kernel/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; j++) {
 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                 fls->flags = be32_to_cpup(p);
 510
 511         ff_layout_sort_mirrors(fls);
 512         rc = ff_layout_check_layout(lgr);
 513         if (rc)
 514                 goto out_err_free;
 515         ff_layout_mark_devices_valid(fls);
 516
 517         ret = &fls->generic_hdr;
 518         dprintk("<-- %s (success)\n", __func__);
 519 out_free_page:
 520         __free_page(scratch);
 521         return ret;
 522 out_err_free:
 523         _ff_layout_free_lseg(fls);
 524         ret = ERR_PTR(rc);
 525         dprintk("<-- %s (%d)\n", __func__, rc);
 526         goto out_free_page;
 527 }
 528
 529 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout)
 530 {
 531         struct pnfs_layout_segment *lseg;
 532
 533         list_for_each_entry(lseg, &layout->plh_segs, pls_list)
 534                 if (lseg->pls_range.iomode == IOMODE_RW)
 535                         return true;
 536
 537         return false;
 538 }
 539
 540 static void
 541 ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
 542 {
 543         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 544
 545         dprintk("--> %s\n", __func__);
 546
 547         if (lseg->pls_range.iomode == IOMODE_RW) {
 548                 struct nfs4_flexfile_layout *ffl;
 549                 struct inode *inode;
 550
 551                 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
 552                 inode = ffl->generic_hdr.plh_inode;
 553                 spin_lock(&inode->i_lock);
 554                 if (!ff_layout_has_rw_segments(lseg->pls_layout)) {
 555                         ffl->commit_info.nbuckets = 0;
 556                         kfree(ffl->commit_info.buckets);
 557                         ffl->commit_info.buckets = NULL;
 558                 }
 559                 spin_unlock(&inode->i_lock);
 560         }
 561         _ff_layout_free_lseg(fls);
 562 }
 563
 564 /* Return 1 until we have multiple lsegs support */
 565 static int
 566 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
 567 {
 568         return 1;
 569 }
 570
 571 static void
 572 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 573 {
 574         /* first IO request? */
 575         if (atomic_inc_return(&timer->n_ops) == 1) {
 576                 timer->start_time = now;
 577         }
 578 }
 579
 580 static ktime_t
 581 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
 582 {
 583         ktime_t start;
 584
 585         if (atomic_dec_return(&timer->n_ops) < 0)
 586                 WARN_ON_ONCE(1);
 587
 588         start = timer->start_time;
 589         timer->start_time = now;
 590         return ktime_sub(now, start);
 591 }
 592
 593 static bool
 594 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
 595                             struct nfs4_ff_layoutstat *layoutstat,
 596                             ktime_t now)
 597 {
 598         static const ktime_t notime = {0};
 599         s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL;
 600
 601         nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now);
 602         if (ktime_equal(mirror->start_time, notime))
 603                 mirror->start_time = now;
 604         if (ktime_equal(mirror->last_report_time, notime))
 605                 mirror->last_report_time = now;
 606         if (layoutstats_timer != 0)
 607                 report_interval = (s64)layoutstats_timer * 1000LL;
 608         if (ktime_to_ms(ktime_sub(now, mirror->last_report_time)) >=
 609                         report_interval) {
 610                 mirror->last_report_time = now;
 611                 return true;
 612         }
 613
 614         return false;
 615 }
 616
 617 static void
 618 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
 619                 __u64 requested)
 620 {
 621         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 622
 623         iostat->ops_requested++;
 624         iostat->bytes_requested += requested;
 625 }
 626
 627 static void
 628 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
 629                 __u64 requested,
 630                 __u64 completed,
 631                 ktime_t time_completed,
 632                 ktime_t time_started)
 633 {
 634         struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
 635         ktime_t completion_time = ktime_sub(time_completed, time_started);
 636         ktime_t timer;
 637
 638         iostat->ops_completed++;
 639         iostat->bytes_completed += completed;
 640         iostat->bytes_not_delivered += requested - completed;
 641
 642         timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed);
 643         iostat->total_busy_time =
 644                         ktime_add(iostat->total_busy_time, timer);
 645         iostat->aggregate_completion_time =
 646                         ktime_add(iostat->aggregate_completion_time,
 647                                         completion_time);
 648 }
 649
 650 static void
 651 nfs4_ff_layout_stat_io_start_read(struct inode *inode,
 652                 struct nfs4_ff_layout_mirror *mirror,
 653                 __u64 requested, ktime_t now)
 654 {
 655         bool report;
 656
 657         spin_lock(&mirror->lock);
 658         report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now);
 659         nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
 660         spin_unlock(&mirror->lock);
 661
 662         if (report)
 663                 pnfs_report_layoutstat(inode, GFP_KERNEL);
 664 }
 665
 666 static void
 667 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
 668                 struct nfs4_ff_layout_mirror *mirror,
 669                 __u64 requested,
 670                 __u64 completed)
 671 {
 672         spin_lock(&mirror->lock);
 673         nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
 674                         requested, completed,
 675                         ktime_get(), task->tk_start);
 676         spin_unlock(&mirror->lock);
 677 }
 678
 679 static void
 680 nfs4_ff_layout_stat_io_start_write(struct inode *inode,
 681                 struct nfs4_ff_layout_mirror *mirror,
 682                 __u64 requested, ktime_t now)
 683 {
 684         bool report;
 685
 686         spin_lock(&mirror->lock);
 687         report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now);
 688         nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
 689         spin_unlock(&mirror->lock);
 690
 691         if (report)
 692                 pnfs_report_layoutstat(inode, GFP_NOIO);
 693 }
 694
 695 static void
 696 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
 697                 struct nfs4_ff_layout_mirror *mirror,
 698                 __u64 requested,
 699                 __u64 completed,
 700                 enum nfs3_stable_how committed)
 701 {
 702         if (committed == NFS_UNSTABLE)
 703                 requested = completed = 0;
 704
 705         spin_lock(&mirror->lock);
 706         nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
 707                         requested, completed, ktime_get(), task->tk_start);
 708         spin_unlock(&mirror->lock);
 709 }
 710
 711 static int
 712 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 713                             struct nfs_commit_info *cinfo,
 714                             gfp_t gfp_flags)
 715 {
 716         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 717         struct pnfs_commit_bucket *buckets;
 718         int size;
 719
 720         if (cinfo->ds->nbuckets != 0) {
 721                 /* This assumes there is only one RW lseg per file.
 722                  * To support multiple lseg per file, we need to
 723                  * change struct pnfs_commit_bucket to allow dynamic
 724                  * increasing nbuckets.
 725                  */
 726                 return 0;
 727         }
 728
 729         size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg);
 730
 731         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 732                           gfp_flags);
 733         if (!buckets)
 734                 return -ENOMEM;
 735         else {
 736                 int i;
 737
 738                 spin_lock(cinfo->lock);
 739                 if (cinfo->ds->nbuckets != 0)
 740                         kfree(buckets);
 741                 else {
 742                         cinfo->ds->buckets = buckets;
 743                         cinfo->ds->nbuckets = size;
 744                         for (i = 0; i < size; i++) {
 745                                 INIT_LIST_HEAD(&buckets[i].written);
 746                                 INIT_LIST_HEAD(&buckets[i].committing);
 747                                 /* mark direct verifier as unset */
 748                                 buckets[i].direct_verf.committed =
 749                                         NFS_INVALID_STABLE_HOW;
 750                         }
 751                 }
 752                 spin_unlock(cinfo->lock);
 753                 return 0;
 754         }
 755 }
 756
 757 static struct nfs4_pnfs_ds *
 758 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg,
 759                                   int start_idx,
 760                                   int *best_idx)
 761 {
 762         struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
 763         struct nfs4_pnfs_ds *ds;
 764         int idx;
 765
 766         /* mirrors are sorted by efficiency */
 767         for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) {
 768                 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
 769                 if (ds) {
 770                         *best_idx = idx;
 771                         return ds;
 772                 }
 773         }
 774
 775         return NULL;
 776 }
 777
 778 static void
 779 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 780                         struct nfs_page *req)
 781 {
 782         struct nfs_pgio_mirror *pgm;
 783         struct nfs4_ff_layout_mirror *mirror;
 784         struct nfs4_pnfs_ds *ds;
 785         int ds_idx;
 786
 787         /* Use full layout for now */
 788         if (!pgio->pg_lseg)
 789                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 790                                                    req->wb_context,
 791                                                    0,
 792                                                    NFS4_MAX_UINT64,
 793                                                    IOMODE_READ,
 794                                                    GFP_KERNEL);
 795         /* If no lseg, fall back to read through mds */
 796         if (pgio->pg_lseg == NULL)
 797                 goto out_mds;
 798
 799         ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx);
 800         if (!ds)
 801                 goto out_mds;
 802         mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
 803
 804         pgio->pg_mirror_idx = ds_idx;
 805
 806         /* read always uses only one mirror - idx 0 for pgio layer */
 807         pgm = &pgio->pg_mirrors[0];
 808         pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
 809
 810         return;
 811 out_mds:
 812         pnfs_put_lseg(pgio->pg_lseg);
 813         pgio->pg_lseg = NULL;
 814         nfs_pageio_reset_read_mds(pgio);
 815 }
 816
 817 static void
 818 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 819                         struct nfs_page *req)
 820 {
 821         struct nfs4_ff_layout_mirror *mirror;
 822         struct nfs_pgio_mirror *pgm;
 823         struct nfs_commit_info cinfo;
 824         struct nfs4_pnfs_ds *ds;
 825         int i;
 826         int status;
 827
 828         if (!pgio->pg_lseg)
 829                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 830                                                    req->wb_context,
 831                                                    0,
 832                                                    NFS4_MAX_UINT64,
 833                                                    IOMODE_RW,
 834                                                    GFP_NOFS);
 835         /* If no lseg, fall back to write through mds */
 836         if (pgio->pg_lseg == NULL)
 837                 goto out_mds;
 838
 839         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 840         status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 841         if (status < 0)
 842                 goto out_mds;
 843
 844         /* Use a direct mapping of ds_idx to pgio mirror_idx */
 845         if (WARN_ON_ONCE(pgio->pg_mirror_count !=
 846             FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)))
 847                 goto out_mds;
 848
 849         for (i = 0; i < pgio->pg_mirror_count; i++) {
 850                 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
 851                 if (!ds)
 852                         goto out_mds;
 853                 pgm = &pgio->pg_mirrors[i];
 854                 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
 855                 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
 856         }
 857
 858         return;
 859
 860 out_mds:
 861         pnfs_put_lseg(pgio->pg_lseg);
 862         pgio->pg_lseg = NULL;
 863         nfs_pageio_reset_write_mds(pgio);
 864 }
 865
 866 static unsigned int
 867 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
 868                                     struct nfs_page *req)
 869 {
 870         if (!pgio->pg_lseg)
 871                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 872                                                    req->wb_context,
 873                                                    0,
 874                                                    NFS4_MAX_UINT64,
 875                                                    IOMODE_RW,
 876                                                    GFP_NOFS);
 877         if (pgio->pg_lseg)
 878                 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
 879
 880         /* no lseg means that pnfs is not in use, so no mirroring here */
 881         nfs_pageio_reset_write_mds(pgio);
 882         return 1;
 883 }
 884
 885 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
 886         .pg_init = ff_layout_pg_init_read,
 887         .pg_test = pnfs_generic_pg_test,
 888         .pg_doio = pnfs_generic_pg_readpages,
 889         .pg_cleanup = pnfs_generic_pg_cleanup,
 890 };
 891
 892 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
 893         .pg_init = ff_layout_pg_init_write,
 894         .pg_test = pnfs_generic_pg_test,
 895         .pg_doio = pnfs_generic_pg_writepages,
 896         .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
 897         .pg_cleanup = pnfs_generic_pg_cleanup,
 898 };
 899
 900 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
 901 {
 902         struct rpc_task *task = &hdr->task;
 903
 904         pnfs_layoutcommit_inode(hdr->inode, false);
 905
 906         if (retry_pnfs) {
 907                 dprintk("%s Reset task %5u for i/o through pNFS "
 908                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 909                         hdr->task.tk_pid,
 910                         hdr->inode->i_sb->s_id,
 911                         (unsigned long long)NFS_FILEID(hdr->inode),
 912                         hdr->args.count,
 913                         (unsigned long long)hdr->args.offset);
 914
 915                 if (!hdr->dreq) {
 916                         struct nfs_open_context *ctx;
 917
 918                         ctx = nfs_list_entry(hdr->pages.next)->wb_context;
 919                         set_bit(NFS_CONTEXT_RESEND_WRITES, &ctx->flags);
 920                         hdr->completion_ops->error_cleanup(&hdr->pages);
 921                 } else {
 922                         nfs_direct_set_resched_writes(hdr->dreq);
 923                         /* fake unstable write to let common nfs resend pages */
 924                         hdr->verf.committed = NFS_UNSTABLE;
 925                         hdr->good_bytes = hdr->args.count;
 926                 }
 927                 return;
 928         }
 929
 930         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 931                 dprintk("%s Reset task %5u for i/o through MDS "
 932                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 933                         hdr->task.tk_pid,
 934                         hdr->inode->i_sb->s_id,
 935                         (unsigned long long)NFS_FILEID(hdr->inode),
 936                         hdr->args.count,
 937                         (unsigned long long)hdr->args.offset);
 938
 939                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
 940         }
 941 }
 942
 943 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
 944 {
 945         struct rpc_task *task = &hdr->task;
 946
 947         pnfs_layoutcommit_inode(hdr->inode, false);
 948
 949         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 950                 dprintk("%s Reset task %5u for i/o through MDS "
 951                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 952                         hdr->task.tk_pid,
 953                         hdr->inode->i_sb->s_id,
 954                         (unsigned long long)NFS_FILEID(hdr->inode),
 955                         hdr->args.count,
 956                         (unsigned long long)hdr->args.offset);
 957
 958                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
 959         }
 960 }
 961
 962 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
 963                                            struct nfs4_state *state,
 964                                            struct nfs_client *clp,
 965                                            struct pnfs_layout_segment *lseg,
 966                                            int idx)
 967 {
 968         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 969         struct inode *inode = lo->plh_inode;
 970         struct nfs_server *mds_server = NFS_SERVER(inode);
 971
 972         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
 973         struct nfs_client *mds_client = mds_server->nfs_client;
 974         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 975
 976         if (task->tk_status >= 0)
 977                 return 0;
 978
 979         switch (task->tk_status) {
 980         /* MDS state errors */
 981         case -NFS4ERR_DELEG_REVOKED:
 982         case -NFS4ERR_ADMIN_REVOKED:
 983         case -NFS4ERR_BAD_STATEID:
 984                 if (state == NULL)
 985                         break;
 986                 nfs_remove_bad_delegation(state->inode);
 987         case -NFS4ERR_OPENMODE:
 988                 if (state == NULL)
 989                         break;
 990                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 991                         goto out_bad_stateid;
 992                 goto wait_on_recovery;
 993         case -NFS4ERR_EXPIRED:
 994                 if (state != NULL) {
 995                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 996                                 goto out_bad_stateid;
 997                 }
 998                 nfs4_schedule_lease_recovery(mds_client);
 999                 goto wait_on_recovery;
1000         /* DS session errors */
1001         case -NFS4ERR_BADSESSION:
1002         case -NFS4ERR_BADSLOT:
1003         case -NFS4ERR_BAD_HIGH_SLOT:
1004         case -NFS4ERR_DEADSESSION:
1005         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1006         case -NFS4ERR_SEQ_FALSE_RETRY:
1007         case -NFS4ERR_SEQ_MISORDERED:
1008                 dprintk("%s ERROR %d, Reset session. Exchangeid "
1009                         "flags 0x%x\n", __func__, task->tk_status,
1010                         clp->cl_exchange_flags);
1011                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1012                 break;
1013         case -NFS4ERR_DELAY:
1014         case -NFS4ERR_GRACE:
1015                 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1016                 break;
1017         case -NFS4ERR_RETRY_UNCACHED_REP:
1018                 break;
1019         /* Invalidate Layout errors */
1020         case -NFS4ERR_PNFS_NO_LAYOUT:
1021         case -ESTALE:           /* mapped NFS4ERR_STALE */
1022         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
1023         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
1024         case -NFS4ERR_FHEXPIRED:
1025         case -NFS4ERR_WRONG_TYPE:
1026                 dprintk("%s Invalid layout error %d\n", __func__,
1027                         task->tk_status);
1028                 /*
1029                  * Destroy layout so new i/o will get a new layout.
1030                  * Layout will not be destroyed until all current lseg
1031                  * references are put. Mark layout as invalid to resend failed
1032                  * i/o and all i/o waiting on the slot table to the MDS until
1033                  * layout is destroyed and a new valid layout is obtained.
1034                  */
1035                 pnfs_destroy_layout(NFS_I(inode));
1036                 rpc_wake_up(&tbl->slot_tbl_waitq);
1037                 goto reset;
1038         /* RPC connection errors */
1039         case -ECONNREFUSED:
1040         case -EHOSTDOWN:
1041         case -EHOSTUNREACH:
1042         case -ENETUNREACH:
1043         case -EIO:
1044         case -ETIMEDOUT:
1045         case -EPIPE:
1046                 dprintk("%s DS connection error %d\n", __func__,
1047                         task->tk_status);
1048                 nfs4_mark_deviceid_unavailable(devid);
1049                 rpc_wake_up(&tbl->slot_tbl_waitq);
1050                 /* fall through */
1051         default:
1052                 if (ff_layout_no_fallback_to_mds(lseg) ||
1053                     ff_layout_has_available_ds(lseg))
1054                         return -NFS4ERR_RESET_TO_PNFS;
1055 reset:
1056                 dprintk("%s Retry through MDS. Error %d\n", __func__,
1057                         task->tk_status);
1058                 return -NFS4ERR_RESET_TO_MDS;
1059         }
1060 out:
1061         task->tk_status = 0;
1062         return -EAGAIN;
1063 out_bad_stateid:
1064         task->tk_status = -EIO;
1065         return 0;
1066 wait_on_recovery:
1067         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
1068         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
1069                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
1070         goto out;
1071 }
1072
1073 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
1074 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1075                                            struct pnfs_layout_segment *lseg,
1076                                            int idx)
1077 {
1078         struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1079
1080         if (task->tk_status >= 0)
1081                 return 0;
1082
1083         switch (task->tk_status) {
1084         /* File access problems. Don't mark the device as unavailable */
1085         case -EACCES:
1086         case -ESTALE:
1087         case -EISDIR:
1088         case -EBADHANDLE:
1089         case -ELOOP:
1090         case -ENOSPC:
1091                 break;
1092         case -EJUKEBOX:
1093                 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1094                 goto out_retry;
1095         default:
1096                 dprintk("%s DS connection error %d\n", __func__,
1097                         task->tk_status);
1098                 nfs4_mark_deviceid_unavailable(devid);
1099         }
1100         /* FIXME: Need to prevent infinite looping here. */
1101         return -NFS4ERR_RESET_TO_PNFS;
1102 out_retry:
1103         task->tk_status = 0;
1104         rpc_restart_call(task);
1105         rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
1106         return -EAGAIN;
1107 }
1108
1109 static int ff_layout_async_handle_error(struct rpc_task *task,
1110                                         struct nfs4_state *state,
1111                                         struct nfs_client *clp,
1112                                         struct pnfs_layout_segment *lseg,
1113                                         int idx)
1114 {
1115         int vers = clp->cl_nfs_mod->rpc_vers->number;
1116
1117         switch (vers) {
1118         case 3:
1119                 return ff_layout_async_handle_error_v3(task, lseg, idx);
1120         case 4:
1121                 return ff_layout_async_handle_error_v4(task, state, clp,
1122                                                        lseg, idx);
1123         default:
1124                 /* should never happen */
1125                 WARN_ON_ONCE(1);
1126                 return 0;
1127         }
1128 }
1129
1130 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1131                                         int idx, u64 offset, u64 length,
1132                                         u32 status, int opnum, int error)
1133 {
1134         struct nfs4_ff_layout_mirror *mirror;
1135         int err;
1136
1137         if (status == 0) {
1138                 switch (error) {
1139                 case -ETIMEDOUT:
1140                 case -EPFNOSUPPORT:
1141                 case -EPROTONOSUPPORT:
1142                 case -EOPNOTSUPP:
1143                 case -ECONNREFUSED:
1144                 case -ECONNRESET:
1145                 case -EHOSTDOWN:
1146                 case -EHOSTUNREACH:
1147                 case -ENETUNREACH:
1148                 case -EADDRINUSE:
1149                 case -ENOBUFS:
1150                 case -EPIPE:
1151                 case -EPERM:
1152                         status = NFS4ERR_NXIO;
1153                         break;
1154                 case -EACCES:
1155                         status = NFS4ERR_ACCESS;
1156                         break;
1157                 default:
1158                         return;
1159                 }
1160         }
1161
1162         mirror = FF_LAYOUT_COMP(lseg, idx);
1163         err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1164                                        mirror, offset, length, status, opnum,
1165                                        GFP_NOIO);
1166         pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
1167         dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1168 }
1169
1170 /* NFS_PROTO call done callback routines */
1171 static int ff_layout_read_done_cb(struct rpc_task *task,
1172                                 struct nfs_pgio_header *hdr)
1173 {
1174         int err;
1175
1176         trace_nfs4_pnfs_read(hdr, task->tk_status);
1177         if (task->tk_status < 0)
1178                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1179                                             hdr->args.offset, hdr->args.count,
1180                                             hdr->res.op_status, OP_READ,
1181                                             task->tk_status);
1182         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1183                                            hdr->ds_clp, hdr->lseg,
1184                                            hdr->pgio_mirror_idx);
1185
1186         switch (err) {
1187         case -NFS4ERR_RESET_TO_PNFS:
1188                 if (ff_layout_choose_best_ds_for_read(hdr->lseg,
1189                                         hdr->pgio_mirror_idx + 1,
1190                                         &hdr->pgio_mirror_idx))
1191                         goto out_eagain;
1192                 set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1193                         &hdr->lseg->pls_layout->plh_flags);
1194                 pnfs_read_resend_pnfs(hdr);
1195                 return task->tk_status;
1196         case -NFS4ERR_RESET_TO_MDS:
1197                 ff_layout_reset_read(hdr);
1198                 return task->tk_status;
1199         case -EAGAIN:
1200                 goto out_eagain;
1201         }
1202
1203         return 0;
1204 out_eagain:
1205         rpc_restart_call_prepare(task);
1206         return -EAGAIN;
1207 }
1208
1209 static bool
1210 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1211 {
1212         return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1213 }
1214
1215 /*
1216  * We reference the rpc_cred of the first WRITE that triggers the need for
1217  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1218  * rfc5661 is not clear about which credential should be used.
1219  *
1220  * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1221  * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1222  * we always send layoutcommit after DS writes.
1223  */
1224 static void
1225 ff_layout_set_layoutcommit(struct nfs_pgio_header *hdr)
1226 {
1227         if (!ff_layout_need_layoutcommit(hdr->lseg))
1228                 return;
1229
1230         pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1231                         hdr->mds_offset + hdr->res.count);
1232         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
1233                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
1234 }
1235
1236 static bool
1237 ff_layout_reset_to_mds(struct pnfs_layout_segment *lseg, int idx)
1238 {
1239         /* No mirroring for now */
1240         struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1241
1242         return ff_layout_test_devid_unavailable(node);
1243 }
1244
1245 static int ff_layout_read_prepare_common(struct rpc_task *task,
1246                                          struct nfs_pgio_header *hdr)
1247 {
1248         nfs4_ff_layout_stat_io_start_read(hdr->inode,
1249                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1250                         hdr->args.count,
1251                         task->tk_start);
1252
1253         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1254                 rpc_exit(task, -EIO);
1255                 return -EIO;
1256         }
1257         if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1258                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
1259                 if (ff_layout_has_available_ds(hdr->lseg))
1260                         pnfs_read_resend_pnfs(hdr);
1261                 else
1262                         ff_layout_reset_read(hdr);
1263                 rpc_exit(task, 0);
1264                 return -EAGAIN;
1265         }
1266         hdr->pgio_done_cb = ff_layout_read_done_cb;
1267
1268         return 0;
1269 }
1270
1271 /*
1272  * Call ops for the async read/write cases
1273  * In the case of dense layouts, the offset needs to be reset to its
1274  * original value.
1275  */
1276 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1277 {
1278         struct nfs_pgio_header *hdr = data;
1279
1280         if (ff_layout_read_prepare_common(task, hdr))
1281                 return;
1282
1283         rpc_call_start(task);
1284 }
1285
1286 static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1287                                     struct nfs4_sequence_args *args,
1288                                     struct nfs4_sequence_res *res,
1289                                     struct rpc_task *task)
1290 {
1291         if (ds_clp->cl_session)
1292                 return nfs41_setup_sequence(ds_clp->cl_session,
1293                                            args,
1294                                            res,
1295                                            task);
1296         return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1297                                    args,
1298                                    res,
1299                                    task);
1300 }
1301
1302 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1303 {
1304         struct nfs_pgio_header *hdr = data;
1305
1306         if (ff_layout_setup_sequence(hdr->ds_clp,
1307                                      &hdr->args.seq_args,
1308                                      &hdr->res.seq_res,
1309                                      task))
1310                 return;
1311
1312         if (ff_layout_read_prepare_common(task, hdr))
1313                 return;
1314
1315         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1316                         hdr->args.lock_context, FMODE_READ) == -EIO)
1317                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1318 }
1319
1320 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1321 {
1322         struct nfs_pgio_header *hdr = data;
1323
1324         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1325
1326         nfs4_ff_layout_stat_io_end_read(task,
1327                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1328                         hdr->args.count, hdr->res.count);
1329
1330         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1331             task->tk_status == 0) {
1332                 nfs4_sequence_done(task, &hdr->res.seq_res);
1333                 return;
1334         }
1335
1336         /* Note this may cause RPC to be resent */
1337         hdr->mds_ops->rpc_call_done(task, hdr);
1338 }
1339
1340 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1341 {
1342         struct nfs_pgio_header *hdr = data;
1343
1344         rpc_count_iostats_metrics(task,
1345             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1346 }
1347
1348 static int ff_layout_write_done_cb(struct rpc_task *task,
1349                                 struct nfs_pgio_header *hdr)
1350 {
1351         int err;
1352
1353         trace_nfs4_pnfs_write(hdr, task->tk_status);
1354         if (task->tk_status < 0)
1355                 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1356                                             hdr->args.offset, hdr->args.count,
1357                                             hdr->res.op_status, OP_WRITE,
1358                                             task->tk_status);
1359         err = ff_layout_async_handle_error(task, hdr->args.context->state,
1360                                            hdr->ds_clp, hdr->lseg,
1361                                            hdr->pgio_mirror_idx);
1362
1363         switch (err) {
1364         case -NFS4ERR_RESET_TO_PNFS:
1365                 pnfs_set_retry_layoutget(hdr->lseg->pls_layout);
1366                 ff_layout_reset_write(hdr, true);
1367                 return task->tk_status;
1368         case -NFS4ERR_RESET_TO_MDS:
1369                 pnfs_clear_retry_layoutget(hdr->lseg->pls_layout);
1370                 ff_layout_reset_write(hdr, false);
1371                 return task->tk_status;
1372         case -EAGAIN:
1373                 rpc_restart_call_prepare(task);
1374                 return -EAGAIN;
1375         }
1376
1377         if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1378             hdr->res.verf->committed == NFS_DATA_SYNC)
1379                 ff_layout_set_layoutcommit(hdr);
1380
1381         /* zero out fattr since we don't care DS attr at all */
1382         hdr->fattr.valid = 0;
1383         if (task->tk_status >= 0)
1384                 nfs_writeback_update_inode(hdr);
1385
1386         return 0;
1387 }
1388
1389 static int ff_layout_commit_done_cb(struct rpc_task *task,
1390                                      struct nfs_commit_data *data)
1391 {
1392         int err;
1393
1394         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1395         if (task->tk_status < 0)
1396                 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1397                                             data->args.offset, data->args.count,
1398                                             data->res.op_status, OP_COMMIT,
1399                                             task->tk_status);
1400         err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1401                                            data->lseg, data->ds_commit_index);
1402
1403         switch (err) {
1404         case -NFS4ERR_RESET_TO_PNFS:
1405                 pnfs_set_retry_layoutget(data->lseg->pls_layout);
1406                 pnfs_generic_prepare_to_resend_writes(data);
1407                 return -EAGAIN;
1408         case -NFS4ERR_RESET_TO_MDS:
1409                 pnfs_clear_retry_layoutget(data->lseg->pls_layout);
1410                 pnfs_generic_prepare_to_resend_writes(data);
1411                 return -EAGAIN;
1412         case -EAGAIN:
1413                 rpc_restart_call_prepare(task);
1414                 return -EAGAIN;
1415         }
1416
1417         if (data->verf.committed == NFS_UNSTABLE
1418             && ff_layout_need_layoutcommit(data->lseg))
1419                 pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
1420
1421         return 0;
1422 }
1423
1424 static int ff_layout_write_prepare_common(struct rpc_task *task,
1425                                           struct nfs_pgio_header *hdr)
1426 {
1427         nfs4_ff_layout_stat_io_start_write(hdr->inode,
1428                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1429                         hdr->args.count,
1430                         task->tk_start);
1431
1432         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1433                 rpc_exit(task, -EIO);
1434                 return -EIO;
1435         }
1436
1437         if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1438                 bool retry_pnfs;
1439
1440                 retry_pnfs = ff_layout_has_available_ds(hdr->lseg);
1441                 dprintk("%s task %u reset io to %s\n", __func__,
1442                         task->tk_pid, retry_pnfs ? "pNFS" : "MDS");
1443                 ff_layout_reset_write(hdr, retry_pnfs);
1444                 rpc_exit(task, 0);
1445                 return -EAGAIN;
1446         }
1447
1448         return 0;
1449 }
1450
1451 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1452 {
1453         struct nfs_pgio_header *hdr = data;
1454
1455         if (ff_layout_write_prepare_common(task, hdr))
1456                 return;
1457
1458         rpc_call_start(task);
1459 }
1460
1461 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1462 {
1463         struct nfs_pgio_header *hdr = data;
1464
1465         if (ff_layout_setup_sequence(hdr->ds_clp,
1466                                      &hdr->args.seq_args,
1467                                      &hdr->res.seq_res,
1468                                      task))
1469                 return;
1470
1471         if (ff_layout_write_prepare_common(task, hdr))
1472                 return;
1473
1474         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1475                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
1476                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1477 }
1478
1479 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1480 {
1481         struct nfs_pgio_header *hdr = data;
1482
1483         nfs4_ff_layout_stat_io_end_write(task,
1484                         FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1485                         hdr->args.count, hdr->res.count,
1486                         hdr->res.verf->committed);
1487
1488         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1489             task->tk_status == 0) {
1490                 nfs4_sequence_done(task, &hdr->res.seq_res);
1491                 return;
1492         }
1493
1494         /* Note this may cause RPC to be resent */
1495         hdr->mds_ops->rpc_call_done(task, hdr);
1496 }
1497
1498 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1499 {
1500         struct nfs_pgio_header *hdr = data;
1501
1502         rpc_count_iostats_metrics(task,
1503             &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1504 }
1505
1506 static void ff_layout_commit_prepare_common(struct rpc_task *task,
1507                 struct nfs_commit_data *cdata)
1508 {
1509         nfs4_ff_layout_stat_io_start_write(cdata->inode,
1510                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1511                         0, task->tk_start);
1512 }
1513
1514 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1515 {
1516         ff_layout_commit_prepare_common(task, data);
1517         rpc_call_start(task);
1518 }
1519
1520 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1521 {
1522         struct nfs_commit_data *wdata = data;
1523
1524         if (ff_layout_setup_sequence(wdata->ds_clp,
1525                                  &wdata->args.seq_args,
1526                                  &wdata->res.seq_res,
1527                                  task))
1528                 return;
1529         ff_layout_commit_prepare_common(task, data);
1530 }
1531
1532 static void ff_layout_commit_done(struct rpc_task *task, void *data)
1533 {
1534         struct nfs_commit_data *cdata = data;
1535         struct nfs_page *req;
1536         __u64 count = 0;
1537
1538         if (task->tk_status == 0) {
1539                 list_for_each_entry(req, &cdata->pages, wb_list)
1540                         count += req->wb_bytes;
1541         }
1542
1543         nfs4_ff_layout_stat_io_end_write(task,
1544                         FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1545                         count, count, NFS_FILE_SYNC);
1546
1547         pnfs_generic_write_commit_done(task, data);
1548 }
1549
1550 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1551 {
1552         struct nfs_commit_data *cdata = data;
1553
1554         rpc_count_iostats_metrics(task,
1555             &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1556 }
1557
1558 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1559         .rpc_call_prepare = ff_layout_read_prepare_v3,
1560         .rpc_call_done = ff_layout_read_call_done,
1561         .rpc_count_stats = ff_layout_read_count_stats,
1562         .rpc_release = pnfs_generic_rw_release,
1563 };
1564
1565 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1566         .rpc_call_prepare = ff_layout_read_prepare_v4,
1567         .rpc_call_done = ff_layout_read_call_done,
1568         .rpc_count_stats = ff_layout_read_count_stats,
1569         .rpc_release = pnfs_generic_rw_release,
1570 };
1571
1572 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1573         .rpc_call_prepare = ff_layout_write_prepare_v3,
1574         .rpc_call_done = ff_layout_write_call_done,
1575         .rpc_count_stats = ff_layout_write_count_stats,
1576         .rpc_release = pnfs_generic_rw_release,
1577 };
1578
1579 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1580         .rpc_call_prepare = ff_layout_write_prepare_v4,
1581         .rpc_call_done = ff_layout_write_call_done,
1582         .rpc_count_stats = ff_layout_write_count_stats,
1583         .rpc_release = pnfs_generic_rw_release,
1584 };
1585
1586 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1587         .rpc_call_prepare = ff_layout_commit_prepare_v3,
1588         .rpc_call_done = ff_layout_commit_done,
1589         .rpc_count_stats = ff_layout_commit_count_stats,
1590         .rpc_release = pnfs_generic_commit_release,
1591 };
1592
1593 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1594         .rpc_call_prepare = ff_layout_commit_prepare_v4,
1595         .rpc_call_done = ff_layout_commit_done,
1596         .rpc_count_stats = ff_layout_commit_count_stats,
1597         .rpc_release = pnfs_generic_commit_release,
1598 };
1599
1600 static enum pnfs_try_status
1601 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1602 {
1603         struct pnfs_layout_segment *lseg = hdr->lseg;
1604         struct nfs4_pnfs_ds *ds;
1605         struct rpc_clnt *ds_clnt;
1606         struct rpc_cred *ds_cred;
1607         loff_t offset = hdr->args.offset;
1608         u32 idx = hdr->pgio_mirror_idx;
1609         int vers;
1610         struct nfs_fh *fh;
1611
1612         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1613                 __func__, hdr->inode->i_ino,
1614                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1615
1616         ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1617         if (!ds)
1618                 goto out_failed;
1619
1620         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1621                                                    hdr->inode);
1622         if (IS_ERR(ds_clnt))
1623                 goto out_failed;
1624
1625         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1626         if (IS_ERR(ds_cred))
1627                 goto out_failed;
1628
1629         vers = nfs4_ff_layout_ds_version(lseg, idx);
1630
1631         dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1632                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1633
1634         atomic_inc(&ds->ds_clp->cl_count);
1635         hdr->ds_clp = ds->ds_clp;
1636         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1637         if (fh)
1638                 hdr->args.fh = fh;
1639         /*
1640          * Note that if we ever decide to split across DSes,
1641          * then we may need to handle dense-like offsets.
1642          */
1643         hdr->args.offset = offset;
1644         hdr->mds_offset = offset;
1645
1646         /* Perform an asynchronous read to ds */
1647         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1648                           vers == 3 ? &ff_layout_read_call_ops_v3 :
1649                                       &ff_layout_read_call_ops_v4,
1650                           0, RPC_TASK_SOFTCONN);
1651
1652         return PNFS_ATTEMPTED;
1653
1654 out_failed:
1655         if (ff_layout_has_available_ds(lseg))
1656                 return PNFS_TRY_AGAIN;
1657         return PNFS_NOT_ATTEMPTED;
1658 }
1659
1660 /* Perform async writes. */
1661 static enum pnfs_try_status
1662 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1663 {
1664         struct pnfs_layout_segment *lseg = hdr->lseg;
1665         struct nfs4_pnfs_ds *ds;
1666         struct rpc_clnt *ds_clnt;
1667         struct rpc_cred *ds_cred;
1668         loff_t offset = hdr->args.offset;
1669         int vers;
1670         struct nfs_fh *fh;
1671         int idx = hdr->pgio_mirror_idx;
1672
1673         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1674         if (!ds)
1675                 return PNFS_NOT_ATTEMPTED;
1676
1677         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1678                                                    hdr->inode);
1679         if (IS_ERR(ds_clnt))
1680                 return PNFS_NOT_ATTEMPTED;
1681
1682         ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1683         if (IS_ERR(ds_cred))
1684                 return PNFS_NOT_ATTEMPTED;
1685
1686         vers = nfs4_ff_layout_ds_version(lseg, idx);
1687
1688         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1689                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1690                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1691                 vers);
1692
1693         hdr->pgio_done_cb = ff_layout_write_done_cb;
1694         atomic_inc(&ds->ds_clp->cl_count);
1695         hdr->ds_clp = ds->ds_clp;
1696         hdr->ds_commit_idx = idx;
1697         fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1698         if (fh)
1699                 hdr->args.fh = fh;
1700
1701         /*
1702          * Note that if we ever decide to split across DSes,
1703          * then we may need to handle dense-like offsets.
1704          */
1705         hdr->args.offset = offset;
1706
1707         /* Perform an asynchronous write */
1708         nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1709                           vers == 3 ? &ff_layout_write_call_ops_v3 :
1710                                       &ff_layout_write_call_ops_v4,
1711                           sync, RPC_TASK_SOFTCONN);
1712         return PNFS_ATTEMPTED;
1713 }
1714
1715 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1716 {
1717         return i;
1718 }
1719
1720 static struct nfs_fh *
1721 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1722 {
1723         struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1724
1725         /* FIXME: Assume that there is only one NFS version available
1726          * for the DS.
1727          */
1728         return &flseg->mirror_array[i]->fh_versions[0];
1729 }
1730
1731 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1732 {
1733         struct pnfs_layout_segment *lseg = data->lseg;
1734         struct nfs4_pnfs_ds *ds;
1735         struct rpc_clnt *ds_clnt;
1736         struct rpc_cred *ds_cred;
1737         u32 idx;
1738         int vers;
1739         struct nfs_fh *fh;
1740
1741         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1742         ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1743         if (!ds)
1744                 goto out_err;
1745
1746         ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1747                                                    data->inode);
1748         if (IS_ERR(ds_clnt))
1749                 goto out_err;
1750
1751         ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1752         if (IS_ERR(ds_cred))
1753                 goto out_err;
1754
1755         vers = nfs4_ff_layout_ds_version(lseg, idx);
1756
1757         dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1758                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1759                 vers);
1760         data->commit_done_cb = ff_layout_commit_done_cb;
1761         data->cred = ds_cred;
1762         atomic_inc(&ds->ds_clp->cl_count);
1763         data->ds_clp = ds->ds_clp;
1764         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1765         if (fh)
1766                 data->args.fh = fh;
1767
1768         return nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1769                                    vers == 3 ? &ff_layout_commit_call_ops_v3 :
1770                                                &ff_layout_commit_call_ops_v4,
1771                                    how, RPC_TASK_SOFTCONN);
1772 out_err:
1773         pnfs_generic_prepare_to_resend_writes(data);
1774         pnfs_generic_commit_release(data);
1775         return -EAGAIN;
1776 }
1777
1778 static int
1779 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1780                            int how, struct nfs_commit_info *cinfo)
1781 {
1782         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1783                                             ff_layout_initiate_commit);
1784 }
1785
1786 static struct pnfs_ds_commit_info *
1787 ff_layout_get_ds_info(struct inode *inode)
1788 {
1789         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1790
1791         if (layout == NULL)
1792                 return NULL;
1793
1794         return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1795 }
1796
1797 static void
1798 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
1799 {
1800         nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1801                                                   id_node));
1802 }
1803
1804 static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1805                                   struct xdr_stream *xdr,
1806                                   const struct nfs4_layoutreturn_args *args)
1807 {
1808         struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1809         __be32 *start;
1810         int count = 0, ret = 0;
1811
1812         start = xdr_reserve_space(xdr, 4);
1813         if (unlikely(!start))
1814                 return -E2BIG;
1815
1816         /* This assume we always return _ALL_ layouts */
1817         spin_lock(&hdr->plh_inode->i_lock);
1818         ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1819         spin_unlock(&hdr->plh_inode->i_lock);
1820
1821         *start = cpu_to_be32(count);
1822
1823         return ret;
1824 }
1825
1826 /* report nothing for now */
1827 static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1828                                      struct xdr_stream *xdr,
1829                                      const struct nfs4_layoutreturn_args *args)
1830 {
1831         __be32 *p;
1832
1833         p = xdr_reserve_space(xdr, 4);
1834         if (likely(p))
1835                 *p = cpu_to_be32(0);
1836 }
1837
1838 static struct nfs4_deviceid_node *
1839 ff_layout_alloc_deviceid_node(struct nfs_server *server,
1840                               struct pnfs_device *pdev, gfp_t gfp_flags)
1841 {
1842         struct nfs4_ff_layout_ds *dsaddr;
1843
1844         dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
1845         if (!dsaddr)
1846                 return NULL;
1847         return &dsaddr->id_node;
1848 }
1849
1850 static void
1851 ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
1852                               struct xdr_stream *xdr,
1853                               const struct nfs4_layoutreturn_args *args)
1854 {
1855         struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
1856         __be32 *start;
1857
1858         dprintk("%s: Begin\n", __func__);
1859         start = xdr_reserve_space(xdr, 4);
1860         BUG_ON(!start);
1861
1862         ff_layout_encode_ioerr(flo, xdr, args);
1863         ff_layout_encode_iostats(flo, xdr, args);
1864
1865         *start = cpu_to_be32((xdr->p - start - 1) * 4);
1866         dprintk("%s: Return\n", __func__);
1867 }
1868
1869 static int
1870 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
1871 {
1872         const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
1873
1874         return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
1875 }
1876
1877 static size_t
1878 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
1879                           const int buflen)
1880 {
1881         const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
1882         const struct in6_addr *addr = &sin6->sin6_addr;
1883
1884         /*
1885          * RFC 4291, Section 2.2.2
1886          *
1887          * Shorthanded ANY address
1888          */
1889         if (ipv6_addr_any(addr))
1890                 return snprintf(buf, buflen, "::");
1891
1892         /*
1893          * RFC 4291, Section 2.2.2
1894          *
1895          * Shorthanded loopback address
1896          */
1897         if (ipv6_addr_loopback(addr))
1898                 return snprintf(buf, buflen, "::1");
1899
1900         /*
1901          * RFC 4291, Section 2.2.3
1902          *
1903          * Special presentation address format for mapped v4
1904          * addresses.
1905          */
1906         if (ipv6_addr_v4mapped(addr))
1907                 return snprintf(buf, buflen, "::ffff:%pI4",
1908                                         &addr->s6_addr32[3]);
1909
1910         /*
1911          * RFC 4291, Section 2.2.1
1912          */
1913         return snprintf(buf, buflen, "%pI6c", addr);
1914 }
1915
1916 /* Derived from rpc_sockaddr2uaddr */
1917 static void
1918 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
1919 {
1920         struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
1921         char portbuf[RPCBIND_MAXUADDRPLEN];
1922         char addrbuf[RPCBIND_MAXUADDRLEN];
1923         char *netid;
1924         unsigned short port;
1925         int len, netid_len;
1926         __be32 *p;
1927
1928         switch (sap->sa_family) {
1929         case AF_INET:
1930                 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
1931                         return;
1932                 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
1933                 netid = "tcp";
1934                 netid_len = 3;
1935                 break;
1936         case AF_INET6:
1937                 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
1938                         return;
1939                 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
1940                 netid = "tcp6";
1941                 netid_len = 4;
1942                 break;
1943         default:
1944                 /* we only support tcp and tcp6 */
1945                 WARN_ON_ONCE(1);
1946                 return;
1947         }
1948
1949         snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
1950         len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
1951
1952         p = xdr_reserve_space(xdr, 4 + netid_len);
1953         xdr_encode_opaque(p, netid, netid_len);
1954
1955         p = xdr_reserve_space(xdr, 4 + len);
1956         xdr_encode_opaque(p, addrbuf, len);
1957 }
1958
1959 static void
1960 ff_layout_encode_nfstime(struct xdr_stream *xdr,
1961                          ktime_t t)
1962 {
1963         struct timespec64 ts;
1964         __be32 *p;
1965
1966         p = xdr_reserve_space(xdr, 12);
1967         ts = ktime_to_timespec64(t);
1968         p = xdr_encode_hyper(p, ts.tv_sec);
1969         *p++ = cpu_to_be32(ts.tv_nsec);
1970 }
1971
1972 static void
1973 ff_layout_encode_io_latency(struct xdr_stream *xdr,
1974                             struct nfs4_ff_io_stat *stat)
1975 {
1976         __be32 *p;
1977
1978         p = xdr_reserve_space(xdr, 5 * 8);
1979         p = xdr_encode_hyper(p, stat->ops_requested);
1980         p = xdr_encode_hyper(p, stat->bytes_requested);
1981         p = xdr_encode_hyper(p, stat->ops_completed);
1982         p = xdr_encode_hyper(p, stat->bytes_completed);
1983         p = xdr_encode_hyper(p, stat->bytes_not_delivered);
1984         ff_layout_encode_nfstime(xdr, stat->total_busy_time);
1985         ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
1986 }
1987
1988 static void
1989 ff_layout_encode_layoutstats(struct xdr_stream *xdr,
1990                              struct nfs42_layoutstat_args *args,
1991                              struct nfs42_layoutstat_devinfo *devinfo)
1992 {
1993         struct nfs4_ff_layout_mirror *mirror = devinfo->layout_private;
1994         struct nfs4_pnfs_ds_addr *da;
1995         struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds;
1996         struct nfs_fh *fh = &mirror->fh_versions[0];
1997         __be32 *p, *start;
1998
1999         da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2000         dprintk("%s: DS %s: encoding address %s\n",
2001                 __func__, ds->ds_remotestr, da->da_remotestr);
2002         /* layoutupdate length */
2003         start = xdr_reserve_space(xdr, 4);
2004         /* netaddr4 */
2005         ff_layout_encode_netaddr(xdr, da);
2006         /* nfs_fh4 */
2007         p = xdr_reserve_space(xdr, 4 + fh->size);
2008         xdr_encode_opaque(p, fh->data, fh->size);
2009         /* ff_io_latency4 read */
2010         spin_lock(&mirror->lock);
2011         ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat);
2012         /* ff_io_latency4 write */
2013         ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat);
2014         spin_unlock(&mirror->lock);
2015         /* nfstime4 */
2016         ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time));
2017         /* bool */
2018         p = xdr_reserve_space(xdr, 4);
2019         *p = cpu_to_be32(false);
2020
2021         *start = cpu_to_be32((xdr->p - start - 1) * 4);
2022 }
2023
2024 static int
2025 ff_layout_mirror_prepare_stats(struct nfs42_layoutstat_args *args,
2026                                struct pnfs_layout_hdr *lo,
2027                                int dev_limit)
2028 {
2029         struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2030         struct nfs4_ff_layout_mirror *mirror;
2031         struct nfs4_deviceid_node *dev;
2032         struct nfs42_layoutstat_devinfo *devinfo;
2033         int i = 0;
2034
2035         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2036                 if (i >= dev_limit)
2037                         break;
2038                 if (!mirror->mirror_ds)
2039                         continue;
2040                 /* mirror refcount put in cleanup_layoutstats */
2041                 if (!atomic_inc_not_zero(&mirror->ref))
2042                         continue;
2043                 dev = &mirror->mirror_ds->id_node;
2044                 devinfo = &args->devinfo[i];
2045                 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE);
2046                 devinfo->offset = 0;
2047                 devinfo->length = NFS4_MAX_UINT64;
2048                 devinfo->read_count = mirror->read_stat.io_stat.ops_completed;
2049                 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed;
2050                 devinfo->write_count = mirror->write_stat.io_stat.ops_completed;
2051                 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed;
2052                 devinfo->layout_type = LAYOUT_FLEX_FILES;
2053                 devinfo->layoutstats_encode = ff_layout_encode_layoutstats;
2054                 devinfo->layout_private = mirror;
2055
2056                 i++;
2057         }
2058         return i;
2059 }
2060
2061 static int
2062 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2063 {
2064         struct nfs4_flexfile_layout *ff_layout;
2065         struct nfs4_ff_layout_mirror *mirror;
2066         int dev_count = 0;
2067
2068         spin_lock(&args->inode->i_lock);
2069         ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout);
2070         list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2071                 if (atomic_read(&mirror->ref) != 0)
2072                         dev_count ++;
2073         }
2074         spin_unlock(&args->inode->i_lock);
2075         /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2076         if (dev_count > PNFS_LAYOUTSTATS_MAXDEV) {
2077                 dprintk("%s: truncating devinfo to limit (%d:%d)\n",
2078                         __func__, dev_count, PNFS_LAYOUTSTATS_MAXDEV);
2079                 dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2080         }
2081         args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO);
2082         if (!args->devinfo)
2083                 return -ENOMEM;
2084
2085         spin_lock(&args->inode->i_lock);
2086         args->num_dev = ff_layout_mirror_prepare_stats(args,
2087                         &ff_layout->generic_hdr, dev_count);
2088         spin_unlock(&args->inode->i_lock);
2089
2090         return 0;
2091 }
2092
2093 static void
2094 ff_layout_cleanup_layoutstats(struct nfs42_layoutstat_data *data)
2095 {
2096         struct nfs4_ff_layout_mirror *mirror;
2097         int i;
2098
2099         for (i = 0; i < data->args.num_dev; i++) {
2100                 mirror = data->args.devinfo[i].layout_private;
2101                 data->args.devinfo[i].layout_private = NULL;
2102                 ff_layout_put_mirror(mirror);
2103         }
2104 }
2105
2106 static struct pnfs_layoutdriver_type flexfilelayout_type = {
2107         .id                     = LAYOUT_FLEX_FILES,
2108         .name                   = "LAYOUT_FLEX_FILES",
2109         .owner                  = THIS_MODULE,
2110         .alloc_layout_hdr       = ff_layout_alloc_layout_hdr,
2111         .free_layout_hdr        = ff_layout_free_layout_hdr,
2112         .alloc_lseg             = ff_layout_alloc_lseg,
2113         .free_lseg              = ff_layout_free_lseg,
2114         .add_lseg               = ff_layout_add_lseg,
2115         .pg_read_ops            = &ff_layout_pg_read_ops,
2116         .pg_write_ops           = &ff_layout_pg_write_ops,
2117         .get_ds_info            = ff_layout_get_ds_info,
2118         .free_deviceid_node     = ff_layout_free_deviceid_node,
2119         .mark_request_commit    = pnfs_layout_mark_request_commit,
2120         .clear_request_commit   = pnfs_generic_clear_request_commit,
2121         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
2122         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
2123         .commit_pagelist        = ff_layout_commit_pagelist,
2124         .read_pagelist          = ff_layout_read_pagelist,
2125         .write_pagelist         = ff_layout_write_pagelist,
2126         .alloc_deviceid_node    = ff_layout_alloc_deviceid_node,
2127         .encode_layoutreturn    = ff_layout_encode_layoutreturn,
2128         .sync                   = pnfs_nfs_generic_sync,
2129         .prepare_layoutstats    = ff_layout_prepare_layoutstats,
2130         .cleanup_layoutstats    = ff_layout_cleanup_layoutstats,
2131 };
2132
2133 static int __init nfs4flexfilelayout_init(void)
2134 {
2135         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
2136                __func__);
2137         return pnfs_register_layoutdriver(&flexfilelayout_type);
2138 }
2139
2140 static void __exit nfs4flexfilelayout_exit(void)
2141 {
2142         printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
2143                __func__);
2144         pnfs_unregister_layoutdriver(&flexfilelayout_type);
2145 }
2146
2147 MODULE_ALIAS("nfs-layouttype4-4");
2148
2149 MODULE_LICENSE("GPL");
2150 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
2151
2152 module_init(nfs4flexfilelayout_init);
2153 module_exit(nfs4flexfilelayout_exit);