inode,
1354                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1355                         hdr->args.count,
1356                         task->tk_start);
1357 }
1358
1359 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
1360                 struct nfs_pgio_header *hdr)
1361 {
1362         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1363                 return;
1364         nfs4_ff_layout_stat_io_end_read(task,
1365                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1366                         hdr->args.count,
1367                         hdr->res.count);
1368 }
1369
1370 static int ff_layout_read_prepare_common(struct rpc_task *task,
1371                                          struct nfs_pgio_header *hdr)
1372 {
1373         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1374                 rpc_exit(task, -EIO);
1375                 return -EIO;
1376         }
1377         if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) {
1378                 rpc_exit(task, -EHOSTDOWN);
1379                 return -EAGAIN;
1380         }
1381
1382         ff_layout_read_record_layoutstats_start(task, hdr);
1383         return 0;
1384 }
1385
1386 /*
1387  * Call ops for the async read/write cases
1388  * In the case of dense layouts, the offset needs to be reset to its
1389  * original value.
1390  */
1391 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1392 {
1393         struct nfs_pgio_header *hdr = data;
1394
1395         if (ff_layout_read_prepare_common(task, hdr))
1396                 return;
1397
1398         rpc_call_start(task);
1399 }
1400
1401 static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1402                                     struct nfs4_sequence_args *args,
1403                                     struct nfs4_sequence_res *res,
1404                                     struct rpc_task *task)
1405 {
1406         if (ds_clp->cl_session)
1407                 return nfs41_setup_sequence(ds_clp->cl_session,
1408                                            args,
1409                                            res,
1410                                            task);
1411         return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1412                                    args,
1413                                    res,
1414                                    task);
1415 }
1416
1417 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1418 {
1419         struct nfs_pgio_header *hdr = data;
1420
1421         if (ff_layout_setup_sequence(hdr->ds_clp,
1422                                      &hdr->args.seq_args,
1423                                      &hdr->res.seq_res,
1424                                      task))
1425                 return;
1426
1427         if (ff_layout_read_prepare_common(task, hdr))
1428                 return;
1429
1430         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1431                         hdr->args.lock_context, FMODE_READ) == -EIO)
1432                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1433 }
1434
1435 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1436 {
1437         struct nfs_pgio_header *hdr = data;
1438
1439         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1440
1441         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1442             task->tk_status == 0) {
1443                 nfs4_sequence_done(task, &hdr->res.seq_res);
1444                 return;
1445         }
1446
1447         /* Note this may cause RPC to be resent */
1448         hdr->mds_ops->rpc_call_done(task, hdr);
1449 }
1450
1451 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1452 {
1453         struct nfs_pgio_header *hdr = data;
1454
1455         ff_layout_read_record_layoutstats_done(task, hdr);
1456         rpc_count_iostats_metrics(task,
1457             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1458 }
1459
1460 static void ff_layout_read_release(void *data)
1461 {
1462         struct nfs_pgio_header *hdr = data;
1463
1464         ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
1465         pnfs_generic_rw_release(data);
1466 }
1467
1468
1469 static int ff_layout_write_done_cb(struct rpc_task *task,
1470                                 struct nfs_pgio_header *hdr)
1471 {
1472         loff_t end_offs = 0;
1473         int err;
1474
1475         trace_nfs4_pnfs_write(hdr, task->tk_status);
1476         if (task->tk_status < 0)
1477                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1478                                             hdr->args.offset, hdr->args.count,
1479                                             hdr->res.op_status, OP_WRITE,
1480                                             task->tk_status);
1481         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1482                                            hdr->ds_clp, hdr->lseg,
1483                                            hdr->pgio_mirror_idx);
1484
1485         switch (err) {
1486         case -NFS4ERR_RESET_TO_PNFS:
1487                 ff_layout_reset_write(hdr, true);
1488                 return task->tk_status;
1489         case -NFS4ERR_RESET_TO_MDS:
1490                 ff_layout_reset_write(hdr, false);
1491                 return task->tk_status;
1492         case -EAGAIN:
1493                 return -EAGAIN;
1494         }
1495
1496         if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1497             hdr->res.verf->committed == NFS_DATA_SYNC)
1498                 end_offs = hdr->mds_offset + (loff_t)hdr->res.count;
1499
1500         /* Note: if the write is unstable, don't set end_offs until commit */
1501         ff_layout_set_layoutcommit(hdr->inode, hdr->lseg, end_offs);
1502
1503         /* zero out fattr since we don't care DS attr at all */
1504         hdr->fattr.valid = 0;
1505         if (task->tk_status >= 0)
1506                 nfs_writeback_update_inode(hdr);
1507
1508         return 0;
1509 }
1510
1511 static int ff_layout_commit_done_cb(struct rpc_task *task,
1512                                      struct nfs_commit_data *data)
1513 {
1514         int err;
1515
1516         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1517         if (task->tk_status < 0)
1518                 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1519                                             data->args.offset, data->args.count,
1520                                             data->res.op_status, OP_COMMIT,
1521                                             task->tk_status);
1522         err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1523                                            data->lseg, data->ds_commit_index);
1524
1525         switch (err) {
1526         case -NFS4ERR_RESET_TO_PNFS:
1527                 pnfs_generic_prepare_to_resend_writes(data);
1528                 return -EAGAIN;
1529         case -NFS4ERR_RESET_TO_MDS:
1530                 pnfs_generic_prepare_to_resend_writes(data);
1531                 return -EAGAIN;
1532         case -EAGAIN:
1533                 rpc_restart_call_prepare(task);
1534                 return -EAGAIN;
1535         }
1536
1537         ff_layout_set_layoutcommit(data->inode, data->lseg, data->lwb);
1538
1539         return 0;
1540 }
1541
1542 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task,
1543                 struct nfs_pgio_header *hdr)
1544 {
1545         if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1546                 return;
1547         nfs4_ff_layout_stat_io_start_write(hdr->inode,
1548                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1549                         hdr->args.count,
1550                         task->tk_start);
1551 }
1552
1553 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task,
1554                 struct nfs_pgio_header *hdr)
1555 {
1556         if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1557                 return;
1558         nfs4_ff_layout_stat_io_end_write(task,
1559                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1560                         hdr->args.count, hdr->res.count,
1561                         hdr->res.verf->committed);
1562 }
1563
1564 static int ff_layout_write_prepare_common(struct rpc_task *task,
1565                                           struct nfs_pgio_header *hdr)
1566 {
1567         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1568                 rpc_exit(task, -EIO);
1569                 return -EIO;
1570         }
1571
1572         if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) {
1573                 rpc_exit(task, -EHOSTDOWN);
1574                 return -EAGAIN;
1575         }
1576
1577         ff_layout_write_record_layoutstats_start(task, hdr);
1578         return 0;
1579 }
1580
1581 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1582 {
1583         struct nfs_pgio_header *hdr = data;
1584
1585         if (ff_layout_write_prepare_common(task, hdr))
1586                 return;
1587
1588         rpc_call_start(task);
1589 }
1590
1591 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1592 {
1593         struct nfs_pgio_header *hdr = data;
1594
1595         if (ff_layout_setup_sequence(hdr->ds_clp,
1596                                      &hdr->args.seq_args,
1597                                      &hdr->res.seq_res,
1598                                      task))
1599                 return;
1600
1601         if (ff_layout_write_prepare_common(task, hdr))
1602                 return;
1603
1604         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1605                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
1606                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1607 }
1608
1609 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1610 {
1611         struct nfs_pgio_header *hdr = data;
1612
1613         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1614             task->tk_status == 0) {
1615                 nfs4_sequence_done(task, &hdr->res.seq_res);
1616                 return;
1617         }
1618
1619         /* Note this may cause RPC to be resent */
1620         hdr->mds_ops->rpc_call_done(task, hdr);
1621 }
1622
1623 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1624 {
1625         struct nfs_pgio_header *hdr = data;
1626
1627         ff_layout_write_record_layoutstats_done(task, hdr);
1628         rpc_count_iostats_metrics(task,
1629             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1630 }
1631
1632 static void ff_layout_write_release(void *data)
1633 {
1634         struct nfs_pgio_header *hdr = data;
1635
1636         ff_layout_write_record_layoutstats_done(&hdr->task, hdr);
1637         pnfs_generic_rw_release(data);
1638 }
1639
1640 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task,
1641                 struct nfs_commit_data *cdata)
1642 {
1643         if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags))
1644                 return;
1645         nfs4_ff_layout_stat_io_start_write(cdata->inode,
1646                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1647                         0, task->tk_start);
1648 }
1649
1650 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task,
1651                 struct nfs_commit_data *cdata)
1652 {
1653         struct nfs_page *req;
1654         __u64 count = 0;
1655
1656         if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags))
1657                 return;
1658
1659         if (task->tk_status == 0) {
1660                 list_for_each_entry(req, &cdata->pages, wb_list)
1661                         count += req->wb_bytes;
1662         }
1663         nfs4_ff_layout_stat_io_end_write(task,
1664                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1665                         count, count, NFS_FILE_SYNC);
1666 }
1667
1668 static void ff_layout_commit_prepare_common(struct rpc_task *task,
1669                 struct nfs_commit_data *cdata)
1670 {
1671         ff_layout_commit_record_layoutstats_start(task, cdata);
1672 }
1673
1674 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1675 {
1676         ff_layout_commit_prepare_common(task, data);
1677         rpc_call_start(task);
1678 }
1679
1680 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1681 {
1682         struct nfs_commit_data *wdata = data;
1683
1684         if (ff_layout_setup_sequence(wdata->ds_clp,
1685                                  &wdata->args.seq_args,
1686                                  &wdata->res.seq_res,
1687                                  task))
1688                 return;
1689         ff_layout_commit_prepare_common(task, data);
1690 }
1691
1692 static void ff_layout_commit_done(struct rpc_task *task, void *data)
1693 {
1694         pnfs_generic_write_commit_done(task, data);
1695 }
1696
1697 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1698 {
1699         struct nfs_commit_data *cdata = data;
1700
1701         ff_layout_commit_record_layoutstats_done(task, cdata);
1702         rpc_count_iostats_metrics(task,
1703             &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1704 }
1705
1706 static void ff_layout_commit_release(void *data)
1707 {
1708         struct nfs_commit_data *cdata = data;
1709
1710         ff_layout_commit_record_layoutstats_done(&cdata->task, cdata);
1711         pnfs_generic_commit_release(data);
1712 }
1713
1714 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1715         .rpc_call_prepare = ff_layout_read_prepare_v3,
1716         .rpc_call_done = ff_layout_read_call_done,
1717         .rpc_count_stats = ff_layout_read_count_stats,
1718         .rpc_release = ff_layout_read_release,
1719 };
1720
1721 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1722         .rpc_call_prepare = ff_layout_read_prepare_v4,
1723         .rpc_call_done = ff_layout_read_call_done,
1724         .rpc_count_stats = ff_layout_read_count_stats,
1725         .rpc_release = ff_layout_read_release,
1726 };
1727
1728 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1729         .rpc_call_prepare = ff_layout_write_prepare_v3,
1730         .rpc_call_done = ff_layout_write_call_done,
1731         .rpc_count_stats = ff_layout_write_count_stats,
1732         .rpc_release = ff_layout_write_release,
1733 };
1734
1735 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1736         .rpc_call_prepare = ff_layout_write_prepare_v4,
1737         .rpc_call_done = ff_layout_write_call_done,
1738         .rpc_count_stats = ff_layout_write_count_stats,
1739         .rpc_release = ff_layout_write_release,
1740 };
1741
1742 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1743         .rpc_call_prepare = ff_layout_commit_prepare_v3,
1744         .rpc_call_done = ff_layout_commit_done,
1745         .rpc_count_stats = ff_layout_commit_count_stats,
1746         .rpc_release = ff_layout_commit_release,
1747 };
1748
1749 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1750         .rpc_call_prepare = ff_layout_commit_prepare_v4,
1751         .rpc_call_done = ff_layout_commit_done,
1752         .rpc_count_stats = ff_layout_commit_count_stats,
1753         .rpc_release = ff_layout_commit_release,
1754 };
1755
1756 static enum pnfs_try_status
1757 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1758 {
1759         struct pnfs_layout_segment *lseg = hdr->lseg;
1760         struct nfs4_pnfs_ds *ds;
1761         struct rpc_clnt *ds_clnt;
1762         struct rpc_cred *ds_cred;
1763         loff_t offset = hdr->args.offset;
1764         u32 idx = hdr->pgio_mirror_idx;
1765         int vers;
1766         struct nfs_fh *fh;
1767
1768         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1769                 __func__, hdr->inode->i_ino,
1770                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1771
1772         ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1773         if (!ds)
1774                 goto out_failed;
1775
1776         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1777                                                    hdr->inode);
1778         if (IS_ERR(ds_clnt))
1779                 goto out_failed;
1780
1781         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1782         if (!ds_cred)
1783                 goto out_failed;
1784
1785         vers = nfs4_ff_layout_ds_version(lseg, idx);
1786
1787         dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1788                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1789
1790         hdr->pgio_done_cb = ff_layout_read_done_cb;
1791         atomic_inc(&ds->ds_clp->cl_count);
1792         hdr->ds_clp = ds->ds_clp;
1793         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1794         if (fh)
1795                 hdr->args.fh = fh;
1796         /*
1797          * Note that if we ever decide to split across DSes,
1798          * then we may need to handle dense-like offsets.
1799          */
1800         hdr->args.offset = offset;
1801         hdr->mds_offset = offset;
1802
1803         /* Perform an asynchronous read to ds */
1804         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1805                           vers == 3 ? &ff_layout_read_call_ops_v3 :
1806                                       &ff_layout_read_call_ops_v4,
1807                           0, RPC_TASK_SOFTCONN);
1808         put_rpccred(ds_cred);
1809         return PNFS_ATTEMPTED;
1810
1811 out_failed:
1812         if (ff_layout_avoid_mds_available_ds(lseg))
1813                 return PNFS_TRY_AGAIN;
1814         return PNFS_NOT_ATTEMPTED;
1815 }
1816
1817 /* Perform async writes. */
1818 static enum pnfs_try_status
1819 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1820 {
1821         struct pnfs_layout_segment *lseg = hdr->lseg;
1822         struct nfs4_pnfs_ds *ds;
1823         struct rpc_clnt *ds_clnt;
1824         struct rpc_cred *ds_cred;
1825         loff_t offset = hdr->args.offset;
1826         int vers;
1827         struct nfs_fh *fh;
1828         int idx = hdr->pgio_mirror_idx;
1829
1830         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1831         if (!ds)
1832                 return PNFS_NOT_ATTEMPTED;
1833
1834         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1835                                                    hdr->inode);
1836         if (IS_ERR(ds_clnt))
1837                 return PNFS_NOT_ATTEMPTED;
1838
1839         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1840         if (!ds_cred)
1841                 return PNFS_NOT_ATTEMPTED;
1842
1843         vers = nfs4_ff_layout_ds_version(lseg, idx);
1844
1845         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1846                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1847                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1848                 vers);
1849
1850         hdr->pgio_done_cb = ff_layout_write_done_cb;
1851         atomic_inc(&ds->ds_clp->cl_count);
1852         hdr->ds_clp = ds->ds_clp;
1853         hdr->ds_commit_idx = idx;
1854         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1855         if (fh)
1856                 hdr->args.fh = fh;
1857
1858         /*
1859          * Note that if we ever decide to split across DSes,
1860          * then we may need to handle dense-like offsets.
1861          */
1862         hdr->args.offset = offset;
1863
1864         /* Perform an asynchronous write */
1865         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1866                           vers == 3 ? &ff_layout_write_call_ops_v3 :
1867                                       &ff_layout_write_call_ops_v4,
1868                           sync, RPC_TASK_SOFTCONN);
1869         put_rpccred(ds_cred);
1870         return PNFS_ATTEMPTED;
1871 }
1872
1873 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1874 {
1875         return i;
1876 }
1877
1878 static struct nfs_fh *
1879 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1880 {
1881         struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1882
1883         /* FIXME: Assume that there is only one NFS version available
1884          * for the DS.
1885          */
1886         return &flseg->mirror_array[i]->fh_versions[0];
1887 }
1888
1889 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1890 {
1891         struct pnfs_layout_segment *lseg = data->lseg;
1892         struct nfs4_pnfs_ds *ds;
1893         struct rpc_clnt *ds_clnt;
1894         struct rpc_cred *ds_cred;
1895         u32 idx;
1896         int vers, ret;
1897         struct nfs_fh *fh;
1898
1899         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1900         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1901         if (!ds)
1902                 goto out_err;
1903
1904         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1905                                                    data->inode);
1906         if (IS_ERR(ds_clnt))
1907                 goto out_err;
1908
1909         ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1910         if (!ds_cred)
1911                 goto out_err;
1912
1913         vers = nfs4_ff_layout_ds_version(lseg, idx);
1914
1915         dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1916                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1917                 vers);
1918         data->commit_done_cb = ff_layout_commit_done_cb;
1919         data->cred = ds_cred;
1920         atomic_inc(&ds->ds_clp->cl_count);
1921         data->ds_clp = ds->ds_clp;
1922         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1923         if (fh)
1924                 data->args.fh = fh;
1925
1926         ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1927                                    vers == 3 ? &ff_layout_commit_call_ops_v3 :
1928                                                &ff_layout_commit_call_ops_v4,
1929                                    how, RPC_TASK_SOFTCONN);
1930         put_rpccred(ds_cred);
1931         return ret;
1932 out_err:
1933         pnfs_generic_prepare_to_resend_writes(data);
1934         pnfs_generic_commit_release(data);
1935         return -EAGAIN;
1936 }
1937
1938 static int
1939 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1940                            int how, struct nfs_commit_info *cinfo)
1941 {
1942         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1943                                             ff_layout_initiate_commit);
1944 }
1945
1946 static struct pnfs_ds_commit_info *
1947 ff_layout_get_ds_info(struct inode *inode)
1948 {
1949         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1950
1951         if (layout == NULL)
1952                 return NULL;
1953
1954         return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1955 }
1956
1957 static void
1958 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
1959 {
1960         nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1961                                                   id_node));
1962 }
1963
1964 static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1965                                   struct xdr_stream *xdr,
1966                                   const struct nfs4_layoutreturn_args *args)
1967 {
1968         struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1969         __be32 *start;
1970         int count = 0, ret = 0;
1971
1972         start = xdr_reserve_space(xdr, 4);
1973         if (unlikely(!start))
1974                 return -E2BIG;
1975
1976         /* This assume we always return _ALL_ layouts */
1977         spin_lock(&hdr->plh_inode->i_lock);
1978         ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1979         spin_unlock(&hdr->plh_inode->i_lock);
1980
1981         *start = cpu_to_be32(count);
1982
1983         return ret;
1984 }
1985
1986 /* report nothing for now */
1987 static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1988                                      struct xdr_stream *xdr,
1989                                      const struct nfs4_layoutreturn_args *args)
1990 {
1991         __be32 *p;
1992
1993         p = xdr_reserve_space(xdr, 4);
1994         if (likely(p))
1995                 *p = cpu_to_be32(0);
1996 }
1997
1998 static struct nfs4_deviceid_node *
1999 ff_layout_alloc_deviceid_node(struct nfs_server *server,
2000                               struct pnfs_device *pdev, gfp_t gfp_flags)
2001 {
2002         struct nfs4_ff_layout_ds *dsaddr;
2003
2004         dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
2005         if (!dsaddr)
2006                 return NULL;
2007         return &dsaddr->id_node;
2008 }
2009
2010 static void
2011 ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
2012                               struct xdr_stream *xdr,
2013                               const struct nfs4_layoutreturn_args *args)
2014 {
2015         struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
2016         __be32 *start;
2017
2018         dprintk("%s: Begin\n", __func__);
2019         start = xdr_reserve_space(xdr, 4);
2020         BUG_ON(!start);
2021
2022         ff_layout_encode_ioerr(flo, xdr, args);
2023         ff_layout_encode_iostats(flo, xdr, args);
2024
2025         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2026         dprintk("%s: Return\n", __func__);
2027 }
2028
2029 static int
2030 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
2031 {
2032         const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
2033
2034         return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
2035 }
2036
2037 static size_t
2038 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
2039                           const int buflen)
2040 {
2041         const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
2042         const struct in6_addr *addr = &sin6->sin6_addr;
2043
2044         /*
2045          * RFC 4291, Section 2.2.2
2046          *
2047          * Shorthanded ANY address
2048          */
2049         if (ipv6_addr_any(addr))
2050                 return snprintf(buf, buflen, "::");
2051
2052         /*
2053          * RFC 4291, Section 2.2.2
2054          *
2055          * Shorthanded loopback address
2056          */
2057         if (ipv6_addr_loopback(addr))
2058                 return snprintf(buf, buflen, "::1");
2059
2060         /*
2061          * RFC 4291, Section 2.2.3
2062          *
2063          * Special presentation address format for mapped v4
2064          * addresses.
2065          */
2066         if (ipv6_addr_v4mapped(addr))
2067                 return snprintf(buf, buflen, "::ffff:%pI4",
2068                                         &addr->s6_addr32[3]);
2069
2070         /*
2071          * RFC 4291, Section 2.2.1
2072          */
2073         return snprintf(buf, buflen, "%pI6c", addr);
2074 }
2075
2076 /* Derived from rpc_sockaddr2uaddr */
2077 static void
2078 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
2079 {
2080         struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
2081         char portbuf[RPCBIND_MAXUADDRPLEN];
2082         char addrbuf[RPCBIND_MAXUADDRLEN];
2083         char *netid;
2084         unsigned short port;
2085         int len, netid_len;
2086         __be32 *p;
2087
2088         switch (sap->sa_family) {
2089         case AF_INET:
2090                 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
2091                         return;
2092                 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
2093                 netid = "tcp";
2094                 netid_len = 3;
2095                 break;
2096         case AF_INET6:
2097                 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
2098                         return;
2099                 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
2100                 netid = "tcp6";
2101                 netid_len = 4;
2102                 break;
2103         default:
2104                 /* we only support tcp and tcp6 */
2105                 WARN_ON_ONCE(1);
2106                 return;
2107         }
2108
2109         snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
2110         len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
2111
2112         p = xdr_reserve_space(xdr, 4 + netid_len);
2113         xdr_encode_opaque(p, netid, netid_len);
2114
2115         p = xdr_reserve_space(xdr, 4 + len);
2116         xdr_encode_opaque(p, addrbuf, len);
2117 }
2118
2119 static void
2120 ff_layout_encode_nfstime(struct xdr_stream *xdr,
2121                          ktime_t t)
2122 {
2123         struct timespec64 ts;
2124         __be32 *p;
2125
2126         p = xdr_reserve_space(xdr, 12);
2127         ts = ktime_to_timespec64(t);
2128         p = xdr_encode_hyper(p, ts.tv_sec);
2129         *p++ = cpu_to_be32(ts.tv_nsec);
2130 }
2131
2132 static void
2133 ff_layout_encode_io_latency(struct xdr_stream *xdr,
2134                             struct nfs4_ff_io_stat *stat)
2135 {
2136         __be32 *p;
2137
2138         p = xdr_reserve_space(xdr, 5 * 8);
2139         p = xdr_encode_hyper(p, stat->ops_requested);
2140         p = xdr_encode_hyper(p, stat->bytes_requested);
2141         p = xdr_encode_hyper(p, stat->ops_completed);
2142         p = xdr_encode_hyper(p, stat->bytes_completed);
2143         p = xdr_encode_hyper(p, stat->bytes_not_delivered);
2144         ff_layout_encode_nfstime(xdr, stat->total_busy_time);
2145         ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
2146 }
2147
2148 static void
2149 ff_layout_encode_layoutstats(struct xdr_stream *xdr,
2150                              struct nfs42_layoutstat_args *args,
2151                              struct nfs42_layoutstat_devinfo *devinfo)
2152 {
2153         struct nfs4_ff_layout_mirror *mirror = devinfo->layout_private;
2154         struct nfs4_pnfs_ds_addr *da;
2155         struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds;
2156         struct nfs_fh *fh = &mirror->fh_versions[0];
2157         __be32 *p, *start;
2158
2159         da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2160         dprintk("%s: DS %s: encoding address %s\n",
2161                 __func__, ds->ds_remotestr, da->da_remotestr);
2162         /* layoutupdate length */
2163         start = xdr_reserve_space(xdr, 4);
2164         /* netaddr4 */
2165         ff_layout_encode_netaddr(xdr, da);
2166         /* nfs_fh4 */
2167         p = xdr_reserve_space(xdr, 4 + fh->size);
2168         xdr_encode_opaque(p, fh->data, fh->size);
2169         /* ff_io_latency4 read */
2170         spin_lock(&mirror->lock);
2171         ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat);
2172         /* ff_io_latency4 write */
2173         ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat);
2174         spin_unlock(&mirror->lock);
2175         /* nfstime4 */
2176         ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time));
2177         /* bool */
2178         p = xdr_reserve_space(xdr, 4);
2179         *p = cpu_to_be32(false);
2180
2181         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2182 }
2183
2184 static int
2185 ff_layout_mirror_prepare_stats(struct nfs42_layoutstat_args *args,
2186                                struct pnfs_layout_hdr *lo,
2187                                int dev_limit)
2188 {
2189         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2190         struct nfs4_ff_layout_mirror *mirror;
2191         struct nfs4_deviceid_node *dev;
2192         struct nfs42_layoutstat_devinfo *devinfo;
2193         int i = 0;
2194
2195         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2196                 if (i >= dev_limit)
2197                         break;
2198                 if (!mirror->mirror_ds)
2199                         continue;
2200                 /* mirror refcount put in cleanup_layoutstats */
2201                 if (!atomic_inc_not_zero(&mirror->ref))
2202                         continue;
2203                 dev = &mirror->mirror_ds->id_node;
2204                 devinfo = &args->devinfo[i];
2205                 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE);
2206                 devinfo->offset = 0;
2207                 devinfo->length = NFS4_MAX_UINT64;
2208                 devinfo->read_count = mirror->read_stat.io_stat.ops_completed;
2209                 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed;
2210                 devinfo->write_count = mirror->write_stat.io_stat.ops_completed;
2211                 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed;
2212                 devinfo->layout_type = LAYOUT_FLEX_FILES;
2213                 devinfo->layoutstats_encode = ff_layout_encode_layoutstats;
2214                 devinfo->layout_private = mirror;
2215
2216                 i++;
2217         }
2218         return i;
2219 }
2220
2221 static int
2222 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2223 {
2224         struct nfs4_flexfile_layout *ff_layout;
2225         struct nfs4_ff_layout_mirror *mirror;
2226         int dev_count = 0;
2227
2228         spin_lock(&args->inode->i_lock);
2229         ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout);
2230         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2231                 if (atomic_read(&mirror->ref) != 0)
2232                         dev_count ++;
2233         }
2234         spin_unlock(&args->inode->i_lock);
2235         /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2236         if (dev_count > PNFS_LAYOUTSTATS_MAXDEV) {
2237                 dprintk("%s: truncating devinfo to limit (%d:%d)\n",
2238                         __func__, dev_count, PNFS_LAYOUTSTATS_MAXDEV);
2239                 dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2240         }
2241         args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO);
2242         if (!args->devinfo)
2243                 return -ENOMEM;
2244
2245         spin_lock(&args->inode->i_lock);
2246         args->num_dev = ff_layout_mirror_prepare_stats(args,
2247                         &ff_layout->generic_hdr, dev_count);
2248         spin_unlock(&args->inode->i_lock);
2249
2250         return 0;
2251 }
2252
2253 static void
2254 ff_layout_cleanup_layoutstats(struct nfs42_layoutstat_data *data)
2255 {
2256         struct nfs4_ff_layout_mirror *mirror;
2257         int i;
2258
2259         for (i = 0; i < data->args.num_dev; i++) {
2260                 mirror = data->args.devinfo[i].layout_private;
2261                 data->args.devinfo[i].layout_private = NULL;
2262                 ff_layout_put_mirror(mirror);
2263         }
2264 }
2265
2266 static struct pnfs_layoutdriver_type flexfilelayout_type = {
2267         .id                     = LAYOUT_FLEX_FILES,
2268         .name                   = "LAYOUT_FLEX_FILES",
2269         .owner                  = THIS_MODULE,
2270         .alloc_layout_hdr       = ff_layout_alloc_layout_hdr,
2271         .free_layout_hdr        = ff_layout_free_layout_hdr,
2272         .alloc_lseg             = ff_layout_alloc_lseg,
2273         .free_lseg              = ff_layout_free_lseg,
2274         .add_lseg               = ff_layout_add_lseg,
2275         .pg_read_ops            = &ff_layout_pg_read_ops,
2276         .pg_write_ops           = &ff_layout_pg_write_ops,
2277         .get_ds_info            = ff_layout_get_ds_info,
2278         .free_deviceid_node     = ff_layout_free_deviceid_node,
2279         .mark_request_commit    = pnfs_layout_mark_request_commit,
2280         .clear_request_commit   = pnfs_generic_clear_request_commit,
2281         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
2282         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
2283         .commit_pagelist        = ff_layout_commit_pagelist,
2284         .read_pagelist          = ff_layout_read_pagelist,
2285         .write_pagelist         = ff_layout_write_pagelist,
2286         .alloc_deviceid_node    = ff_layout_alloc_deviceid_node,
2287         .encode_layoutreturn    = ff_layout_encode_layoutreturn,
2288         .sync                   = pnfs_nfs_generic_sync,
2289         .prepare_layoutstats    = ff_layout_prepare_layoutstats,
2290         .cleanup_layoutstats    = ff_layout_cleanup_layoutstats,
2291 };
2292
2293 static int __init nfs4flexfilelayout_init(void)
2294 {
2295         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
2296                __func__);
2297         if (!ff_zero_group) {
2298                 ff_zero_group = groups_alloc(0);
2299                 if (!ff_zero_group)
2300                         return -ENOMEM;
2301         }
2302         return pnfs_register_layoutdriver(&flexfilelayout_type);
2303 }
2304
2305 static void __exit nfs4flexfilelayout_exit(void)
2306 {
2307         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
2308                __func__);
2309         pnfs_unregister_layoutdriver(&flexfilelayout_type);
2310         if (ff_zero_group) {
2311                 put_group_info(ff_zero_group);
2312                 ff_zero_group = NULL;
2313         }
2314 }
2315
2316 MODULE_ALIAS("nfs-layouttype4-4");
2317
2318 MODULE_LICENSE("GPL");
2319 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
2320
2321 module_init(nfs4flexfilelayout_init);
2322 module_exit(nfs4flexfilelayout_exit);