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 #define FF_LAYOUTRETURN_MAXERR 20 29 30 31 static struct group_info *ff_zero_group; 32 33 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task, 34 struct nfs_pgio_header *hdr); 35 static int ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo, 36 struct nfs42_layoutstat_devinfo *devinfo, 37 int dev_limit); 38 static void ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr, 39 const struct nfs42_layoutstat_devinfo *devinfo, 40 struct nfs4_ff_layout_mirror *mirror); 41 42 static struct pnfs_layout_hdr * 43 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) 44 { 45 struct nfs4_flexfile_layout *ffl; 46 47 ffl = kzalloc(sizeof(*ffl), gfp_flags); 48 if (ffl) { 49 INIT_LIST_HEAD(&ffl->error_list); 50 INIT_LIST_HEAD(&ffl->mirrors); 51 ffl->last_report_time = ktime_get(); 52 return &ffl->generic_hdr; 53 } else 54 return NULL; 55 } 56 57 static void 58 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo) 59 { 60 struct nfs4_ff_layout_ds_err *err, *n; 61 62 list_for_each_entry_safe(err, n, &FF_LAYOUT_FROM_HDR(lo)->error_list, 63 list) { 64 list_del(&err->list); 65 kfree(err); 66 } 67 kfree(FF_LAYOUT_FROM_HDR(lo)); 68 } 69 70 static int decode_pnfs_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid) 71 { 72 __be32 *p; 73 74 p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE); 75 if (unlikely(p == NULL)) 76 return -ENOBUFS; 77 stateid->type = NFS4_PNFS_DS_STATEID_TYPE; 78 memcpy(stateid->data, p, NFS4_STATEID_SIZE); 79 dprintk("%s: stateid id= [%x%x%x%x]\n", __func__, 80 p[0], p[1], p[2], p[3]); 81 return 0; 82 } 83 84 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid) 85 { 86 __be32 *p; 87 88 p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE); 89 if (unlikely(!p)) 90 return -ENOBUFS; 91 memcpy(devid, p, NFS4_DEVICEID4_SIZE); 92 nfs4_print_deviceid(devid); 93 return 0; 94 } 95 96 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh) 97 { 98 __be32 *p; 99 100 p = xdr_inline_decode(xdr, 4); 101 if (unlikely(!p)) 102 return -ENOBUFS; 103 fh->size = be32_to_cpup(p++); 104 if (fh->size > sizeof(struct nfs_fh)) { 105 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n", 106 fh->size); 107 return -EOVERFLOW; 108 } 109 /* fh.data */ 110 p = xdr_inline_decode(xdr, fh->size); 111 if (unlikely(!p)) 112 return -ENOBUFS; 113 memcpy(&fh->data, p, fh->size); 114 dprintk("%s: fh len %d\n", __func__, fh->size); 115 116 return 0; 117 } 118 119 /* 120 * Currently only stringified uids and gids are accepted. 121 * I.e., kerberos is not supported to the DSes, so no pricipals. 122 * 123 * That means that one common function will suffice, but when 124 * principals are added, this should be split to accomodate 125 * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid(). 126 */ 127 static int 128 decode_name(struct xdr_stream *xdr, u32 *id) 129 { 130 __be32 *p; 131 int len; 132 133 /* opaque_length(4)*/ 134 p = xdr_inline_decode(xdr, 4); 135 if (unlikely(!p)) 136 return -ENOBUFS; 137 len = be32_to_cpup(p++); 138 if (len < 0) 139 return -EINVAL; 140 141 dprintk("%s: len %u\n", __func__, len); 142 143 /* opaque body */ 144 p = xdr_inline_decode(xdr, len); 145 if (unlikely(!p)) 146 return -ENOBUFS; 147 148 if (!nfs_map_string_to_numeric((char *)p, len, id)) 149 return -EINVAL; 150 151 return 0; 152 } 153 154 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1, 155 const struct nfs4_ff_layout_mirror *m2) 156 { 157 int i, j; 158 159 if (m1->fh_versions_cnt != m2->fh_versions_cnt) 160 return false; 161 for (i = 0; i < m1->fh_versions_cnt; i++) { 162 bool found_fh = false; 163 for (j = 0; j < m2->fh_versions_cnt; j++) { 164 if (nfs_compare_fh(&m1->fh_versions[i], 165 &m2->fh_versions[j]) == 0) { 166 found_fh = true; 167 break; 168 } 169 } 170 if (!found_fh) 171 return false; 172 } 173 return true; 174 } 175 176 static struct nfs4_ff_layout_mirror * 177 ff_layout_add_mirror(struct pnfs_layout_hdr *lo, 178 struct nfs4_ff_layout_mirror *mirror) 179 { 180 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo); 181 struct nfs4_ff_layout_mirror *pos; 182 struct inode *inode = lo->plh_inode; 183 184 spin_lock(&inode->i_lock); 185 list_for_each_entry(pos, &ff_layout->mirrors, mirrors) { 186 if (memcmp(&mirror->devid, &pos->devid, sizeof(pos->devid)) != 0) 187 continue; 188 if (!ff_mirror_match_fh(mirror, pos)) 189 continue; 190 if (atomic_inc_not_zero(&pos->ref)) { 191 spin_unlock(&inode->i_lock); 192 return pos; 193 } 194 } 195 list_add(&mirror->mirrors, &ff_layout->mirrors); 196 mirror->layout = lo; 197 spin_unlock(&inode->i_lock); 198 return mirror; 199 } 200 201 static void 202 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror) 203 { 204 struct inode *inode; 205 if (mirror->layout == NULL) 206 return; 207 inode = mirror->layout->plh_inode; 208 spin_lock(&inode->i_lock); 209 list_del(&mirror->mirrors); 210 spin_unlock(&inode->i_lock); 211 mirror->layout = NULL; 212 } 213 214 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(gfp_t gfp_flags) 215 { 216 struct nfs4_ff_layout_mirror *mirror; 217 218 mirror = kzalloc(sizeof(*mirror), gfp_flags); 219 if (mirror != NULL) { 220 spin_lock_init(&mirror->lock); 221 atomic_set(&mirror->ref, 1); 222 INIT_LIST_HEAD(&mirror->mirrors); 223 } 224 return mirror; 225 } 226 227 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror) 228 { 229 struct rpc_cred *cred; 230 231 ff_layout_remove_mirror(mirror); 232 kfree(mirror->fh_versions); 233 cred = rcu_access_pointer(mirror->ro_cred); 234 if (cred) 235 put_rpccred(cred); 236 cred = rcu_access_pointer(mirror->rw_cred); 237 if (cred) 238 put_rpccred(cred); 239 nfs4_ff_layout_put_deviceid(mirror->mirror_ds); 240 kfree(mirror); 241 } 242 243 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror) 244 { 245 if (mirror != NULL && atomic_dec_and_test(&mirror->ref)) 246 ff_layout_free_mirror(mirror); 247 } 248 249 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls) 250 { 251 int i; 252 253 if (fls->mirror_array) { 254 for (i = 0; i < fls->mirror_array_cnt; i++) { 255 /* normally mirror_ds is freed in 256 * .free_deviceid_node but we still do it here 257 * for .alloc_lseg error path */ 258 ff_layout_put_mirror(fls->mirror_array[i]); 259 } 260 kfree(fls->mirror_array); 261 fls->mirror_array = NULL; 262 } 263 } 264 265 static int ff_layout_check_layout(struct nfs4_layoutget_res *lgr) 266 { 267 int ret = 0; 268 269 dprintk("--> %s\n", __func__); 270 271 /* FIXME: remove this check when layout segment support is added */ 272 if (lgr->range.offset != 0 || 273 lgr->range.length != NFS4_MAX_UINT64) { 274 dprintk("%s Only whole file layouts supported. Use MDS i/o\n", 275 __func__); 276 ret = -EINVAL; 277 } 278 279 dprintk("--> %s returns %d\n", __func__, ret); 280 return ret; 281 } 282 283 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls) 284 { 285 if (fls) { 286 ff_layout_free_mirror_array(fls); 287 kfree(fls); 288 } 289 } 290 291 static bool 292 ff_lseg_range_is_after(const struct pnfs_layout_range *l1, 293 const struct pnfs_layout_range *l2) 294 { 295 u64 end1, end2; 296 297 if (l1->iomode != l2->iomode) 298 return l1->iomode != IOMODE_READ; 299 end1 = pnfs_calc_offset_end(l1->offset, l1->length); 300 end2 = pnfs_calc_offset_end(l2->offset, l2->length); 301 if (end1 < l2->offset) 302 return false; 303 if (end2 < l1->offset) 304 return true; 305 return l2->offset <= l1->offset; 306 } 307 308 static bool 309 ff_lseg_merge(struct pnfs_layout_segment *new, 310 struct pnfs_layout_segment *old) 311 { 312 u64 new_end, old_end; 313 314 if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags)) 315 return false; 316 if (new->pls_range.iomode != old->pls_range.iomode) 317 return false; 318 old_end = pnfs_calc_offset_end(old->pls_range.offset, 319 old->pls_range.length); 320 if (old_end < new->pls_range.offset) 321 return false; 322 new_end = pnfs_calc_offset_end(new->pls_range.offset, 323 new->pls_range.length); 324 if (new_end < old->pls_range.offset) 325 return false; 326 327 /* Mergeable: copy info from 'old' to 'new' */ 328 if (new_end < old_end) 329 new_end = old_end; 330 if (new->pls_range.offset < old->pls_range.offset) 331 new->pls_range.offset = old->pls_range.offset; 332 new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset, 333 new_end); 334 if (test_bit(NFS_LSEG_ROC, &old->pls_flags)) 335 set_bit(NFS_LSEG_ROC, &new->pls_flags); 336 return true; 337 } 338 339 static void 340 ff_layout_add_lseg(struct pnfs_layout_hdr *lo, 341 struct pnfs_layout_segment *lseg, 342 struct list_head *free_me) 343 { 344 pnfs_generic_layout_insert_lseg(lo, lseg, 345 ff_lseg_range_is_after, 346 ff_lseg_merge, 347 free_me); 348 } 349 350 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls) 351 { 352 int i, j; 353 354 for (i = 0; i < fls->mirror_array_cnt - 1; i++) { 355 for (j = i + 1; j < fls->mirror_array_cnt; j++) 356 if (fls->mirror_array[i]->efficiency < 357 fls->mirror_array[j]->efficiency) 358 swap(fls->mirror_array[i], 359 fls->mirror_array[j]); 360 } 361 } 362 363 static struct pnfs_layout_segment * 364 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh, 365 struct nfs4_layoutget_res *lgr, 366 gfp_t gfp_flags) 367 { 368 struct pnfs_layout_segment *ret; 369 struct nfs4_ff_layout_segment *fls = NULL; 370 struct xdr_stream stream; 371 struct xdr_buf buf; 372 struct page *scratch; 373 u64 stripe_unit; 374 u32 mirror_array_cnt; 375 __be32 *p; 376 int i, rc; 377 378 dprintk("--> %s\n", __func__); 379 scratch = alloc_page(gfp_flags); 380 if (!scratch) 381 return ERR_PTR(-ENOMEM); 382 383 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, 384 lgr->layoutp->len); 385 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE); 386 387 /* stripe unit and mirror_array_cnt */ 388 rc = -EIO; 389 p = xdr_inline_decode(&stream, 8 + 4); 390 if (!p) 391 goto out_err_free; 392 393 p = xdr_decode_hyper(p, &stripe_unit); 394 mirror_array_cnt = be32_to_cpup(p++); 395 dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__, 396 stripe_unit, mirror_array_cnt); 397 398 if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT || 399 mirror_array_cnt == 0) 400 goto out_err_free; 401 402 rc = -ENOMEM; 403 fls = kzalloc(sizeof(*fls), gfp_flags); 404 if (!fls) 405 goto out_err_free; 406 407 fls->mirror_array_cnt = mirror_array_cnt; 408 fls->stripe_unit = stripe_unit; 409 fls->mirror_array = kcalloc(fls->mirror_array_cnt, 410 sizeof(fls->mirror_array[0]), gfp_flags); 411 if (fls->mirror_array == NULL) 412 goto out_err_free; 413 414 for (i = 0; i < fls->mirror_array_cnt; i++) { 415 struct nfs4_ff_layout_mirror *mirror; 416 struct auth_cred acred = { .group_info = ff_zero_group }; 417 struct rpc_cred __rcu *cred; 418 u32 ds_count, fh_count, id; 419 int j; 420 421 rc = -EIO; 422 p = xdr_inline_decode(&stream, 4); 423 if (!p) 424 goto out_err_free; 425 ds_count = be32_to_cpup(p); 426 427 /* FIXME: allow for striping? */ 428 if (ds_count != 1) 429 goto out_err_free; 430 431 fls->mirror_array[i] = ff_layout_alloc_mirror(gfp_flags); 432 if (fls->mirror_array[i] == NULL) { 433 rc = -ENOMEM; 434 goto out_err_free; 435 } 436 437 fls->mirror_array[i]->ds_count = ds_count; 438 439 /* deviceid */ 440 rc = decode_deviceid(&stream, &fls->mirror_array[i]->devid); 441 if (rc) 442 goto out_err_free; 443 444 /* efficiency */ 445 rc = -EIO; 446 p = xdr_inline_decode(&stream, 4); 447 if (!p) 448 goto out_err_free; 449 fls->mirror_array[i]->efficiency = be32_to_cpup(p); 450 451 /* stateid */ 452 rc = decode_pnfs_stateid(&stream, &fls->mirror_array[i]->stateid); 453 if (rc) 454 goto out_err_free; 455 456 /* fh */ 457 p = xdr_inline_decode(&stream, 4); 458 if (!p) 459 goto out_err_free; 460 fh_count = be32_to_cpup(p); 461 462 fls->mirror_array[i]->fh_versions = 463 kzalloc(fh_count * sizeof(struct nfs_fh), 464 gfp_flags); 465 if (fls->mirror_array[i]->fh_versions == NULL) { 466 rc = -ENOMEM; 467 goto out_err_free; 468 } 469 470 for (j = 0; j < fh_count; j++) { 471 rc = decode_nfs_fh(&stream, 472 &fls->mirror_array[i]->fh_versions[j]); 473 if (rc) 474 goto out_err_free; 475 } 476 477 fls->mirror_array[i]->fh_versions_cnt = fh_count; 478 479 /* user */ 480 rc = decode_name(&stream, &id); 481 if (rc) 482 goto out_err_free; 483 484 acred.uid = make_kuid(&init_user_ns, id); 485 486 /* group */ 487 rc = decode_name(&stream, &id); 488 if (rc) 489 goto out_err_free; 490 491 acred.gid = make_kgid(&init_user_ns, id); 492 493 /* find the cred for it */ 494 rcu_assign_pointer(cred, rpc_lookup_generic_cred(&acred, 0, gfp_flags)); 495 if (IS_ERR(cred)) { 496 rc = PTR_ERR(cred); 497 goto out_err_free; 498 } 499 500 if (lgr->range.iomode == IOMODE_READ) 501 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred); 502 else 503 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred); 504 505 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]); 506 if (mirror != fls->mirror_array[i]) { 507 /* swap cred ptrs so free_mirror will clean up old */ 508 if (lgr->range.iomode == IOMODE_READ) { 509 cred = xchg(&mirror->ro_cred, cred); 510 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred); 511 } else { 512 cred = xchg(&mirror->rw_cred, cred); 513 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred); 514 } 515 ff_layout_free_mirror(fls->mirror_array[i]); 516 fls->mirror_array[i] = mirror; 517 } 518 519 dprintk("%s: iomode %s uid %u gid %u\n", __func__, 520 lgr->range.iomode == IOMODE_READ ? "READ" : "RW", 521 from_kuid(&init_user_ns, acred.uid), 522 from_kgid(&init_user_ns, acred.gid)); 523 } 524 525 p = xdr_inline_decode(&stream, 4); 526 if (!p) 527 goto out_sort_mirrors; 528 fls->flags = be32_to_cpup(p); 529 530 p = xdr_inline_decode(&stream, 4); 531 if (!p) 532 goto out_sort_mirrors; 533 for (i=0; i < fls->mirror_array_cnt; i++) 534 fls->mirror_array[i]->report_interval = be32_to_cpup(p); 535 536 out_sort_mirrors: 537 ff_layout_sort_mirrors(fls); 538 rc = ff_layout_check_layout(lgr); 539 if (rc) 540 goto out_err_free; 541 ret = &fls->generic_hdr; 542 dprintk("<-- %s (success)\n", __func__); 543 out_free_page: 544 __free_page(scratch); 545 return ret; 546 out_err_free: 547 _ff_layout_free_lseg(fls); 548 ret = ERR_PTR(rc); 549 dprintk("<-- %s (%d)\n", __func__, rc); 550 goto out_free_page; 551 } 552 553 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout) 554 { 555 struct pnfs_layout_segment *lseg; 556 557 list_for_each_entry(lseg, &layout->plh_segs, pls_list) 558 if (lseg->pls_range.iomode == IOMODE_RW) 559 return true; 560 561 return false; 562 } 563 564 static void 565 ff_layout_free_lseg(struct pnfs_layout_segment *lseg) 566 { 567 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 568 569 dprintk("--> %s\n", __func__); 570 571 if (lseg->pls_range.iomode == IOMODE_RW) { 572 struct nfs4_flexfile_layout *ffl; 573 struct inode *inode; 574 575 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout); 576 inode = ffl->generic_hdr.plh_inode; 577 spin_lock(&inode->i_lock); 578 if (!ff_layout_has_rw_segments(lseg->pls_layout)) { 579 ffl->commit_info.nbuckets = 0; 580 kfree(ffl->commit_info.buckets); 581 ffl->commit_info.buckets = NULL; 582 } 583 spin_unlock(&inode->i_lock); 584 } 585 _ff_layout_free_lseg(fls); 586 } 587 588 /* Return 1 until we have multiple lsegs support */ 589 static int 590 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls) 591 { 592 return 1; 593 } 594 595 static void 596 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now) 597 { 598 /* first IO request? */ 599 if (atomic_inc_return(&timer->n_ops) == 1) { 600 timer->start_time = now; 601 } 602 } 603 604 static ktime_t 605 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now) 606 { 607 ktime_t start; 608 609 if (atomic_dec_return(&timer->n_ops) < 0) 610 WARN_ON_ONCE(1); 611 612 start = timer->start_time; 613 timer->start_time = now; 614 return ktime_sub(now, start); 615 } 616 617 static bool 618 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror, 619 struct nfs4_ff_layoutstat *layoutstat, 620 ktime_t now) 621 { 622 s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL; 623 struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(mirror->layout); 624 625 nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now); 626 if (!mirror->start_time) 627 mirror->start_time = now; 628 if (mirror->report_interval != 0) 629 report_interval = (s64)mirror->report_interval * 1000LL; 630 else if (layoutstats_timer != 0) 631 report_interval = (s64)layoutstats_timer * 1000LL; 632 if (ktime_to_ms(ktime_sub(now, ffl->last_report_time)) >= 633 report_interval) { 634 ffl->last_report_time = now; 635 return true; 636 } 637 638 return false; 639 } 640 641 static void 642 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat, 643 __u64 requested) 644 { 645 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat; 646 647 iostat->ops_requested++; 648 iostat->bytes_requested += requested; 649 } 650 651 static void 652 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat, 653 __u64 requested, 654 __u64 completed, 655 ktime_t time_completed, 656 ktime_t time_started) 657 { 658 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat; 659 ktime_t completion_time = ktime_sub(time_completed, time_started); 660 ktime_t timer; 661 662 iostat->ops_completed++; 663 iostat->bytes_completed += completed; 664 iostat->bytes_not_delivered += requested - completed; 665 666 timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed); 667 iostat->total_busy_time = 668 ktime_add(iostat->total_busy_time, timer); 669 iostat->aggregate_completion_time = 670 ktime_add(iostat->aggregate_completion_time, 671 completion_time); 672 } 673 674 static void 675 nfs4_ff_layout_stat_io_start_read(struct inode *inode, 676 struct nfs4_ff_layout_mirror *mirror, 677 __u64 requested, ktime_t now) 678 { 679 bool report; 680 681 spin_lock(&mirror->lock); 682 report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now); 683 nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested); 684 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 685 spin_unlock(&mirror->lock); 686 687 if (report) 688 pnfs_report_layoutstat(inode, GFP_KERNEL); 689 } 690 691 static void 692 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task, 693 struct nfs4_ff_layout_mirror *mirror, 694 __u64 requested, 695 __u64 completed) 696 { 697 spin_lock(&mirror->lock); 698 nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat, 699 requested, completed, 700 ktime_get(), task->tk_start); 701 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 702 spin_unlock(&mirror->lock); 703 } 704 705 static void 706 nfs4_ff_layout_stat_io_start_write(struct inode *inode, 707 struct nfs4_ff_layout_mirror *mirror, 708 __u64 requested, ktime_t now) 709 { 710 bool report; 711 712 spin_lock(&mirror->lock); 713 report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now); 714 nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested); 715 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 716 spin_unlock(&mirror->lock); 717 718 if (report) 719 pnfs_report_layoutstat(inode, GFP_NOIO); 720 } 721 722 static void 723 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task, 724 struct nfs4_ff_layout_mirror *mirror, 725 __u64 requested, 726 __u64 completed, 727 enum nfs3_stable_how committed) 728 { 729 if (committed == NFS_UNSTABLE) 730 requested = completed = 0; 731 732 spin_lock(&mirror->lock); 733 nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat, 734 requested, completed, ktime_get(), task->tk_start); 735 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 736 spin_unlock(&mirror->lock); 737 } 738 739 static int 740 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg, 741 struct nfs_commit_info *cinfo, 742 gfp_t gfp_flags) 743 { 744 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 745 struct pnfs_commit_bucket *buckets; 746 int size; 747 748 if (cinfo->ds->nbuckets != 0) { 749 /* This assumes there is only one RW lseg per file. 750 * To support multiple lseg per file, we need to 751 * change struct pnfs_commit_bucket to allow dynamic 752 * increasing nbuckets. 753 */ 754 return 0; 755 } 756 757 size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg); 758 759 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), 760 gfp_flags); 761 if (!buckets) 762 return -ENOMEM; 763 else { 764 int i; 765 766 spin_lock(&cinfo->inode->i_lock); 767 if (cinfo->ds->nbuckets != 0) 768 kfree(buckets); 769 else { 770 cinfo->ds->buckets = buckets; 771 cinfo->ds->nbuckets = size; 772 for (i = 0; i < size; i++) { 773 INIT_LIST_HEAD(&buckets[i].written); 774 INIT_LIST_HEAD(&buckets[i].committing); 775 /* mark direct verifier as unset */ 776 buckets[i].direct_verf.committed = 777 NFS_INVALID_STABLE_HOW; 778 } 779 } 780 spin_unlock(&cinfo->inode->i_lock); 781 return 0; 782 } 783 } 784 785 static struct nfs4_pnfs_ds * 786 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg, 787 int start_idx, 788 int *best_idx) 789 { 790 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 791 struct nfs4_pnfs_ds *ds; 792 bool fail_return = false; 793 int idx; 794 795 /* mirrors are sorted by efficiency */ 796 for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) { 797 if (idx+1 == fls->mirror_array_cnt) 798 fail_return = true; 799 ds = nfs4_ff_layout_prepare_ds(lseg, idx, fail_return); 800 if (ds) { 801 *best_idx = idx; 802 return ds; 803 } 804 } 805 806 return NULL; 807 } 808 809 static void 810 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio, 811 struct nfs_page *req, 812 bool strict_iomode) 813 { 814 retry_strict: 815 pnfs_put_lseg(pgio->pg_lseg); 816 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 817 req->wb_context, 818 0, 819 NFS4_MAX_UINT64, 820 IOMODE_READ, 821 strict_iomode, 822 GFP_KERNEL); 823 if (IS_ERR(pgio->pg_lseg)) { 824 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 825 pgio->pg_lseg = NULL; 826 } 827 828 /* If we don't have checking, do get a IOMODE_RW 829 * segment, and the server wants to avoid READs 830 * there, then retry! 831 */ 832 if (pgio->pg_lseg && !strict_iomode && 833 ff_layout_avoid_read_on_rw(pgio->pg_lseg)) { 834 strict_iomode = true; 835 goto retry_strict; 836 } 837 } 838 839 static void 840 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio, 841 struct nfs_page *req) 842 { 843 struct nfs_pgio_mirror *pgm; 844 struct nfs4_ff_layout_mirror *mirror; 845 struct nfs4_pnfs_ds *ds; 846 int ds_idx; 847 848 retry: 849 pnfs_generic_pg_check_layout(pgio); 850 /* Use full layout for now */ 851 if (!pgio->pg_lseg) 852 ff_layout_pg_get_read(pgio, req, false); 853 else if (ff_layout_avoid_read_on_rw(pgio->pg_lseg)) 854 ff_layout_pg_get_read(pgio, req, true); 855 856 /* If no lseg, fall back to read through mds */ 857 if (pgio->pg_lseg == NULL) 858 goto out_mds; 859 860 ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx); 861 if (!ds) { 862 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg)) 863 goto out_mds; 864 pnfs_put_lseg(pgio->pg_lseg); 865 pgio->pg_lseg = NULL; 866 /* Sleep for 1 second before retrying */ 867 ssleep(1); 868 goto retry; 869 } 870 871 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx); 872 873 pgio->pg_mirror_idx = ds_idx; 874 875 /* read always uses only one mirror - idx 0 for pgio layer */ 876 pgm = &pgio->pg_mirrors[0]; 877 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize; 878 879 return; 880 out_mds: 881 pnfs_put_lseg(pgio->pg_lseg); 882 pgio->pg_lseg = NULL; 883 nfs_pageio_reset_read_mds(pgio); 884 } 885 886 static void 887 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio, 888 struct nfs_page *req) 889 { 890 struct nfs4_ff_layout_mirror *mirror; 891 struct nfs_pgio_mirror *pgm; 892 struct nfs_commit_info cinfo; 893 struct nfs4_pnfs_ds *ds; 894 int i; 895 int status; 896 897 retry: 898 pnfs_generic_pg_check_layout(pgio); 899 if (!pgio->pg_lseg) { 900 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 901 req->wb_context, 902 0, 903 NFS4_MAX_UINT64, 904 IOMODE_RW, 905 false, 906 GFP_NOFS); 907 if (IS_ERR(pgio->pg_lseg)) { 908 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 909 pgio->pg_lseg = NULL; 910 return; 911 } 912 } 913 /* If no lseg, fall back to write through mds */ 914 if (pgio->pg_lseg == NULL) 915 goto out_mds; 916 917 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); 918 status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); 919 if (status < 0) 920 goto out_mds; 921 922 /* Use a direct mapping of ds_idx to pgio mirror_idx */ 923 if (WARN_ON_ONCE(pgio->pg_mirror_count != 924 FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg))) 925 goto out_mds; 926 927 for (i = 0; i < pgio->pg_mirror_count; i++) { 928 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true); 929 if (!ds) { 930 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg)) 931 goto out_mds; 932 pnfs_put_lseg(pgio->pg_lseg); 933 pgio->pg_lseg = NULL; 934 /* Sleep for 1 second before retrying */ 935 ssleep(1); 936 goto retry; 937 } 938 pgm = &pgio->pg_mirrors[i]; 939 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i); 940 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize; 941 } 942 943 return; 944 945 out_mds: 946 pnfs_put_lseg(pgio->pg_lseg); 947 pgio->pg_lseg = NULL; 948 nfs_pageio_reset_write_mds(pgio); 949 } 950 951 static unsigned int 952 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio, 953 struct nfs_page *req) 954 { 955 if (!pgio->pg_lseg) { 956 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 957 req->wb_context, 958 0, 959 NFS4_MAX_UINT64, 960 IOMODE_RW, 961 false, 962 GFP_NOFS); 963 if (IS_ERR(pgio->pg_lseg)) { 964 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 965 pgio->pg_lseg = NULL; 966 goto out; 967 } 968 } 969 if (pgio->pg_lseg) 970 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg); 971 972 /* no lseg means that pnfs is not in use, so no mirroring here */ 973 nfs_pageio_reset_write_mds(pgio); 974 out: 975 return 1; 976 } 977 978 static const struct nfs_pageio_ops ff_layout_pg_read_ops = { 979 .pg_init = ff_layout_pg_init_read, 980 .pg_test = pnfs_generic_pg_test, 981 .pg_doio = pnfs_generic_pg_readpages, 982 .pg_cleanup = pnfs_generic_pg_cleanup, 983 }; 984 985 static const struct nfs_pageio_ops ff_layout_pg_write_ops = { 986 .pg_init = ff_layout_pg_init_write, 987 .pg_test = pnfs_generic_pg_test, 988 .pg_doio = pnfs_generic_pg_writepages, 989 .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write, 990 .pg_cleanup = pnfs_generic_pg_cleanup, 991 }; 992 993 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs) 994 { 995 struct rpc_task *task = &hdr->task; 996 997 pnfs_layoutcommit_inode(hdr->inode, false); 998 999 if (retry_pnfs) { 1000 dprintk("%s Reset task %5u for i/o through pNFS " 1001 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1002 hdr->task.tk_pid, 1003 hdr->inode->i_sb->s_id, 1004 (unsigned long long)NFS_FILEID(hdr->inode), 1005 hdr->args.count, 1006 (unsigned long long)hdr->args.offset); 1007 1008 hdr->completion_ops->reschedule_io(hdr); 1009 return; 1010 } 1011 1012 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1013 dprintk("%s Reset task %5u for i/o through MDS " 1014 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1015 hdr->task.tk_pid, 1016 hdr->inode->i_sb->s_id, 1017 (unsigned long long)NFS_FILEID(hdr->inode), 1018 hdr->args.count, 1019 (unsigned long long)hdr->args.offset); 1020 1021 task->tk_status = pnfs_write_done_resend_to_mds(hdr); 1022 } 1023 } 1024 1025 static void ff_layout_reset_read(struct nfs_pgio_header *hdr) 1026 { 1027 struct rpc_task *task = &hdr->task; 1028 1029 pnfs_layoutcommit_inode(hdr->inode, false); 1030 1031 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1032 dprintk("%s Reset task %5u for i/o through MDS " 1033 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1034 hdr->task.tk_pid, 1035 hdr->inode->i_sb->s_id, 1036 (unsigned long long)NFS_FILEID(hdr->inode), 1037 hdr->args.count, 1038 (unsigned long long)hdr->args.offset); 1039 1040 task->tk_status = pnfs_read_done_resend_to_mds(hdr); 1041 } 1042 } 1043 1044 static int ff_layout_async_handle_error_v4(struct rpc_task *task, 1045 struct nfs4_state *state, 1046 struct nfs_client *clp, 1047 struct pnfs_layout_segment *lseg, 1048 int idx) 1049 { 1050 struct pnfs_layout_hdr *lo = lseg->pls_layout; 1051 struct inode *inode = lo->plh_inode; 1052 struct nfs_server *mds_server = NFS_SERVER(inode); 1053 1054 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx); 1055 struct nfs_client *mds_client = mds_server->nfs_client; 1056 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; 1057 1058 switch (task->tk_status) { 1059 /* MDS state errors */ 1060 case -NFS4ERR_DELEG_REVOKED: 1061 case -NFS4ERR_ADMIN_REVOKED: 1062 case -NFS4ERR_BAD_STATEID: 1063 if (state == NULL) 1064 break; 1065 nfs_remove_bad_delegation(state->inode, NULL); 1066 case -NFS4ERR_OPENMODE: 1067 if (state == NULL) 1068 break; 1069 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0) 1070 goto out_bad_stateid; 1071 goto wait_on_recovery; 1072 case -NFS4ERR_EXPIRED: 1073 if (state != NULL) { 1074 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0) 1075 goto out_bad_stateid; 1076 } 1077 nfs4_schedule_lease_recovery(mds_client); 1078 goto wait_on_recovery; 1079 /* DS session errors */ 1080 case -NFS4ERR_BADSESSION: 1081 case -NFS4ERR_BADSLOT: 1082 case -NFS4ERR_BAD_HIGH_SLOT: 1083 case -NFS4ERR_DEADSESSION: 1084 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: 1085 case -NFS4ERR_SEQ_FALSE_RETRY: 1086 case -NFS4ERR_SEQ_MISORDERED: 1087 dprintk("%s ERROR %d, Reset session. Exchangeid " 1088 "flags 0x%x\n", __func__, task->tk_status, 1089 clp->cl_exchange_flags); 1090 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status); 1091 break; 1092 case -NFS4ERR_DELAY: 1093 case -NFS4ERR_GRACE: 1094 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX); 1095 break; 1096 case -NFS4ERR_RETRY_UNCACHED_REP: 1097 break; 1098 /* Invalidate Layout errors */ 1099 case -NFS4ERR_PNFS_NO_LAYOUT: 1100 case -ESTALE: /* mapped NFS4ERR_STALE */ 1101 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */ 1102 case -EISDIR: /* mapped NFS4ERR_ISDIR */ 1103 case -NFS4ERR_FHEXPIRED: 1104 case -NFS4ERR_WRONG_TYPE: 1105 dprintk("%s Invalid layout error %d\n", __func__, 1106 task->tk_status); 1107 /* 1108 * Destroy layout so new i/o will get a new layout. 1109 * Layout will not be destroyed until all current lseg 1110 * references are put. Mark layout as invalid to resend failed 1111 * i/o and all i/o waiting on the slot table to the MDS until 1112 * layout is destroyed and a new valid layout is obtained. 1113 */ 1114 pnfs_destroy_layout(NFS_I(inode)); 1115 rpc_wake_up(&tbl->slot_tbl_waitq); 1116 goto reset; 1117 /* RPC connection errors */ 1118 case -ECONNREFUSED: 1119 case -EHOSTDOWN: 1120 case -EHOSTUNREACH: 1121 case -ENETUNREACH: 1122 case -EIO: 1123 case -ETIMEDOUT: 1124 case -EPIPE: 1125 dprintk("%s DS connection error %d\n", __func__, 1126 task->tk_status); 1127 nfs4_delete_deviceid(devid->ld, devid->nfs_client, 1128 &devid->deviceid); 1129 rpc_wake_up(&tbl->slot_tbl_waitq); 1130 /* fall through */ 1131 default: 1132 if (ff_layout_avoid_mds_available_ds(lseg)) 1133 return -NFS4ERR_RESET_TO_PNFS; 1134 reset: 1135 dprintk("%s Retry through MDS. Error %d\n", __func__, 1136 task->tk_status); 1137 return -NFS4ERR_RESET_TO_MDS; 1138 } 1139 out: 1140 task->tk_status = 0; 1141 return -EAGAIN; 1142 out_bad_stateid: 1143 task->tk_status = -EIO; 1144 return 0; 1145 wait_on_recovery: 1146 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL); 1147 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0) 1148 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task); 1149 goto out; 1150 } 1151 1152 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */ 1153 static int ff_layout_async_handle_error_v3(struct rpc_task *task, 1154 struct pnfs_layout_segment *lseg, 1155 int idx) 1156 { 1157 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx); 1158 1159 switch (task->tk_status) { 1160 /* File access problems. Don't mark the device as unavailable */ 1161 case -EACCES: 1162 case -ESTALE: 1163 case -EISDIR: 1164 case -EBADHANDLE: 1165 case -ELOOP: 1166 case -ENOSPC: 1167 break; 1168 case -EJUKEBOX: 1169 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY); 1170 goto out_retry; 1171 default: 1172 dprintk("%s DS connection error %d\n", __func__, 1173 task->tk_status); 1174 nfs4_delete_deviceid(devid->ld, devid->nfs_client, 1175 &devid->deviceid); 1176 } 1177 /* FIXME: Need to prevent infinite looping here. */ 1178 return -NFS4ERR_RESET_TO_PNFS; 1179 out_retry: 1180 task->tk_status = 0; 1181 rpc_restart_call_prepare(task); 1182 rpc_delay(task, NFS_JUKEBOX_RETRY_TIME); 1183 return -EAGAIN; 1184 } 1185 1186 static int ff_layout_async_handle_error(struct rpc_task *task, 1187 struct nfs4_state *state, 1188 struct nfs_client *clp, 1189 struct pnfs_layout_segment *lseg, 1190 int idx) 1191 { 1192 int vers = clp->cl_nfs_mod->rpc_vers->number; 1193 1194 if (task->tk_status >= 0) 1195 return 0; 1196 1197 /* Handle the case of an invalid layout segment */ 1198 if (!pnfs_is_valid_lseg(lseg)) 1199 return -NFS4ERR_RESET_TO_PNFS; 1200 1201 switch (vers) { 1202 case 3: 1203 return ff_layout_async_handle_error_v3(task, lseg, idx); 1204 case 4: 1205 return ff_layout_async_handle_error_v4(task, state, clp, 1206 lseg, idx); 1207 default: 1208 /* should never happen */ 1209 WARN_ON_ONCE(1); 1210 return 0; 1211 } 1212 } 1213 1214 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg, 1215 int idx, u64 offset, u64 length, 1216 u32 status, int opnum, int error) 1217 { 1218 struct nfs4_ff_layout_mirror *mirror; 1219 int err; 1220 1221 if (status == 0) { 1222 switch (error) { 1223 case -ETIMEDOUT: 1224 case -EPFNOSUPPORT: 1225 case -EPROTONOSUPPORT: 1226 case -EOPNOTSUPP: 1227 case -ECONNREFUSED: 1228 case -ECONNRESET: 1229 case -EHOSTDOWN: 1230 case -EHOSTUNREACH: 1231 case -ENETUNREACH: 1232 case -EADDRINUSE: 1233 case -ENOBUFS: 1234 case -EPIPE: 1235 case -EPERM: 1236 status = NFS4ERR_NXIO; 1237 break; 1238 case -EACCES: 1239 status = NFS4ERR_ACCESS; 1240 break; 1241 default: 1242 return; 1243 } 1244 } 1245 1246 switch (status) { 1247 case NFS4ERR_DELAY: 1248 case NFS4ERR_GRACE: 1249 return; 1250 default: 1251 break; 1252 } 1253 1254 mirror = FF_LAYOUT_COMP(lseg, idx); 1255 err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout), 1256 mirror, offset, length, status, opnum, 1257 GFP_NOIO); 1258 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg); 1259 dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status); 1260 } 1261 1262 /* NFS_PROTO call done callback routines */ 1263 static int ff_layout_read_done_cb(struct rpc_task *task, 1264 struct nfs_pgio_header *hdr) 1265 { 1266 int err; 1267 1268 trace_nfs4_pnfs_read(hdr, task->tk_status); 1269 if (task->tk_status < 0) 1270 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx, 1271 hdr->args.offset, hdr->args.count, 1272 hdr->res.op_status, OP_READ, 1273 task->tk_status); 1274 err = ff_layout_async_handle_error(task, hdr->args.context->state, 1275 hdr->ds_clp, hdr->lseg, 1276 hdr->pgio_mirror_idx); 1277 1278 switch (err) { 1279 case -NFS4ERR_RESET_TO_PNFS: 1280 if (ff_layout_choose_best_ds_for_read(hdr->lseg, 1281 hdr->pgio_mirror_idx + 1, 1282 &hdr->pgio_mirror_idx)) 1283 goto out_eagain; 1284 ff_layout_read_record_layoutstats_done(task, hdr); 1285 pnfs_read_resend_pnfs(hdr); 1286 return task->tk_status; 1287 case -NFS4ERR_RESET_TO_MDS: 1288 ff_layout_reset_read(hdr); 1289 return task->tk_status; 1290 case -EAGAIN: 1291 goto out_eagain; 1292 } 1293 1294 return 0; 1295 out_eagain: 1296 rpc_restart_call_prepare(task); 1297 return -EAGAIN; 1298 } 1299 1300 static bool 1301 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg) 1302 { 1303 return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT); 1304 } 1305 1306 /* 1307 * We reference the rpc_cred of the first WRITE that triggers the need for 1308 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. 1309 * rfc5661 is not clear about which credential should be used. 1310 * 1311 * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so 1312 * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751 1313 * we always send layoutcommit after DS writes. 1314 */ 1315 static void 1316 ff_layout_set_layoutcommit(struct inode *inode, 1317 struct pnfs_layout_segment *lseg, 1318 loff_t end_offset) 1319 { 1320 if (!ff_layout_need_layoutcommit(lseg)) 1321 return; 1322 1323 pnfs_set_layoutcommit(inode, lseg, end_offset); 1324 dprintk("%s inode %lu pls_end_pos %llu\n", __func__, inode->i_ino, 1325 (unsigned long long) NFS_I(inode)->layout->plh_lwb); 1326 } 1327 1328 static bool 1329 ff_layout_device_unavailable(struct pnfs_layout_segment *lseg, int idx) 1330 { 1331 /* No mirroring for now */ 1332 struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx); 1333 1334 return ff_layout_test_devid_unavailable(node); 1335 } 1336 1337 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task, 1338 struct nfs_pgio_header *hdr) 1339 { 1340 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags)) 1341 return; 1342 nfs4_ff_layout_stat_io_start_read(hdr->inode, 1343 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1344 hdr->args.count, 1345 task->tk_start); 1346 } 1347 1348 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task, 1349 struct nfs_pgio_header *hdr) 1350 { 1351 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags)) 1352 return; 1353 nfs4_ff_layout_stat_io_end_read(task, 1354 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1355 hdr->args.count, 1356 hdr->res.count); 1357 } 1358 1359 static int ff_layout_read_prepare_common(struct rpc_task *task, 1360 struct nfs_pgio_header *hdr) 1361 { 1362 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { 1363 rpc_exit(task, -EIO); 1364 return -EIO; 1365 } 1366 if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) { 1367 rpc_exit(task, -EHOSTDOWN); 1368 return -EAGAIN; 1369 } 1370 1371 ff_layout_read_record_layoutstats_start(task, hdr); 1372 return 0; 1373 } 1374 1375 /* 1376 * Call ops for the async read/write cases 1377 * In the case of dense layouts, the offset needs to be reset to its 1378 * original value. 1379 */ 1380 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data) 1381 { 1382 struct nfs_pgio_header *hdr = data; 1383 1384 if (ff_layout_read_prepare_common(task, hdr)) 1385 return; 1386 1387 rpc_call_start(task); 1388 } 1389 1390 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data) 1391 { 1392 struct nfs_pgio_header *hdr = data; 1393 1394 if (nfs4_setup_sequence(hdr->ds_clp, 1395 &hdr->args.seq_args, 1396 &hdr->res.seq_res, 1397 task)) 1398 return; 1399 1400 if (ff_layout_read_prepare_common(task, hdr)) 1401 return; 1402 1403 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context, 1404 hdr->args.lock_context, FMODE_READ) == -EIO) 1405 rpc_exit(task, -EIO); /* lost lock, terminate I/O */ 1406 } 1407 1408 static void ff_layout_read_call_done(struct rpc_task *task, void *data) 1409 { 1410 struct nfs_pgio_header *hdr = data; 1411 1412 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); 1413 1414 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && 1415 task->tk_status == 0) { 1416 nfs4_sequence_done(task, &hdr->res.seq_res); 1417 return; 1418 } 1419 1420 /* Note this may cause RPC to be resent */ 1421 hdr->mds_ops->rpc_call_done(task, hdr); 1422 } 1423 1424 static void ff_layout_read_count_stats(struct rpc_task *task, void *data) 1425 { 1426 struct nfs_pgio_header *hdr = data; 1427 1428 ff_layout_read_record_layoutstats_done(task, hdr); 1429 rpc_count_iostats_metrics(task, 1430 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]); 1431 } 1432 1433 static void ff_layout_read_release(void *data) 1434 { 1435 struct nfs_pgio_header *hdr = data; 1436 1437 ff_layout_read_record_layoutstats_done(&hdr->task, hdr); 1438 pnfs_generic_rw_release(data); 1439 } 1440 1441 1442 static int ff_layout_write_done_cb(struct rpc_task *task, 1443 struct nfs_pgio_header *hdr) 1444 { 1445 loff_t end_offs = 0; 1446 int err; 1447 1448 trace_nfs4_pnfs_write(hdr, task->tk_status); 1449 if (task->tk_status < 0) 1450 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx, 1451 hdr->args.offset, hdr->args.count, 1452 hdr->res.op_status, OP_WRITE, 1453 task->tk_status); 1454 err = ff_layout_async_handle_error(task, hdr->args.context->state, 1455 hdr->ds_clp, hdr->lseg, 1456 hdr->pgio_mirror_idx); 1457 1458 switch (err) { 1459 case -NFS4ERR_RESET_TO_PNFS: 1460 ff_layout_reset_write(hdr, true); 1461 return task->tk_status; 1462 case -NFS4ERR_RESET_TO_MDS: 1463 ff_layout_reset_write(hdr, false); 1464 return task->tk_status; 1465 case -EAGAIN: 1466 return -EAGAIN; 1467 } 1468 1469 if (hdr->res.verf->committed == NFS_FILE_SYNC || 1470 hdr->res.verf->committed == NFS_DATA_SYNC) 1471 end_offs = hdr->mds_offset + (loff_t)hdr->res.count; 1472 1473 /* Note: if the write is unstable, don't set end_offs until commit */ 1474 ff_layout_set_layoutcommit(hdr->inode, hdr->lseg, end_offs); 1475 1476 /* zero out fattr since we don't care DS attr at all */ 1477 hdr->fattr.valid = 0; 1478 if (task->tk_status >= 0) 1479 nfs_writeback_update_inode(hdr); 1480 1481 return 0; 1482 } 1483 1484 static int ff_layout_commit_done_cb(struct rpc_task *task, 1485 struct nfs_commit_data *data) 1486 { 1487 int err; 1488 1489 trace_nfs4_pnfs_commit_ds(data, task->tk_status); 1490 if (task->tk_status < 0) 1491 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index, 1492 data->args.offset, data->args.count, 1493 data->res.op_status, OP_COMMIT, 1494 task->tk_status); 1495 err = ff_layout_async_handle_error(task, NULL, data->ds_clp, 1496 data->lseg, data->ds_commit_index); 1497 1498 switch (err) { 1499 case -NFS4ERR_RESET_TO_PNFS: 1500 pnfs_generic_prepare_to_resend_writes(data); 1501 return -EAGAIN; 1502 case -NFS4ERR_RESET_TO_MDS: 1503 pnfs_generic_prepare_to_resend_writes(data); 1504 return -EAGAIN; 1505 case -EAGAIN: 1506 rpc_restart_call_prepare(task); 1507 return -EAGAIN; 1508 } 1509 1510 ff_layout_set_layoutcommit(data->inode, data->lseg, data->lwb); 1511 1512 return 0; 1513 } 1514 1515 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task, 1516 struct nfs_pgio_header *hdr) 1517 { 1518 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags)) 1519 return; 1520 nfs4_ff_layout_stat_io_start_write(hdr->inode, 1521 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1522 hdr->args.count, 1523 task->tk_start); 1524 } 1525 1526 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task, 1527 struct nfs_pgio_header *hdr) 1528 { 1529 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags)) 1530 return; 1531 nfs4_ff_layout_stat_io_end_write(task, 1532 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1533 hdr->args.count, hdr->res.count, 1534 hdr->res.verf->committed); 1535 } 1536 1537 static int ff_layout_write_prepare_common(struct rpc_task *task, 1538 struct nfs_pgio_header *hdr) 1539 { 1540 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { 1541 rpc_exit(task, -EIO); 1542 return -EIO; 1543 } 1544 1545 if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) { 1546 rpc_exit(task, -EHOSTDOWN); 1547 return -EAGAIN; 1548 } 1549 1550 ff_layout_write_record_layoutstats_start(task, hdr); 1551 return 0; 1552 } 1553 1554 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data) 1555 { 1556 struct nfs_pgio_header *hdr = data; 1557 1558 if (ff_layout_write_prepare_common(task, hdr)) 1559 return; 1560 1561 rpc_call_start(task); 1562 } 1563 1564 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data) 1565 { 1566 struct nfs_pgio_header *hdr = data; 1567 1568 if (nfs4_setup_sequence(hdr->ds_clp, 1569 &hdr->args.seq_args, 1570 &hdr->res.seq_res, 1571 task)) 1572 return; 1573 1574 if (ff_layout_write_prepare_common(task, hdr)) 1575 return; 1576 1577 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context, 1578 hdr->args.lock_context, FMODE_WRITE) == -EIO) 1579 rpc_exit(task, -EIO); /* lost lock, terminate I/O */ 1580 } 1581 1582 static void ff_layout_write_call_done(struct rpc_task *task, void *data) 1583 { 1584 struct nfs_pgio_header *hdr = data; 1585 1586 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && 1587 task->tk_status == 0) { 1588 nfs4_sequence_done(task, &hdr->res.seq_res); 1589 return; 1590 } 1591 1592 /* Note this may cause RPC to be resent */ 1593 hdr->mds_ops->rpc_call_done(task, hdr); 1594 } 1595 1596 static void ff_layout_write_count_stats(struct rpc_task *task, void *data) 1597 { 1598 struct nfs_pgio_header *hdr = data; 1599 1600 ff_layout_write_record_layoutstats_done(task, hdr); 1601 rpc_count_iostats_metrics(task, 1602 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]); 1603 } 1604 1605 static void ff_layout_write_release(void *data) 1606 { 1607 struct nfs_pgio_header *hdr = data; 1608 1609 ff_layout_write_record_layoutstats_done(&hdr->task, hdr); 1610 pnfs_generic_rw_release(data); 1611 } 1612 1613 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task, 1614 struct nfs_commit_data *cdata) 1615 { 1616 if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags)) 1617 return; 1618 nfs4_ff_layout_stat_io_start_write(cdata->inode, 1619 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index), 1620 0, task->tk_start); 1621 } 1622 1623 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task, 1624 struct nfs_commit_data *cdata) 1625 { 1626 struct nfs_page *req; 1627 __u64 count = 0; 1628 1629 if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags)) 1630 return; 1631 1632 if (task->tk_status == 0) { 1633 list_for_each_entry(req, &cdata->pages, wb_list) 1634 count += req->wb_bytes; 1635 } 1636 nfs4_ff_layout_stat_io_end_write(task, 1637 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index), 1638 count, count, NFS_FILE_SYNC); 1639 } 1640 1641 static void ff_layout_commit_prepare_common(struct rpc_task *task, 1642 struct nfs_commit_data *cdata) 1643 { 1644 ff_layout_commit_record_layoutstats_start(task, cdata); 1645 } 1646 1647 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data) 1648 { 1649 ff_layout_commit_prepare_common(task, data); 1650 rpc_call_start(task); 1651 } 1652 1653 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data) 1654 { 1655 struct nfs_commit_data *wdata = data; 1656 1657 if (nfs4_setup_sequence(wdata->ds_clp, 1658 &wdata->args.seq_args, 1659 &wdata->res.seq_res, 1660 task)) 1661 return; 1662 ff_layout_commit_prepare_common(task, data); 1663 } 1664 1665 static void ff_layout_commit_done(struct rpc_task *task, void *data) 1666 { 1667 pnfs_generic_write_commit_done(task, data); 1668 } 1669 1670 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data) 1671 { 1672 struct nfs_commit_data *cdata = data; 1673 1674 ff_layout_commit_record_layoutstats_done(task, cdata); 1675 rpc_count_iostats_metrics(task, 1676 &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]); 1677 } 1678 1679 static void ff_layout_commit_release(void *data) 1680 { 1681 struct nfs_commit_data *cdata = data; 1682 1683 ff_layout_commit_record_layoutstats_done(&cdata->task, cdata); 1684 pnfs_generic_commit_release(data); 1685 } 1686 1687 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = { 1688 .rpc_call_prepare = ff_layout_read_prepare_v3, 1689 .rpc_call_done = ff_layout_read_call_done, 1690 .rpc_count_stats = ff_layout_read_count_stats, 1691 .rpc_release = ff_layout_read_release, 1692 }; 1693 1694 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = { 1695 .rpc_call_prepare = ff_layout_read_prepare_v4, 1696 .rpc_call_done = ff_layout_read_call_done, 1697 .rpc_count_stats = ff_layout_read_count_stats, 1698 .rpc_release = ff_layout_read_release, 1699 }; 1700 1701 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = { 1702 .rpc_call_prepare = ff_layout_write_prepare_v3, 1703 .rpc_call_done = ff_layout_write_call_done, 1704 .rpc_count_stats = ff_layout_write_count_stats, 1705 .rpc_release = ff_layout_write_release, 1706 }; 1707 1708 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = { 1709 .rpc_call_prepare = ff_layout_write_prepare_v4, 1710 .rpc_call_done = ff_layout_write_call_done, 1711 .rpc_count_stats = ff_layout_write_count_stats, 1712 .rpc_release = ff_layout_write_release, 1713 }; 1714 1715 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = { 1716 .rpc_call_prepare = ff_layout_commit_prepare_v3, 1717 .rpc_call_done = ff_layout_commit_done, 1718 .rpc_count_stats = ff_layout_commit_count_stats, 1719 .rpc_release = ff_layout_commit_release, 1720 }; 1721 1722 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = { 1723 .rpc_call_prepare = ff_layout_commit_prepare_v4, 1724 .rpc_call_done = ff_layout_commit_done, 1725 .rpc_count_stats = ff_layout_commit_count_stats, 1726 .rpc_release = ff_layout_commit_release, 1727 }; 1728 1729 static enum pnfs_try_status 1730 ff_layout_read_pagelist(struct nfs_pgio_header *hdr) 1731 { 1732 struct pnfs_layout_segment *lseg = hdr->lseg; 1733 struct nfs4_pnfs_ds *ds; 1734 struct rpc_clnt *ds_clnt; 1735 struct rpc_cred *ds_cred; 1736 loff_t offset = hdr->args.offset; 1737 u32 idx = hdr->pgio_mirror_idx; 1738 int vers; 1739 struct nfs_fh *fh; 1740 1741 dprintk("--> %s ino %lu pgbase %u req %zu@%llu\n", 1742 __func__, hdr->inode->i_ino, 1743 hdr->args.pgbase, (size_t)hdr->args.count, offset); 1744 1745 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false); 1746 if (!ds) 1747 goto out_failed; 1748 1749 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp, 1750 hdr->inode); 1751 if (IS_ERR(ds_clnt)) 1752 goto out_failed; 1753 1754 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred); 1755 if (!ds_cred) 1756 goto out_failed; 1757 1758 vers = nfs4_ff_layout_ds_version(lseg, idx); 1759 1760 dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__, 1761 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers); 1762 1763 hdr->pgio_done_cb = ff_layout_read_done_cb; 1764 atomic_inc(&ds->ds_clp->cl_count); 1765 hdr->ds_clp = ds->ds_clp; 1766 fh = nfs4_ff_layout_select_ds_fh(lseg, idx); 1767 if (fh) 1768 hdr->args.fh = fh; 1769 /* 1770 * Note that if we ever decide to split across DSes, 1771 * then we may need to handle dense-like offsets. 1772 */ 1773 hdr->args.offset = offset; 1774 hdr->mds_offset = offset; 1775 1776 /* Perform an asynchronous read to ds */ 1777 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops, 1778 vers == 3 ? &ff_layout_read_call_ops_v3 : 1779 &ff_layout_read_call_ops_v4, 1780 0, RPC_TASK_SOFTCONN); 1781 put_rpccred(ds_cred); 1782 return PNFS_ATTEMPTED; 1783 1784 out_failed: 1785 if (ff_layout_avoid_mds_available_ds(lseg)) 1786 return PNFS_TRY_AGAIN; 1787 return PNFS_NOT_ATTEMPTED; 1788 } 1789 1790 /* Perform async writes. */ 1791 static enum pnfs_try_status 1792 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync) 1793 { 1794 struct pnfs_layout_segment *lseg = hdr->lseg; 1795 struct nfs4_pnfs_ds *ds; 1796 struct rpc_clnt *ds_clnt; 1797 struct rpc_cred *ds_cred; 1798 loff_t offset = hdr->args.offset; 1799 int vers; 1800 struct nfs_fh *fh; 1801 int idx = hdr->pgio_mirror_idx; 1802 1803 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true); 1804 if (!ds) 1805 goto out_failed; 1806 1807 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp, 1808 hdr->inode); 1809 if (IS_ERR(ds_clnt)) 1810 goto out_failed; 1811 1812 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred); 1813 if (!ds_cred) 1814 goto out_failed; 1815 1816 vers = nfs4_ff_layout_ds_version(lseg, idx); 1817 1818 dprintk("%s ino %lu sync %d req %zu@%llu DS: %s cl_count %d vers %d\n", 1819 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count, 1820 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), 1821 vers); 1822 1823 hdr->pgio_done_cb = ff_layout_write_done_cb; 1824 atomic_inc(&ds->ds_clp->cl_count); 1825 hdr->ds_clp = ds->ds_clp; 1826 hdr->ds_commit_idx = idx; 1827 fh = nfs4_ff_layout_select_ds_fh(lseg, idx); 1828 if (fh) 1829 hdr->args.fh = fh; 1830 1831 /* 1832 * Note that if we ever decide to split across DSes, 1833 * then we may need to handle dense-like offsets. 1834 */ 1835 hdr->args.offset = offset; 1836 1837 /* Perform an asynchronous write */ 1838 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops, 1839 vers == 3 ? &ff_layout_write_call_ops_v3 : 1840 &ff_layout_write_call_ops_v4, 1841 sync, RPC_TASK_SOFTCONN); 1842 put_rpccred(ds_cred); 1843 return PNFS_ATTEMPTED; 1844 1845 out_failed: 1846 if (ff_layout_avoid_mds_available_ds(lseg)) 1847 return PNFS_TRY_AGAIN; 1848 return PNFS_NOT_ATTEMPTED; 1849 } 1850 1851 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) 1852 { 1853 return i; 1854 } 1855 1856 static struct nfs_fh * 1857 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) 1858 { 1859 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg); 1860 1861 /* FIXME: Assume that there is only one NFS version available 1862 * for the DS. 1863 */ 1864 return &flseg->mirror_array[i]->fh_versions[0]; 1865 } 1866 1867 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how) 1868 { 1869 struct pnfs_layout_segment *lseg = data->lseg; 1870 struct nfs4_pnfs_ds *ds; 1871 struct rpc_clnt *ds_clnt; 1872 struct rpc_cred *ds_cred; 1873 u32 idx; 1874 int vers, ret; 1875 struct nfs_fh *fh; 1876 1877 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); 1878 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true); 1879 if (!ds) 1880 goto out_err; 1881 1882 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp, 1883 data->inode); 1884 if (IS_ERR(ds_clnt)) 1885 goto out_err; 1886 1887 ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred); 1888 if (!ds_cred) 1889 goto out_err; 1890 1891 vers = nfs4_ff_layout_ds_version(lseg, idx); 1892 1893 dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__, 1894 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count), 1895 vers); 1896 data->commit_done_cb = ff_layout_commit_done_cb; 1897 data->cred = ds_cred; 1898 atomic_inc(&ds->ds_clp->cl_count); 1899 data->ds_clp = ds->ds_clp; 1900 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); 1901 if (fh) 1902 data->args.fh = fh; 1903 1904 ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops, 1905 vers == 3 ? &ff_layout_commit_call_ops_v3 : 1906 &ff_layout_commit_call_ops_v4, 1907 how, RPC_TASK_SOFTCONN); 1908 put_rpccred(ds_cred); 1909 return ret; 1910 out_err: 1911 pnfs_generic_prepare_to_resend_writes(data); 1912 pnfs_generic_commit_release(data); 1913 return -EAGAIN; 1914 } 1915 1916 static int 1917 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, 1918 int how, struct nfs_commit_info *cinfo) 1919 { 1920 return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo, 1921 ff_layout_initiate_commit); 1922 } 1923 1924 static struct pnfs_ds_commit_info * 1925 ff_layout_get_ds_info(struct inode *inode) 1926 { 1927 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; 1928 1929 if (layout == NULL) 1930 return NULL; 1931 1932 return &FF_LAYOUT_FROM_HDR(layout)->commit_info; 1933 } 1934 1935 static void 1936 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d) 1937 { 1938 nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds, 1939 id_node)); 1940 } 1941 1942 static int ff_layout_encode_ioerr(struct xdr_stream *xdr, 1943 const struct nfs4_layoutreturn_args *args, 1944 const struct nfs4_flexfile_layoutreturn_args *ff_args) 1945 { 1946 __be32 *start; 1947 1948 start = xdr_reserve_space(xdr, 4); 1949 if (unlikely(!start)) 1950 return -E2BIG; 1951 1952 *start = cpu_to_be32(ff_args->num_errors); 1953 /* This assume we always return _ALL_ layouts */ 1954 return ff_layout_encode_ds_ioerr(xdr, &ff_args->errors); 1955 } 1956 1957 static void 1958 encode_opaque_fixed(struct xdr_stream *xdr, const void *buf, size_t len) 1959 { 1960 WARN_ON_ONCE(xdr_stream_encode_opaque_fixed(xdr, buf, len) < 0); 1961 } 1962 1963 static void 1964 ff_layout_encode_ff_iostat_head(struct xdr_stream *xdr, 1965 const nfs4_stateid *stateid, 1966 const struct nfs42_layoutstat_devinfo *devinfo) 1967 { 1968 __be32 *p; 1969 1970 p = xdr_reserve_space(xdr, 8 + 8); 1971 p = xdr_encode_hyper(p, devinfo->offset); 1972 p = xdr_encode_hyper(p, devinfo->length); 1973 encode_opaque_fixed(xdr, stateid->data, NFS4_STATEID_SIZE); 1974 p = xdr_reserve_space(xdr, 4*8); 1975 p = xdr_encode_hyper(p, devinfo->read_count); 1976 p = xdr_encode_hyper(p, devinfo->read_bytes); 1977 p = xdr_encode_hyper(p, devinfo->write_count); 1978 p = xdr_encode_hyper(p, devinfo->write_bytes); 1979 encode_opaque_fixed(xdr, devinfo->dev_id.data, NFS4_DEVICEID4_SIZE); 1980 } 1981 1982 static void 1983 ff_layout_encode_ff_iostat(struct xdr_stream *xdr, 1984 const nfs4_stateid *stateid, 1985 const struct nfs42_layoutstat_devinfo *devinfo) 1986 { 1987 ff_layout_encode_ff_iostat_head(xdr, stateid, devinfo); 1988 ff_layout_encode_ff_layoutupdate(xdr, devinfo, 1989 devinfo->ld_private.data); 1990 } 1991 1992 /* report nothing for now */ 1993 static void ff_layout_encode_iostats_array(struct xdr_stream *xdr, 1994 const struct nfs4_layoutreturn_args *args, 1995 struct nfs4_flexfile_layoutreturn_args *ff_args) 1996 { 1997 __be32 *p; 1998 int i; 1999 2000 p = xdr_reserve_space(xdr, 4); 2001 *p = cpu_to_be32(ff_args->num_dev); 2002 for (i = 0; i < ff_args->num_dev; i++) 2003 ff_layout_encode_ff_iostat(xdr, 2004 &args->layout->plh_stateid, 2005 &ff_args->devinfo[i]); 2006 } 2007 2008 static void 2009 ff_layout_free_iostats_array(struct nfs42_layoutstat_devinfo *devinfo, 2010 unsigned int num_entries) 2011 { 2012 unsigned int i; 2013 2014 for (i = 0; i < num_entries; i++) { 2015 if (!devinfo[i].ld_private.ops) 2016 continue; 2017 if (!devinfo[i].ld_private.ops->free) 2018 continue; 2019 devinfo[i].ld_private.ops->free(&devinfo[i].ld_private); 2020 } 2021 } 2022 2023 static struct nfs4_deviceid_node * 2024 ff_layout_alloc_deviceid_node(struct nfs_server *server, 2025 struct pnfs_device *pdev, gfp_t gfp_flags) 2026 { 2027 struct nfs4_ff_layout_ds *dsaddr; 2028 2029 dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags); 2030 if (!dsaddr) 2031 return NULL; 2032 return &dsaddr->id_node; 2033 } 2034 2035 static void 2036 ff_layout_encode_layoutreturn(struct xdr_stream *xdr, 2037 const void *voidargs, 2038 const struct nfs4_xdr_opaque_data *ff_opaque) 2039 { 2040 const struct nfs4_layoutreturn_args *args = voidargs; 2041 struct nfs4_flexfile_layoutreturn_args *ff_args = ff_opaque->data; 2042 struct xdr_buf tmp_buf = { 2043 .head = { 2044 [0] = { 2045 .iov_base = page_address(ff_args->pages[0]), 2046 }, 2047 }, 2048 .buflen = PAGE_SIZE, 2049 }; 2050 struct xdr_stream tmp_xdr; 2051 __be32 *start; 2052 2053 dprintk("%s: Begin\n", __func__); 2054 2055 xdr_init_encode(&tmp_xdr, &tmp_buf, NULL); 2056 2057 ff_layout_encode_ioerr(&tmp_xdr, args, ff_args); 2058 ff_layout_encode_iostats_array(&tmp_xdr, args, ff_args); 2059 2060 start = xdr_reserve_space(xdr, 4); 2061 *start = cpu_to_be32(tmp_buf.len); 2062 xdr_write_pages(xdr, ff_args->pages, 0, tmp_buf.len); 2063 2064 dprintk("%s: Return\n", __func__); 2065 } 2066 2067 static void 2068 ff_layout_free_layoutreturn(struct nfs4_xdr_opaque_data *args) 2069 { 2070 struct nfs4_flexfile_layoutreturn_args *ff_args; 2071 2072 if (!args->data) 2073 return; 2074 ff_args = args->data; 2075 args->data = NULL; 2076 2077 ff_layout_free_ds_ioerr(&ff_args->errors); 2078 ff_layout_free_iostats_array(ff_args->devinfo, ff_args->num_dev); 2079 2080 put_page(ff_args->pages[0]); 2081 kfree(ff_args); 2082 } 2083 2084 static const struct nfs4_xdr_opaque_ops layoutreturn_ops = { 2085 .encode = ff_layout_encode_layoutreturn, 2086 .free = ff_layout_free_layoutreturn, 2087 }; 2088 2089 static int 2090 ff_layout_prepare_layoutreturn(struct nfs4_layoutreturn_args *args) 2091 { 2092 struct nfs4_flexfile_layoutreturn_args *ff_args; 2093 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(args->layout); 2094 2095 ff_args = kmalloc(sizeof(*ff_args), GFP_KERNEL); 2096 if (!ff_args) 2097 goto out_nomem; 2098 ff_args->pages[0] = alloc_page(GFP_KERNEL); 2099 if (!ff_args->pages[0]) 2100 goto out_nomem_free; 2101 2102 INIT_LIST_HEAD(&ff_args->errors); 2103 ff_args->num_errors = ff_layout_fetch_ds_ioerr(args->layout, 2104 &args->range, &ff_args->errors, 2105 FF_LAYOUTRETURN_MAXERR); 2106 2107 spin_lock(&args->inode->i_lock); 2108 ff_args->num_dev = ff_layout_mirror_prepare_stats(&ff_layout->generic_hdr, 2109 &ff_args->devinfo[0], ARRAY_SIZE(ff_args->devinfo)); 2110 spin_unlock(&args->inode->i_lock); 2111 2112 args->ld_private->ops = &layoutreturn_ops; 2113 args->ld_private->data = ff_args; 2114 return 0; 2115 out_nomem_free: 2116 kfree(ff_args); 2117 out_nomem: 2118 return -ENOMEM; 2119 } 2120 2121 static int 2122 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen) 2123 { 2124 const struct sockaddr_in *sin = (struct sockaddr_in *)sap; 2125 2126 return snprintf(buf, buflen, "%pI4", &sin->sin_addr); 2127 } 2128 2129 static size_t 2130 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf, 2131 const int buflen) 2132 { 2133 const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap; 2134 const struct in6_addr *addr = &sin6->sin6_addr; 2135 2136 /* 2137 * RFC 4291, Section 2.2.2 2138 * 2139 * Shorthanded ANY address 2140 */ 2141 if (ipv6_addr_any(addr)) 2142 return snprintf(buf, buflen, "::"); 2143 2144 /* 2145 * RFC 4291, Section 2.2.2 2146 * 2147 * Shorthanded loopback address 2148 */ 2149 if (ipv6_addr_loopback(addr)) 2150 return snprintf(buf, buflen, "::1"); 2151 2152 /* 2153 * RFC 4291, Section 2.2.3 2154 * 2155 * Special presentation address format for mapped v4 2156 * addresses. 2157 */ 2158 if (ipv6_addr_v4mapped(addr)) 2159 return snprintf(buf, buflen, "::ffff:%pI4", 2160 &addr->s6_addr32[3]); 2161 2162 /* 2163 * RFC 4291, Section 2.2.1 2164 */ 2165 return snprintf(buf, buflen, "%pI6c", addr); 2166 } 2167 2168 /* Derived from rpc_sockaddr2uaddr */ 2169 static void 2170 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da) 2171 { 2172 struct sockaddr *sap = (struct sockaddr *)&da->da_addr; 2173 char portbuf[RPCBIND_MAXUADDRPLEN]; 2174 char addrbuf[RPCBIND_MAXUADDRLEN]; 2175 char *netid; 2176 unsigned short port; 2177 int len, netid_len; 2178 __be32 *p; 2179 2180 switch (sap->sa_family) { 2181 case AF_INET: 2182 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0) 2183 return; 2184 port = ntohs(((struct sockaddr_in *)sap)->sin_port); 2185 netid = "tcp"; 2186 netid_len = 3; 2187 break; 2188 case AF_INET6: 2189 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0) 2190 return; 2191 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port); 2192 netid = "tcp6"; 2193 netid_len = 4; 2194 break; 2195 default: 2196 /* we only support tcp and tcp6 */ 2197 WARN_ON_ONCE(1); 2198 return; 2199 } 2200 2201 snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff); 2202 len = strlcat(addrbuf, portbuf, sizeof(addrbuf)); 2203 2204 p = xdr_reserve_space(xdr, 4 + netid_len); 2205 xdr_encode_opaque(p, netid, netid_len); 2206 2207 p = xdr_reserve_space(xdr, 4 + len); 2208 xdr_encode_opaque(p, addrbuf, len); 2209 } 2210 2211 static void 2212 ff_layout_encode_nfstime(struct xdr_stream *xdr, 2213 ktime_t t) 2214 { 2215 struct timespec64 ts; 2216 __be32 *p; 2217 2218 p = xdr_reserve_space(xdr, 12); 2219 ts = ktime_to_timespec64(t); 2220 p = xdr_encode_hyper(p, ts.tv_sec); 2221 *p++ = cpu_to_be32(ts.tv_nsec); 2222 } 2223 2224 static void 2225 ff_layout_encode_io_latency(struct xdr_stream *xdr, 2226 struct nfs4_ff_io_stat *stat) 2227 { 2228 __be32 *p; 2229 2230 p = xdr_reserve_space(xdr, 5 * 8); 2231 p = xdr_encode_hyper(p, stat->ops_requested); 2232 p = xdr_encode_hyper(p, stat->bytes_requested); 2233 p = xdr_encode_hyper(p, stat->ops_completed); 2234 p = xdr_encode_hyper(p, stat->bytes_completed); 2235 p = xdr_encode_hyper(p, stat->bytes_not_delivered); 2236 ff_layout_encode_nfstime(xdr, stat->total_busy_time); 2237 ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time); 2238 } 2239 2240 static void 2241 ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr, 2242 const struct nfs42_layoutstat_devinfo *devinfo, 2243 struct nfs4_ff_layout_mirror *mirror) 2244 { 2245 struct nfs4_pnfs_ds_addr *da; 2246 struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds; 2247 struct nfs_fh *fh = &mirror->fh_versions[0]; 2248 __be32 *p; 2249 2250 da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node); 2251 dprintk("%s: DS %s: encoding address %s\n", 2252 __func__, ds->ds_remotestr, da->da_remotestr); 2253 /* netaddr4 */ 2254 ff_layout_encode_netaddr(xdr, da); 2255 /* nfs_fh4 */ 2256 p = xdr_reserve_space(xdr, 4 + fh->size); 2257 xdr_encode_opaque(p, fh->data, fh->size); 2258 /* ff_io_latency4 read */ 2259 spin_lock(&mirror->lock); 2260 ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat); 2261 /* ff_io_latency4 write */ 2262 ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat); 2263 spin_unlock(&mirror->lock); 2264 /* nfstime4 */ 2265 ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time)); 2266 /* bool */ 2267 p = xdr_reserve_space(xdr, 4); 2268 *p = cpu_to_be32(false); 2269 } 2270 2271 static void 2272 ff_layout_encode_layoutstats(struct xdr_stream *xdr, const void *args, 2273 const struct nfs4_xdr_opaque_data *opaque) 2274 { 2275 struct nfs42_layoutstat_devinfo *devinfo = container_of(opaque, 2276 struct nfs42_layoutstat_devinfo, ld_private); 2277 __be32 *start; 2278 2279 /* layoutupdate length */ 2280 start = xdr_reserve_space(xdr, 4); 2281 ff_layout_encode_ff_layoutupdate(xdr, devinfo, opaque->data); 2282 2283 *start = cpu_to_be32((xdr->p - start - 1) * 4); 2284 } 2285 2286 static void 2287 ff_layout_free_layoutstats(struct nfs4_xdr_opaque_data *opaque) 2288 { 2289 struct nfs4_ff_layout_mirror *mirror = opaque->data; 2290 2291 ff_layout_put_mirror(mirror); 2292 } 2293 2294 static const struct nfs4_xdr_opaque_ops layoutstat_ops = { 2295 .encode = ff_layout_encode_layoutstats, 2296 .free = ff_layout_free_layoutstats, 2297 }; 2298 2299 static int 2300 ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo, 2301 struct nfs42_layoutstat_devinfo *devinfo, 2302 int dev_limit) 2303 { 2304 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo); 2305 struct nfs4_ff_layout_mirror *mirror; 2306 struct nfs4_deviceid_node *dev; 2307 int i = 0; 2308 2309 list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) { 2310 if (i >= dev_limit) 2311 break; 2312 if (IS_ERR_OR_NULL(mirror->mirror_ds)) 2313 continue; 2314 if (!test_and_clear_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags)) 2315 continue; 2316 /* mirror refcount put in cleanup_layoutstats */ 2317 if (!atomic_inc_not_zero(&mirror->ref)) 2318 continue; 2319 dev = &mirror->mirror_ds->id_node; 2320 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE); 2321 devinfo->offset = 0; 2322 devinfo->length = NFS4_MAX_UINT64; 2323 spin_lock(&mirror->lock); 2324 devinfo->read_count = mirror->read_stat.io_stat.ops_completed; 2325 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed; 2326 devinfo->write_count = mirror->write_stat.io_stat.ops_completed; 2327 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed; 2328 spin_unlock(&mirror->lock); 2329 devinfo->layout_type = LAYOUT_FLEX_FILES; 2330 devinfo->ld_private.ops = &layoutstat_ops; 2331 devinfo->ld_private.data = mirror; 2332 2333 devinfo++; 2334 i++; 2335 } 2336 return i; 2337 } 2338 2339 static int 2340 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args) 2341 { 2342 struct nfs4_flexfile_layout *ff_layout; 2343 const int dev_count = PNFS_LAYOUTSTATS_MAXDEV; 2344 2345 /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */ 2346 args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO); 2347 if (!args->devinfo) 2348 return -ENOMEM; 2349 2350 spin_lock(&args->inode->i_lock); 2351 ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout); 2352 args->num_dev = ff_layout_mirror_prepare_stats(&ff_layout->generic_hdr, 2353 &args->devinfo[0], dev_count); 2354 spin_unlock(&args->inode->i_lock); 2355 if (!args->num_dev) { 2356 kfree(args->devinfo); 2357 args->devinfo = NULL; 2358 return -ENOENT; 2359 } 2360 2361 return 0; 2362 } 2363 2364 static int 2365 ff_layout_set_layoutdriver(struct nfs_server *server, 2366 const struct nfs_fh *dummy) 2367 { 2368 #if IS_ENABLED(CONFIG_NFS_V4_2) 2369 server->caps |= NFS_CAP_LAYOUTSTATS; 2370 #endif 2371 return 0; 2372 } 2373 2374 static struct pnfs_layoutdriver_type flexfilelayout_type = { 2375 .id = LAYOUT_FLEX_FILES, 2376 .name = "LAYOUT_FLEX_FILES", 2377 .owner = THIS_MODULE, 2378 .set_layoutdriver = ff_layout_set_layoutdriver, 2379 .alloc_layout_hdr = ff_layout_alloc_layout_hdr, 2380 .free_layout_hdr = ff_layout_free_layout_hdr, 2381 .alloc_lseg = ff_layout_alloc_lseg, 2382 .free_lseg = ff_layout_free_lseg, 2383 .add_lseg = ff_layout_add_lseg, 2384 .pg_read_ops = &ff_layout_pg_read_ops, 2385 .pg_write_ops = &ff_layout_pg_write_ops, 2386 .get_ds_info = ff_layout_get_ds_info, 2387 .free_deviceid_node = ff_layout_free_deviceid_node, 2388 .mark_request_commit = pnfs_layout_mark_request_commit, 2389 .clear_request_commit = pnfs_generic_clear_request_commit, 2390 .scan_commit_lists = pnfs_generic_scan_commit_lists, 2391 .recover_commit_reqs = pnfs_generic_recover_commit_reqs, 2392 .commit_pagelist = ff_layout_commit_pagelist, 2393 .read_pagelist = ff_layout_read_pagelist, 2394 .write_pagelist = ff_layout_write_pagelist, 2395 .alloc_deviceid_node = ff_layout_alloc_deviceid_node, 2396 .prepare_layoutreturn = ff_layout_prepare_layoutreturn, 2397 .sync = pnfs_nfs_generic_sync, 2398 .prepare_layoutstats = ff_layout_prepare_layoutstats, 2399 }; 2400 2401 static int __init nfs4flexfilelayout_init(void) 2402 { 2403 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n", 2404 __func__); 2405 if (!ff_zero_group) { 2406 ff_zero_group = groups_alloc(0); 2407 if (!ff_zero_group) 2408 return -ENOMEM; 2409 } 2410 return pnfs_register_layoutdriver(&flexfilelayout_type); 2411 } 2412 2413 static void __exit nfs4flexfilelayout_exit(void) 2414 { 2415 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n", 2416 __func__); 2417 pnfs_unregister_layoutdriver(&flexfilelayout_type); 2418 if (ff_zero_group) { 2419 put_group_info(ff_zero_group); 2420 ff_zero_group = NULL; 2421 } 2422 } 2423 2424 MODULE_ALIAS("nfs-layouttype4-4"); 2425 2426 MODULE_LICENSE("GPL"); 2427 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver"); 2428 2429 module_init(nfs4flexfilelayout_init); 2430 module_exit(nfs4flexfilelayout_exit); 2431