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