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 rc = -EIO; 458 p = xdr_inline_decode(&stream, 4); 459 if (!p) 460 goto out_err_free; 461 fh_count = be32_to_cpup(p); 462 463 fls->mirror_array[i]->fh_versions = 464 kzalloc(fh_count * sizeof(struct nfs_fh), 465 gfp_flags); 466 if (fls->mirror_array[i]->fh_versions == NULL) { 467 rc = -ENOMEM; 468 goto out_err_free; 469 } 470 471 for (j = 0; j < fh_count; j++) { 472 rc = decode_nfs_fh(&stream, 473 &fls->mirror_array[i]->fh_versions[j]); 474 if (rc) 475 goto out_err_free; 476 } 477 478 fls->mirror_array[i]->fh_versions_cnt = fh_count; 479 480 /* user */ 481 rc = decode_name(&stream, &id); 482 if (rc) 483 goto out_err_free; 484 485 acred.uid = make_kuid(&init_user_ns, id); 486 487 /* group */ 488 rc = decode_name(&stream, &id); 489 if (rc) 490 goto out_err_free; 491 492 acred.gid = make_kgid(&init_user_ns, id); 493 494 /* find the cred for it */ 495 rcu_assign_pointer(cred, rpc_lookup_generic_cred(&acred, 0, gfp_flags)); 496 if (IS_ERR(cred)) { 497 rc = PTR_ERR(cred); 498 goto out_err_free; 499 } 500 501 if (lgr->range.iomode == IOMODE_READ) 502 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred); 503 else 504 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred); 505 506 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]); 507 if (mirror != fls->mirror_array[i]) { 508 /* swap cred ptrs so free_mirror will clean up old */ 509 if (lgr->range.iomode == IOMODE_READ) { 510 cred = xchg(&mirror->ro_cred, cred); 511 rcu_assign_pointer(fls->mirror_array[i]->ro_cred, cred); 512 } else { 513 cred = xchg(&mirror->rw_cred, cred); 514 rcu_assign_pointer(fls->mirror_array[i]->rw_cred, cred); 515 } 516 ff_layout_free_mirror(fls->mirror_array[i]); 517 fls->mirror_array[i] = mirror; 518 } 519 520 dprintk("%s: iomode %s uid %u gid %u\n", __func__, 521 lgr->range.iomode == IOMODE_READ ? "READ" : "RW", 522 from_kuid(&init_user_ns, acred.uid), 523 from_kgid(&init_user_ns, acred.gid)); 524 } 525 526 p = xdr_inline_decode(&stream, 4); 527 if (!p) 528 goto out_sort_mirrors; 529 fls->flags = be32_to_cpup(p); 530 531 p = xdr_inline_decode(&stream, 4); 532 if (!p) 533 goto out_sort_mirrors; 534 for (i=0; i < fls->mirror_array_cnt; i++) 535 fls->mirror_array[i]->report_interval = be32_to_cpup(p); 536 537 out_sort_mirrors: 538 ff_layout_sort_mirrors(fls); 539 rc = ff_layout_check_layout(lgr); 540 if (rc) 541 goto out_err_free; 542 ret = &fls->generic_hdr; 543 dprintk("<-- %s (success)\n", __func__); 544 out_free_page: 545 __free_page(scratch); 546 return ret; 547 out_err_free: 548 _ff_layout_free_lseg(fls); 549 ret = ERR_PTR(rc); 550 dprintk("<-- %s (%d)\n", __func__, rc); 551 goto out_free_page; 552 } 553 554 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout) 555 { 556 struct pnfs_layout_segment *lseg; 557 558 list_for_each_entry(lseg, &layout->plh_segs, pls_list) 559 if (lseg->pls_range.iomode == IOMODE_RW) 560 return true; 561 562 return false; 563 } 564 565 static void 566 ff_layout_free_lseg(struct pnfs_layout_segment *lseg) 567 { 568 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 569 570 dprintk("--> %s\n", __func__); 571 572 if (lseg->pls_range.iomode == IOMODE_RW) { 573 struct nfs4_flexfile_layout *ffl; 574 struct inode *inode; 575 576 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout); 577 inode = ffl->generic_hdr.plh_inode; 578 spin_lock(&inode->i_lock); 579 if (!ff_layout_has_rw_segments(lseg->pls_layout)) { 580 ffl->commit_info.nbuckets = 0; 581 kfree(ffl->commit_info.buckets); 582 ffl->commit_info.buckets = NULL; 583 } 584 spin_unlock(&inode->i_lock); 585 } 586 _ff_layout_free_lseg(fls); 587 } 588 589 /* Return 1 until we have multiple lsegs support */ 590 static int 591 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls) 592 { 593 return 1; 594 } 595 596 static void 597 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now) 598 { 599 /* first IO request? */ 600 if (atomic_inc_return(&timer->n_ops) == 1) { 601 timer->start_time = now; 602 } 603 } 604 605 static ktime_t 606 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now) 607 { 608 ktime_t start; 609 610 if (atomic_dec_return(&timer->n_ops) < 0) 611 WARN_ON_ONCE(1); 612 613 start = timer->start_time; 614 timer->start_time = now; 615 return ktime_sub(now, start); 616 } 617 618 static bool 619 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror, 620 struct nfs4_ff_layoutstat *layoutstat, 621 ktime_t now) 622 { 623 s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL; 624 struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(mirror->layout); 625 626 nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now); 627 if (!mirror->start_time) 628 mirror->start_time = now; 629 if (mirror->report_interval != 0) 630 report_interval = (s64)mirror->report_interval * 1000LL; 631 else if (layoutstats_timer != 0) 632 report_interval = (s64)layoutstats_timer * 1000LL; 633 if (ktime_to_ms(ktime_sub(now, ffl->last_report_time)) >= 634 report_interval) { 635 ffl->last_report_time = now; 636 return true; 637 } 638 639 return false; 640 } 641 642 static void 643 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat, 644 __u64 requested) 645 { 646 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat; 647 648 iostat->ops_requested++; 649 iostat->bytes_requested += requested; 650 } 651 652 static void 653 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat, 654 __u64 requested, 655 __u64 completed, 656 ktime_t time_completed, 657 ktime_t time_started) 658 { 659 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat; 660 ktime_t completion_time = ktime_sub(time_completed, time_started); 661 ktime_t timer; 662 663 iostat->ops_completed++; 664 iostat->bytes_completed += completed; 665 iostat->bytes_not_delivered += requested - completed; 666 667 timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed); 668 iostat->total_busy_time = 669 ktime_add(iostat->total_busy_time, timer); 670 iostat->aggregate_completion_time = 671 ktime_add(iostat->aggregate_completion_time, 672 completion_time); 673 } 674 675 static void 676 nfs4_ff_layout_stat_io_start_read(struct inode *inode, 677 struct nfs4_ff_layout_mirror *mirror, 678 __u64 requested, ktime_t now) 679 { 680 bool report; 681 682 spin_lock(&mirror->lock); 683 report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now); 684 nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested); 685 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 686 spin_unlock(&mirror->lock); 687 688 if (report) 689 pnfs_report_layoutstat(inode, GFP_KERNEL); 690 } 691 692 static void 693 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task, 694 struct nfs4_ff_layout_mirror *mirror, 695 __u64 requested, 696 __u64 completed) 697 { 698 spin_lock(&mirror->lock); 699 nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat, 700 requested, completed, 701 ktime_get(), task->tk_start); 702 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 703 spin_unlock(&mirror->lock); 704 } 705 706 static void 707 nfs4_ff_layout_stat_io_start_write(struct inode *inode, 708 struct nfs4_ff_layout_mirror *mirror, 709 __u64 requested, ktime_t now) 710 { 711 bool report; 712 713 spin_lock(&mirror->lock); 714 report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now); 715 nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested); 716 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 717 spin_unlock(&mirror->lock); 718 719 if (report) 720 pnfs_report_layoutstat(inode, GFP_NOIO); 721 } 722 723 static void 724 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task, 725 struct nfs4_ff_layout_mirror *mirror, 726 __u64 requested, 727 __u64 completed, 728 enum nfs3_stable_how committed) 729 { 730 if (committed == NFS_UNSTABLE) 731 requested = completed = 0; 732 733 spin_lock(&mirror->lock); 734 nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat, 735 requested, completed, ktime_get(), task->tk_start); 736 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 737 spin_unlock(&mirror->lock); 738 } 739 740 static int 741 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg, 742 struct nfs_commit_info *cinfo, 743 gfp_t gfp_flags) 744 { 745 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 746 struct pnfs_commit_bucket *buckets; 747 int size; 748 749 if (cinfo->ds->nbuckets != 0) { 750 /* This assumes there is only one RW lseg per file. 751 * To support multiple lseg per file, we need to 752 * change struct pnfs_commit_bucket to allow dynamic 753 * increasing nbuckets. 754 */ 755 return 0; 756 } 757 758 size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg); 759 760 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), 761 gfp_flags); 762 if (!buckets) 763 return -ENOMEM; 764 else { 765 int i; 766 767 spin_lock(&cinfo->inode->i_lock); 768 if (cinfo->ds->nbuckets != 0) 769 kfree(buckets); 770 else { 771 cinfo->ds->buckets = buckets; 772 cinfo->ds->nbuckets = size; 773 for (i = 0; i < size; i++) { 774 INIT_LIST_HEAD(&buckets[i].written); 775 INIT_LIST_HEAD(&buckets[i].committing); 776 /* mark direct verifier as unset */ 777 buckets[i].direct_verf.committed = 778 NFS_INVALID_STABLE_HOW; 779 } 780 } 781 spin_unlock(&cinfo->inode->i_lock); 782 return 0; 783 } 784 } 785 786 static struct nfs4_pnfs_ds * 787 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg, 788 int start_idx, 789 int *best_idx) 790 { 791 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 792 struct nfs4_pnfs_ds *ds; 793 bool fail_return = false; 794 int idx; 795 796 /* mirrors are sorted by efficiency */ 797 for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) { 798 if (idx+1 == fls->mirror_array_cnt) 799 fail_return = true; 800 ds = nfs4_ff_layout_prepare_ds(lseg, idx, fail_return); 801 if (ds) { 802 *best_idx = idx; 803 return ds; 804 } 805 } 806 807 return NULL; 808 } 809 810 static void 811 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio, 812 struct nfs_page *req, 813 bool strict_iomode) 814 { 815 retry_strict: 816 pnfs_put_lseg(pgio->pg_lseg); 817 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 818 req->wb_context, 819 0, 820 NFS4_MAX_UINT64, 821 IOMODE_READ, 822 strict_iomode, 823 GFP_KERNEL); 824 if (IS_ERR(pgio->pg_lseg)) { 825 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 826 pgio->pg_lseg = NULL; 827 } 828 829 /* If we don't have checking, do get a IOMODE_RW 830 * segment, and the server wants to avoid READs 831 * there, then retry! 832 */ 833 if (pgio->pg_lseg && !strict_iomode && 834 ff_layout_avoid_read_on_rw(pgio->pg_lseg)) { 835 strict_iomode = true; 836 goto retry_strict; 837 } 838 } 839 840 static void 841 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio, 842 struct nfs_page *req) 843 { 844 struct nfs_pgio_mirror *pgm; 845 struct nfs4_ff_layout_mirror *mirror; 846 struct nfs4_pnfs_ds *ds; 847 int ds_idx; 848 849 retry: 850 pnfs_generic_pg_check_layout(pgio); 851 /* Use full layout for now */ 852 if (!pgio->pg_lseg) 853 ff_layout_pg_get_read(pgio, req, false); 854 else if (ff_layout_avoid_read_on_rw(pgio->pg_lseg)) 855 ff_layout_pg_get_read(pgio, req, true); 856 857 /* If no lseg, fall back to read through mds */ 858 if (pgio->pg_lseg == NULL) 859 goto out_mds; 860 861 ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx); 862 if (!ds) { 863 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg)) 864 goto out_mds; 865 pnfs_put_lseg(pgio->pg_lseg); 866 pgio->pg_lseg = NULL; 867 /* Sleep for 1 second before retrying */ 868 ssleep(1); 869 goto retry; 870 } 871 872 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx); 873 874 pgio->pg_mirror_idx = ds_idx; 875 876 /* read always uses only one mirror - idx 0 for pgio layer */ 877 pgm = &pgio->pg_mirrors[0]; 878 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize; 879 880 return; 881 out_mds: 882 pnfs_put_lseg(pgio->pg_lseg); 883 pgio->pg_lseg = NULL; 884 nfs_pageio_reset_read_mds(pgio); 885 } 886 887 static void 888 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio, 889 struct nfs_page *req) 890 { 891 struct nfs4_ff_layout_mirror *mirror; 892 struct nfs_pgio_mirror *pgm; 893 struct nfs_commit_info cinfo; 894 struct nfs4_pnfs_ds *ds; 895 int i; 896 int status; 897 898 retry: 899 pnfs_generic_pg_check_layout(pgio); 900 if (!pgio->pg_lseg) { 901 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 902 req->wb_context, 903 0, 904 NFS4_MAX_UINT64, 905 IOMODE_RW, 906 false, 907 GFP_NOFS); 908 if (IS_ERR(pgio->pg_lseg)) { 909 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 910 pgio->pg_lseg = NULL; 911 return; 912 } 913 } 914 /* If no lseg, fall back to write through mds */ 915 if (pgio->pg_lseg == NULL) 916 goto out_mds; 917 918 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); 919 status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); 920 if (status < 0) 921 goto out_mds; 922 923 /* Use a direct mapping of ds_idx to pgio mirror_idx */ 924 if (WARN_ON_ONCE(pgio->pg_mirror_count != 925 FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg))) 926 goto out_mds; 927 928 for (i = 0; i < pgio->pg_mirror_count; i++) { 929 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true); 930 if (!ds) { 931 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg)) 932 goto out_mds; 933 pnfs_put_lseg(pgio->pg_lseg); 934 pgio->pg_lseg = NULL; 935 /* Sleep for 1 second before retrying */ 936 ssleep(1); 937 goto retry; 938 } 939 pgm = &pgio->pg_mirrors[i]; 940 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i); 941 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize; 942 } 943 944 return; 945 946 out_mds: 947 pnfs_put_lseg(pgio->pg_lseg); 948 pgio->pg_lseg = NULL; 949 nfs_pageio_reset_write_mds(pgio); 950 } 951 952 static unsigned int 953 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio, 954 struct nfs_page *req) 955 { 956 if (!pgio->pg_lseg) { 957 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 958 req->wb_context, 959 0, 960 NFS4_MAX_UINT64, 961 IOMODE_RW, 962 false, 963 GFP_NOFS); 964 if (IS_ERR(pgio->pg_lseg)) { 965 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 966 pgio->pg_lseg = NULL; 967 goto out; 968 } 969 } 970 if (pgio->pg_lseg) 971 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg); 972 973 /* no lseg means that pnfs is not in use, so no mirroring here */ 974 nfs_pageio_reset_write_mds(pgio); 975 out: 976 return 1; 977 } 978 979 static const struct nfs_pageio_ops ff_layout_pg_read_ops = { 980 .pg_init = ff_layout_pg_init_read, 981 .pg_test = pnfs_generic_pg_test, 982 .pg_doio = pnfs_generic_pg_readpages, 983 .pg_cleanup = pnfs_generic_pg_cleanup, 984 }; 985 986 static const struct nfs_pageio_ops ff_layout_pg_write_ops = { 987 .pg_init = ff_layout_pg_init_write, 988 .pg_test = pnfs_generic_pg_test, 989 .pg_doio = pnfs_generic_pg_writepages, 990 .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write, 991 .pg_cleanup = pnfs_generic_pg_cleanup, 992 }; 993 994 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs) 995 { 996 struct rpc_task *task = &hdr->task; 997 998 pnfs_layoutcommit_inode(hdr->inode, false); 999 1000 if (retry_pnfs) { 1001 dprintk("%s Reset task %5u for i/o through pNFS " 1002 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1003 hdr->task.tk_pid, 1004 hdr->inode->i_sb->s_id, 1005 (unsigned long long)NFS_FILEID(hdr->inode), 1006 hdr->args.count, 1007 (unsigned long long)hdr->args.offset); 1008 1009 hdr->completion_ops->reschedule_io(hdr); 1010 return; 1011 } 1012 1013 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1014 dprintk("%s Reset task %5u for i/o through MDS " 1015 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1016 hdr->task.tk_pid, 1017 hdr->inode->i_sb->s_id, 1018 (unsigned long long)NFS_FILEID(hdr->inode), 1019 hdr->args.count, 1020 (unsigned long long)hdr->args.offset); 1021 1022 task->tk_status = pnfs_write_done_resend_to_mds(hdr); 1023 } 1024 } 1025 1026 static void ff_layout_reset_read(struct nfs_pgio_header *hdr) 1027 { 1028 struct rpc_task *task = &hdr->task; 1029 1030 pnfs_layoutcommit_inode(hdr->inode, false); 1031 1032 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1033 dprintk("%s Reset task %5u for i/o through MDS " 1034 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1035 hdr->task.tk_pid, 1036 hdr->inode->i_sb->s_id, 1037 (unsigned long long)NFS_FILEID(hdr->inode), 1038 hdr->args.count, 1039 (unsigned long long)hdr->args.offset); 1040 1041 task->tk_status = pnfs_read_done_resend_to_mds(hdr); 1042 } 1043 } 1044 1045 static int ff_layout_async_handle_error_v4(struct rpc_task *task, 1046 struct nfs4_state *state, 1047 struct nfs_client *clp, 1048 struct pnfs_layout_segment *lseg, 1049 int idx) 1050 { 1051 struct pnfs_layout_hdr *lo = lseg->pls_layout; 1052 struct inode *inode = lo->plh_inode; 1053 struct nfs_server *mds_server = NFS_SERVER(inode); 1054 1055 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx); 1056 struct nfs_client *mds_client = mds_server->nfs_client; 1057 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; 1058 1059 switch (task->tk_status) { 1060 /* MDS state errors */ 1061 case -NFS4ERR_DELEG_REVOKED: 1062 case -NFS4ERR_ADMIN_REVOKED: 1063 case -NFS4ERR_BAD_STATEID: 1064 if (state == NULL) 1065 break; 1066 nfs_remove_bad_delegation(state->inode, NULL); 1067 case -NFS4ERR_OPENMODE: 1068 if (state == NULL) 1069 break; 1070 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0) 1071 goto out_bad_stateid; 1072 goto wait_on_recovery; 1073 case -NFS4ERR_EXPIRED: 1074 if (state != NULL) { 1075 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0) 1076 goto out_bad_stateid; 1077 } 1078 nfs4_schedule_lease_recovery(mds_client); 1079 goto wait_on_recovery; 1080 /* DS session errors */ 1081 case -NFS4ERR_BADSESSION: 1082 case -NFS4ERR_BADSLOT: 1083 case -NFS4ERR_BAD_HIGH_SLOT: 1084 case -NFS4ERR_DEADSESSION: 1085 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: 1086 case -NFS4ERR_SEQ_FALSE_RETRY: 1087 case -NFS4ERR_SEQ_MISORDERED: 1088 dprintk("%s ERROR %d, Reset session. Exchangeid " 1089 "flags 0x%x\n", __func__, task->tk_status, 1090 clp->cl_exchange_flags); 1091 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status); 1092 break; 1093 case -NFS4ERR_DELAY: 1094 case -NFS4ERR_GRACE: 1095 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX); 1096 break; 1097 case -NFS4ERR_RETRY_UNCACHED_REP: 1098 break; 1099 /* Invalidate Layout errors */ 1100 case -NFS4ERR_PNFS_NO_LAYOUT: 1101 case -ESTALE: /* mapped NFS4ERR_STALE */ 1102 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */ 1103 case -EISDIR: /* mapped NFS4ERR_ISDIR */ 1104 case -NFS4ERR_FHEXPIRED: 1105 case -NFS4ERR_WRONG_TYPE: 1106 dprintk("%s Invalid layout error %d\n", __func__, 1107 task->tk_status); 1108 /* 1109 * Destroy layout so new i/o will get a new layout. 1110 * Layout will not be destroyed until all current lseg 1111 * references are put. Mark layout as invalid to resend failed 1112 * i/o and all i/o waiting on the slot table to the MDS until 1113 * layout is destroyed and a new valid layout is obtained. 1114 */ 1115 pnfs_destroy_layout(NFS_I(inode)); 1116 rpc_wake_up(&tbl->slot_tbl_waitq); 1117 goto reset; 1118 /* RPC connection errors */ 1119 case -ECONNREFUSED: 1120 case -EHOSTDOWN: 1121 case -EHOSTUNREACH: 1122 case -ENETUNREACH: 1123 case -EIO: 1124 case -ETIMEDOUT: 1125 case -EPIPE: 1126 dprintk("%s DS connection error %d\n", __func__, 1127 task->tk_status); 1128 nfs4_delete_deviceid(devid->ld, devid->nfs_client, 1129 &devid->deviceid); 1130 rpc_wake_up(&tbl->slot_tbl_waitq); 1131 /* fall through */ 1132 default: 1133 if (ff_layout_avoid_mds_available_ds(lseg)) 1134 return -NFS4ERR_RESET_TO_PNFS; 1135 reset: 1136 dprintk("%s Retry through MDS. Error %d\n", __func__, 1137 task->tk_status); 1138 return -NFS4ERR_RESET_TO_MDS; 1139 } 1140 out: 1141 task->tk_status = 0; 1142 return -EAGAIN; 1143 out_bad_stateid: 1144 task->tk_status = -EIO; 1145 return 0; 1146 wait_on_recovery: 1147 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL); 1148 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0) 1149 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task); 1150 goto out; 1151 } 1152 1153 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */ 1154 static int ff_layout_async_handle_error_v3(struct rpc_task *task, 1155 struct pnfs_layout_segment *lseg, 1156 int idx) 1157 { 1158 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx); 1159 1160 switch (task->tk_status) { 1161 /* File access problems. Don't mark the device as unavailable */ 1162 case -EACCES: 1163 case -ESTALE: 1164 case -EISDIR: 1165 case -EBADHANDLE: 1166 case -ELOOP: 1167 case -ENOSPC: 1168 break; 1169 case -EJUKEBOX: 1170 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY); 1171 goto out_retry; 1172 default: 1173 dprintk("%s DS connection error %d\n", __func__, 1174 task->tk_status); 1175 nfs4_delete_deviceid(devid->ld, devid->nfs_client, 1176 &devid->deviceid); 1177 } 1178 /* FIXME: Need to prevent infinite looping here. */ 1179 return -NFS4ERR_RESET_TO_PNFS; 1180 out_retry: 1181 task->tk_status = 0; 1182 rpc_restart_call_prepare(task); 1183 rpc_delay(task, NFS_JUKEBOX_RETRY_TIME); 1184 return -EAGAIN; 1185 } 1186 1187 static int ff_layout_async_handle_error(struct rpc_task *task, 1188 struct nfs4_state *state, 1189 struct nfs_client *clp, 1190 struct pnfs_layout_segment *lseg, 1191 int idx) 1192 { 1193 int vers = clp->cl_nfs_mod->rpc_vers->number; 1194 1195 if (task->tk_status >= 0) 1196 return 0; 1197 1198 /* Handle the case of an invalid layout segment */ 1199 if (!pnfs_is_valid_lseg(lseg)) 1200 return -NFS4ERR_RESET_TO_PNFS; 1201 1202 switch (vers) { 1203 case 3: 1204 return ff_layout_async_handle_error_v3(task, lseg, idx); 1205 case 4: 1206 return ff_layout_async_handle_error_v4(task, state, clp, 1207 lseg, idx); 1208 default: 1209 /* should never happen */ 1210 WARN_ON_ONCE(1); 1211 return 0; 1212 } 1213 } 1214 1215 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg, 1216 int idx, u64 offset, u64 length, 1217 u32 status, int opnum, int error) 1218 { 1219 struct nfs4_ff_layout_mirror *mirror; 1220 int err; 1221 1222 if (status == 0) { 1223 switch (error) { 1224 case -ETIMEDOUT: 1225 case -EPFNOSUPPORT: 1226 case -EPROTONOSUPPORT: 1227 case -EOPNOTSUPP: 1228 case -ECONNREFUSED: 1229 case -ECONNRESET: 1230 case -EHOSTDOWN: 1231 case -EHOSTUNREACH: 1232 case -ENETUNREACH: 1233 case -EADDRINUSE: 1234 case -ENOBUFS: 1235 case -EPIPE: 1236 case -EPERM: 1237 status = NFS4ERR_NXIO; 1238 break; 1239 case -EACCES: 1240 status = NFS4ERR_ACCESS; 1241 break; 1242 default: 1243 return; 1244 } 1245 } 1246 1247 switch (status) { 1248 case NFS4ERR_DELAY: 1249 case NFS4ERR_GRACE: 1250 return; 1251 default: 1252 break; 1253 } 1254 1255 mirror = FF_LAYOUT_COMP(lseg, idx); 1256 err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout), 1257 mirror, offset, length, status, opnum, 1258 GFP_NOIO); 1259 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg); 1260 dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status); 1261 } 1262 1263 /* NFS_PROTO call done callback routines */ 1264 static int ff_layout_read_done_cb(struct rpc_task *task, 1265 struct nfs_pgio_header *hdr) 1266 { 1267 int err; 1268 1269 trace_nfs4_pnfs_read(hdr, task->tk_status); 1270 if (task->tk_status < 0) 1271 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx, 1272 hdr->args.offset, hdr->args.count, 1273 hdr->res.op_status, OP_READ, 1274 task->tk_status); 1275 err = ff_layout_async_handle_error(task, hdr->args.context->state, 1276 hdr->ds_clp, hdr->lseg, 1277 hdr->pgio_mirror_idx); 1278 1279 switch (err) { 1280 case -NFS4ERR_RESET_TO_PNFS: 1281 if (ff_layout_choose_best_ds_for_read(hdr->lseg, 1282 hdr->pgio_mirror_idx + 1, 1283 &hdr->pgio_mirror_idx)) 1284 goto out_eagain; 1285 ff_layout_read_record_layoutstats_done(task, hdr); 1286 pnfs_read_resend_pnfs(hdr); 1287 return task->tk_status; 1288 case -NFS4ERR_RESET_TO_MDS: 1289 ff_layout_reset_read(hdr); 1290 return task->tk_status; 1291 case -EAGAIN: 1292 goto out_eagain; 1293 } 1294 1295 return 0; 1296 out_eagain: 1297 rpc_restart_call_prepare(task); 1298 return -EAGAIN; 1299 } 1300 1301 static bool 1302 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg) 1303 { 1304 return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT); 1305 } 1306 1307 /* 1308 * We reference the rpc_cred of the first WRITE that triggers the need for 1309 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. 1310 * rfc5661 is not clear about which credential should be used. 1311 * 1312 * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so 1313 * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751 1314 * we always send layoutcommit after DS writes. 1315 */ 1316 static void 1317 ff_layout_set_layoutcommit(struct inode *inode, 1318 struct pnfs_layout_segment *lseg, 1319 loff_t end_offset) 1320 { 1321 if (!ff_layout_need_layoutcommit(lseg)) 1322 return; 1323 1324 pnfs_set_layoutcommit(inode, lseg, end_offset); 1325 dprintk("%s inode %lu pls_end_pos %llu\n", __func__, inode->i_ino, 1326 (unsigned long long) NFS_I(inode)->layout->plh_lwb); 1327 } 1328 1329 static bool 1330 ff_layout_device_unavailable(struct pnfs_layout_segment *lseg, int idx) 1331 { 1332 /* No mirroring for now */ 1333 struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx); 1334 1335 return ff_layout_test_devid_unavailable(node); 1336 } 1337 1338 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task, 1339 struct nfs_pgio_header *hdr) 1340 { 1341 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags)) 1342 return; 1343 nfs4_ff_layout_stat_io_start_read(hdr->inode, 1344 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1345 hdr->args.count, 1346 task->tk_start); 1347 } 1348 1349 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task, 1350 struct nfs_pgio_header *hdr) 1351 { 1352 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags)) 1353 return; 1354 nfs4_ff_layout_stat_io_end_read(task, 1355 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1356 hdr->args.count, 1357 hdr->res.count); 1358 } 1359 1360 static int ff_layout_read_prepare_common(struct rpc_task *task, 1361 struct nfs_pgio_header *hdr) 1362 { 1363 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { 1364 rpc_exit(task, -EIO); 1365 return -EIO; 1366 } 1367 if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) { 1368 rpc_exit(task, -EHOSTDOWN); 1369 return -EAGAIN; 1370 } 1371 1372 ff_layout_read_record_layoutstats_start(task, hdr); 1373 return 0; 1374 } 1375 1376 /* 1377 * Call ops for the async read/write cases 1378 * In the case of dense layouts, the offset needs to be reset to its 1379 * original value. 1380 */ 1381 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data) 1382 { 1383 struct nfs_pgio_header *hdr = data; 1384 1385 if (ff_layout_read_prepare_common(task, hdr)) 1386 return; 1387 1388 rpc_call_start(task); 1389 } 1390 1391 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data) 1392 { 1393 struct nfs_pgio_header *hdr = data; 1394 1395 if (nfs4_setup_sequence(hdr->ds_clp, 1396 &hdr->args.seq_args, 1397 &hdr->res.seq_res, 1398 task)) 1399 return; 1400 1401 if (ff_layout_read_prepare_common(task, hdr)) 1402 return; 1403 1404 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context, 1405 hdr->args.lock_context, FMODE_READ) == -EIO) 1406 rpc_exit(task, -EIO); /* lost lock, terminate I/O */ 1407 } 1408 1409 static void ff_layout_read_call_done(struct rpc_task *task, void *data) 1410 { 1411 struct nfs_pgio_header *hdr = data; 1412 1413 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); 1414 1415 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && 1416 task->tk_status == 0) { 1417 nfs4_sequence_done(task, &hdr->res.seq_res); 1418 return; 1419 } 1420 1421 /* Note this may cause RPC to be resent */ 1422 hdr->mds_ops->rpc_call_done(task, hdr); 1423 } 1424 1425 static void ff_layout_read_count_stats(struct rpc_task *task, void *data) 1426 { 1427 struct nfs_pgio_header *hdr = data; 1428 1429 ff_layout_read_record_layoutstats_done(task, hdr); 1430 rpc_count_iostats_metrics(task, 1431 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]); 1432 } 1433 1434 static void ff_layout_read_release(void *data) 1435 { 1436 struct nfs_pgio_header *hdr = data; 1437 1438 ff_layout_read_record_layoutstats_done(&hdr->task, hdr); 1439 pnfs_generic_rw_release(data); 1440 } 1441 1442 1443 static int ff_layout_write_done_cb(struct rpc_task *task, 1444 struct nfs_pgio_header *hdr) 1445 { 1446 loff_t end_offs = 0; 1447 int err; 1448 1449 trace_nfs4_pnfs_write(hdr, task->tk_status); 1450 if (task->tk_status < 0) 1451 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx, 1452 hdr->args.offset, hdr->args.count, 1453 hdr->res.op_status, OP_WRITE, 1454 task->tk_status); 1455 err = ff_layout_async_handle_error(task, hdr->args.context->state, 1456 hdr->ds_clp, hdr->lseg, 1457 hdr->pgio_mirror_idx); 1458 1459 switch (err) { 1460 case -NFS4ERR_RESET_TO_PNFS: 1461 ff_layout_reset_write(hdr, true); 1462 return task->tk_status; 1463 case -NFS4ERR_RESET_TO_MDS: 1464 ff_layout_reset_write(hdr, false); 1465 return task->tk_status; 1466 case -EAGAIN: 1467 return -EAGAIN; 1468 } 1469 1470 if (hdr->res.verf->committed == NFS_FILE_SYNC || 1471 hdr->res.verf->committed == NFS_DATA_SYNC) 1472 end_offs = hdr->mds_offset + (loff_t)hdr->res.count; 1473 1474 /* Note: if the write is unstable, don't set end_offs until commit */ 1475 ff_layout_set_layoutcommit(hdr->inode, hdr->lseg, end_offs); 1476 1477 /* zero out fattr since we don't care DS attr at all */ 1478 hdr->fattr.valid = 0; 1479 if (task->tk_status >= 0) 1480 nfs_writeback_update_inode(hdr); 1481 1482 return 0; 1483 } 1484 1485 static int ff_layout_commit_done_cb(struct rpc_task *task, 1486 struct nfs_commit_data *data) 1487 { 1488 int err; 1489 1490 trace_nfs4_pnfs_commit_ds(data, task->tk_status); 1491 if (task->tk_status < 0) 1492 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index, 1493 data->args.offset, data->args.count, 1494 data->res.op_status, OP_COMMIT, 1495 task->tk_status); 1496 err = ff_layout_async_handle_error(task, NULL, data->ds_clp, 1497 data->lseg, data->ds_commit_index); 1498 1499 switch (err) { 1500 case -NFS4ERR_RESET_TO_PNFS: 1501 pnfs_generic_prepare_to_resend_writes(data); 1502 return -EAGAIN; 1503 case -NFS4ERR_RESET_TO_MDS: 1504 pnfs_generic_prepare_to_resend_writes(data); 1505 return -EAGAIN; 1506 case -EAGAIN: 1507 rpc_restart_call_prepare(task); 1508 return -EAGAIN; 1509 } 1510 1511 ff_layout_set_layoutcommit(data->inode, data->lseg, data->lwb); 1512 1513 return 0; 1514 } 1515 1516 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task, 1517 struct nfs_pgio_header *hdr) 1518 { 1519 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags)) 1520 return; 1521 nfs4_ff_layout_stat_io_start_write(hdr->inode, 1522 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1523 hdr->args.count, 1524 task->tk_start); 1525 } 1526 1527 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task, 1528 struct nfs_pgio_header *hdr) 1529 { 1530 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags)) 1531 return; 1532 nfs4_ff_layout_stat_io_end_write(task, 1533 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx), 1534 hdr->args.count, hdr->res.count, 1535 hdr->res.verf->committed); 1536 } 1537 1538 static int ff_layout_write_prepare_common(struct rpc_task *task, 1539 struct nfs_pgio_header *hdr) 1540 { 1541 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { 1542 rpc_exit(task, -EIO); 1543 return -EIO; 1544 } 1545 1546 if (ff_layout_device_unavailable(hdr->lseg, hdr->pgio_mirror_idx)) { 1547 rpc_exit(task, -EHOSTDOWN); 1548 return -EAGAIN; 1549 } 1550 1551 ff_layout_write_record_layoutstats_start(task, hdr); 1552 return 0; 1553 } 1554 1555 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data) 1556 { 1557 struct nfs_pgio_header *hdr = data; 1558 1559 if (ff_layout_write_prepare_common(task, hdr)) 1560 return; 1561 1562 rpc_call_start(task); 1563 } 1564 1565 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data) 1566 { 1567 struct nfs_pgio_header *hdr = data; 1568 1569 if (nfs4_setup_sequence(hdr->ds_clp, 1570 &hdr->args.seq_args, 1571 &hdr->res.seq_res, 1572 task)) 1573 return; 1574 1575 if (ff_layout_write_prepare_common(task, hdr)) 1576 return; 1577 1578 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context, 1579 hdr->args.lock_context, FMODE_WRITE) == -EIO) 1580 rpc_exit(task, -EIO); /* lost lock, terminate I/O */ 1581 } 1582 1583 static void ff_layout_write_call_done(struct rpc_task *task, void *data) 1584 { 1585 struct nfs_pgio_header *hdr = data; 1586 1587 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && 1588 task->tk_status == 0) { 1589 nfs4_sequence_done(task, &hdr->res.seq_res); 1590 return; 1591 } 1592 1593 /* Note this may cause RPC to be resent */ 1594 hdr->mds_ops->rpc_call_done(task, hdr); 1595 } 1596 1597 static void ff_layout_write_count_stats(struct rpc_task *task, void *data) 1598 { 1599 struct nfs_pgio_header *hdr = data; 1600 1601 ff_layout_write_record_layoutstats_done(task, hdr); 1602 rpc_count_iostats_metrics(task, 1603 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]); 1604 } 1605 1606 static void ff_layout_write_release(void *data) 1607 { 1608 struct nfs_pgio_header *hdr = data; 1609 1610 ff_layout_write_record_layoutstats_done(&hdr->task, hdr); 1611 pnfs_generic_rw_release(data); 1612 } 1613 1614 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task, 1615 struct nfs_commit_data *cdata) 1616 { 1617 if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags)) 1618 return; 1619 nfs4_ff_layout_stat_io_start_write(cdata->inode, 1620 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index), 1621 0, task->tk_start); 1622 } 1623 1624 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task, 1625 struct nfs_commit_data *cdata) 1626 { 1627 struct nfs_page *req; 1628 __u64 count = 0; 1629 1630 if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags)) 1631 return; 1632 1633 if (task->tk_status == 0) { 1634 list_for_each_entry(req, &cdata->pages, wb_list) 1635 count += req->wb_bytes; 1636 } 1637 nfs4_ff_layout_stat_io_end_write(task, 1638 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index), 1639 count, count, NFS_FILE_SYNC); 1640 } 1641 1642 static void ff_layout_commit_prepare_common(struct rpc_task *task, 1643 struct nfs_commit_data *cdata) 1644 { 1645 ff_layout_commit_record_layoutstats_start(task, cdata); 1646 } 1647 1648 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data) 1649 { 1650 ff_layout_commit_prepare_common(task, data); 1651 rpc_call_start(task); 1652 } 1653 1654 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data) 1655 { 1656 struct nfs_commit_data *wdata = data; 1657 1658 if (nfs4_setup_sequence(wdata->ds_clp, 1659 &wdata->args.seq_args, 1660 &wdata->res.seq_res, 1661 task)) 1662 return; 1663 ff_layout_commit_prepare_common(task, data); 1664 } 1665 1666 static void ff_layout_commit_done(struct rpc_task *task, void *data) 1667 { 1668 pnfs_generic_write_commit_done(task, data); 1669 } 1670 1671 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data) 1672 { 1673 struct nfs_commit_data *cdata = data; 1674 1675 ff_layout_commit_record_layoutstats_done(task, cdata); 1676 rpc_count_iostats_metrics(task, 1677 &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]); 1678 } 1679 1680 static void ff_layout_commit_release(void *data) 1681 { 1682 struct nfs_commit_data *cdata = data; 1683 1684 ff_layout_commit_record_layoutstats_done(&cdata->task, cdata); 1685 pnfs_generic_commit_release(data); 1686 } 1687 1688 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = { 1689 .rpc_call_prepare = ff_layout_read_prepare_v3, 1690 .rpc_call_done = ff_layout_read_call_done, 1691 .rpc_count_stats = ff_layout_read_count_stats, 1692 .rpc_release = ff_layout_read_release, 1693 }; 1694 1695 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = { 1696 .rpc_call_prepare = ff_layout_read_prepare_v4, 1697 .rpc_call_done = ff_layout_read_call_done, 1698 .rpc_count_stats = ff_layout_read_count_stats, 1699 .rpc_release = ff_layout_read_release, 1700 }; 1701 1702 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = { 1703 .rpc_call_prepare = ff_layout_write_prepare_v3, 1704 .rpc_call_done = ff_layout_write_call_done, 1705 .rpc_count_stats = ff_layout_write_count_stats, 1706 .rpc_release = ff_layout_write_release, 1707 }; 1708 1709 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = { 1710 .rpc_call_prepare = ff_layout_write_prepare_v4, 1711 .rpc_call_done = ff_layout_write_call_done, 1712 .rpc_count_stats = ff_layout_write_count_stats, 1713 .rpc_release = ff_layout_write_release, 1714 }; 1715 1716 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = { 1717 .rpc_call_prepare = ff_layout_commit_prepare_v3, 1718 .rpc_call_done = ff_layout_commit_done, 1719 .rpc_count_stats = ff_layout_commit_count_stats, 1720 .rpc_release = ff_layout_commit_release, 1721 }; 1722 1723 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = { 1724 .rpc_call_prepare = ff_layout_commit_prepare_v4, 1725 .rpc_call_done = ff_layout_commit_done, 1726 .rpc_count_stats = ff_layout_commit_count_stats, 1727 .rpc_release = ff_layout_commit_release, 1728 }; 1729 1730 static enum pnfs_try_status 1731 ff_layout_read_pagelist(struct nfs_pgio_header *hdr) 1732 { 1733 struct pnfs_layout_segment *lseg = hdr->lseg; 1734 struct nfs4_pnfs_ds *ds; 1735 struct rpc_clnt *ds_clnt; 1736 struct rpc_cred *ds_cred; 1737 loff_t offset = hdr->args.offset; 1738 u32 idx = hdr->pgio_mirror_idx; 1739 int vers; 1740 struct nfs_fh *fh; 1741 1742 dprintk("--> %s ino %lu pgbase %u req %zu@%llu\n", 1743 __func__, hdr->inode->i_ino, 1744 hdr->args.pgbase, (size_t)hdr->args.count, offset); 1745 1746 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false); 1747 if (!ds) 1748 goto out_failed; 1749 1750 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp, 1751 hdr->inode); 1752 if (IS_ERR(ds_clnt)) 1753 goto out_failed; 1754 1755 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred); 1756 if (!ds_cred) 1757 goto out_failed; 1758 1759 vers = nfs4_ff_layout_ds_version(lseg, idx); 1760 1761 dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__, 1762 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers); 1763 1764 hdr->pgio_done_cb = ff_layout_read_done_cb; 1765 atomic_inc(&ds->ds_clp->cl_count); 1766 hdr->ds_clp = ds->ds_clp; 1767 fh = nfs4_ff_layout_select_ds_fh(lseg, idx); 1768 if (fh) 1769 hdr->args.fh = fh; 1770 /* 1771 * Note that if we ever decide to split across DSes, 1772 * then we may need to handle dense-like offsets. 1773 */ 1774 hdr->args.offset = offset; 1775 hdr->mds_offset = offset; 1776 1777 /* Perform an asynchronous read to ds */ 1778 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops, 1779 vers == 3 ? &ff_layout_read_call_ops_v3 : 1780 &ff_layout_read_call_ops_v4, 1781 0, RPC_TASK_SOFTCONN); 1782 put_rpccred(ds_cred); 1783 return PNFS_ATTEMPTED; 1784 1785 out_failed: 1786 if (ff_layout_avoid_mds_available_ds(lseg)) 1787 return PNFS_TRY_AGAIN; 1788 return PNFS_NOT_ATTEMPTED; 1789 } 1790 1791 /* Perform async writes. */ 1792 static enum pnfs_try_status 1793 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync) 1794 { 1795 struct pnfs_layout_segment *lseg = hdr->lseg; 1796 struct nfs4_pnfs_ds *ds; 1797 struct rpc_clnt *ds_clnt; 1798 struct rpc_cred *ds_cred; 1799 loff_t offset = hdr->args.offset; 1800 int vers; 1801 struct nfs_fh *fh; 1802 int idx = hdr->pgio_mirror_idx; 1803 1804 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true); 1805 if (!ds) 1806 goto out_failed; 1807 1808 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp, 1809 hdr->inode); 1810 if (IS_ERR(ds_clnt)) 1811 goto out_failed; 1812 1813 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred); 1814 if (!ds_cred) 1815 goto out_failed; 1816 1817 vers = nfs4_ff_layout_ds_version(lseg, idx); 1818 1819 dprintk("%s ino %lu sync %d req %zu@%llu DS: %s cl_count %d vers %d\n", 1820 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count, 1821 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), 1822 vers); 1823 1824 hdr->pgio_done_cb = ff_layout_write_done_cb; 1825 atomic_inc(&ds->ds_clp->cl_count); 1826 hdr->ds_clp = ds->ds_clp; 1827 hdr->ds_commit_idx = idx; 1828 fh = nfs4_ff_layout_select_ds_fh(lseg, idx); 1829 if (fh) 1830 hdr->args.fh = fh; 1831 1832 /* 1833 * Note that if we ever decide to split across DSes, 1834 * then we may need to handle dense-like offsets. 1835 */ 1836 hdr->args.offset = offset; 1837 1838 /* Perform an asynchronous write */ 1839 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops, 1840 vers == 3 ? &ff_layout_write_call_ops_v3 : 1841 &ff_layout_write_call_ops_v4, 1842 sync, RPC_TASK_SOFTCONN); 1843 put_rpccred(ds_cred); 1844 return PNFS_ATTEMPTED; 1845 1846 out_failed: 1847 if (ff_layout_avoid_mds_available_ds(lseg)) 1848 return PNFS_TRY_AGAIN; 1849 return PNFS_NOT_ATTEMPTED; 1850 } 1851 1852 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) 1853 { 1854 return i; 1855 } 1856 1857 static struct nfs_fh * 1858 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) 1859 { 1860 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg); 1861 1862 /* FIXME: Assume that there is only one NFS version available 1863 * for the DS. 1864 */ 1865 return &flseg->mirror_array[i]->fh_versions[0]; 1866 } 1867 1868 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how) 1869 { 1870 struct pnfs_layout_segment *lseg = data->lseg; 1871 struct nfs4_pnfs_ds *ds; 1872 struct rpc_clnt *ds_clnt; 1873 struct rpc_cred *ds_cred; 1874 u32 idx; 1875 int vers, ret; 1876 struct nfs_fh *fh; 1877 1878 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); 1879 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true); 1880 if (!ds) 1881 goto out_err; 1882 1883 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp, 1884 data->inode); 1885 if (IS_ERR(ds_clnt)) 1886 goto out_err; 1887 1888 ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred); 1889 if (!ds_cred) 1890 goto out_err; 1891 1892 vers = nfs4_ff_layout_ds_version(lseg, idx); 1893 1894 dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__, 1895 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count), 1896 vers); 1897 data->commit_done_cb = ff_layout_commit_done_cb; 1898 data->cred = ds_cred; 1899 atomic_inc(&ds->ds_clp->cl_count); 1900 data->ds_clp = ds->ds_clp; 1901 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); 1902 if (fh) 1903 data->args.fh = fh; 1904 1905 ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops, 1906 vers == 3 ? &ff_layout_commit_call_ops_v3 : 1907 &ff_layout_commit_call_ops_v4, 1908 how, RPC_TASK_SOFTCONN); 1909 put_rpccred(ds_cred); 1910 return ret; 1911 out_err: 1912 pnfs_generic_prepare_to_resend_writes(data); 1913 pnfs_generic_commit_release(data); 1914 return -EAGAIN; 1915 } 1916 1917 static int 1918 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, 1919 int how, struct nfs_commit_info *cinfo) 1920 { 1921 return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo, 1922 ff_layout_initiate_commit); 1923 } 1924 1925 static struct pnfs_ds_commit_info * 1926 ff_layout_get_ds_info(struct inode *inode) 1927 { 1928 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; 1929 1930 if (layout == NULL) 1931 return NULL; 1932 1933 return &FF_LAYOUT_FROM_HDR(layout)->commit_info; 1934 } 1935 1936 static void 1937 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d) 1938 { 1939 nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds, 1940 id_node)); 1941 } 1942 1943 static int ff_layout_encode_ioerr(struct xdr_stream *xdr, 1944 const struct nfs4_layoutreturn_args *args, 1945 const struct nfs4_flexfile_layoutreturn_args *ff_args) 1946 { 1947 __be32 *start; 1948 1949 start = xdr_reserve_space(xdr, 4); 1950 if (unlikely(!start)) 1951 return -E2BIG; 1952 1953 *start = cpu_to_be32(ff_args->num_errors); 1954 /* This assume we always return _ALL_ layouts */ 1955 return ff_layout_encode_ds_ioerr(xdr, &ff_args->errors); 1956 } 1957 1958 static void 1959 encode_opaque_fixed(struct xdr_stream *xdr, const void *buf, size_t len) 1960 { 1961 WARN_ON_ONCE(xdr_stream_encode_opaque_fixed(xdr, buf, len) < 0); 1962 } 1963 1964 static void 1965 ff_layout_encode_ff_iostat_head(struct xdr_stream *xdr, 1966 const nfs4_stateid *stateid, 1967 const struct nfs42_layoutstat_devinfo *devinfo) 1968 { 1969 __be32 *p; 1970 1971 p = xdr_reserve_space(xdr, 8 + 8); 1972 p = xdr_encode_hyper(p, devinfo->offset); 1973 p = xdr_encode_hyper(p, devinfo->length); 1974 encode_opaque_fixed(xdr, stateid->data, NFS4_STATEID_SIZE); 1975 p = xdr_reserve_space(xdr, 4*8); 1976 p = xdr_encode_hyper(p, devinfo->read_count); 1977 p = xdr_encode_hyper(p, devinfo->read_bytes); 1978 p = xdr_encode_hyper(p, devinfo->write_count); 1979 p = xdr_encode_hyper(p, devinfo->write_bytes); 1980 encode_opaque_fixed(xdr, devinfo->dev_id.data, NFS4_DEVICEID4_SIZE); 1981 } 1982 1983 static void 1984 ff_layout_encode_ff_iostat(struct xdr_stream *xdr, 1985 const nfs4_stateid *stateid, 1986 const struct nfs42_layoutstat_devinfo *devinfo) 1987 { 1988 ff_layout_encode_ff_iostat_head(xdr, stateid, devinfo); 1989 ff_layout_encode_ff_layoutupdate(xdr, devinfo, 1990 devinfo->ld_private.data); 1991 } 1992 1993 /* report nothing for now */ 1994 static void ff_layout_encode_iostats_array(struct xdr_stream *xdr, 1995 const struct nfs4_layoutreturn_args *args, 1996 struct nfs4_flexfile_layoutreturn_args *ff_args) 1997 { 1998 __be32 *p; 1999 int i; 2000 2001 p = xdr_reserve_space(xdr, 4); 2002 *p = cpu_to_be32(ff_args->num_dev); 2003 for (i = 0; i < ff_args->num_dev; i++) 2004 ff_layout_encode_ff_iostat(xdr, 2005 &args->layout->plh_stateid, 2006 &ff_args->devinfo[i]); 2007 } 2008 2009 static void 2010 ff_layout_free_iostats_array(struct nfs42_layoutstat_devinfo *devinfo, 2011 unsigned int num_entries) 2012 { 2013 unsigned int i; 2014 2015 for (i = 0; i < num_entries; i++) { 2016 if (!devinfo[i].ld_private.ops) 2017 continue; 2018 if (!devinfo[i].ld_private.ops->free) 2019 continue; 2020 devinfo[i].ld_private.ops->free(&devinfo[i].ld_private); 2021 } 2022 } 2023 2024 static struct nfs4_deviceid_node * 2025 ff_layout_alloc_deviceid_node(struct nfs_server *server, 2026 struct pnfs_device *pdev, gfp_t gfp_flags) 2027 { 2028 struct nfs4_ff_layout_ds *dsaddr; 2029 2030 dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags); 2031 if (!dsaddr) 2032 return NULL; 2033 return &dsaddr->id_node; 2034 } 2035 2036 static void 2037 ff_layout_encode_layoutreturn(struct xdr_stream *xdr, 2038 const void *voidargs, 2039 const struct nfs4_xdr_opaque_data *ff_opaque) 2040 { 2041 const struct nfs4_layoutreturn_args *args = voidargs; 2042 struct nfs4_flexfile_layoutreturn_args *ff_args = ff_opaque->data; 2043 struct xdr_buf tmp_buf = { 2044 .head = { 2045 [0] = { 2046 .iov_base = page_address(ff_args->pages[0]), 2047 }, 2048 }, 2049 .buflen = PAGE_SIZE, 2050 }; 2051 struct xdr_stream tmp_xdr; 2052 __be32 *start; 2053 2054 dprintk("%s: Begin\n", __func__); 2055 2056 xdr_init_encode(&tmp_xdr, &tmp_buf, NULL); 2057 2058 ff_layout_encode_ioerr(&tmp_xdr, args, ff_args); 2059 ff_layout_encode_iostats_array(&tmp_xdr, args, ff_args); 2060 2061 start = xdr_reserve_space(xdr, 4); 2062 *start = cpu_to_be32(tmp_buf.len); 2063 xdr_write_pages(xdr, ff_args->pages, 0, tmp_buf.len); 2064 2065 dprintk("%s: Return\n", __func__); 2066 } 2067 2068 static void 2069 ff_layout_free_layoutreturn(struct nfs4_xdr_opaque_data *args) 2070 { 2071 struct nfs4_flexfile_layoutreturn_args *ff_args; 2072 2073 if (!args->data) 2074 return; 2075 ff_args = args->data; 2076 args->data = NULL; 2077 2078 ff_layout_free_ds_ioerr(&ff_args->errors); 2079 ff_layout_free_iostats_array(ff_args->devinfo, ff_args->num_dev); 2080 2081 put_page(ff_args->pages[0]); 2082 kfree(ff_args); 2083 } 2084 2085 static const struct nfs4_xdr_opaque_ops layoutreturn_ops = { 2086 .encode = ff_layout_encode_layoutreturn, 2087 .free = ff_layout_free_layoutreturn, 2088 }; 2089 2090 static int 2091 ff_layout_prepare_layoutreturn(struct nfs4_layoutreturn_args *args) 2092 { 2093 struct nfs4_flexfile_layoutreturn_args *ff_args; 2094 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(args->layout); 2095 2096 ff_args = kmalloc(sizeof(*ff_args), GFP_KERNEL); 2097 if (!ff_args) 2098 goto out_nomem; 2099 ff_args->pages[0] = alloc_page(GFP_KERNEL); 2100 if (!ff_args->pages[0]) 2101 goto out_nomem_free; 2102 2103 INIT_LIST_HEAD(&ff_args->errors); 2104 ff_args->num_errors = ff_layout_fetch_ds_ioerr(args->layout, 2105 &args->range, &ff_args->errors, 2106 FF_LAYOUTRETURN_MAXERR); 2107 2108 spin_lock(&args->inode->i_lock); 2109 ff_args->num_dev = ff_layout_mirror_prepare_stats(&ff_layout->generic_hdr, 2110 &ff_args->devinfo[0], ARRAY_SIZE(ff_args->devinfo)); 2111 spin_unlock(&args->inode->i_lock); 2112 2113 args->ld_private->ops = &layoutreturn_ops; 2114 args->ld_private->data = ff_args; 2115 return 0; 2116 out_nomem_free: 2117 kfree(ff_args); 2118 out_nomem: 2119 return -ENOMEM; 2120 } 2121 2122 static int 2123 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen) 2124 { 2125 const struct sockaddr_in *sin = (struct sockaddr_in *)sap; 2126 2127 return snprintf(buf, buflen, "%pI4", &sin->sin_addr); 2128 } 2129 2130 static size_t 2131 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf, 2132 const int buflen) 2133 { 2134 const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap; 2135 const struct in6_addr *addr = &sin6->sin6_addr; 2136 2137 /* 2138 * RFC 4291, Section 2.2.2 2139 * 2140 * Shorthanded ANY address 2141 */ 2142 if (ipv6_addr_any(addr)) 2143 return snprintf(buf, buflen, "::"); 2144 2145 /* 2146 * RFC 4291, Section 2.2.2 2147 * 2148 * Shorthanded loopback address 2149 */ 2150 if (ipv6_addr_loopback(addr)) 2151 return snprintf(buf, buflen, "::1"); 2152 2153 /* 2154 * RFC 4291, Section 2.2.3 2155 * 2156 * Special presentation address format for mapped v4 2157 * addresses. 2158 */ 2159 if (ipv6_addr_v4mapped(addr)) 2160 return snprintf(buf, buflen, "::ffff:%pI4", 2161 &addr->s6_addr32[3]); 2162 2163 /* 2164 * RFC 4291, Section 2.2.1 2165 */ 2166 return snprintf(buf, buflen, "%pI6c", addr); 2167 } 2168 2169 /* Derived from rpc_sockaddr2uaddr */ 2170 static void 2171 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da) 2172 { 2173 struct sockaddr *sap = (struct sockaddr *)&da->da_addr; 2174 char portbuf[RPCBIND_MAXUADDRPLEN]; 2175 char addrbuf[RPCBIND_MAXUADDRLEN]; 2176 char *netid; 2177 unsigned short port; 2178 int len, netid_len; 2179 __be32 *p; 2180 2181 switch (sap->sa_family) { 2182 case AF_INET: 2183 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0) 2184 return; 2185 port = ntohs(((struct sockaddr_in *)sap)->sin_port); 2186 netid = "tcp"; 2187 netid_len = 3; 2188 break; 2189 case AF_INET6: 2190 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0) 2191 return; 2192 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port); 2193 netid = "tcp6"; 2194 netid_len = 4; 2195 break; 2196 default: 2197 /* we only support tcp and tcp6 */ 2198 WARN_ON_ONCE(1); 2199 return; 2200 } 2201 2202 snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff); 2203 len = strlcat(addrbuf, portbuf, sizeof(addrbuf)); 2204 2205 p = xdr_reserve_space(xdr, 4 + netid_len); 2206 xdr_encode_opaque(p, netid, netid_len); 2207 2208 p = xdr_reserve_space(xdr, 4 + len); 2209 xdr_encode_opaque(p, addrbuf, len); 2210 } 2211 2212 static void 2213 ff_layout_encode_nfstime(struct xdr_stream *xdr, 2214 ktime_t t) 2215 { 2216 struct timespec64 ts; 2217 __be32 *p; 2218 2219 p = xdr_reserve_space(xdr, 12); 2220 ts = ktime_to_timespec64(t); 2221 p = xdr_encode_hyper(p, ts.tv_sec); 2222 *p++ = cpu_to_be32(ts.tv_nsec); 2223 } 2224 2225 static void 2226 ff_layout_encode_io_latency(struct xdr_stream *xdr, 2227 struct nfs4_ff_io_stat *stat) 2228 { 2229 __be32 *p; 2230 2231 p = xdr_reserve_space(xdr, 5 * 8); 2232 p = xdr_encode_hyper(p, stat->ops_requested); 2233 p = xdr_encode_hyper(p, stat->bytes_requested); 2234 p = xdr_encode_hyper(p, stat->ops_completed); 2235 p = xdr_encode_hyper(p, stat->bytes_completed); 2236 p = xdr_encode_hyper(p, stat->bytes_not_delivered); 2237 ff_layout_encode_nfstime(xdr, stat->total_busy_time); 2238 ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time); 2239 } 2240 2241 static void 2242 ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr, 2243 const struct nfs42_layoutstat_devinfo *devinfo, 2244 struct nfs4_ff_layout_mirror *mirror) 2245 { 2246 struct nfs4_pnfs_ds_addr *da; 2247 struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds; 2248 struct nfs_fh *fh = &mirror->fh_versions[0]; 2249 __be32 *p; 2250 2251 da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node); 2252 dprintk("%s: DS %s: encoding address %s\n", 2253 __func__, ds->ds_remotestr, da->da_remotestr); 2254 /* netaddr4 */ 2255 ff_layout_encode_netaddr(xdr, da); 2256 /* nfs_fh4 */ 2257 p = xdr_reserve_space(xdr, 4 + fh->size); 2258 xdr_encode_opaque(p, fh->data, fh->size); 2259 /* ff_io_latency4 read */ 2260 spin_lock(&mirror->lock); 2261 ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat); 2262 /* ff_io_latency4 write */ 2263 ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat); 2264 spin_unlock(&mirror->lock); 2265 /* nfstime4 */ 2266 ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time)); 2267 /* bool */ 2268 p = xdr_reserve_space(xdr, 4); 2269 *p = cpu_to_be32(false); 2270 } 2271 2272 static void 2273 ff_layout_encode_layoutstats(struct xdr_stream *xdr, const void *args, 2274 const struct nfs4_xdr_opaque_data *opaque) 2275 { 2276 struct nfs42_layoutstat_devinfo *devinfo = container_of(opaque, 2277 struct nfs42_layoutstat_devinfo, ld_private); 2278 __be32 *start; 2279 2280 /* layoutupdate length */ 2281 start = xdr_reserve_space(xdr, 4); 2282 ff_layout_encode_ff_layoutupdate(xdr, devinfo, opaque->data); 2283 2284 *start = cpu_to_be32((xdr->p - start - 1) * 4); 2285 } 2286 2287 static void 2288 ff_layout_free_layoutstats(struct nfs4_xdr_opaque_data *opaque) 2289 { 2290 struct nfs4_ff_layout_mirror *mirror = opaque->data; 2291 2292 ff_layout_put_mirror(mirror); 2293 } 2294 2295 static const struct nfs4_xdr_opaque_ops layoutstat_ops = { 2296 .encode = ff_layout_encode_layoutstats, 2297 .free = ff_layout_free_layoutstats, 2298 }; 2299 2300 static int 2301 ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo, 2302 struct nfs42_layoutstat_devinfo *devinfo, 2303 int dev_limit) 2304 { 2305 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo); 2306 struct nfs4_ff_layout_mirror *mirror; 2307 struct nfs4_deviceid_node *dev; 2308 int i = 0; 2309 2310 list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) { 2311 if (i >= dev_limit) 2312 break; 2313 if (IS_ERR_OR_NULL(mirror->mirror_ds)) 2314 continue; 2315 if (!test_and_clear_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags)) 2316 continue; 2317 /* mirror refcount put in cleanup_layoutstats */ 2318 if (!atomic_inc_not_zero(&mirror->ref)) 2319 continue; 2320 dev = &mirror->mirror_ds->id_node; 2321 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE); 2322 devinfo->offset = 0; 2323 devinfo->length = NFS4_MAX_UINT64; 2324 spin_lock(&mirror->lock); 2325 devinfo->read_count = mirror->read_stat.io_stat.ops_completed; 2326 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed; 2327 devinfo->write_count = mirror->write_stat.io_stat.ops_completed; 2328 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed; 2329 spin_unlock(&mirror->lock); 2330 devinfo->layout_type = LAYOUT_FLEX_FILES; 2331 devinfo->ld_private.ops = &layoutstat_ops; 2332 devinfo->ld_private.data = mirror; 2333 2334 devinfo++; 2335 i++; 2336 } 2337 return i; 2338 } 2339 2340 static int 2341 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args) 2342 { 2343 struct nfs4_flexfile_layout *ff_layout; 2344 const int dev_count = PNFS_LAYOUTSTATS_MAXDEV; 2345 2346 /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */ 2347 args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO); 2348 if (!args->devinfo) 2349 return -ENOMEM; 2350 2351 spin_lock(&args->inode->i_lock); 2352 ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout); 2353 args->num_dev = ff_layout_mirror_prepare_stats(&ff_layout->generic_hdr, 2354 &args->devinfo[0], dev_count); 2355 spin_unlock(&args->inode->i_lock); 2356 if (!args->num_dev) { 2357 kfree(args->devinfo); 2358 args->devinfo = NULL; 2359 return -ENOENT; 2360 } 2361 2362 return 0; 2363 } 2364 2365 static int 2366 ff_layout_set_layoutdriver(struct nfs_server *server, 2367 const struct nfs_fh *dummy) 2368 { 2369 #if IS_ENABLED(CONFIG_NFS_V4_2) 2370 server->caps |= NFS_CAP_LAYOUTSTATS; 2371 #endif 2372 return 0; 2373 } 2374 2375 static struct pnfs_layoutdriver_type flexfilelayout_type = { 2376 .id = LAYOUT_FLEX_FILES, 2377 .name = "LAYOUT_FLEX_FILES", 2378 .owner = THIS_MODULE, 2379 .set_layoutdriver = ff_layout_set_layoutdriver, 2380 .alloc_layout_hdr = ff_layout_alloc_layout_hdr, 2381 .free_layout_hdr = ff_layout_free_layout_hdr, 2382 .alloc_lseg = ff_layout_alloc_lseg, 2383 .free_lseg = ff_layout_free_lseg, 2384 .add_lseg = ff_layout_add_lseg, 2385 .pg_read_ops = &ff_layout_pg_read_ops, 2386 .pg_write_ops = &ff_layout_pg_write_ops, 2387 .get_ds_info = ff_layout_get_ds_info, 2388 .free_deviceid_node = ff_layout_free_deviceid_node, 2389 .mark_request_commit = pnfs_layout_mark_request_commit, 2390 .clear_request_commit = pnfs_generic_clear_request_commit, 2391 .scan_commit_lists = pnfs_generic_scan_commit_lists, 2392 .recover_commit_reqs = pnfs_generic_recover_commit_reqs, 2393 .commit_pagelist = ff_layout_commit_pagelist, 2394 .read_pagelist = ff_layout_read_pagelist, 2395 .write_pagelist = ff_layout_write_pagelist, 2396 .alloc_deviceid_node = ff_layout_alloc_deviceid_node, 2397 .prepare_layoutreturn = ff_layout_prepare_layoutreturn, 2398 .sync = pnfs_nfs_generic_sync, 2399 .prepare_layoutstats = ff_layout_prepare_layoutstats, 2400 }; 2401 2402 static int __init nfs4flexfilelayout_init(void) 2403 { 2404 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n", 2405 __func__); 2406 if (!ff_zero_group) { 2407 ff_zero_group = groups_alloc(0); 2408 if (!ff_zero_group) 2409 return -ENOMEM; 2410 } 2411 return pnfs_register_layoutdriver(&flexfilelayout_type); 2412 } 2413 2414 static void __exit nfs4flexfilelayout_exit(void) 2415 { 2416 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n", 2417 __func__); 2418 pnfs_unregister_layoutdriver(&flexfilelayout_type); 2419 if (ff_zero_group) { 2420 put_group_info(ff_zero_group); 2421 ff_zero_group = NULL; 2422 } 2423 } 2424 2425 MODULE_ALIAS("nfs-layouttype4-4"); 2426 2427 MODULE_LICENSE("GPL"); 2428 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver"); 2429 2430 module_init(nfs4flexfilelayout_init); 2431 module_exit(nfs4flexfilelayout_exit); 2432