1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Module for pnfs flexfile layout driver. 4 * 5 * Copyright (c) 2014, Primary Data, Inc. All rights reserved. 6 * 7 * Tao Peng <bergwolf@primarydata.com> 8 */ 9 10 #include <linux/nfs_fs.h> 11 #include <linux/nfs_mount.h> 12 #include <linux/nfs_page.h> 13 #include <linux/module.h> 14 #include <linux/file.h> 15 #include <linux/sched/mm.h> 16 17 #include <linux/sunrpc/metrics.h> 18 19 #include "flexfilelayout.h" 20 #include "../nfs4session.h" 21 #include "../nfs4idmap.h" 22 #include "../internal.h" 23 #include "../delegation.h" 24 #include "../nfs4trace.h" 25 #include "../iostat.h" 26 #include "../nfs.h" 27 #include "../nfs42.h" 28 29 #define NFSDBG_FACILITY NFSDBG_PNFS_LD 30 31 #define FF_LAYOUT_POLL_RETRY_MAX (15*HZ) 32 #define FF_LAYOUTRETURN_MAXERR 20 33 34 enum nfs4_ff_op_type { 35 NFS4_FF_OP_LAYOUTSTATS, 36 NFS4_FF_OP_LAYOUTRETURN, 37 }; 38 39 static unsigned short io_maxretrans; 40 41 static const struct pnfs_commit_ops ff_layout_commit_ops; 42 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task, 43 struct nfs_pgio_header *hdr); 44 static int 45 ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo, 46 struct nfs42_layoutstat_devinfo *devinfo, 47 int dev_limit, enum nfs4_ff_op_type type); 48 static void ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr, 49 const struct nfs42_layoutstat_devinfo *devinfo, 50 struct nfs4_ff_layout_ds_stripe *dss_info); 51 52 static struct pnfs_layout_hdr * 53 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) 54 { 55 struct nfs4_flexfile_layout *ffl; 56 57 ffl = kzalloc_obj(*ffl, gfp_flags); 58 if (ffl) { 59 pnfs_init_ds_commit_info(&ffl->commit_info); 60 INIT_LIST_HEAD(&ffl->error_list); 61 INIT_LIST_HEAD(&ffl->mirrors); 62 ffl->last_report_time = ktime_get(); 63 ffl->commit_info.ops = &ff_layout_commit_ops; 64 return &ffl->generic_hdr; 65 } else 66 return NULL; 67 } 68 69 static void 70 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo) 71 { 72 struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(lo); 73 struct nfs4_ff_layout_ds_err *err, *n; 74 75 list_for_each_entry_safe(err, n, &ffl->error_list, list) { 76 list_del(&err->list); 77 kfree(err); 78 } 79 kfree_rcu(ffl, generic_hdr.plh_rcu); 80 } 81 82 static int decode_pnfs_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid) 83 { 84 __be32 *p; 85 86 p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE); 87 if (unlikely(p == NULL)) 88 return -ENOBUFS; 89 stateid->type = NFS4_PNFS_DS_STATEID_TYPE; 90 memcpy(stateid->data, p, NFS4_STATEID_SIZE); 91 dprintk("%s: stateid id= [%x%x%x%x]\n", __func__, 92 p[0], p[1], p[2], p[3]); 93 return 0; 94 } 95 96 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid) 97 { 98 __be32 *p; 99 100 p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE); 101 if (unlikely(!p)) 102 return -ENOBUFS; 103 memcpy(devid, p, NFS4_DEVICEID4_SIZE); 104 nfs4_print_deviceid(devid); 105 return 0; 106 } 107 108 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh) 109 { 110 __be32 *p; 111 112 p = xdr_inline_decode(xdr, 4); 113 if (unlikely(!p)) 114 return -ENOBUFS; 115 fh->size = be32_to_cpup(p++); 116 if (fh->size > NFS_MAXFHSIZE) { 117 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n", 118 fh->size); 119 return -EOVERFLOW; 120 } 121 /* fh.data */ 122 p = xdr_inline_decode(xdr, fh->size); 123 if (unlikely(!p)) 124 return -ENOBUFS; 125 memcpy(&fh->data, p, fh->size); 126 dprintk("%s: fh len %d\n", __func__, fh->size); 127 128 return 0; 129 } 130 131 /* 132 * Currently only stringified uids and gids are accepted. 133 * I.e., kerberos is not supported to the DSes, so no pricipals. 134 * 135 * That means that one common function will suffice, but when 136 * principals are added, this should be split to accomodate 137 * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid(). 138 */ 139 static int 140 decode_name(struct xdr_stream *xdr, u32 *id) 141 { 142 __be32 *p; 143 int len; 144 145 /* opaque_length(4)*/ 146 p = xdr_inline_decode(xdr, 4); 147 if (unlikely(!p)) 148 return -ENOBUFS; 149 len = be32_to_cpup(p++); 150 if (len < 0) 151 return -EINVAL; 152 153 dprintk("%s: len %u\n", __func__, len); 154 155 /* opaque body */ 156 p = xdr_inline_decode(xdr, len); 157 if (unlikely(!p)) 158 return -ENOBUFS; 159 160 if (!nfs_map_string_to_numeric((char *)p, len, id)) 161 return -EINVAL; 162 163 return 0; 164 } 165 166 static struct nfsd_file * 167 ff_local_open_fh(struct pnfs_layout_segment *lseg, u32 ds_idx, u32 dss_id, 168 struct nfs_client *clp, const struct cred *cred, 169 struct nfs_fh *fh, fmode_t mode) 170 { 171 #if IS_ENABLED(CONFIG_NFS_LOCALIO) 172 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx); 173 174 return nfs_local_open_fh(clp, cred, fh, &mirror->dss[dss_id].nfl, mode); 175 #else 176 return NULL; 177 #endif 178 } 179 180 static bool ff_dss_match_fh(const struct nfs4_ff_layout_ds_stripe *dss1, 181 const struct nfs4_ff_layout_ds_stripe *dss2) 182 { 183 int i, j; 184 185 if (dss1->fh_versions_cnt != dss2->fh_versions_cnt) 186 return false; 187 188 for (i = 0; i < dss1->fh_versions_cnt; i++) { 189 bool found_fh = false; 190 for (j = 0; j < dss2->fh_versions_cnt; j++) { 191 if (nfs_compare_fh(&dss1->fh_versions[i], 192 &dss2->fh_versions[j]) == 0) { 193 found_fh = true; 194 break; 195 } 196 } 197 if (!found_fh) 198 return false; 199 } 200 return true; 201 } 202 203 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1, 204 const struct nfs4_ff_layout_mirror *m2) 205 { 206 u32 dss_id; 207 208 if (m1->dss_count != m2->dss_count) 209 return false; 210 211 for (dss_id = 0; dss_id < m1->dss_count; dss_id++) 212 if (!ff_dss_match_fh(&m1->dss[dss_id], &m2->dss[dss_id])) 213 return false; 214 215 return true; 216 } 217 218 static bool ff_mirror_match_devid(const struct nfs4_ff_layout_mirror *m1, 219 const struct nfs4_ff_layout_mirror *m2) 220 { 221 u32 dss_id; 222 223 if (m1->dss_count != m2->dss_count) 224 return false; 225 226 for (dss_id = 0; dss_id < m1->dss_count; dss_id++) 227 if (memcmp(&m1->dss[dss_id].devid, 228 &m2->dss[dss_id].devid, 229 sizeof(m1->dss[dss_id].devid)) != 0) 230 return false; 231 232 return true; 233 } 234 235 static struct nfs4_ff_layout_mirror * 236 ff_layout_add_mirror(struct pnfs_layout_hdr *lo, 237 struct nfs4_ff_layout_mirror *mirror) 238 { 239 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo); 240 struct nfs4_ff_layout_mirror *pos; 241 struct inode *inode = lo->plh_inode; 242 243 spin_lock(&inode->i_lock); 244 list_for_each_entry(pos, &ff_layout->mirrors, mirrors) { 245 if (!ff_mirror_match_devid(mirror, pos)) 246 continue; 247 if (!ff_mirror_match_fh(mirror, pos)) 248 continue; 249 if (refcount_inc_not_zero(&pos->ref)) { 250 spin_unlock(&inode->i_lock); 251 return pos; 252 } 253 } 254 list_add(&mirror->mirrors, &ff_layout->mirrors); 255 mirror->layout = lo; 256 spin_unlock(&inode->i_lock); 257 return mirror; 258 } 259 260 static void 261 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror) 262 { 263 struct inode *inode; 264 if (mirror->layout == NULL) 265 return; 266 inode = mirror->layout->plh_inode; 267 spin_lock(&inode->i_lock); 268 list_del(&mirror->mirrors); 269 spin_unlock(&inode->i_lock); 270 mirror->layout = NULL; 271 } 272 273 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(u32 dss_count, 274 gfp_t gfp_flags) 275 { 276 struct nfs4_ff_layout_mirror *mirror; 277 278 mirror = kzalloc_obj(*mirror, gfp_flags); 279 if (mirror == NULL) 280 return NULL; 281 282 spin_lock_init(&mirror->lock); 283 refcount_set(&mirror->ref, 1); 284 INIT_LIST_HEAD(&mirror->mirrors); 285 286 mirror->dss_count = dss_count; 287 mirror->dss = 288 kzalloc_objs(struct nfs4_ff_layout_ds_stripe, dss_count, 289 gfp_flags); 290 if (mirror->dss == NULL) { 291 kfree(mirror); 292 return NULL; 293 } 294 295 for (u32 dss_id = 0; dss_id < mirror->dss_count; dss_id++) 296 nfs_localio_file_init(&mirror->dss[dss_id].nfl); 297 298 return mirror; 299 } 300 301 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror) 302 { 303 const struct cred *cred; 304 u32 dss_id; 305 306 ff_layout_remove_mirror(mirror); 307 308 for (dss_id = 0; dss_id < mirror->dss_count; dss_id++) { 309 kfree(mirror->dss[dss_id].fh_versions); 310 cred = rcu_access_pointer(mirror->dss[dss_id].ro_cred); 311 put_cred(cred); 312 cred = rcu_access_pointer(mirror->dss[dss_id].rw_cred); 313 put_cred(cred); 314 nfs_close_local_fh(&mirror->dss[dss_id].nfl); 315 nfs4_ff_layout_put_deviceid(mirror->dss[dss_id].mirror_ds); 316 } 317 318 kfree(mirror->dss); 319 kfree(mirror); 320 } 321 322 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror) 323 { 324 if (mirror != NULL && refcount_dec_and_test(&mirror->ref)) 325 ff_layout_free_mirror(mirror); 326 } 327 328 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls) 329 { 330 u32 i; 331 332 for (i = 0; i < fls->mirror_array_cnt; i++) 333 ff_layout_put_mirror(fls->mirror_array[i]); 334 } 335 336 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls) 337 { 338 if (fls) { 339 ff_layout_free_mirror_array(fls); 340 kfree(fls); 341 } 342 } 343 344 static bool 345 ff_lseg_match_mirrors(struct pnfs_layout_segment *l1, 346 struct pnfs_layout_segment *l2) 347 { 348 const struct nfs4_ff_layout_segment *fl1 = FF_LAYOUT_LSEG(l1); 349 const struct nfs4_ff_layout_segment *fl2 = FF_LAYOUT_LSEG(l2); 350 u32 i; 351 352 if (fl1->mirror_array_cnt != fl2->mirror_array_cnt) 353 return false; 354 for (i = 0; i < fl1->mirror_array_cnt; i++) { 355 if (fl1->mirror_array[i] != fl2->mirror_array[i]) 356 return false; 357 } 358 return true; 359 } 360 361 static bool 362 ff_lseg_range_is_after(const struct pnfs_layout_range *l1, 363 const struct pnfs_layout_range *l2) 364 { 365 u64 end1, end2; 366 367 if (l1->iomode != l2->iomode) 368 return l1->iomode != IOMODE_READ; 369 end1 = pnfs_calc_offset_end(l1->offset, l1->length); 370 end2 = pnfs_calc_offset_end(l2->offset, l2->length); 371 if (end1 < l2->offset) 372 return false; 373 if (end2 < l1->offset) 374 return true; 375 return l2->offset <= l1->offset; 376 } 377 378 static bool 379 ff_lseg_merge(struct pnfs_layout_segment *new, 380 struct pnfs_layout_segment *old) 381 { 382 u64 new_end, old_end; 383 384 if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags)) 385 return false; 386 if (new->pls_range.iomode != old->pls_range.iomode) 387 return false; 388 old_end = pnfs_calc_offset_end(old->pls_range.offset, 389 old->pls_range.length); 390 if (old_end < new->pls_range.offset) 391 return false; 392 new_end = pnfs_calc_offset_end(new->pls_range.offset, 393 new->pls_range.length); 394 if (new_end < old->pls_range.offset) 395 return false; 396 if (!ff_lseg_match_mirrors(new, old)) 397 return false; 398 399 /* Mergeable: copy info from 'old' to 'new' */ 400 if (new_end < old_end) 401 new_end = old_end; 402 if (new->pls_range.offset < old->pls_range.offset) 403 new->pls_range.offset = old->pls_range.offset; 404 new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset, 405 new_end); 406 if (test_bit(NFS_LSEG_ROC, &old->pls_flags)) 407 set_bit(NFS_LSEG_ROC, &new->pls_flags); 408 return true; 409 } 410 411 static void 412 ff_layout_add_lseg(struct pnfs_layout_hdr *lo, 413 struct pnfs_layout_segment *lseg, 414 struct list_head *free_me) 415 { 416 pnfs_generic_layout_insert_lseg(lo, lseg, 417 ff_lseg_range_is_after, 418 ff_lseg_merge, 419 free_me); 420 } 421 422 static u32 ff_mirror_efficiency_sum(const struct nfs4_ff_layout_mirror *mirror) 423 { 424 u32 dss_id, sum = 0; 425 426 for (dss_id = 0; dss_id < mirror->dss_count; dss_id++) 427 sum += mirror->dss[dss_id].efficiency; 428 429 return sum; 430 } 431 432 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls) 433 { 434 int i, j; 435 436 for (i = 0; i < fls->mirror_array_cnt - 1; i++) { 437 for (j = i + 1; j < fls->mirror_array_cnt; j++) 438 if (ff_mirror_efficiency_sum(fls->mirror_array[i]) < 439 ff_mirror_efficiency_sum(fls->mirror_array[j])) 440 swap(fls->mirror_array[i], 441 fls->mirror_array[j]); 442 } 443 } 444 445 static struct pnfs_layout_segment * 446 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh, 447 struct nfs4_layoutget_res *lgr, 448 gfp_t gfp_flags) 449 { 450 struct pnfs_layout_segment *ret; 451 struct nfs4_ff_layout_segment *fls = NULL; 452 struct xdr_stream stream; 453 struct xdr_buf buf; 454 struct folio *scratch; 455 u64 stripe_unit; 456 u32 mirror_array_cnt; 457 __be32 *p; 458 int i, rc; 459 struct nfs4_ff_layout_ds_stripe *dss_info; 460 461 dprintk("--> %s\n", __func__); 462 scratch = folio_alloc(gfp_flags, 0); 463 if (!scratch) 464 return ERR_PTR(-ENOMEM); 465 466 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, 467 lgr->layoutp->len); 468 xdr_set_scratch_folio(&stream, scratch); 469 470 /* stripe unit and mirror_array_cnt */ 471 rc = -EIO; 472 p = xdr_inline_decode(&stream, 8 + 4); 473 if (!p) 474 goto out_err_free; 475 476 p = xdr_decode_hyper(p, &stripe_unit); 477 mirror_array_cnt = be32_to_cpup(p++); 478 dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__, 479 stripe_unit, mirror_array_cnt); 480 481 if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT || 482 mirror_array_cnt == 0) 483 goto out_err_free; 484 485 rc = -ENOMEM; 486 fls = kzalloc_flex(*fls, mirror_array, mirror_array_cnt, gfp_flags); 487 if (!fls) 488 goto out_err_free; 489 490 fls->mirror_array_cnt = mirror_array_cnt; 491 fls->stripe_unit = stripe_unit; 492 493 u32 dss_count = 0; 494 for (i = 0; i < fls->mirror_array_cnt; i++) { 495 struct nfs4_ff_layout_mirror *mirror; 496 struct cred *kcred; 497 const struct cred __rcu *cred; 498 kuid_t uid; 499 kgid_t gid; 500 u32 fh_count, id; 501 int j, dss_id; 502 503 rc = -EIO; 504 p = xdr_inline_decode(&stream, 4); 505 if (!p) 506 goto out_err_free; 507 508 // Ensure all mirrors have same stripe count. 509 if (dss_count == 0) 510 dss_count = be32_to_cpup(p); 511 else if (dss_count != be32_to_cpup(p)) 512 goto out_err_free; 513 514 if (dss_count > NFS4_FLEXFILE_LAYOUT_MAX_STRIPE_CNT || 515 dss_count == 0) 516 goto out_err_free; 517 518 if (dss_count > 1 && stripe_unit == 0) 519 goto out_err_free; 520 521 fls->mirror_array[i] = ff_layout_alloc_mirror(dss_count, gfp_flags); 522 if (fls->mirror_array[i] == NULL) { 523 rc = -ENOMEM; 524 goto out_err_free; 525 } 526 527 for (dss_id = 0; dss_id < dss_count; dss_id++) { 528 dss_info = &fls->mirror_array[i]->dss[dss_id]; 529 dss_info->mirror = fls->mirror_array[i]; 530 531 /* deviceid */ 532 rc = decode_deviceid(&stream, &dss_info->devid); 533 if (rc) 534 goto out_err_free; 535 536 /* efficiency */ 537 rc = -EIO; 538 p = xdr_inline_decode(&stream, 4); 539 if (!p) 540 goto out_err_free; 541 dss_info->efficiency = be32_to_cpup(p); 542 543 /* stateid */ 544 rc = decode_pnfs_stateid(&stream, &dss_info->stateid); 545 if (rc) 546 goto out_err_free; 547 548 /* fh */ 549 rc = -EIO; 550 p = xdr_inline_decode(&stream, 4); 551 if (!p) 552 goto out_err_free; 553 fh_count = be32_to_cpup(p); 554 if (fh_count == 0) { 555 rc = -EINVAL; 556 goto out_err_free; 557 } 558 559 dss_info->fh_versions = 560 kzalloc_objs(struct nfs_fh, fh_count, gfp_flags); 561 if (dss_info->fh_versions == NULL) { 562 rc = -ENOMEM; 563 goto out_err_free; 564 } 565 566 for (j = 0; j < fh_count; j++) { 567 rc = decode_nfs_fh(&stream, 568 &dss_info->fh_versions[j]); 569 if (rc) 570 goto out_err_free; 571 } 572 573 dss_info->fh_versions_cnt = fh_count; 574 575 /* user */ 576 rc = decode_name(&stream, &id); 577 if (rc) 578 goto out_err_free; 579 580 uid = make_kuid(&init_user_ns, id); 581 582 /* group */ 583 rc = decode_name(&stream, &id); 584 if (rc) 585 goto out_err_free; 586 587 gid = make_kgid(&init_user_ns, id); 588 589 if (gfp_flags & __GFP_FS) 590 kcred = prepare_kernel_cred(&init_task); 591 else { 592 unsigned int nofs_flags = memalloc_nofs_save(); 593 594 kcred = prepare_kernel_cred(&init_task); 595 memalloc_nofs_restore(nofs_flags); 596 } 597 rc = -ENOMEM; 598 if (!kcred) 599 goto out_err_free; 600 kcred->fsuid = uid; 601 kcred->fsgid = gid; 602 cred = RCU_INITIALIZER(kcred); 603 604 if (lgr->range.iomode == IOMODE_READ) 605 rcu_assign_pointer(dss_info->ro_cred, cred); 606 else 607 rcu_assign_pointer(dss_info->rw_cred, cred); 608 } 609 610 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]); 611 if (mirror != fls->mirror_array[i]) { 612 for (dss_id = 0; dss_id < dss_count; dss_id++) { 613 dss_info = &fls->mirror_array[i]->dss[dss_id]; 614 /* swap cred ptrs so free_mirror will clean up old */ 615 if (lgr->range.iomode == IOMODE_READ) { 616 cred = xchg(&mirror->dss[dss_id].ro_cred, 617 dss_info->ro_cred); 618 rcu_assign_pointer(dss_info->ro_cred, cred); 619 } else { 620 cred = xchg(&mirror->dss[dss_id].rw_cred, 621 dss_info->rw_cred); 622 rcu_assign_pointer(dss_info->rw_cred, cred); 623 } 624 } 625 ff_layout_free_mirror(fls->mirror_array[i]); 626 fls->mirror_array[i] = mirror; 627 } 628 629 dprintk("%s: iomode %s uid %u gid %u\n", __func__, 630 lgr->range.iomode == IOMODE_READ ? "READ" : "RW", 631 from_kuid(&init_user_ns, uid), 632 from_kgid(&init_user_ns, gid)); 633 } 634 635 p = xdr_inline_decode(&stream, 4); 636 if (!p) 637 goto out_sort_mirrors; 638 fls->flags = be32_to_cpup(p); 639 if (fls->flags & FF_FLAGS_NO_IO_THRU_MDS) 640 set_bit(NFS4_FF_HDR_NO_IO_THRU_MDS, 641 &FF_LAYOUT_FROM_HDR(lh)->flags); 642 643 p = xdr_inline_decode(&stream, 4); 644 if (!p) 645 goto out_sort_mirrors; 646 for (i=0; i < fls->mirror_array_cnt; i++) 647 fls->mirror_array[i]->report_interval = be32_to_cpup(p); 648 649 out_sort_mirrors: 650 ff_layout_sort_mirrors(fls); 651 ret = &fls->generic_hdr; 652 dprintk("<-- %s (success)\n", __func__); 653 out_free_page: 654 folio_put(scratch); 655 return ret; 656 out_err_free: 657 _ff_layout_free_lseg(fls); 658 ret = ERR_PTR(rc); 659 dprintk("<-- %s (%d)\n", __func__, rc); 660 goto out_free_page; 661 } 662 663 static void 664 ff_layout_free_lseg(struct pnfs_layout_segment *lseg) 665 { 666 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 667 668 dprintk("--> %s\n", __func__); 669 670 if (lseg->pls_range.iomode == IOMODE_RW) { 671 struct nfs4_flexfile_layout *ffl; 672 struct inode *inode; 673 674 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout); 675 inode = ffl->generic_hdr.plh_inode; 676 spin_lock(&inode->i_lock); 677 pnfs_generic_ds_cinfo_release_lseg(&ffl->commit_info, lseg); 678 spin_unlock(&inode->i_lock); 679 } 680 _ff_layout_free_lseg(fls); 681 } 682 683 static u32 calc_commit_idx(struct pnfs_layout_segment *lseg, 684 u32 mirror_idx, u32 dss_id) 685 { 686 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg); 687 688 return (mirror_idx * flseg->mirror_array[0]->dss_count) + dss_id; 689 } 690 691 static u32 calc_mirror_idx_from_commit(struct pnfs_layout_segment *lseg, 692 u32 commit_index) 693 { 694 return commit_index / FF_LAYOUT_LSEG(lseg)->mirror_array[0]->dss_count; 695 } 696 697 static u32 calc_dss_id_from_commit(struct pnfs_layout_segment *lseg, 698 u32 commit_index) 699 { 700 return commit_index % FF_LAYOUT_LSEG(lseg)->mirror_array[0]->dss_count; 701 } 702 703 static void 704 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now) 705 { 706 /* first IO request? */ 707 if (atomic_inc_return(&timer->n_ops) == 1) { 708 timer->start_time = now; 709 } 710 } 711 712 static ktime_t 713 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now) 714 { 715 ktime_t start; 716 717 if (atomic_dec_return(&timer->n_ops) < 0) 718 WARN_ON_ONCE(1); 719 720 start = timer->start_time; 721 timer->start_time = now; 722 return ktime_sub(now, start); 723 } 724 725 static bool 726 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror, 727 u32 dss_id, 728 struct nfs4_ff_layoutstat *layoutstat, 729 ktime_t now) 730 { 731 s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL; 732 struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(mirror->layout); 733 734 nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now); 735 if (!mirror->dss[dss_id].start_time) 736 mirror->dss[dss_id].start_time = now; 737 if (mirror->report_interval != 0) 738 report_interval = (s64)mirror->report_interval * 1000LL; 739 else if (layoutstats_timer != 0) 740 report_interval = (s64)layoutstats_timer * 1000LL; 741 if (ktime_to_ms(ktime_sub(now, ffl->last_report_time)) >= 742 report_interval) { 743 ffl->last_report_time = now; 744 return true; 745 } 746 747 return false; 748 } 749 750 static void 751 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat, 752 __u64 requested) 753 { 754 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat; 755 756 iostat->ops_requested++; 757 iostat->bytes_requested += requested; 758 } 759 760 static void 761 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat, 762 __u64 requested, 763 __u64 completed, 764 ktime_t time_completed, 765 ktime_t time_started) 766 { 767 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat; 768 ktime_t completion_time = ktime_sub(time_completed, time_started); 769 ktime_t timer; 770 771 iostat->ops_completed++; 772 iostat->bytes_completed += completed; 773 iostat->bytes_not_delivered += requested - completed; 774 775 timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed); 776 iostat->total_busy_time = 777 ktime_add(iostat->total_busy_time, timer); 778 iostat->aggregate_completion_time = 779 ktime_add(iostat->aggregate_completion_time, 780 completion_time); 781 } 782 783 static void 784 nfs4_ff_layout_stat_io_start_read(struct inode *inode, 785 struct nfs4_ff_layout_mirror *mirror, 786 u32 dss_id, 787 __u64 requested, ktime_t now) 788 { 789 bool report; 790 791 spin_lock(&mirror->lock); 792 report = nfs4_ff_layoutstat_start_io( 793 mirror, dss_id, &mirror->dss[dss_id].read_stat, now); 794 nfs4_ff_layout_stat_io_update_requested( 795 &mirror->dss[dss_id].read_stat, requested); 796 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 797 spin_unlock(&mirror->lock); 798 799 if (report) 800 pnfs_report_layoutstat(inode, nfs_io_gfp_mask()); 801 } 802 803 static void 804 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task, 805 struct nfs4_ff_layout_mirror *mirror, 806 u32 dss_id, 807 __u64 requested, 808 __u64 completed) 809 { 810 spin_lock(&mirror->lock); 811 nfs4_ff_layout_stat_io_update_completed(&mirror->dss[dss_id].read_stat, 812 requested, completed, 813 ktime_get(), task->tk_start); 814 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 815 spin_unlock(&mirror->lock); 816 } 817 818 static void 819 nfs4_ff_layout_stat_io_start_write(struct inode *inode, 820 struct nfs4_ff_layout_mirror *mirror, 821 u32 dss_id, 822 __u64 requested, ktime_t now) 823 { 824 bool report; 825 826 spin_lock(&mirror->lock); 827 report = nfs4_ff_layoutstat_start_io( 828 mirror, 829 dss_id, 830 &mirror->dss[dss_id].write_stat, 831 now); 832 nfs4_ff_layout_stat_io_update_requested( 833 &mirror->dss[dss_id].write_stat, 834 requested); 835 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 836 spin_unlock(&mirror->lock); 837 838 if (report) 839 pnfs_report_layoutstat(inode, nfs_io_gfp_mask()); 840 } 841 842 static void 843 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task, 844 struct nfs4_ff_layout_mirror *mirror, 845 u32 dss_id, 846 __u64 requested, 847 __u64 completed, 848 enum nfs3_stable_how committed) 849 { 850 if (committed == NFS_UNSTABLE) 851 requested = completed = 0; 852 853 spin_lock(&mirror->lock); 854 nfs4_ff_layout_stat_io_update_completed(&mirror->dss[dss_id].write_stat, 855 requested, completed, ktime_get(), task->tk_start); 856 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags); 857 spin_unlock(&mirror->lock); 858 } 859 860 static void 861 ff_layout_mark_ds_unreachable(struct pnfs_layout_segment *lseg, u32 idx, u32 dss_id) 862 { 863 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id); 864 865 if (devid) 866 nfs4_mark_deviceid_unavailable(devid); 867 } 868 869 static void 870 ff_layout_mark_ds_reachable(struct pnfs_layout_segment *lseg, u32 idx, u32 dss_id) 871 { 872 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id); 873 874 if (devid) 875 nfs4_mark_deviceid_available(devid); 876 } 877 878 static struct nfs4_pnfs_ds * 879 ff_layout_choose_ds_for_read(struct pnfs_layout_segment *lseg, 880 u32 start_idx, u32 *best_idx, 881 u32 offset, u32 *dss_id, 882 bool check_device) 883 { 884 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg); 885 struct nfs4_ff_layout_mirror *mirror; 886 struct nfs4_pnfs_ds *ds = ERR_PTR(-EAGAIN); 887 u32 idx; 888 889 /* mirrors are initially sorted by efficiency */ 890 for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) { 891 mirror = FF_LAYOUT_COMP(lseg, idx); 892 *dss_id = nfs4_ff_layout_calc_dss_id( 893 fls->stripe_unit, 894 fls->mirror_array[idx]->dss_count, 895 offset); 896 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, *dss_id, false); 897 if (IS_ERR(ds)) 898 continue; 899 900 if (check_device && 901 nfs4_test_deviceid_unavailable(&mirror->dss[*dss_id].mirror_ds->id_node)) { 902 // reinitialize the error state in case if this is the last iteration 903 ds = ERR_PTR(-EINVAL); 904 continue; 905 } 906 907 *best_idx = idx; 908 break; 909 } 910 911 return ds; 912 } 913 914 static struct nfs4_pnfs_ds * 915 ff_layout_choose_any_ds_for_read(struct pnfs_layout_segment *lseg, 916 u32 start_idx, u32 *best_idx, 917 u32 offset, u32 *dss_id) 918 { 919 return ff_layout_choose_ds_for_read(lseg, start_idx, best_idx, 920 offset, dss_id, false); 921 } 922 923 static struct nfs4_pnfs_ds * 924 ff_layout_choose_valid_ds_for_read(struct pnfs_layout_segment *lseg, 925 u32 start_idx, u32 *best_idx, 926 u32 offset, u32 *dss_id) 927 { 928 return ff_layout_choose_ds_for_read(lseg, start_idx, best_idx, 929 offset, dss_id, true); 930 } 931 932 static struct nfs4_pnfs_ds * 933 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg, 934 u32 start_idx, u32 *best_idx, 935 u32 offset, u32 *dss_id) 936 { 937 struct nfs4_pnfs_ds *ds; 938 939 ds = ff_layout_choose_valid_ds_for_read(lseg, start_idx, best_idx, 940 offset, dss_id); 941 if (!IS_ERR(ds)) 942 return ds; 943 return ff_layout_choose_any_ds_for_read(lseg, start_idx, best_idx, 944 offset, dss_id); 945 } 946 947 static struct nfs4_pnfs_ds * 948 ff_layout_get_ds_for_read(struct nfs_pageio_descriptor *pgio, 949 u32 *best_idx, 950 u32 offset, 951 u32 *dss_id) 952 { 953 struct pnfs_layout_segment *lseg = pgio->pg_lseg; 954 struct nfs4_pnfs_ds *ds; 955 956 ds = ff_layout_choose_best_ds_for_read(lseg, pgio->pg_mirror_idx, 957 best_idx, offset, dss_id); 958 if (!IS_ERR(ds) || !pgio->pg_mirror_idx) 959 return ds; 960 return ff_layout_choose_best_ds_for_read(lseg, 0, best_idx, 961 offset, dss_id); 962 } 963 964 static void 965 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio, 966 struct nfs_page *req, 967 bool strict_iomode) 968 { 969 pnfs_put_lseg(pgio->pg_lseg); 970 pgio->pg_lseg = 971 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req), 972 req_offset(req), req->wb_bytes, IOMODE_READ, 973 strict_iomode, nfs_io_gfp_mask()); 974 if (IS_ERR(pgio->pg_lseg)) { 975 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 976 pgio->pg_lseg = NULL; 977 } 978 } 979 980 static bool 981 ff_layout_lseg_is_striped(const struct nfs4_ff_layout_segment *fls) 982 { 983 return fls->mirror_array[0]->dss_count > 1; 984 } 985 986 /* 987 * ff_layout_pg_test(). Called by nfs_can_coalesce_requests() 988 * 989 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number 990 * of bytes (maximum @req->wb_bytes) that can be coalesced. 991 */ 992 static size_t 993 ff_layout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, 994 struct nfs_page *req) 995 { 996 unsigned int size; 997 u64 p_stripe, r_stripe; 998 u32 stripe_offset; 999 u64 segment_offset = pgio->pg_lseg->pls_range.offset; 1000 u32 stripe_unit = FF_LAYOUT_LSEG(pgio->pg_lseg)->stripe_unit; 1001 1002 /* calls nfs_generic_pg_test */ 1003 size = pnfs_generic_pg_test(pgio, prev, req); 1004 if (!size) 1005 return 0; 1006 else if (!ff_layout_lseg_is_striped(FF_LAYOUT_LSEG(pgio->pg_lseg))) 1007 return size; 1008 1009 /* see if req and prev are in the same stripe */ 1010 if (prev) { 1011 p_stripe = (u64)req_offset(prev) - segment_offset; 1012 r_stripe = (u64)req_offset(req) - segment_offset; 1013 do_div(p_stripe, stripe_unit); 1014 do_div(r_stripe, stripe_unit); 1015 1016 if (p_stripe != r_stripe) 1017 return 0; 1018 } 1019 1020 /* calculate remaining bytes in the current stripe */ 1021 div_u64_rem((u64)req_offset(req) - segment_offset, 1022 stripe_unit, 1023 &stripe_offset); 1024 WARN_ON_ONCE(stripe_offset > stripe_unit); 1025 if (stripe_offset >= stripe_unit) 1026 return 0; 1027 return min(stripe_unit - (unsigned int)stripe_offset, size); 1028 } 1029 1030 static void 1031 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio, 1032 struct nfs_page *req) 1033 { 1034 struct nfs_pgio_mirror *pgm; 1035 struct nfs4_ff_layout_mirror *mirror; 1036 struct nfs4_pnfs_ds *ds; 1037 u32 ds_idx, dss_id; 1038 1039 if (NFS_SERVER(pgio->pg_inode)->flags & 1040 (NFS_MOUNT_SOFT|NFS_MOUNT_SOFTERR)) 1041 pgio->pg_maxretrans = io_maxretrans; 1042 retry: 1043 pnfs_generic_pg_check_layout(pgio, req); 1044 /* Use full layout for now */ 1045 if (!pgio->pg_lseg) { 1046 ff_layout_pg_get_read(pgio, req, false); 1047 if (!pgio->pg_lseg) 1048 goto out_nolseg; 1049 } 1050 if (ff_layout_avoid_read_on_rw(pgio->pg_lseg)) { 1051 ff_layout_pg_get_read(pgio, req, true); 1052 if (!pgio->pg_lseg) 1053 goto out_nolseg; 1054 } 1055 /* Reset wb_nio, since getting layout segment was successful */ 1056 req->wb_nio = 0; 1057 1058 ds = ff_layout_get_ds_for_read(pgio, &ds_idx, 1059 req_offset(req), &dss_id); 1060 if (IS_ERR(ds)) { 1061 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg)) 1062 goto out_mds; 1063 pnfs_generic_pg_cleanup(pgio); 1064 /* Sleep for 1 second before retrying */ 1065 ssleep(1); 1066 goto retry; 1067 } 1068 1069 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx); 1070 pgm = &pgio->pg_mirrors[0]; 1071 pgm->pg_bsize = mirror->dss[dss_id].mirror_ds->ds_versions[0].rsize; 1072 1073 pgio->pg_mirror_idx = ds_idx; 1074 return; 1075 out_nolseg: 1076 if (pgio->pg_error < 0) { 1077 if (pgio->pg_error != -EAGAIN) 1078 return; 1079 /* Retry getting layout segment if lower layer returned -EAGAIN */ 1080 if (pgio->pg_maxretrans && req->wb_nio++ > pgio->pg_maxretrans) { 1081 if (NFS_SERVER(pgio->pg_inode)->flags & NFS_MOUNT_SOFTERR) 1082 pgio->pg_error = -ETIMEDOUT; 1083 else 1084 pgio->pg_error = -EIO; 1085 return; 1086 } 1087 pgio->pg_error = 0; 1088 /* Sleep for 1 second before retrying */ 1089 ssleep(1); 1090 goto retry; 1091 } 1092 out_mds: 1093 trace_pnfs_mds_fallback_pg_init_read(pgio->pg_inode, 1094 0, NFS4_MAX_UINT64, IOMODE_READ, 1095 NFS_I(pgio->pg_inode)->layout, 1096 pgio->pg_lseg); 1097 pgio->pg_maxretrans = 0; 1098 nfs_pageio_reset_read_mds(pgio); 1099 } 1100 1101 static void 1102 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio, 1103 struct nfs_page *req) 1104 { 1105 struct nfs4_ff_layout_mirror *mirror; 1106 struct nfs_pgio_mirror *pgm; 1107 struct nfs4_pnfs_ds *ds; 1108 u32 i, dss_id; 1109 1110 retry: 1111 pnfs_generic_pg_check_layout(pgio, req); 1112 if (!pgio->pg_lseg) { 1113 pgio->pg_lseg = 1114 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req), 1115 req_offset(req), req->wb_bytes, 1116 IOMODE_RW, false, nfs_io_gfp_mask()); 1117 if (IS_ERR(pgio->pg_lseg)) { 1118 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 1119 pgio->pg_lseg = NULL; 1120 return; 1121 } 1122 } 1123 /* If no lseg, fall back to write through mds */ 1124 if (pgio->pg_lseg == NULL) 1125 goto out_mds; 1126 1127 /* Use a direct mapping of ds_idx to pgio mirror_idx */ 1128 if (pgio->pg_mirror_count != FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)) 1129 goto out_eagain; 1130 1131 for (i = 0; i < pgio->pg_mirror_count; i++) { 1132 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i); 1133 dss_id = nfs4_ff_layout_calc_dss_id( 1134 FF_LAYOUT_LSEG(pgio->pg_lseg)->stripe_unit, 1135 mirror->dss_count, 1136 req_offset(req)); 1137 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, mirror, 1138 dss_id, true); 1139 if (IS_ERR(ds)) { 1140 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg)) 1141 goto out_mds; 1142 pnfs_generic_pg_cleanup(pgio); 1143 /* Sleep for 1 second before retrying */ 1144 ssleep(1); 1145 goto retry; 1146 } 1147 pgm = &pgio->pg_mirrors[i]; 1148 pgm->pg_bsize = mirror->dss[dss_id].mirror_ds->ds_versions[0].wsize; 1149 } 1150 1151 if (NFS_SERVER(pgio->pg_inode)->flags & 1152 (NFS_MOUNT_SOFT|NFS_MOUNT_SOFTERR)) 1153 pgio->pg_maxretrans = io_maxretrans; 1154 return; 1155 out_eagain: 1156 pnfs_generic_pg_cleanup(pgio); 1157 pgio->pg_error = -EAGAIN; 1158 return; 1159 out_mds: 1160 trace_pnfs_mds_fallback_pg_init_write(pgio->pg_inode, 1161 0, NFS4_MAX_UINT64, IOMODE_RW, 1162 NFS_I(pgio->pg_inode)->layout, 1163 pgio->pg_lseg); 1164 pgio->pg_maxretrans = 0; 1165 nfs_pageio_reset_write_mds(pgio); 1166 pgio->pg_error = -EAGAIN; 1167 } 1168 1169 static unsigned int 1170 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio, 1171 struct nfs_page *req) 1172 { 1173 if (!pgio->pg_lseg) { 1174 pgio->pg_lseg = 1175 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req), 1176 req_offset(req), req->wb_bytes, 1177 IOMODE_RW, false, nfs_io_gfp_mask()); 1178 if (IS_ERR(pgio->pg_lseg)) { 1179 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 1180 pgio->pg_lseg = NULL; 1181 goto out; 1182 } 1183 } 1184 if (pgio->pg_lseg) 1185 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg); 1186 1187 trace_pnfs_mds_fallback_pg_get_mirror_count(pgio->pg_inode, 1188 0, NFS4_MAX_UINT64, IOMODE_RW, 1189 NFS_I(pgio->pg_inode)->layout, 1190 pgio->pg_lseg); 1191 if (NFS_I(pgio->pg_inode)->layout && 1192 ff_layout_hdr_no_fallback_to_mds(NFS_I(pgio->pg_inode)->layout)) { 1193 /* 1194 * FF_FLAGS_NO_IO_THRU_MDS: no current lseg but the server's 1195 * policy forbids MDS fallback. Surface -EAGAIN so writeback 1196 * retries rather than silently issuing the WRITE via MDS. 1197 */ 1198 pgio->pg_error = -EAGAIN; 1199 goto out; 1200 } 1201 /* no lseg means that pnfs is not in use, so no mirroring here */ 1202 nfs_pageio_reset_write_mds(pgio); 1203 out: 1204 return 1; 1205 } 1206 1207 static u32 1208 ff_layout_pg_set_mirror_write(struct nfs_pageio_descriptor *desc, u32 idx) 1209 { 1210 u32 old = desc->pg_mirror_idx; 1211 1212 desc->pg_mirror_idx = idx; 1213 return old; 1214 } 1215 1216 static struct nfs_pgio_mirror * 1217 ff_layout_pg_get_mirror_write(struct nfs_pageio_descriptor *desc, u32 idx) 1218 { 1219 return &desc->pg_mirrors[idx]; 1220 } 1221 1222 static const struct nfs_pageio_ops ff_layout_pg_read_ops = { 1223 .pg_init = ff_layout_pg_init_read, 1224 .pg_test = ff_layout_pg_test, 1225 .pg_doio = pnfs_generic_pg_readpages, 1226 .pg_cleanup = pnfs_generic_pg_cleanup, 1227 }; 1228 1229 static const struct nfs_pageio_ops ff_layout_pg_write_ops = { 1230 .pg_init = ff_layout_pg_init_write, 1231 .pg_test = ff_layout_pg_test, 1232 .pg_doio = pnfs_generic_pg_writepages, 1233 .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write, 1234 .pg_cleanup = pnfs_generic_pg_cleanup, 1235 .pg_get_mirror = ff_layout_pg_get_mirror_write, 1236 .pg_set_mirror = ff_layout_pg_set_mirror_write, 1237 }; 1238 1239 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs) 1240 { 1241 struct rpc_task *task = &hdr->task; 1242 1243 pnfs_layoutcommit_inode(hdr->inode, false); 1244 1245 if (retry_pnfs) { 1246 dprintk("%s Reset task %5u for i/o through pNFS " 1247 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1248 hdr->task.tk_pid, 1249 hdr->inode->i_sb->s_id, 1250 (unsigned long long)hdr->inode->i_ino, 1251 hdr->args.count, 1252 (unsigned long long)hdr->args.offset); 1253 1254 hdr->completion_ops->reschedule_io(hdr); 1255 return; 1256 } 1257 1258 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1259 dprintk("%s Reset task %5u for i/o through MDS " 1260 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1261 hdr->task.tk_pid, 1262 hdr->inode->i_sb->s_id, 1263 (unsigned long long)hdr->inode->i_ino, 1264 hdr->args.count, 1265 (unsigned long long)hdr->args.offset); 1266 1267 trace_pnfs_mds_fallback_write_done(hdr->inode, 1268 hdr->args.offset, hdr->args.count, 1269 IOMODE_RW, NFS_I(hdr->inode)->layout, 1270 hdr->lseg); 1271 task->tk_status = pnfs_write_done_resend_to_mds(hdr); 1272 } 1273 } 1274 1275 static void ff_layout_resend_pnfs_read(struct nfs_pgio_header *hdr) 1276 { 1277 u32 idx = hdr->pgio_mirror_idx + 1; 1278 u32 new_idx = 0; 1279 u32 dss_id = 0; 1280 struct nfs4_pnfs_ds *ds; 1281 1282 ds = ff_layout_choose_any_ds_for_read(hdr->lseg, idx, &new_idx, 1283 hdr->args.offset, &dss_id); 1284 if (IS_ERR(ds)) 1285 pnfs_error_mark_layout_for_return(hdr->inode, hdr->lseg); 1286 else 1287 ff_layout_send_layouterror(hdr->lseg); 1288 pnfs_read_resend_pnfs(hdr, new_idx); 1289 } 1290 1291 static void ff_layout_reset_read(struct nfs_pgio_header *hdr) 1292 { 1293 struct rpc_task *task = &hdr->task; 1294 1295 pnfs_layoutcommit_inode(hdr->inode, false); 1296 pnfs_error_mark_layout_for_return(hdr->inode, hdr->lseg); 1297 1298 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1299 dprintk("%s Reset task %5u for i/o through MDS " 1300 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, 1301 hdr->task.tk_pid, 1302 hdr->inode->i_sb->s_id, 1303 (unsigned long long)hdr->inode->i_ino, 1304 hdr->args.count, 1305 (unsigned long long)hdr->args.offset); 1306 1307 trace_pnfs_mds_fallback_read_done(hdr->inode, 1308 hdr->args.offset, hdr->args.count, 1309 IOMODE_READ, NFS_I(hdr->inode)->layout, 1310 hdr->lseg); 1311 task->tk_status = pnfs_read_done_resend_to_mds(hdr); 1312 } 1313 } 1314 1315 static int ff_layout_async_handle_error_v4(struct rpc_task *task, 1316 u32 op_status, 1317 struct nfs4_state *state, 1318 struct nfs_client *clp, 1319 struct pnfs_layout_segment *lseg, 1320 u32 idx, u32 dss_id) 1321 { 1322 struct pnfs_layout_hdr *lo = lseg->pls_layout; 1323 struct inode *inode = lo->plh_inode; 1324 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id); 1325 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; 1326 1327 switch (op_status) { 1328 case NFS4_OK: 1329 case NFS4ERR_NXIO: 1330 break; 1331 case NFSERR_PERM: 1332 if (!task->tk_xprt) 1333 break; 1334 xprt_force_disconnect(task->tk_xprt); 1335 goto out_retry; 1336 case NFS4ERR_BADSESSION: 1337 case NFS4ERR_BADSLOT: 1338 case NFS4ERR_BAD_HIGH_SLOT: 1339 case NFS4ERR_DEADSESSION: 1340 case NFS4ERR_CONN_NOT_BOUND_TO_SESSION: 1341 case NFS4ERR_SEQ_FALSE_RETRY: 1342 case NFS4ERR_SEQ_MISORDERED: 1343 dprintk("%s ERROR %d, Reset session. Exchangeid " 1344 "flags 0x%x\n", __func__, task->tk_status, 1345 clp->cl_exchange_flags); 1346 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status); 1347 goto out_retry; 1348 case NFS4ERR_DELAY: 1349 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY); 1350 fallthrough; 1351 case NFS4ERR_GRACE: 1352 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX); 1353 goto out_retry; 1354 case NFS4ERR_RETRY_UNCACHED_REP: 1355 goto out_retry; 1356 /* Invalidate Layout errors */ 1357 case NFS4ERR_PNFS_NO_LAYOUT: 1358 case NFS4ERR_STALE: 1359 case NFS4ERR_BADHANDLE: 1360 case NFS4ERR_ISDIR: 1361 case NFS4ERR_FHEXPIRED: 1362 case NFS4ERR_WRONG_TYPE: 1363 dprintk("%s Invalid layout error %d\n", __func__, 1364 task->tk_status); 1365 /* 1366 * Destroy layout so new i/o will get a new layout. 1367 * Layout will not be destroyed until all current lseg 1368 * references are put. Mark layout as invalid to resend failed 1369 * i/o and all i/o waiting on the slot table to the MDS until 1370 * layout is destroyed and a new valid layout is obtained. 1371 */ 1372 pnfs_destroy_layout(NFS_I(inode)); 1373 rpc_wake_up(&tbl->slot_tbl_waitq); 1374 goto reset; 1375 default: 1376 break; 1377 } 1378 1379 switch (task->tk_status) { 1380 /* RPC connection errors */ 1381 case -ENETDOWN: 1382 case -ENETUNREACH: 1383 if (test_bit(NFS_CS_NETUNREACH_FATAL, &clp->cl_flags)) 1384 return -NFS4ERR_FATAL_IOERROR; 1385 fallthrough; 1386 case -ECONNREFUSED: 1387 case -EHOSTDOWN: 1388 case -EHOSTUNREACH: 1389 case -EIO: 1390 case -ETIMEDOUT: 1391 case -EPIPE: 1392 case -EPROTO: 1393 case -ENODEV: 1394 dprintk("%s DS connection error %d\n", __func__, 1395 task->tk_status); 1396 nfs4_delete_deviceid(devid->ld, devid->nfs_client, 1397 &devid->deviceid); 1398 rpc_wake_up(&tbl->slot_tbl_waitq); 1399 break; 1400 default: 1401 break; 1402 } 1403 1404 if (ff_layout_avoid_mds_available_ds(lseg)) 1405 return -NFS4ERR_RESET_TO_PNFS; 1406 reset: 1407 dprintk("%s Retry through MDS. Error %d\n", __func__, 1408 task->tk_status); 1409 return -NFS4ERR_RESET_TO_MDS; 1410 1411 out_retry: 1412 task->tk_status = 0; 1413 return -EAGAIN; 1414 } 1415 1416 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */ 1417 static int ff_layout_async_handle_error_v3(struct rpc_task *task, 1418 u32 op_status, 1419 struct nfs_client *clp, 1420 struct pnfs_layout_segment *lseg, 1421 u32 idx, u32 dss_id) 1422 { 1423 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id); 1424 1425 switch (op_status) { 1426 case NFS_OK: 1427 case NFSERR_NXIO: 1428 break; 1429 case NFSERR_PERM: 1430 if (!task->tk_xprt) 1431 break; 1432 xprt_force_disconnect(task->tk_xprt); 1433 goto out_retry; 1434 case NFSERR_ACCES: 1435 case NFSERR_BADHANDLE: 1436 case NFSERR_FBIG: 1437 case NFSERR_IO: 1438 case NFSERR_NOSPC: 1439 case NFSERR_ROFS: 1440 case NFSERR_STALE: 1441 goto out_reset_to_pnfs; 1442 case NFSERR_JUKEBOX: 1443 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY); 1444 goto out_retry; 1445 default: 1446 break; 1447 } 1448 1449 switch (task->tk_status) { 1450 /* File access problems. Don't mark the device as unavailable */ 1451 case -EACCES: 1452 case -ESTALE: 1453 case -EISDIR: 1454 case -EBADHANDLE: 1455 case -ELOOP: 1456 case -ENOSPC: 1457 break; 1458 case -EJUKEBOX: 1459 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY); 1460 goto out_retry; 1461 case -ENETDOWN: 1462 case -ENETUNREACH: 1463 if (test_bit(NFS_CS_NETUNREACH_FATAL, &clp->cl_flags)) 1464 return -NFS4ERR_FATAL_IOERROR; 1465 fallthrough; 1466 default: 1467 dprintk("%s DS connection error %d\n", __func__, 1468 task->tk_status); 1469 nfs4_delete_deviceid(devid->ld, devid->nfs_client, 1470 &devid->deviceid); 1471 } 1472 out_reset_to_pnfs: 1473 /* FIXME: Need to prevent infinite looping here. */ 1474 return -NFS4ERR_RESET_TO_PNFS; 1475 out_retry: 1476 task->tk_status = 0; 1477 rpc_restart_call_prepare(task); 1478 rpc_delay(task, NFS_JUKEBOX_RETRY_TIME); 1479 return -EAGAIN; 1480 } 1481 1482 static int ff_layout_async_handle_error(struct rpc_task *task, 1483 u32 op_status, 1484 struct nfs4_state *state, 1485 struct nfs_client *clp, 1486 struct pnfs_layout_segment *lseg, 1487 u32 idx, u32 dss_id) 1488 { 1489 int vers = clp->cl_nfs_mod->rpc_vers->number; 1490 1491 if (task->tk_status >= 0) { 1492 ff_layout_mark_ds_reachable(lseg, idx, dss_id); 1493 return 0; 1494 } 1495 1496 /* Handle the case of an invalid layout segment */ 1497 if (!pnfs_is_valid_lseg(lseg)) 1498 return -NFS4ERR_RESET_TO_PNFS; 1499 1500 switch (vers) { 1501 case 3: 1502 return ff_layout_async_handle_error_v3(task, op_status, clp, 1503 lseg, idx, dss_id); 1504 case 4: 1505 return ff_layout_async_handle_error_v4(task, op_status, state, 1506 clp, lseg, idx, dss_id); 1507 default: 1508 /* should never happen */ 1509 WARN_ON_ONCE(1); 1510 return 0; 1511 } 1512 } 1513 1514 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg, 1515 u32 idx, u32 dss_id, u64 offset, u64 length, 1516 u32 *op_status, int opnum, int error) 1517 { 1518 struct nfs4_ff_layout_mirror *mirror; 1519 u32 status = *op_status; 1520 int err; 1521 1522 if (status == 0) { 1523 switch (error) { 1524 case -ETIMEDOUT: 1525 case -EPFNOSUPPORT: 1526 case -EPROTONOSUPPORT: 1527 case -EOPNOTSUPP: 1528 case -EINVAL: 1529 case -ECONNREFUSED: 1530 case -ECONNRESET: 1531 case -EHOSTDOWN: 1532 case -EHOSTUNREACH: 1533 case -ENETDOWN: 1534 case -ENETUNREACH: 1535 case -EADDRINUSE: 1536 case -ENOBUFS: 1537 case -EPIPE: 1538 case -EPERM: 1539 case -EPROTO: 1540 case -ENODEV: 1541 *op_status = status = NFS4ERR_NXIO; 1542 break; 1543 case -EACCES: 1544 *op_status = status = NFS4ERR_ACCESS; 1545 break; 1546 default: 1547 return; 1548 } 1549 } 1550 1551 mirror = FF_LAYOUT_COMP(lseg, idx); 1552 err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout), 1553 mirror, dss_id, offset, length, status, opnum, 1554 nfs_io_gfp_mask()); 1555 1556 switch (status) { 1557 case NFS4ERR_DELAY: 1558 case NFS4ERR_GRACE: 1559 case NFS4ERR_PERM: 1560 break; 1561 case NFS4ERR_NXIO: 1562 ff_layout_mark_ds_unreachable(lseg, idx, dss_id); 1563 /* 1564 * Don't return the layout if this is a read and we still 1565 * have layouts to try 1566 */ 1567 if (opnum == OP_READ) 1568 break; 1569 fallthrough; 1570 default: 1571 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, 1572 lseg); 1573 } 1574 1575 dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status); 1576 } 1577 1578 /* NFS_PROTO call done callback routines */ 1579 static int ff_layout_read_done_cb(struct rpc_task *task, 1580 struct nfs_pgio_header *hdr) 1581 { 1582 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(hdr->lseg); 1583 u32 dss_id = nfs4_ff_layout_calc_dss_id( 1584 flseg->stripe_unit, 1585 flseg->mirror_array[hdr->pgio_mirror_idx]->dss_count, 1586 hdr->args.offset); 1587 int err; 1588 1589 if (task->tk_status < 0) { 1590 ff_layout_io_track_ds_error(hdr->lseg, 1591 hdr->pgio_mirror_idx, dss_id, 1592 hdr->args.offset, hdr->args.count, 1593 &hdr->res.op_status, OP_READ, 1594 task->tk_status); 1595 trace_ff_layout_read_error(hdr, task->tk_status); 1596 } 1597 1598 err = ff_layout_async_handle_error(task, hdr->res.op_status, 1599 hdr->args.context->state, 1600 hdr->ds_clp, hdr->lseg, 1601 hdr->pgio_mirror_idx, 1602 dss_id); 1603 1604 trace_nfs4_pnfs_read(hdr, err); 1605 clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags); 1606 clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags); 1607 switch (err) { 1608 case -NFS4ERR_RESET_TO_PNFS: 1609 set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags); 1610 return task->tk_status; 1611 case -NFS4ERR_RESET_TO_MDS: 1612 set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags); 1613 return task->tk_status; 1614 case -EAGAIN: 1615 goto out_eagain; 1616 case -NFS4ERR_FATAL_IOERROR: 1617 task->tk_status = -EIO; 1618 return 0; 1619 } 1620 1621 return 0; 1622 out_eagain: 1623 rpc_restart_call_prepare(task); 1624 return -EAGAIN; 1625 } 1626 1627 static bool 1628 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg) 1629 { 1630 return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT); 1631 } 1632 1633 /* 1634 * We reference the rpc_cred of the first WRITE that triggers the need for 1635 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. 1636 * rfc5661 is not clear about which credential should be used. 1637 * 1638 * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so 1639 * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751 1640 * we always send layoutcommit after DS writes. 1641 */ 1642 static void 1643 ff_layout_set_layoutcommit(struct inode *inode, 1644 struct pnfs_layout_segment *lseg, 1645 loff_t end_offset) 1646 { 1647 if (!ff_layout_need_layoutcommit(lseg)) 1648 return; 1649 1650 pnfs_set_layoutcommit(inode, lseg, end_offset); 1651 dprintk("%s inode %llu pls_end_pos %llu\n", __func__, inode->i_ino, 1652 (unsigned long long) NFS_I(inode)->layout->plh_lwb); 1653 } 1654 1655 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task, 1656 struct nfs_pgio_header *hdr) 1657 { 1658 struct nfs4_ff_layout_mirror *mirror; 1659 u32 dss_id; 1660 1661 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags)) 1662 return; 1663 1664 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx); 1665 dss_id = nfs4_ff_layout_calc_dss_id( 1666 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit, 1667 mirror->dss_count, 1668 hdr->args.offset); 1669 1670 nfs4_ff_layout_stat_io_start_read( 1671 hdr->inode, 1672 mirror, 1673 dss_id, 1674 hdr->args.count, 1675 task->tk_start); 1676 } 1677 1678 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task, 1679 struct nfs_pgio_header *hdr) 1680 { 1681 struct nfs4_ff_layout_mirror *mirror; 1682 u32 dss_id; 1683 1684 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags)) 1685 return; 1686 1687 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx); 1688 dss_id = nfs4_ff_layout_calc_dss_id( 1689 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit, 1690 mirror->dss_count, 1691 hdr->args.offset); 1692 1693 nfs4_ff_layout_stat_io_end_read( 1694 task, 1695 mirror, 1696 dss_id, 1697 hdr->args.count, 1698 hdr->res.count); 1699 set_bit(NFS_LSEG_LAYOUTRETURN, &hdr->lseg->pls_flags); 1700 } 1701 1702 static int ff_layout_read_prepare_common(struct rpc_task *task, 1703 struct nfs_pgio_header *hdr) 1704 { 1705 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { 1706 rpc_exit(task, -EIO); 1707 return -EIO; 1708 } 1709 1710 if (!pnfs_is_valid_lseg(hdr->lseg)) { 1711 rpc_exit(task, -EAGAIN); 1712 return -EAGAIN; 1713 } 1714 1715 ff_layout_read_record_layoutstats_start(task, hdr); 1716 return 0; 1717 } 1718 1719 /* 1720 * Call ops for the async read/write cases 1721 * In the case of dense layouts, the offset needs to be reset to its 1722 * original value. 1723 */ 1724 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data) 1725 { 1726 struct nfs_pgio_header *hdr = data; 1727 1728 if (ff_layout_read_prepare_common(task, hdr)) 1729 return; 1730 1731 rpc_call_start(task); 1732 } 1733 1734 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data) 1735 { 1736 struct nfs_pgio_header *hdr = data; 1737 1738 if (nfs4_setup_sequence(hdr->ds_clp, 1739 &hdr->args.seq_args, 1740 &hdr->res.seq_res, 1741 task)) 1742 return; 1743 1744 ff_layout_read_prepare_common(task, hdr); 1745 } 1746 1747 static void ff_layout_read_call_done(struct rpc_task *task, void *data) 1748 { 1749 struct nfs_pgio_header *hdr = data; 1750 1751 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && 1752 task->tk_status == 0) { 1753 nfs4_sequence_done(task, &hdr->res.seq_res); 1754 return; 1755 } 1756 1757 /* Note this may cause RPC to be resent */ 1758 hdr->mds_ops->rpc_call_done(task, hdr); 1759 } 1760 1761 static void ff_layout_read_count_stats(struct rpc_task *task, void *data) 1762 { 1763 struct nfs_pgio_header *hdr = data; 1764 1765 ff_layout_read_record_layoutstats_done(task, hdr); 1766 rpc_count_iostats_metrics(task, 1767 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]); 1768 } 1769 1770 static void ff_layout_read_release(void *data) 1771 { 1772 struct nfs_pgio_header *hdr = data; 1773 1774 ff_layout_read_record_layoutstats_done(&hdr->task, hdr); 1775 if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags)) 1776 ff_layout_resend_pnfs_read(hdr); 1777 else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags)) 1778 ff_layout_reset_read(hdr); 1779 pnfs_generic_rw_release(data); 1780 } 1781 1782 1783 static int ff_layout_write_done_cb(struct rpc_task *task, 1784 struct nfs_pgio_header *hdr) 1785 { 1786 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(hdr->lseg); 1787 u32 dss_id = nfs4_ff_layout_calc_dss_id( 1788 flseg->stripe_unit, 1789 flseg->mirror_array[hdr->pgio_mirror_idx]->dss_count, 1790 hdr->args.offset); 1791 loff_t end_offs = 0; 1792 int err; 1793 1794 if (task->tk_status < 0) { 1795 ff_layout_io_track_ds_error(hdr->lseg, 1796 hdr->pgio_mirror_idx, dss_id, 1797 hdr->args.offset, hdr->args.count, 1798 &hdr->res.op_status, OP_WRITE, 1799 task->tk_status); 1800 trace_ff_layout_write_error(hdr, task->tk_status); 1801 } 1802 1803 err = ff_layout_async_handle_error(task, hdr->res.op_status, 1804 hdr->args.context->state, 1805 hdr->ds_clp, hdr->lseg, 1806 hdr->pgio_mirror_idx, 1807 dss_id); 1808 1809 trace_nfs4_pnfs_write(hdr, err); 1810 clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags); 1811 clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags); 1812 switch (err) { 1813 case -NFS4ERR_RESET_TO_PNFS: 1814 set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags); 1815 return task->tk_status; 1816 case -NFS4ERR_RESET_TO_MDS: 1817 set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags); 1818 return task->tk_status; 1819 case -EAGAIN: 1820 return -EAGAIN; 1821 case -NFS4ERR_FATAL_IOERROR: 1822 task->tk_status = -EIO; 1823 return 0; 1824 } 1825 1826 if (hdr->res.verf->committed == NFS_FILE_SYNC || 1827 hdr->res.verf->committed == NFS_DATA_SYNC) 1828 end_offs = hdr->mds_offset + (loff_t)hdr->res.count; 1829 1830 /* Note: if the write is unstable, don't set end_offs until commit */ 1831 ff_layout_set_layoutcommit(hdr->inode, hdr->lseg, end_offs); 1832 1833 /* zero out fattr since we don't care DS attr at all */ 1834 hdr->fattr.valid = 0; 1835 if (task->tk_status >= 0) 1836 nfs_writeback_update_inode(hdr); 1837 1838 return 0; 1839 } 1840 1841 static int ff_layout_commit_done_cb(struct rpc_task *task, 1842 struct nfs_commit_data *data) 1843 { 1844 int err; 1845 u32 idx = calc_mirror_idx_from_commit(data->lseg, data->ds_commit_index); 1846 u32 dss_id = calc_dss_id_from_commit(data->lseg, data->ds_commit_index); 1847 1848 if (task->tk_status < 0) { 1849 ff_layout_io_track_ds_error(data->lseg, idx, dss_id, 1850 data->args.offset, data->args.count, 1851 &data->res.op_status, OP_COMMIT, 1852 task->tk_status); 1853 trace_ff_layout_commit_error(data, task->tk_status); 1854 } 1855 1856 err = ff_layout_async_handle_error(task, data->res.op_status, 1857 NULL, data->ds_clp, data->lseg, idx, 1858 dss_id); 1859 1860 trace_nfs4_pnfs_commit_ds(data, err); 1861 switch (err) { 1862 case -NFS4ERR_RESET_TO_PNFS: 1863 pnfs_generic_prepare_to_resend_writes(data); 1864 return -EAGAIN; 1865 case -NFS4ERR_RESET_TO_MDS: 1866 pnfs_generic_prepare_to_resend_writes(data); 1867 return -EAGAIN; 1868 case -EAGAIN: 1869 rpc_restart_call_prepare(task); 1870 return -EAGAIN; 1871 case -NFS4ERR_FATAL_IOERROR: 1872 task->tk_status = -EIO; 1873 return 0; 1874 } 1875 1876 ff_layout_set_layoutcommit(data->inode, data->lseg, data->lwb); 1877 return 0; 1878 } 1879 1880 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task, 1881 struct nfs_pgio_header *hdr) 1882 { 1883 struct nfs4_ff_layout_mirror *mirror; 1884 u32 dss_id; 1885 1886 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags)) 1887 return; 1888 1889 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx); 1890 dss_id = nfs4_ff_layout_calc_dss_id( 1891 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit, 1892 mirror->dss_count, 1893 hdr->args.offset); 1894 1895 nfs4_ff_layout_stat_io_start_write( 1896 hdr->inode, 1897 mirror, 1898 dss_id, 1899 hdr->args.count, 1900 task->tk_start); 1901 } 1902 1903 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task, 1904 struct nfs_pgio_header *hdr) 1905 { 1906 struct nfs4_ff_layout_mirror *mirror; 1907 u32 dss_id; 1908 1909 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags)) 1910 return; 1911 1912 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx); 1913 dss_id = nfs4_ff_layout_calc_dss_id( 1914 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit, 1915 mirror->dss_count, 1916 hdr->args.offset); 1917 1918 nfs4_ff_layout_stat_io_end_write( 1919 task, 1920 mirror, 1921 dss_id, 1922 hdr->args.count, 1923 hdr->res.count, 1924 hdr->res.verf->committed); 1925 set_bit(NFS_LSEG_LAYOUTRETURN, &hdr->lseg->pls_flags); 1926 } 1927 1928 static int ff_layout_write_prepare_common(struct rpc_task *task, 1929 struct nfs_pgio_header *hdr) 1930 { 1931 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { 1932 rpc_exit(task, -EIO); 1933 return -EIO; 1934 } 1935 1936 if (!pnfs_is_valid_lseg(hdr->lseg)) { 1937 rpc_exit(task, -EAGAIN); 1938 return -EAGAIN; 1939 } 1940 1941 ff_layout_write_record_layoutstats_start(task, hdr); 1942 return 0; 1943 } 1944 1945 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data) 1946 { 1947 struct nfs_pgio_header *hdr = data; 1948 1949 if (ff_layout_write_prepare_common(task, hdr)) 1950 return; 1951 1952 rpc_call_start(task); 1953 } 1954 1955 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data) 1956 { 1957 struct nfs_pgio_header *hdr = data; 1958 1959 if (nfs4_setup_sequence(hdr->ds_clp, 1960 &hdr->args.seq_args, 1961 &hdr->res.seq_res, 1962 task)) 1963 return; 1964 1965 ff_layout_write_prepare_common(task, hdr); 1966 } 1967 1968 static void ff_layout_write_call_done(struct rpc_task *task, void *data) 1969 { 1970 struct nfs_pgio_header *hdr = data; 1971 1972 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && 1973 task->tk_status == 0) { 1974 nfs4_sequence_done(task, &hdr->res.seq_res); 1975 return; 1976 } 1977 1978 /* Note this may cause RPC to be resent */ 1979 hdr->mds_ops->rpc_call_done(task, hdr); 1980 } 1981 1982 static void ff_layout_write_count_stats(struct rpc_task *task, void *data) 1983 { 1984 struct nfs_pgio_header *hdr = data; 1985 1986 ff_layout_write_record_layoutstats_done(task, hdr); 1987 rpc_count_iostats_metrics(task, 1988 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]); 1989 } 1990 1991 static void ff_layout_write_release(void *data) 1992 { 1993 struct nfs_pgio_header *hdr = data; 1994 1995 ff_layout_write_record_layoutstats_done(&hdr->task, hdr); 1996 if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags)) { 1997 ff_layout_send_layouterror(hdr->lseg); 1998 ff_layout_reset_write(hdr, true); 1999 } else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags)) 2000 ff_layout_reset_write(hdr, false); 2001 pnfs_generic_rw_release(data); 2002 } 2003 2004 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task, 2005 struct nfs_commit_data *cdata) 2006 { 2007 u32 idx, dss_id; 2008 2009 if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags)) 2010 return; 2011 2012 idx = calc_mirror_idx_from_commit(cdata->lseg, cdata->ds_commit_index); 2013 dss_id = calc_dss_id_from_commit(cdata->lseg, cdata->ds_commit_index); 2014 nfs4_ff_layout_stat_io_start_write(cdata->inode, 2015 FF_LAYOUT_COMP(cdata->lseg, idx), 2016 dss_id, 2017 0, task->tk_start); 2018 } 2019 2020 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task, 2021 struct nfs_commit_data *cdata) 2022 { 2023 struct nfs_page *req; 2024 __u64 count = 0; 2025 u32 idx, dss_id; 2026 2027 if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags)) 2028 return; 2029 2030 if (task->tk_status == 0) { 2031 list_for_each_entry(req, &cdata->pages, wb_list) 2032 count += req->wb_bytes; 2033 } 2034 2035 idx = calc_mirror_idx_from_commit(cdata->lseg, cdata->ds_commit_index); 2036 dss_id = calc_dss_id_from_commit(cdata->lseg, cdata->ds_commit_index); 2037 nfs4_ff_layout_stat_io_end_write(task, 2038 FF_LAYOUT_COMP(cdata->lseg, idx), 2039 dss_id, 2040 count, count, NFS_FILE_SYNC); 2041 set_bit(NFS_LSEG_LAYOUTRETURN, &cdata->lseg->pls_flags); 2042 } 2043 2044 static int ff_layout_commit_prepare_common(struct rpc_task *task, 2045 struct nfs_commit_data *cdata) 2046 { 2047 if (!pnfs_is_valid_lseg(cdata->lseg)) { 2048 rpc_exit(task, -EAGAIN); 2049 return -EAGAIN; 2050 } 2051 2052 ff_layout_commit_record_layoutstats_start(task, cdata); 2053 return 0; 2054 } 2055 2056 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data) 2057 { 2058 if (ff_layout_commit_prepare_common(task, data)) 2059 return; 2060 2061 rpc_call_start(task); 2062 } 2063 2064 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data) 2065 { 2066 struct nfs_commit_data *wdata = data; 2067 2068 if (nfs4_setup_sequence(wdata->ds_clp, 2069 &wdata->args.seq_args, 2070 &wdata->res.seq_res, 2071 task)) 2072 return; 2073 ff_layout_commit_prepare_common(task, data); 2074 } 2075 2076 static void ff_layout_commit_done(struct rpc_task *task, void *data) 2077 { 2078 pnfs_generic_write_commit_done(task, data); 2079 } 2080 2081 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data) 2082 { 2083 struct nfs_commit_data *cdata = data; 2084 2085 ff_layout_commit_record_layoutstats_done(task, cdata); 2086 rpc_count_iostats_metrics(task, 2087 &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]); 2088 } 2089 2090 static void ff_layout_commit_release(void *data) 2091 { 2092 struct nfs_commit_data *cdata = data; 2093 2094 ff_layout_commit_record_layoutstats_done(&cdata->task, cdata); 2095 pnfs_generic_commit_release(data); 2096 } 2097 2098 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = { 2099 .rpc_call_prepare = ff_layout_read_prepare_v3, 2100 .rpc_call_done = ff_layout_read_call_done, 2101 .rpc_count_stats = ff_layout_read_count_stats, 2102 .rpc_release = ff_layout_read_release, 2103 }; 2104 2105 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = { 2106 .rpc_call_prepare = ff_layout_read_prepare_v4, 2107 .rpc_call_done = ff_layout_read_call_done, 2108 .rpc_count_stats = ff_layout_read_count_stats, 2109 .rpc_release = ff_layout_read_release, 2110 }; 2111 2112 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = { 2113 .rpc_call_prepare = ff_layout_write_prepare_v3, 2114 .rpc_call_done = ff_layout_write_call_done, 2115 .rpc_count_stats = ff_layout_write_count_stats, 2116 .rpc_release = ff_layout_write_release, 2117 }; 2118 2119 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = { 2120 .rpc_call_prepare = ff_layout_write_prepare_v4, 2121 .rpc_call_done = ff_layout_write_call_done, 2122 .rpc_count_stats = ff_layout_write_count_stats, 2123 .rpc_release = ff_layout_write_release, 2124 }; 2125 2126 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = { 2127 .rpc_call_prepare = ff_layout_commit_prepare_v3, 2128 .rpc_call_done = ff_layout_commit_done, 2129 .rpc_count_stats = ff_layout_commit_count_stats, 2130 .rpc_release = ff_layout_commit_release, 2131 }; 2132 2133 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = { 2134 .rpc_call_prepare = ff_layout_commit_prepare_v4, 2135 .rpc_call_done = ff_layout_commit_done, 2136 .rpc_count_stats = ff_layout_commit_count_stats, 2137 .rpc_release = ff_layout_commit_release, 2138 }; 2139 2140 static enum pnfs_try_status 2141 ff_layout_read_pagelist(struct nfs_pgio_header *hdr) 2142 { 2143 struct pnfs_layout_segment *lseg = hdr->lseg; 2144 struct nfs4_pnfs_ds *ds; 2145 struct rpc_clnt *ds_clnt; 2146 struct nfsd_file *localio; 2147 struct nfs4_ff_layout_mirror *mirror; 2148 const struct cred *ds_cred; 2149 loff_t offset = hdr->args.offset; 2150 u32 idx = hdr->pgio_mirror_idx; 2151 int vers; 2152 struct nfs_fh *fh; 2153 u32 dss_id; 2154 bool ds_fatal_error = false; 2155 2156 dprintk("--> %s ino %llu pgbase %u req %zu@%llu\n", 2157 __func__, hdr->inode->i_ino, 2158 hdr->args.pgbase, (size_t)hdr->args.count, offset); 2159 2160 mirror = FF_LAYOUT_COMP(lseg, idx); 2161 dss_id = nfs4_ff_layout_calc_dss_id( 2162 FF_LAYOUT_LSEG(lseg)->stripe_unit, 2163 mirror->dss_count, 2164 offset); 2165 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, dss_id, false); 2166 if (IS_ERR(ds)) { 2167 ds_fatal_error = nfs_error_is_fatal(PTR_ERR(ds)); 2168 goto out_failed; 2169 } 2170 2171 ds_clnt = nfs4_ff_find_or_create_ds_client(mirror, ds->ds_clp, 2172 hdr->inode, dss_id); 2173 if (IS_ERR(ds_clnt)) 2174 goto out_failed; 2175 2176 ds_cred = ff_layout_get_ds_cred(mirror, &lseg->pls_range, hdr->cred, dss_id); 2177 if (!ds_cred) 2178 goto out_failed; 2179 2180 vers = nfs4_ff_layout_ds_version(mirror, dss_id); 2181 2182 dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__, 2183 ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count), vers); 2184 2185 hdr->pgio_done_cb = ff_layout_read_done_cb; 2186 refcount_inc(&ds->ds_clp->cl_count); 2187 hdr->ds_clp = ds->ds_clp; 2188 fh = nfs4_ff_layout_select_ds_fh(mirror, dss_id); 2189 if (fh) 2190 hdr->args.fh = fh; 2191 2192 nfs4_ff_layout_select_ds_stateid(mirror, dss_id, &hdr->args.stateid); 2193 2194 /* 2195 * Note that if we ever decide to split across DSes, 2196 * then we may need to handle dense-like offsets. 2197 */ 2198 hdr->args.offset = offset; 2199 hdr->mds_offset = offset; 2200 2201 /* Start IO accounting for local read */ 2202 localio = ff_local_open_fh(lseg, idx, dss_id, ds->ds_clp, ds_cred, fh, 2203 FMODE_READ); 2204 if (localio) { 2205 hdr->task.tk_start = ktime_get(); 2206 ff_layout_read_record_layoutstats_start(&hdr->task, hdr); 2207 } 2208 2209 /* Perform an asynchronous read to ds */ 2210 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops, 2211 vers == 3 ? &ff_layout_read_call_ops_v3 : 2212 &ff_layout_read_call_ops_v4, 2213 0, RPC_TASK_SOFTCONN, localio); 2214 put_cred(ds_cred); 2215 return PNFS_ATTEMPTED; 2216 2217 out_failed: 2218 if (ff_layout_avoid_mds_available_ds(lseg) && !ds_fatal_error) 2219 return PNFS_TRY_AGAIN; 2220 if (ff_layout_no_fallback_to_mds(lseg)) { 2221 /* 2222 * FF_FLAGS_NO_IO_THRU_MDS: force fresh LAYOUTGET, 2223 * never fall through to MDS I/O. 2224 */ 2225 pnfs_error_mark_layout_for_return(hdr->inode, lseg); 2226 return PNFS_TRY_AGAIN; 2227 } 2228 trace_pnfs_mds_fallback_read_pagelist(hdr->inode, 2229 hdr->args.offset, hdr->args.count, 2230 IOMODE_READ, NFS_I(hdr->inode)->layout, lseg); 2231 return PNFS_NOT_ATTEMPTED; 2232 } 2233 2234 /* Perform async writes. */ 2235 static enum pnfs_try_status 2236 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync) 2237 { 2238 struct pnfs_layout_segment *lseg = hdr->lseg; 2239 struct nfs4_pnfs_ds *ds; 2240 struct rpc_clnt *ds_clnt; 2241 struct nfsd_file *localio; 2242 struct nfs4_ff_layout_mirror *mirror; 2243 const struct cred *ds_cred; 2244 loff_t offset = hdr->args.offset; 2245 int vers; 2246 struct nfs_fh *fh; 2247 u32 idx = hdr->pgio_mirror_idx; 2248 u32 dss_id; 2249 bool ds_fatal_error = false; 2250 2251 mirror = FF_LAYOUT_COMP(lseg, idx); 2252 dss_id = nfs4_ff_layout_calc_dss_id( 2253 FF_LAYOUT_LSEG(lseg)->stripe_unit, 2254 mirror->dss_count, 2255 offset); 2256 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, dss_id, true); 2257 if (IS_ERR(ds)) { 2258 ds_fatal_error = nfs_error_is_fatal(PTR_ERR(ds)); 2259 goto out_failed; 2260 } 2261 2262 ds_clnt = nfs4_ff_find_or_create_ds_client(mirror, ds->ds_clp, 2263 hdr->inode, dss_id); 2264 if (IS_ERR(ds_clnt)) 2265 goto out_failed; 2266 2267 ds_cred = ff_layout_get_ds_cred(mirror, &lseg->pls_range, hdr->cred, dss_id); 2268 if (!ds_cred) 2269 goto out_failed; 2270 2271 vers = nfs4_ff_layout_ds_version(mirror, dss_id); 2272 2273 dprintk("%s ino %llu sync %d req %zu@%llu DS: %s cl_count %d vers %d\n", 2274 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count, 2275 offset, ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count), 2276 vers); 2277 2278 hdr->pgio_done_cb = ff_layout_write_done_cb; 2279 refcount_inc(&ds->ds_clp->cl_count); 2280 hdr->ds_clp = ds->ds_clp; 2281 hdr->ds_commit_idx = calc_commit_idx(lseg, idx, dss_id); 2282 fh = nfs4_ff_layout_select_ds_fh(mirror, dss_id); 2283 if (fh) 2284 hdr->args.fh = fh; 2285 2286 nfs4_ff_layout_select_ds_stateid(mirror, dss_id, &hdr->args.stateid); 2287 2288 /* 2289 * Note that if we ever decide to split across DSes, 2290 * then we may need to handle dense-like offsets. 2291 */ 2292 hdr->args.offset = offset; 2293 2294 /* Start IO accounting for local write */ 2295 localio = ff_local_open_fh(lseg, idx, dss_id, ds->ds_clp, ds_cred, fh, 2296 FMODE_READ|FMODE_WRITE); 2297 if (localio) { 2298 hdr->task.tk_start = ktime_get(); 2299 ff_layout_write_record_layoutstats_start(&hdr->task, hdr); 2300 } 2301 2302 /* Perform an asynchronous write */ 2303 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops, 2304 vers == 3 ? &ff_layout_write_call_ops_v3 : 2305 &ff_layout_write_call_ops_v4, 2306 sync, RPC_TASK_SOFTCONN, localio); 2307 put_cred(ds_cred); 2308 return PNFS_ATTEMPTED; 2309 2310 out_failed: 2311 if (ff_layout_avoid_mds_available_ds(lseg) && !ds_fatal_error) 2312 return PNFS_TRY_AGAIN; 2313 if (ff_layout_no_fallback_to_mds(lseg)) { 2314 /* 2315 * FF_FLAGS_NO_IO_THRU_MDS: force fresh LAYOUTGET, 2316 * never fall through to MDS I/O. 2317 */ 2318 pnfs_error_mark_layout_for_return(hdr->inode, lseg); 2319 return PNFS_TRY_AGAIN; 2320 } 2321 trace_pnfs_mds_fallback_write_pagelist(hdr->inode, 2322 hdr->args.offset, hdr->args.count, 2323 IOMODE_RW, NFS_I(hdr->inode)->layout, lseg); 2324 return PNFS_NOT_ATTEMPTED; 2325 } 2326 2327 static struct nfs_fh * 2328 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i, u32 dss_id) 2329 { 2330 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg); 2331 2332 /* FIXME: Assume that there is only one NFS version available 2333 * for the DS. 2334 */ 2335 return &flseg->mirror_array[i]->dss[dss_id].fh_versions[0]; 2336 } 2337 2338 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how) 2339 { 2340 struct pnfs_layout_segment *lseg = data->lseg; 2341 struct nfs4_pnfs_ds *ds; 2342 struct rpc_clnt *ds_clnt; 2343 struct nfsd_file *localio; 2344 struct nfs4_ff_layout_mirror *mirror; 2345 const struct cred *ds_cred; 2346 u32 idx, dss_id; 2347 int vers, ret; 2348 struct nfs_fh *fh; 2349 2350 if (!lseg || !(pnfs_is_valid_lseg(lseg) || 2351 test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))) 2352 goto out_err; 2353 2354 idx = calc_mirror_idx_from_commit(lseg, data->ds_commit_index); 2355 mirror = FF_LAYOUT_COMP(lseg, idx); 2356 dss_id = calc_dss_id_from_commit(lseg, data->ds_commit_index); 2357 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, dss_id, true); 2358 if (IS_ERR(ds)) 2359 goto out_err; 2360 2361 ds_clnt = nfs4_ff_find_or_create_ds_client(mirror, ds->ds_clp, 2362 data->inode, dss_id); 2363 if (IS_ERR(ds_clnt)) 2364 goto out_err; 2365 2366 ds_cred = ff_layout_get_ds_cred(mirror, &lseg->pls_range, data->cred, dss_id); 2367 if (!ds_cred) 2368 goto out_err; 2369 2370 vers = nfs4_ff_layout_ds_version(mirror, dss_id); 2371 2372 dprintk("%s ino %llu, how %d cl_count %d vers %d\n", __func__, 2373 data->inode->i_ino, how, refcount_read(&ds->ds_clp->cl_count), 2374 vers); 2375 data->commit_done_cb = ff_layout_commit_done_cb; 2376 data->cred = ds_cred; 2377 refcount_inc(&ds->ds_clp->cl_count); 2378 data->ds_clp = ds->ds_clp; 2379 fh = select_ds_fh_from_commit(lseg, idx, dss_id); 2380 if (fh) 2381 data->args.fh = fh; 2382 2383 /* Start IO accounting for local commit */ 2384 localio = ff_local_open_fh(lseg, idx, dss_id, ds->ds_clp, ds_cred, fh, 2385 FMODE_READ|FMODE_WRITE); 2386 if (localio) { 2387 data->task.tk_start = ktime_get(); 2388 ff_layout_commit_record_layoutstats_start(&data->task, data); 2389 } 2390 2391 ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops, 2392 vers == 3 ? &ff_layout_commit_call_ops_v3 : 2393 &ff_layout_commit_call_ops_v4, 2394 how, RPC_TASK_SOFTCONN, localio); 2395 put_cred(ds_cred); 2396 return ret; 2397 out_err: 2398 pnfs_generic_prepare_to_resend_writes(data); 2399 pnfs_generic_commit_release(data); 2400 return -EAGAIN; 2401 } 2402 2403 static int 2404 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, 2405 int how, struct nfs_commit_info *cinfo) 2406 { 2407 return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo, 2408 ff_layout_initiate_commit); 2409 } 2410 2411 static bool ff_layout_match_rw(const struct rpc_task *task, 2412 const struct nfs_pgio_header *hdr, 2413 const struct pnfs_layout_segment *lseg) 2414 { 2415 return hdr->lseg == lseg; 2416 } 2417 2418 static bool ff_layout_match_commit(const struct rpc_task *task, 2419 const struct nfs_commit_data *cdata, 2420 const struct pnfs_layout_segment *lseg) 2421 { 2422 return cdata->lseg == lseg; 2423 } 2424 2425 static bool ff_layout_match_io(const struct rpc_task *task, const void *data) 2426 { 2427 const struct rpc_call_ops *ops = task->tk_ops; 2428 2429 if (ops == &ff_layout_read_call_ops_v3 || 2430 ops == &ff_layout_read_call_ops_v4 || 2431 ops == &ff_layout_write_call_ops_v3 || 2432 ops == &ff_layout_write_call_ops_v4) 2433 return ff_layout_match_rw(task, task->tk_calldata, data); 2434 if (ops == &ff_layout_commit_call_ops_v3 || 2435 ops == &ff_layout_commit_call_ops_v4) 2436 return ff_layout_match_commit(task, task->tk_calldata, data); 2437 return false; 2438 } 2439 2440 static void ff_layout_cancel_io(struct pnfs_layout_segment *lseg) 2441 { 2442 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg); 2443 struct nfs4_ff_layout_mirror *mirror; 2444 struct nfs4_ff_layout_ds *mirror_ds; 2445 struct nfs4_pnfs_ds *ds; 2446 struct nfs_client *ds_clp; 2447 struct rpc_clnt *clnt; 2448 u32 idx, dss_id; 2449 2450 for (idx = 0; idx < flseg->mirror_array_cnt; idx++) { 2451 mirror = flseg->mirror_array[idx]; 2452 for (dss_id = 0; dss_id < mirror->dss_count; dss_id++) { 2453 mirror_ds = mirror->dss[dss_id].mirror_ds; 2454 if (IS_ERR_OR_NULL(mirror_ds)) 2455 continue; 2456 ds = mirror->dss[dss_id].mirror_ds->ds; 2457 if (!ds) 2458 continue; 2459 ds_clp = ds->ds_clp; 2460 if (!ds_clp) 2461 continue; 2462 clnt = ds_clp->cl_rpcclient; 2463 if (!clnt) 2464 continue; 2465 if (!rpc_cancel_tasks(clnt, -EAGAIN, 2466 ff_layout_match_io, lseg)) 2467 continue; 2468 rpc_clnt_disconnect(clnt); 2469 } 2470 } 2471 } 2472 2473 static struct pnfs_ds_commit_info * 2474 ff_layout_get_ds_info(struct inode *inode) 2475 { 2476 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; 2477 2478 if (layout == NULL) 2479 return NULL; 2480 2481 return &FF_LAYOUT_FROM_HDR(layout)->commit_info; 2482 } 2483 2484 static void 2485 ff_layout_setup_ds_info(struct pnfs_ds_commit_info *fl_cinfo, 2486 struct pnfs_layout_segment *lseg) 2487 { 2488 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg); 2489 struct inode *inode = lseg->pls_layout->plh_inode; 2490 struct pnfs_commit_array *array, *new; 2491 u32 size = flseg->mirror_array_cnt * flseg->mirror_array[0]->dss_count; 2492 2493 new = pnfs_alloc_commit_array(size, 2494 nfs_io_gfp_mask()); 2495 if (new) { 2496 spin_lock(&inode->i_lock); 2497 array = pnfs_add_commit_array(fl_cinfo, new, lseg); 2498 spin_unlock(&inode->i_lock); 2499 if (array != new) 2500 pnfs_free_commit_array(new); 2501 } 2502 } 2503 2504 static void 2505 ff_layout_release_ds_info(struct pnfs_ds_commit_info *fl_cinfo, 2506 struct inode *inode) 2507 { 2508 spin_lock(&inode->i_lock); 2509 pnfs_generic_ds_cinfo_destroy(fl_cinfo); 2510 spin_unlock(&inode->i_lock); 2511 } 2512 2513 static void 2514 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d) 2515 { 2516 nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds, 2517 id_node)); 2518 } 2519 2520 static int ff_layout_encode_ioerr(struct xdr_stream *xdr, 2521 const struct nfs4_layoutreturn_args *args, 2522 const struct nfs4_flexfile_layoutreturn_args *ff_args) 2523 { 2524 __be32 *start; 2525 2526 start = xdr_reserve_space(xdr, 4); 2527 if (unlikely(!start)) 2528 return -E2BIG; 2529 2530 *start = cpu_to_be32(ff_args->num_errors); 2531 /* This assume we always return _ALL_ layouts */ 2532 return ff_layout_encode_ds_ioerr(xdr, &ff_args->errors); 2533 } 2534 2535 static void 2536 ff_layout_encode_ff_iostat_head(struct xdr_stream *xdr, 2537 const nfs4_stateid *stateid, 2538 const struct nfs42_layoutstat_devinfo *devinfo) 2539 { 2540 __be32 *p; 2541 2542 p = xdr_reserve_space(xdr, 8 + 8); 2543 p = xdr_encode_hyper(p, devinfo->offset); 2544 p = xdr_encode_hyper(p, devinfo->length); 2545 encode_opaque_fixed(xdr, stateid->data, NFS4_STATEID_SIZE); 2546 p = xdr_reserve_space(xdr, 4*8); 2547 p = xdr_encode_hyper(p, devinfo->read_count); 2548 p = xdr_encode_hyper(p, devinfo->read_bytes); 2549 p = xdr_encode_hyper(p, devinfo->write_count); 2550 p = xdr_encode_hyper(p, devinfo->write_bytes); 2551 encode_opaque_fixed(xdr, devinfo->dev_id.data, NFS4_DEVICEID4_SIZE); 2552 } 2553 2554 static void 2555 ff_layout_encode_ff_iostat(struct xdr_stream *xdr, 2556 const nfs4_stateid *stateid, 2557 const struct nfs42_layoutstat_devinfo *devinfo) 2558 { 2559 ff_layout_encode_ff_iostat_head(xdr, stateid, devinfo); 2560 ff_layout_encode_ff_layoutupdate(xdr, devinfo, 2561 devinfo->ld_private.data); 2562 } 2563 2564 /* report nothing for now */ 2565 static void ff_layout_encode_iostats_array(struct xdr_stream *xdr, 2566 const struct nfs4_layoutreturn_args *args, 2567 struct nfs4_flexfile_layoutreturn_args *ff_args) 2568 { 2569 __be32 *p; 2570 int i; 2571 2572 p = xdr_reserve_space(xdr, 4); 2573 *p = cpu_to_be32(ff_args->num_dev); 2574 for (i = 0; i < ff_args->num_dev; i++) 2575 ff_layout_encode_ff_iostat(xdr, 2576 &args->layout->plh_stateid, 2577 &ff_args->devinfo[i]); 2578 } 2579 2580 static void 2581 ff_layout_free_iostats_array(struct nfs42_layoutstat_devinfo *devinfo, 2582 unsigned int num_entries) 2583 { 2584 unsigned int i; 2585 2586 for (i = 0; i < num_entries; i++) { 2587 if (!devinfo[i].ld_private.ops) 2588 continue; 2589 if (!devinfo[i].ld_private.ops->free) 2590 continue; 2591 devinfo[i].ld_private.ops->free(&devinfo[i].ld_private); 2592 } 2593 } 2594 2595 static struct nfs4_deviceid_node * 2596 ff_layout_alloc_deviceid_node(struct nfs_server *server, 2597 struct pnfs_device *pdev, gfp_t gfp_flags) 2598 { 2599 struct nfs4_ff_layout_ds *dsaddr; 2600 2601 dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags); 2602 if (!dsaddr) 2603 return NULL; 2604 return &dsaddr->id_node; 2605 } 2606 2607 static void 2608 ff_layout_encode_layoutreturn(struct xdr_stream *xdr, 2609 const void *voidargs, 2610 const struct nfs4_xdr_opaque_data *ff_opaque) 2611 { 2612 const struct nfs4_layoutreturn_args *args = voidargs; 2613 struct nfs4_flexfile_layoutreturn_args *ff_args = ff_opaque->data; 2614 struct xdr_buf tmp_buf = { 2615 .head = { 2616 [0] = { 2617 .iov_base = page_address(ff_args->pages[0]), 2618 }, 2619 }, 2620 .buflen = PAGE_SIZE, 2621 }; 2622 struct xdr_stream tmp_xdr; 2623 __be32 *start; 2624 2625 dprintk("%s: Begin\n", __func__); 2626 2627 xdr_init_encode(&tmp_xdr, &tmp_buf, NULL, NULL); 2628 2629 ff_layout_encode_ioerr(&tmp_xdr, args, ff_args); 2630 ff_layout_encode_iostats_array(&tmp_xdr, args, ff_args); 2631 2632 start = xdr_reserve_space(xdr, 4); 2633 *start = cpu_to_be32(tmp_buf.len); 2634 xdr_write_pages(xdr, ff_args->pages, 0, tmp_buf.len); 2635 2636 dprintk("%s: Return\n", __func__); 2637 } 2638 2639 static void 2640 ff_layout_free_layoutreturn(struct nfs4_xdr_opaque_data *args) 2641 { 2642 struct nfs4_flexfile_layoutreturn_args *ff_args; 2643 2644 if (!args->data) 2645 return; 2646 ff_args = args->data; 2647 args->data = NULL; 2648 2649 ff_layout_free_ds_ioerr(&ff_args->errors); 2650 ff_layout_free_iostats_array(ff_args->devinfo, ff_args->num_dev); 2651 2652 put_page(ff_args->pages[0]); 2653 kfree(ff_args); 2654 } 2655 2656 static const struct nfs4_xdr_opaque_ops layoutreturn_ops = { 2657 .encode = ff_layout_encode_layoutreturn, 2658 .free = ff_layout_free_layoutreturn, 2659 }; 2660 2661 static int 2662 ff_layout_prepare_layoutreturn(struct nfs4_layoutreturn_args *args) 2663 { 2664 struct nfs4_flexfile_layoutreturn_args *ff_args; 2665 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(args->layout); 2666 2667 ff_args = kmalloc_obj(*ff_args, nfs_io_gfp_mask()); 2668 if (!ff_args) 2669 goto out_nomem; 2670 ff_args->pages[0] = alloc_page(nfs_io_gfp_mask()); 2671 if (!ff_args->pages[0]) 2672 goto out_nomem_free; 2673 2674 INIT_LIST_HEAD(&ff_args->errors); 2675 ff_args->num_errors = ff_layout_fetch_ds_ioerr(args->layout, 2676 &args->range, &ff_args->errors, 2677 FF_LAYOUTRETURN_MAXERR); 2678 2679 spin_lock(&args->inode->i_lock); 2680 ff_args->num_dev = ff_layout_mirror_prepare_stats( 2681 &ff_layout->generic_hdr, &ff_args->devinfo[0], 2682 ARRAY_SIZE(ff_args->devinfo), NFS4_FF_OP_LAYOUTRETURN); 2683 spin_unlock(&args->inode->i_lock); 2684 2685 args->ld_private->ops = &layoutreturn_ops; 2686 args->ld_private->data = ff_args; 2687 return 0; 2688 out_nomem_free: 2689 kfree(ff_args); 2690 out_nomem: 2691 return -ENOMEM; 2692 } 2693 2694 #ifdef CONFIG_NFS_V4_2 2695 void 2696 ff_layout_send_layouterror(struct pnfs_layout_segment *lseg) 2697 { 2698 struct pnfs_layout_hdr *lo = lseg->pls_layout; 2699 struct nfs42_layout_error *errors; 2700 LIST_HEAD(head); 2701 2702 if (!nfs_server_capable(lo->plh_inode, NFS_CAP_LAYOUTERROR)) 2703 return; 2704 ff_layout_fetch_ds_ioerr(lo, &lseg->pls_range, &head, -1); 2705 if (list_empty(&head)) 2706 return; 2707 2708 errors = kmalloc_objs(*errors, NFS42_LAYOUTERROR_MAX, nfs_io_gfp_mask()); 2709 if (errors != NULL) { 2710 const struct nfs4_ff_layout_ds_err *pos; 2711 size_t n = 0; 2712 2713 list_for_each_entry(pos, &head, list) { 2714 errors[n].offset = pos->offset; 2715 errors[n].length = pos->length; 2716 nfs4_stateid_copy(&errors[n].stateid, &pos->stateid); 2717 errors[n].errors[0].dev_id = pos->deviceid; 2718 errors[n].errors[0].status = pos->status; 2719 errors[n].errors[0].opnum = pos->opnum; 2720 n++; 2721 if (!list_is_last(&pos->list, &head) && 2722 n < NFS42_LAYOUTERROR_MAX) 2723 continue; 2724 if (nfs42_proc_layouterror(lseg, errors, n) < 0) 2725 break; 2726 n = 0; 2727 } 2728 kfree(errors); 2729 } 2730 ff_layout_free_ds_ioerr(&head); 2731 } 2732 #else 2733 void 2734 ff_layout_send_layouterror(struct pnfs_layout_segment *lseg) 2735 { 2736 } 2737 #endif 2738 2739 static int 2740 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen) 2741 { 2742 const struct sockaddr_in *sin = (struct sockaddr_in *)sap; 2743 2744 return snprintf(buf, buflen, "%pI4", &sin->sin_addr); 2745 } 2746 2747 static size_t 2748 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf, 2749 const int buflen) 2750 { 2751 const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap; 2752 const struct in6_addr *addr = &sin6->sin6_addr; 2753 2754 /* 2755 * RFC 4291, Section 2.2.2 2756 * 2757 * Shorthanded ANY address 2758 */ 2759 if (ipv6_addr_any(addr)) 2760 return snprintf(buf, buflen, "::"); 2761 2762 /* 2763 * RFC 4291, Section 2.2.2 2764 * 2765 * Shorthanded loopback address 2766 */ 2767 if (ipv6_addr_loopback(addr)) 2768 return snprintf(buf, buflen, "::1"); 2769 2770 /* 2771 * RFC 4291, Section 2.2.3 2772 * 2773 * Special presentation address format for mapped v4 2774 * addresses. 2775 */ 2776 if (ipv6_addr_v4mapped(addr)) 2777 return snprintf(buf, buflen, "::ffff:%pI4", 2778 &addr->s6_addr32[3]); 2779 2780 /* 2781 * RFC 4291, Section 2.2.1 2782 */ 2783 return snprintf(buf, buflen, "%pI6c", addr); 2784 } 2785 2786 /* Derived from rpc_sockaddr2uaddr */ 2787 static void 2788 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da) 2789 { 2790 struct sockaddr *sap = (struct sockaddr *)&da->da_addr; 2791 char portbuf[RPCBIND_MAXUADDRPLEN]; 2792 char addrbuf[RPCBIND_MAXUADDRLEN]; 2793 unsigned short port; 2794 int len, netid_len; 2795 __be32 *p; 2796 2797 switch (sap->sa_family) { 2798 case AF_INET: 2799 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0) 2800 return; 2801 port = ntohs(((struct sockaddr_in *)sap)->sin_port); 2802 break; 2803 case AF_INET6: 2804 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0) 2805 return; 2806 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port); 2807 break; 2808 default: 2809 WARN_ON_ONCE(1); 2810 return; 2811 } 2812 2813 snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff); 2814 len = strlcat(addrbuf, portbuf, sizeof(addrbuf)); 2815 2816 netid_len = strlen(da->da_netid); 2817 p = xdr_reserve_space(xdr, 4 + netid_len); 2818 xdr_encode_opaque(p, da->da_netid, netid_len); 2819 2820 p = xdr_reserve_space(xdr, 4 + len); 2821 xdr_encode_opaque(p, addrbuf, len); 2822 } 2823 2824 static void 2825 ff_layout_encode_nfstime(struct xdr_stream *xdr, 2826 ktime_t t) 2827 { 2828 struct timespec64 ts; 2829 __be32 *p; 2830 2831 p = xdr_reserve_space(xdr, 12); 2832 ts = ktime_to_timespec64(t); 2833 p = xdr_encode_hyper(p, ts.tv_sec); 2834 *p++ = cpu_to_be32(ts.tv_nsec); 2835 } 2836 2837 static void 2838 ff_layout_encode_io_latency(struct xdr_stream *xdr, 2839 struct nfs4_ff_io_stat *stat) 2840 { 2841 __be32 *p; 2842 2843 p = xdr_reserve_space(xdr, 5 * 8); 2844 p = xdr_encode_hyper(p, stat->ops_requested); 2845 p = xdr_encode_hyper(p, stat->bytes_requested); 2846 p = xdr_encode_hyper(p, stat->ops_completed); 2847 p = xdr_encode_hyper(p, stat->bytes_completed); 2848 p = xdr_encode_hyper(p, stat->bytes_not_delivered); 2849 ff_layout_encode_nfstime(xdr, stat->total_busy_time); 2850 ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time); 2851 } 2852 2853 static void 2854 ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr, 2855 const struct nfs42_layoutstat_devinfo *devinfo, 2856 struct nfs4_ff_layout_ds_stripe *dss_info) 2857 { 2858 struct nfs4_pnfs_ds_addr *da; 2859 struct nfs4_pnfs_ds *ds = dss_info->mirror_ds->ds; 2860 struct nfs_fh *fh = &dss_info->fh_versions[0]; 2861 __be32 *p; 2862 2863 da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node); 2864 dprintk("%s: DS %s: encoding address %s\n", 2865 __func__, ds->ds_remotestr, da->da_remotestr); 2866 /* netaddr4 */ 2867 ff_layout_encode_netaddr(xdr, da); 2868 /* nfs_fh4 */ 2869 p = xdr_reserve_space(xdr, 4 + fh->size); 2870 xdr_encode_opaque(p, fh->data, fh->size); 2871 /* ff_io_latency4 read */ 2872 spin_lock(&dss_info->mirror->lock); 2873 ff_layout_encode_io_latency(xdr, 2874 &dss_info->read_stat.io_stat); 2875 /* ff_io_latency4 write */ 2876 ff_layout_encode_io_latency(xdr, 2877 &dss_info->write_stat.io_stat); 2878 spin_unlock(&dss_info->mirror->lock); 2879 /* nfstime4 */ 2880 ff_layout_encode_nfstime(xdr, 2881 ktime_sub(ktime_get(), 2882 dss_info->start_time)); 2883 /* bool */ 2884 p = xdr_reserve_space(xdr, 4); 2885 *p = cpu_to_be32(false); 2886 } 2887 2888 static void 2889 ff_layout_encode_layoutstats(struct xdr_stream *xdr, const void *args, 2890 const struct nfs4_xdr_opaque_data *opaque) 2891 { 2892 struct nfs42_layoutstat_devinfo *devinfo = container_of(opaque, 2893 struct nfs42_layoutstat_devinfo, ld_private); 2894 __be32 *start; 2895 2896 /* layoutupdate length */ 2897 start = xdr_reserve_space(xdr, 4); 2898 ff_layout_encode_ff_layoutupdate(xdr, devinfo, opaque->data); 2899 2900 *start = cpu_to_be32((xdr->p - start - 1) * 4); 2901 } 2902 2903 static void 2904 ff_layout_free_layoutstats(struct nfs4_xdr_opaque_data *opaque) 2905 { 2906 struct nfs4_ff_layout_ds_stripe *dss_info = opaque->data; 2907 struct nfs4_ff_layout_mirror *mirror = dss_info->mirror; 2908 2909 ff_layout_put_mirror(mirror); 2910 } 2911 2912 static const struct nfs4_xdr_opaque_ops layoutstat_ops = { 2913 .encode = ff_layout_encode_layoutstats, 2914 .free = ff_layout_free_layoutstats, 2915 }; 2916 2917 static int 2918 ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo, 2919 struct nfs42_layoutstat_devinfo *devinfo, 2920 int dev_limit, enum nfs4_ff_op_type type) 2921 { 2922 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo); 2923 struct nfs4_ff_layout_mirror *mirror; 2924 struct nfs4_ff_layout_ds_stripe *dss_info; 2925 struct nfs4_deviceid_node *dev; 2926 int i = 0, dss_id; 2927 2928 list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) { 2929 for (dss_id = 0; dss_id < mirror->dss_count; ++dss_id) { 2930 dss_info = &mirror->dss[dss_id]; 2931 if (i >= dev_limit) 2932 break; 2933 if (IS_ERR_OR_NULL(dss_info->mirror_ds)) 2934 continue; 2935 if (!test_and_clear_bit(NFS4_FF_MIRROR_STAT_AVAIL, 2936 &mirror->flags) && 2937 type != NFS4_FF_OP_LAYOUTRETURN) 2938 continue; 2939 /* mirror refcount put in cleanup_layoutstats */ 2940 if (!refcount_inc_not_zero(&mirror->ref)) 2941 continue; 2942 dev = &dss_info->mirror_ds->id_node; 2943 memcpy(&devinfo->dev_id, 2944 &dev->deviceid, 2945 NFS4_DEVICEID4_SIZE); 2946 devinfo->offset = 0; 2947 devinfo->length = NFS4_MAX_UINT64; 2948 spin_lock(&mirror->lock); 2949 devinfo->read_count = 2950 dss_info->read_stat.io_stat.ops_completed; 2951 devinfo->read_bytes = 2952 dss_info->read_stat.io_stat.bytes_completed; 2953 devinfo->write_count = 2954 dss_info->write_stat.io_stat.ops_completed; 2955 devinfo->write_bytes = 2956 dss_info->write_stat.io_stat.bytes_completed; 2957 spin_unlock(&mirror->lock); 2958 devinfo->layout_type = LAYOUT_FLEX_FILES; 2959 devinfo->ld_private.ops = &layoutstat_ops; 2960 devinfo->ld_private.data = &mirror->dss[dss_id]; 2961 2962 devinfo++; 2963 i++; 2964 } 2965 } 2966 return i; 2967 } 2968 2969 static int ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args) 2970 { 2971 struct pnfs_layout_hdr *lo; 2972 struct nfs4_flexfile_layout *ff_layout; 2973 const int dev_count = PNFS_LAYOUTSTATS_MAXDEV; 2974 2975 /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */ 2976 args->devinfo = kmalloc_objs(*args->devinfo, dev_count, 2977 nfs_io_gfp_mask()); 2978 if (!args->devinfo) 2979 return -ENOMEM; 2980 2981 spin_lock(&args->inode->i_lock); 2982 lo = NFS_I(args->inode)->layout; 2983 if (lo && pnfs_layout_is_valid(lo)) { 2984 ff_layout = FF_LAYOUT_FROM_HDR(lo); 2985 args->num_dev = ff_layout_mirror_prepare_stats( 2986 &ff_layout->generic_hdr, &args->devinfo[0], dev_count, 2987 NFS4_FF_OP_LAYOUTSTATS); 2988 } else 2989 args->num_dev = 0; 2990 spin_unlock(&args->inode->i_lock); 2991 if (!args->num_dev) { 2992 kfree(args->devinfo); 2993 args->devinfo = NULL; 2994 return -ENOENT; 2995 } 2996 2997 return 0; 2998 } 2999 3000 static int 3001 ff_layout_set_layoutdriver(struct nfs_server *server, 3002 const struct nfs_fh *dummy) 3003 { 3004 #if IS_ENABLED(CONFIG_NFS_V4_2) 3005 server->caps |= NFS_CAP_LAYOUTSTATS | NFS_CAP_REBOOT_LAYOUTRETURN; 3006 #endif 3007 return 0; 3008 } 3009 3010 static const struct pnfs_commit_ops ff_layout_commit_ops = { 3011 .setup_ds_info = ff_layout_setup_ds_info, 3012 .release_ds_info = ff_layout_release_ds_info, 3013 .mark_request_commit = pnfs_layout_mark_request_commit, 3014 .clear_request_commit = pnfs_generic_clear_request_commit, 3015 .scan_commit_lists = pnfs_generic_scan_commit_lists, 3016 .recover_commit_reqs = pnfs_generic_recover_commit_reqs, 3017 .commit_pagelist = ff_layout_commit_pagelist, 3018 }; 3019 3020 static struct pnfs_layoutdriver_type flexfilelayout_type = { 3021 .id = LAYOUT_FLEX_FILES, 3022 .name = "LAYOUT_FLEX_FILES", 3023 .owner = THIS_MODULE, 3024 .flags = PNFS_LAYOUTGET_ON_OPEN, 3025 .max_layoutget_response = 4096, /* 1 page or so... */ 3026 .set_layoutdriver = ff_layout_set_layoutdriver, 3027 .alloc_layout_hdr = ff_layout_alloc_layout_hdr, 3028 .free_layout_hdr = ff_layout_free_layout_hdr, 3029 .alloc_lseg = ff_layout_alloc_lseg, 3030 .free_lseg = ff_layout_free_lseg, 3031 .add_lseg = ff_layout_add_lseg, 3032 .pg_read_ops = &ff_layout_pg_read_ops, 3033 .pg_write_ops = &ff_layout_pg_write_ops, 3034 .get_ds_info = ff_layout_get_ds_info, 3035 .free_deviceid_node = ff_layout_free_deviceid_node, 3036 .read_pagelist = ff_layout_read_pagelist, 3037 .write_pagelist = ff_layout_write_pagelist, 3038 .alloc_deviceid_node = ff_layout_alloc_deviceid_node, 3039 .prepare_layoutreturn = ff_layout_prepare_layoutreturn, 3040 .sync = pnfs_nfs_generic_sync, 3041 .prepare_layoutstats = ff_layout_prepare_layoutstats, 3042 .cancel_io = ff_layout_cancel_io, 3043 }; 3044 3045 static int __init nfs4flexfilelayout_init(void) 3046 { 3047 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n", 3048 __func__); 3049 return pnfs_register_layoutdriver(&flexfilelayout_type); 3050 } 3051 3052 static void __exit nfs4flexfilelayout_exit(void) 3053 { 3054 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n", 3055 __func__); 3056 pnfs_unregister_layoutdriver(&flexfilelayout_type); 3057 } 3058 3059 MODULE_ALIAS("nfs-layouttype4-4"); 3060 3061 MODULE_LICENSE("GPL"); 3062 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver"); 3063 3064 module_init(nfs4flexfilelayout_init); 3065 module_exit(nfs4flexfilelayout_exit); 3066 3067 module_param(io_maxretrans, ushort, 0644); 3068 MODULE_PARM_DESC(io_maxretrans, "The number of times the NFSv4.1 client " 3069 "retries an I/O request before returning an error. "); 3070