1 /* 2 * pNFS functions to call and manage layout drivers. 3 * 4 * Copyright (c) 2002 [year of first publication] 5 * The Regents of the University of Michigan 6 * All Rights Reserved 7 * 8 * Dean Hildebrand <dhildebz@umich.edu> 9 * 10 * Permission is granted to use, copy, create derivative works, and 11 * redistribute this software and such derivative works for any purpose, 12 * so long as the name of the University of Michigan is not used in 13 * any advertising or publicity pertaining to the use or distribution 14 * of this software without specific, written prior authorization. If 15 * the above copyright notice or any other identification of the 16 * University of Michigan is included in any copy of any portion of 17 * this software, then the disclaimer below must also be included. 18 * 19 * This software is provided as is, without representation or warranty 20 * of any kind either express or implied, including without limitation 21 * the implied warranties of merchantability, fitness for a particular 22 * purpose, or noninfringement. The Regents of the University of 23 * Michigan shall not be liable for any damages, including special, 24 * indirect, incidental, or consequential damages, with respect to any 25 * claim arising out of or in connection with the use of the software, 26 * even if it has been or is hereafter advised of the possibility of 27 * such damages. 28 */ 29 30 #include <linux/nfs_fs.h> 31 #include <linux/nfs_page.h> 32 #include <linux/module.h> 33 #include "internal.h" 34 #include "pnfs.h" 35 #include "iostat.h" 36 #include "nfs4trace.h" 37 #include "delegation.h" 38 #include "nfs42.h" 39 40 #define NFSDBG_FACILITY NFSDBG_PNFS 41 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ) 42 43 /* Locking: 44 * 45 * pnfs_spinlock: 46 * protects pnfs_modules_tbl. 47 */ 48 static DEFINE_SPINLOCK(pnfs_spinlock); 49 50 /* 51 * pnfs_modules_tbl holds all pnfs modules 52 */ 53 static LIST_HEAD(pnfs_modules_tbl); 54 55 static int 56 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid, 57 enum pnfs_iomode iomode, bool sync); 58 59 /* Return the registered pnfs layout driver module matching given id */ 60 static struct pnfs_layoutdriver_type * 61 find_pnfs_driver_locked(u32 id) 62 { 63 struct pnfs_layoutdriver_type *local; 64 65 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid) 66 if (local->id == id) 67 goto out; 68 local = NULL; 69 out: 70 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local); 71 return local; 72 } 73 74 static struct pnfs_layoutdriver_type * 75 find_pnfs_driver(u32 id) 76 { 77 struct pnfs_layoutdriver_type *local; 78 79 spin_lock(&pnfs_spinlock); 80 local = find_pnfs_driver_locked(id); 81 if (local != NULL && !try_module_get(local->owner)) { 82 dprintk("%s: Could not grab reference on module\n", __func__); 83 local = NULL; 84 } 85 spin_unlock(&pnfs_spinlock); 86 return local; 87 } 88 89 void 90 unset_pnfs_layoutdriver(struct nfs_server *nfss) 91 { 92 if (nfss->pnfs_curr_ld) { 93 if (nfss->pnfs_curr_ld->clear_layoutdriver) 94 nfss->pnfs_curr_ld->clear_layoutdriver(nfss); 95 /* Decrement the MDS count. Purge the deviceid cache if zero */ 96 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count)) 97 nfs4_deviceid_purge_client(nfss->nfs_client); 98 module_put(nfss->pnfs_curr_ld->owner); 99 } 100 nfss->pnfs_curr_ld = NULL; 101 } 102 103 /* 104 * Try to set the server's pnfs module to the pnfs layout type specified by id. 105 * Currently only one pNFS layout driver per filesystem is supported. 106 * 107 * @id layout type. Zero (illegal layout type) indicates pNFS not in use. 108 */ 109 void 110 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh, 111 u32 id) 112 { 113 struct pnfs_layoutdriver_type *ld_type = NULL; 114 115 if (id == 0) 116 goto out_no_driver; 117 if (!(server->nfs_client->cl_exchange_flags & 118 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) { 119 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n", 120 __func__, id, server->nfs_client->cl_exchange_flags); 121 goto out_no_driver; 122 } 123 ld_type = find_pnfs_driver(id); 124 if (!ld_type) { 125 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id); 126 ld_type = find_pnfs_driver(id); 127 if (!ld_type) { 128 dprintk("%s: No pNFS module found for %u.\n", 129 __func__, id); 130 goto out_no_driver; 131 } 132 } 133 server->pnfs_curr_ld = ld_type; 134 if (ld_type->set_layoutdriver 135 && ld_type->set_layoutdriver(server, mntfh)) { 136 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout " 137 "driver %u.\n", __func__, id); 138 module_put(ld_type->owner); 139 goto out_no_driver; 140 } 141 /* Bump the MDS count */ 142 atomic_inc(&server->nfs_client->cl_mds_count); 143 144 dprintk("%s: pNFS module for %u set\n", __func__, id); 145 return; 146 147 out_no_driver: 148 dprintk("%s: Using NFSv4 I/O\n", __func__); 149 server->pnfs_curr_ld = NULL; 150 } 151 152 int 153 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type) 154 { 155 int status = -EINVAL; 156 struct pnfs_layoutdriver_type *tmp; 157 158 if (ld_type->id == 0) { 159 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__); 160 return status; 161 } 162 if (!ld_type->alloc_lseg || !ld_type->free_lseg) { 163 printk(KERN_ERR "NFS: %s Layout driver must provide " 164 "alloc_lseg and free_lseg.\n", __func__); 165 return status; 166 } 167 168 spin_lock(&pnfs_spinlock); 169 tmp = find_pnfs_driver_locked(ld_type->id); 170 if (!tmp) { 171 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl); 172 status = 0; 173 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id, 174 ld_type->name); 175 } else { 176 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n", 177 __func__, ld_type->id); 178 } 179 spin_unlock(&pnfs_spinlock); 180 181 return status; 182 } 183 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver); 184 185 void 186 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type) 187 { 188 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id); 189 spin_lock(&pnfs_spinlock); 190 list_del(&ld_type->pnfs_tblid); 191 spin_unlock(&pnfs_spinlock); 192 } 193 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver); 194 195 /* 196 * pNFS client layout cache 197 */ 198 199 /* Need to hold i_lock if caller does not already hold reference */ 200 void 201 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo) 202 { 203 atomic_inc(&lo->plh_refcount); 204 } 205 206 static struct pnfs_layout_hdr * 207 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags) 208 { 209 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld; 210 return ld->alloc_layout_hdr(ino, gfp_flags); 211 } 212 213 static void 214 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo) 215 { 216 struct nfs_server *server = NFS_SERVER(lo->plh_inode); 217 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld; 218 219 if (!list_empty(&lo->plh_layouts)) { 220 struct nfs_client *clp = server->nfs_client; 221 222 spin_lock(&clp->cl_lock); 223 list_del_init(&lo->plh_layouts); 224 spin_unlock(&clp->cl_lock); 225 } 226 put_rpccred(lo->plh_lc_cred); 227 return ld->free_layout_hdr(lo); 228 } 229 230 static void 231 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo) 232 { 233 struct nfs_inode *nfsi = NFS_I(lo->plh_inode); 234 dprintk("%s: freeing layout cache %p\n", __func__, lo); 235 nfsi->layout = NULL; 236 /* Reset MDS Threshold I/O counters */ 237 nfsi->write_io = 0; 238 nfsi->read_io = 0; 239 } 240 241 void 242 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo) 243 { 244 struct inode *inode = lo->plh_inode; 245 246 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) { 247 if (!list_empty(&lo->plh_segs)) 248 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n"); 249 pnfs_detach_layout_hdr(lo); 250 spin_unlock(&inode->i_lock); 251 pnfs_free_layout_hdr(lo); 252 } 253 } 254 255 static int 256 pnfs_iomode_to_fail_bit(u32 iomode) 257 { 258 return iomode == IOMODE_RW ? 259 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED; 260 } 261 262 static void 263 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit) 264 { 265 lo->plh_retry_timestamp = jiffies; 266 if (!test_and_set_bit(fail_bit, &lo->plh_flags)) 267 atomic_inc(&lo->plh_refcount); 268 } 269 270 static void 271 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit) 272 { 273 if (test_and_clear_bit(fail_bit, &lo->plh_flags)) 274 atomic_dec(&lo->plh_refcount); 275 } 276 277 static void 278 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode) 279 { 280 struct inode *inode = lo->plh_inode; 281 struct pnfs_layout_range range = { 282 .iomode = iomode, 283 .offset = 0, 284 .length = NFS4_MAX_UINT64, 285 }; 286 LIST_HEAD(head); 287 288 spin_lock(&inode->i_lock); 289 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode)); 290 pnfs_mark_matching_lsegs_invalid(lo, &head, &range); 291 spin_unlock(&inode->i_lock); 292 pnfs_free_lseg_list(&head); 293 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__, 294 iomode == IOMODE_RW ? "RW" : "READ"); 295 } 296 297 static bool 298 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode) 299 { 300 unsigned long start, end; 301 int fail_bit = pnfs_iomode_to_fail_bit(iomode); 302 303 if (test_bit(fail_bit, &lo->plh_flags) == 0) 304 return false; 305 end = jiffies; 306 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT; 307 if (!time_in_range(lo->plh_retry_timestamp, start, end)) { 308 /* It is time to retry the failed layoutgets */ 309 pnfs_layout_clear_fail_bit(lo, fail_bit); 310 return false; 311 } 312 return true; 313 } 314 315 static void 316 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg) 317 { 318 INIT_LIST_HEAD(&lseg->pls_list); 319 INIT_LIST_HEAD(&lseg->pls_lc_list); 320 atomic_set(&lseg->pls_refcount, 1); 321 smp_mb(); 322 set_bit(NFS_LSEG_VALID, &lseg->pls_flags); 323 lseg->pls_layout = lo; 324 } 325 326 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg) 327 { 328 struct inode *ino = lseg->pls_layout->plh_inode; 329 330 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg); 331 } 332 333 static void 334 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo, 335 struct pnfs_layout_segment *lseg) 336 { 337 struct inode *inode = lo->plh_inode; 338 339 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags)); 340 list_del_init(&lseg->pls_list); 341 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */ 342 atomic_dec(&lo->plh_refcount); 343 if (list_empty(&lo->plh_segs)) 344 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); 345 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq); 346 } 347 348 /* Return true if layoutreturn is needed */ 349 static bool 350 pnfs_layout_need_return(struct pnfs_layout_hdr *lo, 351 struct pnfs_layout_segment *lseg) 352 { 353 struct pnfs_layout_segment *s; 354 355 if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags)) 356 return false; 357 358 list_for_each_entry(s, &lo->plh_segs, pls_list) 359 if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags)) 360 return false; 361 362 return true; 363 } 364 365 static bool 366 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo) 367 { 368 if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) 369 return false; 370 lo->plh_return_iomode = 0; 371 pnfs_get_layout_hdr(lo); 372 clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags); 373 return true; 374 } 375 376 static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg, 377 struct pnfs_layout_hdr *lo, struct inode *inode) 378 { 379 lo = lseg->pls_layout; 380 inode = lo->plh_inode; 381 382 spin_lock(&inode->i_lock); 383 if (pnfs_layout_need_return(lo, lseg)) { 384 nfs4_stateid stateid; 385 enum pnfs_iomode iomode; 386 bool send; 387 388 nfs4_stateid_copy(&stateid, &lo->plh_stateid); 389 iomode = lo->plh_return_iomode; 390 send = pnfs_prepare_layoutreturn(lo); 391 spin_unlock(&inode->i_lock); 392 if (send) { 393 /* Send an async layoutreturn so we dont deadlock */ 394 pnfs_send_layoutreturn(lo, &stateid, iomode, false); 395 } 396 } else 397 spin_unlock(&inode->i_lock); 398 } 399 400 void 401 pnfs_put_lseg(struct pnfs_layout_segment *lseg) 402 { 403 struct pnfs_layout_hdr *lo; 404 struct inode *inode; 405 406 if (!lseg) 407 return; 408 409 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg, 410 atomic_read(&lseg->pls_refcount), 411 test_bit(NFS_LSEG_VALID, &lseg->pls_flags)); 412 413 /* Handle the case where refcount != 1 */ 414 if (atomic_add_unless(&lseg->pls_refcount, -1, 1)) 415 return; 416 417 lo = lseg->pls_layout; 418 inode = lo->plh_inode; 419 /* Do we need a layoutreturn? */ 420 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags)) 421 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode); 422 423 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) { 424 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) { 425 spin_unlock(&inode->i_lock); 426 return; 427 } 428 pnfs_get_layout_hdr(lo); 429 pnfs_layout_remove_lseg(lo, lseg); 430 spin_unlock(&inode->i_lock); 431 pnfs_free_lseg(lseg); 432 pnfs_put_layout_hdr(lo); 433 } 434 } 435 EXPORT_SYMBOL_GPL(pnfs_put_lseg); 436 437 static void pnfs_free_lseg_async_work(struct work_struct *work) 438 { 439 struct pnfs_layout_segment *lseg; 440 struct pnfs_layout_hdr *lo; 441 442 lseg = container_of(work, struct pnfs_layout_segment, pls_work); 443 lo = lseg->pls_layout; 444 445 pnfs_free_lseg(lseg); 446 pnfs_put_layout_hdr(lo); 447 } 448 449 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg) 450 { 451 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work); 452 schedule_work(&lseg->pls_work); 453 } 454 455 void 456 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg) 457 { 458 if (!lseg) 459 return; 460 461 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock); 462 463 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg, 464 atomic_read(&lseg->pls_refcount), 465 test_bit(NFS_LSEG_VALID, &lseg->pls_flags)); 466 if (atomic_dec_and_test(&lseg->pls_refcount)) { 467 struct pnfs_layout_hdr *lo = lseg->pls_layout; 468 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) 469 return; 470 pnfs_get_layout_hdr(lo); 471 pnfs_layout_remove_lseg(lo, lseg); 472 pnfs_free_lseg_async(lseg); 473 } 474 } 475 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked); 476 477 static u64 478 end_offset(u64 start, u64 len) 479 { 480 u64 end; 481 482 end = start + len; 483 return end >= start ? end : NFS4_MAX_UINT64; 484 } 485 486 /* 487 * is l2 fully contained in l1? 488 * start1 end1 489 * [----------------------------------) 490 * start2 end2 491 * [----------------) 492 */ 493 static bool 494 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1, 495 const struct pnfs_layout_range *l2) 496 { 497 u64 start1 = l1->offset; 498 u64 end1 = end_offset(start1, l1->length); 499 u64 start2 = l2->offset; 500 u64 end2 = end_offset(start2, l2->length); 501 502 return (start1 <= start2) && (end1 >= end2); 503 } 504 505 /* 506 * is l1 and l2 intersecting? 507 * start1 end1 508 * [----------------------------------) 509 * start2 end2 510 * [----------------) 511 */ 512 static bool 513 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1, 514 const struct pnfs_layout_range *l2) 515 { 516 u64 start1 = l1->offset; 517 u64 end1 = end_offset(start1, l1->length); 518 u64 start2 = l2->offset; 519 u64 end2 = end_offset(start2, l2->length); 520 521 return (end1 == NFS4_MAX_UINT64 || end1 > start2) && 522 (end2 == NFS4_MAX_UINT64 || end2 > start1); 523 } 524 525 static bool 526 should_free_lseg(const struct pnfs_layout_range *lseg_range, 527 const struct pnfs_layout_range *recall_range) 528 { 529 return (recall_range->iomode == IOMODE_ANY || 530 lseg_range->iomode == recall_range->iomode) && 531 pnfs_lseg_range_intersecting(lseg_range, recall_range); 532 } 533 534 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg, 535 struct list_head *tmp_list) 536 { 537 if (!atomic_dec_and_test(&lseg->pls_refcount)) 538 return false; 539 pnfs_layout_remove_lseg(lseg->pls_layout, lseg); 540 list_add(&lseg->pls_list, tmp_list); 541 return true; 542 } 543 544 /* Returns 1 if lseg is removed from list, 0 otherwise */ 545 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg, 546 struct list_head *tmp_list) 547 { 548 int rv = 0; 549 550 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) { 551 /* Remove the reference keeping the lseg in the 552 * list. It will now be removed when all 553 * outstanding io is finished. 554 */ 555 dprintk("%s: lseg %p ref %d\n", __func__, lseg, 556 atomic_read(&lseg->pls_refcount)); 557 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list)) 558 rv = 1; 559 } 560 return rv; 561 } 562 563 /* Returns count of number of matching invalid lsegs remaining in list 564 * after call. 565 */ 566 int 567 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo, 568 struct list_head *tmp_list, 569 const struct pnfs_layout_range *recall_range) 570 { 571 struct pnfs_layout_segment *lseg, *next; 572 int remaining = 0; 573 574 dprintk("%s:Begin lo %p\n", __func__, lo); 575 576 if (list_empty(&lo->plh_segs)) 577 return 0; 578 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) 579 if (!recall_range || 580 should_free_lseg(&lseg->pls_range, recall_range)) { 581 dprintk("%s: freeing lseg %p iomode %d " 582 "offset %llu length %llu\n", __func__, 583 lseg, lseg->pls_range.iomode, lseg->pls_range.offset, 584 lseg->pls_range.length); 585 if (!mark_lseg_invalid(lseg, tmp_list)) 586 remaining++; 587 } 588 dprintk("%s:Return %i\n", __func__, remaining); 589 return remaining; 590 } 591 592 /* note free_me must contain lsegs from a single layout_hdr */ 593 void 594 pnfs_free_lseg_list(struct list_head *free_me) 595 { 596 struct pnfs_layout_segment *lseg, *tmp; 597 598 if (list_empty(free_me)) 599 return; 600 601 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) { 602 list_del(&lseg->pls_list); 603 pnfs_free_lseg(lseg); 604 } 605 } 606 607 void 608 pnfs_destroy_layout(struct nfs_inode *nfsi) 609 { 610 struct pnfs_layout_hdr *lo; 611 LIST_HEAD(tmp_list); 612 613 spin_lock(&nfsi->vfs_inode.i_lock); 614 lo = nfsi->layout; 615 if (lo) { 616 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */ 617 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL); 618 pnfs_get_layout_hdr(lo); 619 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED); 620 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED); 621 spin_unlock(&nfsi->vfs_inode.i_lock); 622 pnfs_free_lseg_list(&tmp_list); 623 pnfs_put_layout_hdr(lo); 624 } else 625 spin_unlock(&nfsi->vfs_inode.i_lock); 626 } 627 EXPORT_SYMBOL_GPL(pnfs_destroy_layout); 628 629 static bool 630 pnfs_layout_add_bulk_destroy_list(struct inode *inode, 631 struct list_head *layout_list) 632 { 633 struct pnfs_layout_hdr *lo; 634 bool ret = false; 635 636 spin_lock(&inode->i_lock); 637 lo = NFS_I(inode)->layout; 638 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) { 639 pnfs_get_layout_hdr(lo); 640 list_add(&lo->plh_bulk_destroy, layout_list); 641 ret = true; 642 } 643 spin_unlock(&inode->i_lock); 644 return ret; 645 } 646 647 /* Caller must hold rcu_read_lock and clp->cl_lock */ 648 static int 649 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp, 650 struct nfs_server *server, 651 struct list_head *layout_list) 652 { 653 struct pnfs_layout_hdr *lo, *next; 654 struct inode *inode; 655 656 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) { 657 inode = igrab(lo->plh_inode); 658 if (inode == NULL) 659 continue; 660 list_del_init(&lo->plh_layouts); 661 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list)) 662 continue; 663 rcu_read_unlock(); 664 spin_unlock(&clp->cl_lock); 665 iput(inode); 666 spin_lock(&clp->cl_lock); 667 rcu_read_lock(); 668 return -EAGAIN; 669 } 670 return 0; 671 } 672 673 static int 674 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list, 675 bool is_bulk_recall) 676 { 677 struct pnfs_layout_hdr *lo; 678 struct inode *inode; 679 struct pnfs_layout_range range = { 680 .iomode = IOMODE_ANY, 681 .offset = 0, 682 .length = NFS4_MAX_UINT64, 683 }; 684 LIST_HEAD(lseg_list); 685 int ret = 0; 686 687 while (!list_empty(layout_list)) { 688 lo = list_entry(layout_list->next, struct pnfs_layout_hdr, 689 plh_bulk_destroy); 690 dprintk("%s freeing layout for inode %lu\n", __func__, 691 lo->plh_inode->i_ino); 692 inode = lo->plh_inode; 693 694 pnfs_layoutcommit_inode(inode, false); 695 696 spin_lock(&inode->i_lock); 697 list_del_init(&lo->plh_bulk_destroy); 698 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */ 699 if (is_bulk_recall) 700 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); 701 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range)) 702 ret = -EAGAIN; 703 spin_unlock(&inode->i_lock); 704 pnfs_free_lseg_list(&lseg_list); 705 /* Free all lsegs that are attached to commit buckets */ 706 nfs_commit_inode(inode, 0); 707 pnfs_put_layout_hdr(lo); 708 iput(inode); 709 } 710 return ret; 711 } 712 713 int 714 pnfs_destroy_layouts_byfsid(struct nfs_client *clp, 715 struct nfs_fsid *fsid, 716 bool is_recall) 717 { 718 struct nfs_server *server; 719 LIST_HEAD(layout_list); 720 721 spin_lock(&clp->cl_lock); 722 rcu_read_lock(); 723 restart: 724 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) { 725 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0) 726 continue; 727 if (pnfs_layout_bulk_destroy_byserver_locked(clp, 728 server, 729 &layout_list) != 0) 730 goto restart; 731 } 732 rcu_read_unlock(); 733 spin_unlock(&clp->cl_lock); 734 735 if (list_empty(&layout_list)) 736 return 0; 737 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall); 738 } 739 740 int 741 pnfs_destroy_layouts_byclid(struct nfs_client *clp, 742 bool is_recall) 743 { 744 struct nfs_server *server; 745 LIST_HEAD(layout_list); 746 747 spin_lock(&clp->cl_lock); 748 rcu_read_lock(); 749 restart: 750 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) { 751 if (pnfs_layout_bulk_destroy_byserver_locked(clp, 752 server, 753 &layout_list) != 0) 754 goto restart; 755 } 756 rcu_read_unlock(); 757 spin_unlock(&clp->cl_lock); 758 759 if (list_empty(&layout_list)) 760 return 0; 761 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall); 762 } 763 764 /* 765 * Called by the state manger to remove all layouts established under an 766 * expired lease. 767 */ 768 void 769 pnfs_destroy_all_layouts(struct nfs_client *clp) 770 { 771 nfs4_deviceid_mark_client_invalid(clp); 772 nfs4_deviceid_purge_client(clp); 773 774 pnfs_destroy_layouts_byclid(clp, false); 775 } 776 777 /* 778 * Compare 2 layout stateid sequence ids, to see which is newer, 779 * taking into account wraparound issues. 780 */ 781 static bool pnfs_seqid_is_newer(u32 s1, u32 s2) 782 { 783 return (s32)(s1 - s2) > 0; 784 } 785 786 /* update lo->plh_stateid with new if is more recent */ 787 void 788 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new, 789 bool update_barrier) 790 { 791 u32 oldseq, newseq, new_barrier; 792 int empty = list_empty(&lo->plh_segs); 793 794 oldseq = be32_to_cpu(lo->plh_stateid.seqid); 795 newseq = be32_to_cpu(new->seqid); 796 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) { 797 nfs4_stateid_copy(&lo->plh_stateid, new); 798 if (update_barrier) { 799 new_barrier = be32_to_cpu(new->seqid); 800 } else { 801 /* Because of wraparound, we want to keep the barrier 802 * "close" to the current seqids. 803 */ 804 new_barrier = newseq - atomic_read(&lo->plh_outstanding); 805 } 806 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier)) 807 lo->plh_barrier = new_barrier; 808 } 809 } 810 811 static bool 812 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo, 813 const nfs4_stateid *stateid) 814 { 815 u32 seqid = be32_to_cpu(stateid->seqid); 816 817 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier); 818 } 819 820 /* lget is set to 1 if called from inside send_layoutget call chain */ 821 static bool 822 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo) 823 { 824 return lo->plh_block_lgets || 825 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); 826 } 827 828 int 829 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo, 830 const struct pnfs_layout_range *range, 831 struct nfs4_state *open_state) 832 { 833 int status = 0; 834 835 dprintk("--> %s\n", __func__); 836 spin_lock(&lo->plh_inode->i_lock); 837 if (pnfs_layoutgets_blocked(lo)) { 838 status = -EAGAIN; 839 } else if (!nfs4_valid_open_stateid(open_state)) { 840 status = -EBADF; 841 } else if (list_empty(&lo->plh_segs) || 842 test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) { 843 int seq; 844 845 do { 846 seq = read_seqbegin(&open_state->seqlock); 847 nfs4_stateid_copy(dst, &open_state->stateid); 848 } while (read_seqretry(&open_state->seqlock, seq)); 849 } else 850 nfs4_stateid_copy(dst, &lo->plh_stateid); 851 spin_unlock(&lo->plh_inode->i_lock); 852 dprintk("<-- %s\n", __func__); 853 return status; 854 } 855 856 /* 857 * Get layout from server. 858 * for now, assume that whole file layouts are requested. 859 * arg->offset: 0 860 * arg->length: all ones 861 */ 862 static struct pnfs_layout_segment * 863 send_layoutget(struct pnfs_layout_hdr *lo, 864 struct nfs_open_context *ctx, 865 const struct pnfs_layout_range *range, 866 gfp_t gfp_flags) 867 { 868 struct inode *ino = lo->plh_inode; 869 struct nfs_server *server = NFS_SERVER(ino); 870 struct nfs4_layoutget *lgp; 871 struct pnfs_layout_segment *lseg; 872 loff_t i_size; 873 874 dprintk("--> %s\n", __func__); 875 876 /* 877 * Synchronously retrieve layout information from server and 878 * store in lseg. If we race with a concurrent seqid morphing 879 * op, then re-send the LAYOUTGET. 880 */ 881 do { 882 lgp = kzalloc(sizeof(*lgp), gfp_flags); 883 if (lgp == NULL) 884 return NULL; 885 886 i_size = i_size_read(ino); 887 888 lgp->args.minlength = PAGE_CACHE_SIZE; 889 if (lgp->args.minlength > range->length) 890 lgp->args.minlength = range->length; 891 if (range->iomode == IOMODE_READ) { 892 if (range->offset >= i_size) 893 lgp->args.minlength = 0; 894 else if (i_size - range->offset < lgp->args.minlength) 895 lgp->args.minlength = i_size - range->offset; 896 } 897 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE; 898 pnfs_copy_range(&lgp->args.range, range); 899 lgp->args.type = server->pnfs_curr_ld->id; 900 lgp->args.inode = ino; 901 lgp->args.ctx = get_nfs_open_context(ctx); 902 lgp->gfp_flags = gfp_flags; 903 lgp->cred = lo->plh_lc_cred; 904 905 lseg = nfs4_proc_layoutget(lgp, gfp_flags); 906 } while (lseg == ERR_PTR(-EAGAIN)); 907 908 if (IS_ERR(lseg) && !nfs_error_is_fatal(PTR_ERR(lseg))) 909 lseg = NULL; 910 else 911 pnfs_layout_clear_fail_bit(lo, 912 pnfs_iomode_to_fail_bit(range->iomode)); 913 914 return lseg; 915 } 916 917 static void pnfs_clear_layoutcommit(struct inode *inode, 918 struct list_head *head) 919 { 920 struct nfs_inode *nfsi = NFS_I(inode); 921 struct pnfs_layout_segment *lseg, *tmp; 922 923 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) 924 return; 925 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) { 926 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) 927 continue; 928 pnfs_lseg_dec_and_remove_zero(lseg, head); 929 } 930 } 931 932 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo) 933 { 934 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags); 935 smp_mb__after_atomic(); 936 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN); 937 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq); 938 } 939 940 static int 941 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid, 942 enum pnfs_iomode iomode, bool sync) 943 { 944 struct inode *ino = lo->plh_inode; 945 struct nfs4_layoutreturn *lrp; 946 int status = 0; 947 948 lrp = kzalloc(sizeof(*lrp), GFP_NOFS); 949 if (unlikely(lrp == NULL)) { 950 status = -ENOMEM; 951 spin_lock(&ino->i_lock); 952 pnfs_clear_layoutreturn_waitbit(lo); 953 spin_unlock(&ino->i_lock); 954 pnfs_put_layout_hdr(lo); 955 goto out; 956 } 957 958 nfs4_stateid_copy(&lrp->args.stateid, stateid); 959 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id; 960 lrp->args.inode = ino; 961 lrp->args.range.iomode = iomode; 962 lrp->args.range.offset = 0; 963 lrp->args.range.length = NFS4_MAX_UINT64; 964 lrp->args.layout = lo; 965 lrp->clp = NFS_SERVER(ino)->nfs_client; 966 lrp->cred = lo->plh_lc_cred; 967 968 status = nfs4_proc_layoutreturn(lrp, sync); 969 out: 970 dprintk("<-- %s status: %d\n", __func__, status); 971 return status; 972 } 973 974 /* 975 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr 976 * when the layout segment list is empty. 977 * 978 * Note that a pnfs_layout_hdr can exist with an empty layout segment 979 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the 980 * deviceid is marked invalid. 981 */ 982 int 983 _pnfs_return_layout(struct inode *ino) 984 { 985 struct pnfs_layout_hdr *lo = NULL; 986 struct nfs_inode *nfsi = NFS_I(ino); 987 LIST_HEAD(tmp_list); 988 nfs4_stateid stateid; 989 int status = 0, empty; 990 bool send; 991 992 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino); 993 994 spin_lock(&ino->i_lock); 995 lo = nfsi->layout; 996 if (!lo) { 997 spin_unlock(&ino->i_lock); 998 dprintk("NFS: %s no layout to return\n", __func__); 999 goto out; 1000 } 1001 nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid); 1002 /* Reference matched in nfs4_layoutreturn_release */ 1003 pnfs_get_layout_hdr(lo); 1004 empty = list_empty(&lo->plh_segs); 1005 pnfs_clear_layoutcommit(ino, &tmp_list); 1006 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL); 1007 1008 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) { 1009 struct pnfs_layout_range range = { 1010 .iomode = IOMODE_ANY, 1011 .offset = 0, 1012 .length = NFS4_MAX_UINT64, 1013 }; 1014 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range); 1015 } 1016 1017 /* Don't send a LAYOUTRETURN if list was initially empty */ 1018 if (empty) { 1019 spin_unlock(&ino->i_lock); 1020 dprintk("NFS: %s no layout segments to return\n", __func__); 1021 goto out_put_layout_hdr; 1022 } 1023 1024 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags); 1025 send = pnfs_prepare_layoutreturn(lo); 1026 spin_unlock(&ino->i_lock); 1027 pnfs_free_lseg_list(&tmp_list); 1028 if (send) 1029 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true); 1030 out_put_layout_hdr: 1031 pnfs_put_layout_hdr(lo); 1032 out: 1033 dprintk("<-- %s status: %d\n", __func__, status); 1034 return status; 1035 } 1036 EXPORT_SYMBOL_GPL(_pnfs_return_layout); 1037 1038 int 1039 pnfs_commit_and_return_layout(struct inode *inode) 1040 { 1041 struct pnfs_layout_hdr *lo; 1042 int ret; 1043 1044 spin_lock(&inode->i_lock); 1045 lo = NFS_I(inode)->layout; 1046 if (lo == NULL) { 1047 spin_unlock(&inode->i_lock); 1048 return 0; 1049 } 1050 pnfs_get_layout_hdr(lo); 1051 /* Block new layoutgets and read/write to ds */ 1052 lo->plh_block_lgets++; 1053 spin_unlock(&inode->i_lock); 1054 filemap_fdatawait(inode->i_mapping); 1055 ret = pnfs_layoutcommit_inode(inode, true); 1056 if (ret == 0) 1057 ret = _pnfs_return_layout(inode); 1058 spin_lock(&inode->i_lock); 1059 lo->plh_block_lgets--; 1060 spin_unlock(&inode->i_lock); 1061 pnfs_put_layout_hdr(lo); 1062 return ret; 1063 } 1064 1065 bool pnfs_roc(struct inode *ino) 1066 { 1067 struct nfs_inode *nfsi = NFS_I(ino); 1068 struct nfs_open_context *ctx; 1069 struct nfs4_state *state; 1070 struct pnfs_layout_hdr *lo; 1071 struct pnfs_layout_segment *lseg, *tmp; 1072 nfs4_stateid stateid; 1073 LIST_HEAD(tmp_list); 1074 bool found = false, layoutreturn = false, roc = false; 1075 1076 spin_lock(&ino->i_lock); 1077 lo = nfsi->layout; 1078 if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) 1079 goto out_noroc; 1080 1081 /* no roc if we hold a delegation */ 1082 if (nfs4_check_delegation(ino, FMODE_READ)) 1083 goto out_noroc; 1084 1085 list_for_each_entry(ctx, &nfsi->open_files, list) { 1086 state = ctx->state; 1087 /* Don't return layout if there is open file state */ 1088 if (state != NULL && state->state != 0) 1089 goto out_noroc; 1090 } 1091 1092 nfs4_stateid_copy(&stateid, &lo->plh_stateid); 1093 /* always send layoutreturn if being marked so */ 1094 if (test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, 1095 &lo->plh_flags)) 1096 layoutreturn = pnfs_prepare_layoutreturn(lo); 1097 1098 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list) 1099 /* If we are sending layoutreturn, invalidate all valid lsegs */ 1100 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) { 1101 mark_lseg_invalid(lseg, &tmp_list); 1102 found = true; 1103 } 1104 /* ROC in two conditions: 1105 * 1. there are ROC lsegs 1106 * 2. we don't send layoutreturn 1107 */ 1108 if (found && !layoutreturn) { 1109 /* lo ref dropped in pnfs_roc_release() */ 1110 pnfs_get_layout_hdr(lo); 1111 roc = true; 1112 } 1113 1114 out_noroc: 1115 spin_unlock(&ino->i_lock); 1116 pnfs_free_lseg_list(&tmp_list); 1117 pnfs_layoutcommit_inode(ino, true); 1118 if (layoutreturn) 1119 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true); 1120 return roc; 1121 } 1122 1123 void pnfs_roc_release(struct inode *ino) 1124 { 1125 struct pnfs_layout_hdr *lo; 1126 1127 spin_lock(&ino->i_lock); 1128 lo = NFS_I(ino)->layout; 1129 pnfs_clear_layoutreturn_waitbit(lo); 1130 if (atomic_dec_and_test(&lo->plh_refcount)) { 1131 pnfs_detach_layout_hdr(lo); 1132 spin_unlock(&ino->i_lock); 1133 pnfs_free_layout_hdr(lo); 1134 } else 1135 spin_unlock(&ino->i_lock); 1136 } 1137 1138 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier) 1139 { 1140 struct pnfs_layout_hdr *lo; 1141 1142 spin_lock(&ino->i_lock); 1143 lo = NFS_I(ino)->layout; 1144 pnfs_mark_layout_returned_if_empty(lo); 1145 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier)) 1146 lo->plh_barrier = barrier; 1147 spin_unlock(&ino->i_lock); 1148 trace_nfs4_layoutreturn_on_close(ino, 0); 1149 } 1150 1151 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier) 1152 { 1153 struct nfs_inode *nfsi = NFS_I(ino); 1154 struct pnfs_layout_hdr *lo; 1155 u32 current_seqid; 1156 1157 spin_lock(&ino->i_lock); 1158 lo = nfsi->layout; 1159 current_seqid = be32_to_cpu(lo->plh_stateid.seqid); 1160 1161 /* Since close does not return a layout stateid for use as 1162 * a barrier, we choose the worst-case barrier. 1163 */ 1164 *barrier = current_seqid + atomic_read(&lo->plh_outstanding); 1165 spin_unlock(&ino->i_lock); 1166 } 1167 1168 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task) 1169 { 1170 struct nfs_inode *nfsi = NFS_I(ino); 1171 struct pnfs_layout_hdr *lo; 1172 bool sleep = false; 1173 1174 /* we might not have grabbed lo reference. so need to check under 1175 * i_lock */ 1176 spin_lock(&ino->i_lock); 1177 lo = nfsi->layout; 1178 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) 1179 sleep = true; 1180 spin_unlock(&ino->i_lock); 1181 1182 if (sleep) 1183 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL); 1184 1185 return sleep; 1186 } 1187 1188 /* 1189 * Compare two layout segments for sorting into layout cache. 1190 * We want to preferentially return RW over RO layouts, so ensure those 1191 * are seen first. 1192 */ 1193 static s64 1194 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1, 1195 const struct pnfs_layout_range *l2) 1196 { 1197 s64 d; 1198 1199 /* high offset > low offset */ 1200 d = l1->offset - l2->offset; 1201 if (d) 1202 return d; 1203 1204 /* short length > long length */ 1205 d = l2->length - l1->length; 1206 if (d) 1207 return d; 1208 1209 /* read > read/write */ 1210 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ); 1211 } 1212 1213 static bool 1214 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1, 1215 const struct pnfs_layout_range *l2) 1216 { 1217 return pnfs_lseg_range_cmp(l1, l2) > 0; 1218 } 1219 1220 static bool 1221 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg, 1222 struct pnfs_layout_segment *old) 1223 { 1224 return false; 1225 } 1226 1227 void 1228 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo, 1229 struct pnfs_layout_segment *lseg, 1230 bool (*is_after)(const struct pnfs_layout_range *, 1231 const struct pnfs_layout_range *), 1232 bool (*do_merge)(struct pnfs_layout_segment *, 1233 struct pnfs_layout_segment *), 1234 struct list_head *free_me) 1235 { 1236 struct pnfs_layout_segment *lp, *tmp; 1237 1238 dprintk("%s:Begin\n", __func__); 1239 1240 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) { 1241 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0) 1242 continue; 1243 if (do_merge(lseg, lp)) { 1244 mark_lseg_invalid(lp, free_me); 1245 continue; 1246 } 1247 if (is_after(&lseg->pls_range, &lp->pls_range)) 1248 continue; 1249 list_add_tail(&lseg->pls_list, &lp->pls_list); 1250 dprintk("%s: inserted lseg %p " 1251 "iomode %d offset %llu length %llu before " 1252 "lp %p iomode %d offset %llu length %llu\n", 1253 __func__, lseg, lseg->pls_range.iomode, 1254 lseg->pls_range.offset, lseg->pls_range.length, 1255 lp, lp->pls_range.iomode, lp->pls_range.offset, 1256 lp->pls_range.length); 1257 goto out; 1258 } 1259 list_add_tail(&lseg->pls_list, &lo->plh_segs); 1260 dprintk("%s: inserted lseg %p " 1261 "iomode %d offset %llu length %llu at tail\n", 1262 __func__, lseg, lseg->pls_range.iomode, 1263 lseg->pls_range.offset, lseg->pls_range.length); 1264 out: 1265 pnfs_get_layout_hdr(lo); 1266 1267 dprintk("%s:Return\n", __func__); 1268 } 1269 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg); 1270 1271 static void 1272 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo, 1273 struct pnfs_layout_segment *lseg, 1274 struct list_head *free_me) 1275 { 1276 struct inode *inode = lo->plh_inode; 1277 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld; 1278 1279 if (ld->add_lseg != NULL) 1280 ld->add_lseg(lo, lseg, free_me); 1281 else 1282 pnfs_generic_layout_insert_lseg(lo, lseg, 1283 pnfs_lseg_range_is_after, 1284 pnfs_lseg_no_merge, 1285 free_me); 1286 } 1287 1288 static struct pnfs_layout_hdr * 1289 alloc_init_layout_hdr(struct inode *ino, 1290 struct nfs_open_context *ctx, 1291 gfp_t gfp_flags) 1292 { 1293 struct pnfs_layout_hdr *lo; 1294 1295 lo = pnfs_alloc_layout_hdr(ino, gfp_flags); 1296 if (!lo) 1297 return NULL; 1298 atomic_set(&lo->plh_refcount, 1); 1299 INIT_LIST_HEAD(&lo->plh_layouts); 1300 INIT_LIST_HEAD(&lo->plh_segs); 1301 INIT_LIST_HEAD(&lo->plh_bulk_destroy); 1302 lo->plh_inode = ino; 1303 lo->plh_lc_cred = get_rpccred(ctx->cred); 1304 return lo; 1305 } 1306 1307 static struct pnfs_layout_hdr * 1308 pnfs_find_alloc_layout(struct inode *ino, 1309 struct nfs_open_context *ctx, 1310 gfp_t gfp_flags) 1311 { 1312 struct nfs_inode *nfsi = NFS_I(ino); 1313 struct pnfs_layout_hdr *new = NULL; 1314 1315 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout); 1316 1317 if (nfsi->layout != NULL) 1318 goto out_existing; 1319 spin_unlock(&ino->i_lock); 1320 new = alloc_init_layout_hdr(ino, ctx, gfp_flags); 1321 spin_lock(&ino->i_lock); 1322 1323 if (likely(nfsi->layout == NULL)) { /* Won the race? */ 1324 nfsi->layout = new; 1325 return new; 1326 } else if (new != NULL) 1327 pnfs_free_layout_hdr(new); 1328 out_existing: 1329 pnfs_get_layout_hdr(nfsi->layout); 1330 return nfsi->layout; 1331 } 1332 1333 /* 1334 * iomode matching rules: 1335 * iomode lseg match 1336 * ----- ----- ----- 1337 * ANY READ true 1338 * ANY RW true 1339 * RW READ false 1340 * RW RW true 1341 * READ READ true 1342 * READ RW true 1343 */ 1344 static bool 1345 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range, 1346 const struct pnfs_layout_range *range) 1347 { 1348 struct pnfs_layout_range range1; 1349 1350 if ((range->iomode == IOMODE_RW && 1351 ls_range->iomode != IOMODE_RW) || 1352 !pnfs_lseg_range_intersecting(ls_range, range)) 1353 return 0; 1354 1355 /* range1 covers only the first byte in the range */ 1356 range1 = *range; 1357 range1.length = 1; 1358 return pnfs_lseg_range_contained(ls_range, &range1); 1359 } 1360 1361 /* 1362 * lookup range in layout 1363 */ 1364 static struct pnfs_layout_segment * 1365 pnfs_find_lseg(struct pnfs_layout_hdr *lo, 1366 struct pnfs_layout_range *range) 1367 { 1368 struct pnfs_layout_segment *lseg, *ret = NULL; 1369 1370 dprintk("%s:Begin\n", __func__); 1371 1372 list_for_each_entry(lseg, &lo->plh_segs, pls_list) { 1373 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) && 1374 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) && 1375 pnfs_lseg_range_match(&lseg->pls_range, range)) { 1376 ret = pnfs_get_lseg(lseg); 1377 break; 1378 } 1379 } 1380 1381 dprintk("%s:Return lseg %p ref %d\n", 1382 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0); 1383 return ret; 1384 } 1385 1386 /* 1387 * Use mdsthreshold hints set at each OPEN to determine if I/O should go 1388 * to the MDS or over pNFS 1389 * 1390 * The nfs_inode read_io and write_io fields are cumulative counters reset 1391 * when there are no layout segments. Note that in pnfs_update_layout iomode 1392 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a 1393 * WRITE request. 1394 * 1395 * A return of true means use MDS I/O. 1396 * 1397 * From rfc 5661: 1398 * If a file's size is smaller than the file size threshold, data accesses 1399 * SHOULD be sent to the metadata server. If an I/O request has a length that 1400 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata 1401 * server. If both file size and I/O size are provided, the client SHOULD 1402 * reach or exceed both thresholds before sending its read or write 1403 * requests to the data server. 1404 */ 1405 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx, 1406 struct inode *ino, int iomode) 1407 { 1408 struct nfs4_threshold *t = ctx->mdsthreshold; 1409 struct nfs_inode *nfsi = NFS_I(ino); 1410 loff_t fsize = i_size_read(ino); 1411 bool size = false, size_set = false, io = false, io_set = false, ret = false; 1412 1413 if (t == NULL) 1414 return ret; 1415 1416 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n", 1417 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz); 1418 1419 switch (iomode) { 1420 case IOMODE_READ: 1421 if (t->bm & THRESHOLD_RD) { 1422 dprintk("%s fsize %llu\n", __func__, fsize); 1423 size_set = true; 1424 if (fsize < t->rd_sz) 1425 size = true; 1426 } 1427 if (t->bm & THRESHOLD_RD_IO) { 1428 dprintk("%s nfsi->read_io %llu\n", __func__, 1429 nfsi->read_io); 1430 io_set = true; 1431 if (nfsi->read_io < t->rd_io_sz) 1432 io = true; 1433 } 1434 break; 1435 case IOMODE_RW: 1436 if (t->bm & THRESHOLD_WR) { 1437 dprintk("%s fsize %llu\n", __func__, fsize); 1438 size_set = true; 1439 if (fsize < t->wr_sz) 1440 size = true; 1441 } 1442 if (t->bm & THRESHOLD_WR_IO) { 1443 dprintk("%s nfsi->write_io %llu\n", __func__, 1444 nfsi->write_io); 1445 io_set = true; 1446 if (nfsi->write_io < t->wr_io_sz) 1447 io = true; 1448 } 1449 break; 1450 } 1451 if (size_set && io_set) { 1452 if (size && io) 1453 ret = true; 1454 } else if (size || io) 1455 ret = true; 1456 1457 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret); 1458 return ret; 1459 } 1460 1461 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo) 1462 { 1463 /* 1464 * send layoutcommit as it can hold up layoutreturn due to lseg 1465 * reference 1466 */ 1467 pnfs_layoutcommit_inode(lo->plh_inode, false); 1468 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN, 1469 nfs_wait_bit_killable, 1470 TASK_UNINTERRUPTIBLE); 1471 } 1472 1473 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo) 1474 { 1475 unsigned long *bitlock = &lo->plh_flags; 1476 1477 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock); 1478 smp_mb__after_atomic(); 1479 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET); 1480 } 1481 1482 /* 1483 * Layout segment is retreived from the server if not cached. 1484 * The appropriate layout segment is referenced and returned to the caller. 1485 */ 1486 struct pnfs_layout_segment * 1487 pnfs_update_layout(struct inode *ino, 1488 struct nfs_open_context *ctx, 1489 loff_t pos, 1490 u64 count, 1491 enum pnfs_iomode iomode, 1492 gfp_t gfp_flags) 1493 { 1494 struct pnfs_layout_range arg = { 1495 .iomode = iomode, 1496 .offset = pos, 1497 .length = count, 1498 }; 1499 unsigned pg_offset; 1500 struct nfs_server *server = NFS_SERVER(ino); 1501 struct nfs_client *clp = server->nfs_client; 1502 struct pnfs_layout_hdr *lo; 1503 struct pnfs_layout_segment *lseg = NULL; 1504 bool first; 1505 1506 if (!pnfs_enabled_sb(NFS_SERVER(ino))) { 1507 trace_pnfs_update_layout(ino, pos, count, iomode, NULL, 1508 PNFS_UPDATE_LAYOUT_NO_PNFS); 1509 goto out; 1510 } 1511 1512 if (iomode == IOMODE_READ && i_size_read(ino) == 0) { 1513 trace_pnfs_update_layout(ino, pos, count, iomode, NULL, 1514 PNFS_UPDATE_LAYOUT_RD_ZEROLEN); 1515 goto out; 1516 } 1517 1518 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) { 1519 trace_pnfs_update_layout(ino, pos, count, iomode, NULL, 1520 PNFS_UPDATE_LAYOUT_MDSTHRESH); 1521 goto out; 1522 } 1523 1524 lookup_again: 1525 first = false; 1526 spin_lock(&ino->i_lock); 1527 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags); 1528 if (lo == NULL) { 1529 spin_unlock(&ino->i_lock); 1530 trace_pnfs_update_layout(ino, pos, count, iomode, NULL, 1531 PNFS_UPDATE_LAYOUT_NOMEM); 1532 goto out; 1533 } 1534 1535 /* Do we even need to bother with this? */ 1536 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) { 1537 trace_pnfs_update_layout(ino, pos, count, iomode, lo, 1538 PNFS_UPDATE_LAYOUT_BULK_RECALL); 1539 dprintk("%s matches recall, use MDS\n", __func__); 1540 goto out_unlock; 1541 } 1542 1543 /* if LAYOUTGET already failed once we don't try again */ 1544 if (pnfs_layout_io_test_failed(lo, iomode)) { 1545 trace_pnfs_update_layout(ino, pos, count, iomode, lo, 1546 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL); 1547 goto out_unlock; 1548 } 1549 1550 first = list_empty(&lo->plh_segs); 1551 if (first) { 1552 /* The first layoutget for the file. Need to serialize per 1553 * RFC 5661 Errata 3208. 1554 */ 1555 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, 1556 &lo->plh_flags)) { 1557 spin_unlock(&ino->i_lock); 1558 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET, 1559 TASK_UNINTERRUPTIBLE); 1560 pnfs_put_layout_hdr(lo); 1561 goto lookup_again; 1562 } 1563 } else { 1564 /* Check to see if the layout for the given range 1565 * already exists 1566 */ 1567 lseg = pnfs_find_lseg(lo, &arg); 1568 if (lseg) { 1569 trace_pnfs_update_layout(ino, pos, count, iomode, lo, 1570 PNFS_UPDATE_LAYOUT_FOUND_CACHED); 1571 goto out_unlock; 1572 } 1573 } 1574 1575 /* 1576 * Because we free lsegs before sending LAYOUTRETURN, we need to wait 1577 * for LAYOUTRETURN even if first is true. 1578 */ 1579 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) { 1580 spin_unlock(&ino->i_lock); 1581 dprintk("%s wait for layoutreturn\n", __func__); 1582 if (pnfs_prepare_to_retry_layoutget(lo)) { 1583 if (first) 1584 pnfs_clear_first_layoutget(lo); 1585 pnfs_put_layout_hdr(lo); 1586 dprintk("%s retrying\n", __func__); 1587 goto lookup_again; 1588 } 1589 trace_pnfs_update_layout(ino, pos, count, iomode, lo, 1590 PNFS_UPDATE_LAYOUT_RETURN); 1591 goto out_put_layout_hdr; 1592 } 1593 1594 if (pnfs_layoutgets_blocked(lo)) { 1595 trace_pnfs_update_layout(ino, pos, count, iomode, lo, 1596 PNFS_UPDATE_LAYOUT_BLOCKED); 1597 goto out_unlock; 1598 } 1599 atomic_inc(&lo->plh_outstanding); 1600 spin_unlock(&ino->i_lock); 1601 1602 if (list_empty(&lo->plh_layouts)) { 1603 /* The lo must be on the clp list if there is any 1604 * chance of a CB_LAYOUTRECALL(FILE) coming in. 1605 */ 1606 spin_lock(&clp->cl_lock); 1607 if (list_empty(&lo->plh_layouts)) 1608 list_add_tail(&lo->plh_layouts, &server->layouts); 1609 spin_unlock(&clp->cl_lock); 1610 } 1611 1612 pg_offset = arg.offset & ~PAGE_CACHE_MASK; 1613 if (pg_offset) { 1614 arg.offset -= pg_offset; 1615 arg.length += pg_offset; 1616 } 1617 if (arg.length != NFS4_MAX_UINT64) 1618 arg.length = PAGE_CACHE_ALIGN(arg.length); 1619 1620 lseg = send_layoutget(lo, ctx, &arg, gfp_flags); 1621 atomic_dec(&lo->plh_outstanding); 1622 trace_pnfs_update_layout(ino, pos, count, iomode, lo, 1623 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET); 1624 out_put_layout_hdr: 1625 if (first) 1626 pnfs_clear_first_layoutget(lo); 1627 pnfs_put_layout_hdr(lo); 1628 out: 1629 dprintk("%s: inode %s/%llu pNFS layout segment %s for " 1630 "(%s, offset: %llu, length: %llu)\n", 1631 __func__, ino->i_sb->s_id, 1632 (unsigned long long)NFS_FILEID(ino), 1633 IS_ERR_OR_NULL(lseg) ? "not found" : "found", 1634 iomode==IOMODE_RW ? "read/write" : "read-only", 1635 (unsigned long long)pos, 1636 (unsigned long long)count); 1637 return lseg; 1638 out_unlock: 1639 spin_unlock(&ino->i_lock); 1640 goto out_put_layout_hdr; 1641 } 1642 EXPORT_SYMBOL_GPL(pnfs_update_layout); 1643 1644 static bool 1645 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range) 1646 { 1647 switch (range->iomode) { 1648 case IOMODE_READ: 1649 case IOMODE_RW: 1650 break; 1651 default: 1652 return false; 1653 } 1654 if (range->offset == NFS4_MAX_UINT64) 1655 return false; 1656 if (range->length == 0) 1657 return false; 1658 if (range->length != NFS4_MAX_UINT64 && 1659 range->length > NFS4_MAX_UINT64 - range->offset) 1660 return false; 1661 return true; 1662 } 1663 1664 struct pnfs_layout_segment * 1665 pnfs_layout_process(struct nfs4_layoutget *lgp) 1666 { 1667 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout; 1668 struct nfs4_layoutget_res *res = &lgp->res; 1669 struct pnfs_layout_segment *lseg; 1670 struct inode *ino = lo->plh_inode; 1671 LIST_HEAD(free_me); 1672 int status = -EINVAL; 1673 1674 if (!pnfs_sanity_check_layout_range(&res->range)) 1675 goto out; 1676 1677 /* Inject layout blob into I/O device driver */ 1678 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags); 1679 if (!lseg || IS_ERR(lseg)) { 1680 if (!lseg) 1681 status = -ENOMEM; 1682 else 1683 status = PTR_ERR(lseg); 1684 dprintk("%s: Could not allocate layout: error %d\n", 1685 __func__, status); 1686 goto out; 1687 } 1688 1689 init_lseg(lo, lseg); 1690 lseg->pls_range = res->range; 1691 1692 spin_lock(&ino->i_lock); 1693 if (pnfs_layoutgets_blocked(lo)) { 1694 dprintk("%s forget reply due to state\n", __func__); 1695 goto out_forget_reply; 1696 } 1697 1698 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) { 1699 /* existing state ID, make sure the sequence number matches. */ 1700 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) { 1701 dprintk("%s forget reply due to sequence\n", __func__); 1702 status = -EAGAIN; 1703 goto out_forget_reply; 1704 } 1705 pnfs_set_layout_stateid(lo, &res->stateid, false); 1706 } else { 1707 /* 1708 * We got an entirely new state ID. Mark all segments for the 1709 * inode invalid, and don't bother validating the stateid 1710 * sequence number. 1711 */ 1712 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL); 1713 1714 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid); 1715 lo->plh_barrier = be32_to_cpu(res->stateid.seqid); 1716 } 1717 1718 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags); 1719 1720 pnfs_get_lseg(lseg); 1721 pnfs_layout_insert_lseg(lo, lseg, &free_me); 1722 1723 if (res->return_on_close) 1724 set_bit(NFS_LSEG_ROC, &lseg->pls_flags); 1725 1726 spin_unlock(&ino->i_lock); 1727 pnfs_free_lseg_list(&free_me); 1728 return lseg; 1729 out: 1730 return ERR_PTR(status); 1731 1732 out_forget_reply: 1733 spin_unlock(&ino->i_lock); 1734 lseg->pls_layout = lo; 1735 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg); 1736 goto out; 1737 } 1738 1739 static void 1740 pnfs_set_plh_return_iomode(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode) 1741 { 1742 if (lo->plh_return_iomode == iomode) 1743 return; 1744 if (lo->plh_return_iomode != 0) 1745 iomode = IOMODE_ANY; 1746 lo->plh_return_iomode = iomode; 1747 } 1748 1749 int 1750 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo, 1751 struct list_head *tmp_list, 1752 const struct pnfs_layout_range *return_range) 1753 { 1754 struct pnfs_layout_segment *lseg, *next; 1755 int remaining = 0; 1756 1757 dprintk("%s:Begin lo %p\n", __func__, lo); 1758 1759 if (list_empty(&lo->plh_segs)) 1760 return 0; 1761 1762 assert_spin_locked(&lo->plh_inode->i_lock); 1763 1764 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) 1765 if (should_free_lseg(&lseg->pls_range, return_range)) { 1766 dprintk("%s: marking lseg %p iomode %d " 1767 "offset %llu length %llu\n", __func__, 1768 lseg, lseg->pls_range.iomode, 1769 lseg->pls_range.offset, 1770 lseg->pls_range.length); 1771 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags); 1772 pnfs_set_plh_return_iomode(lo, return_range->iomode); 1773 if (!mark_lseg_invalid(lseg, tmp_list)) 1774 remaining++; 1775 set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, 1776 &lo->plh_flags); 1777 } 1778 return remaining; 1779 } 1780 1781 void pnfs_error_mark_layout_for_return(struct inode *inode, 1782 struct pnfs_layout_segment *lseg) 1783 { 1784 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout; 1785 struct pnfs_layout_range range = { 1786 .iomode = lseg->pls_range.iomode, 1787 .offset = 0, 1788 .length = NFS4_MAX_UINT64, 1789 }; 1790 LIST_HEAD(free_me); 1791 bool return_now = false; 1792 1793 spin_lock(&inode->i_lock); 1794 pnfs_set_plh_return_iomode(lo, range.iomode); 1795 /* 1796 * mark all matching lsegs so that we are sure to have no live 1797 * segments at hand when sending layoutreturn. See pnfs_put_lseg() 1798 * for how it works. 1799 */ 1800 if (!pnfs_mark_matching_lsegs_return(lo, &free_me, &range)) { 1801 nfs4_stateid stateid; 1802 enum pnfs_iomode iomode = lo->plh_return_iomode; 1803 1804 nfs4_stateid_copy(&stateid, &lo->plh_stateid); 1805 return_now = pnfs_prepare_layoutreturn(lo); 1806 spin_unlock(&inode->i_lock); 1807 if (return_now) 1808 pnfs_send_layoutreturn(lo, &stateid, iomode, false); 1809 } else { 1810 spin_unlock(&inode->i_lock); 1811 nfs_commit_inode(inode, 0); 1812 } 1813 pnfs_free_lseg_list(&free_me); 1814 } 1815 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return); 1816 1817 void 1818 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req) 1819 { 1820 u64 rd_size = req->wb_bytes; 1821 1822 if (pgio->pg_lseg == NULL) { 1823 if (pgio->pg_dreq == NULL) 1824 rd_size = i_size_read(pgio->pg_inode) - req_offset(req); 1825 else 1826 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq); 1827 1828 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 1829 req->wb_context, 1830 req_offset(req), 1831 rd_size, 1832 IOMODE_READ, 1833 GFP_KERNEL); 1834 if (IS_ERR(pgio->pg_lseg)) { 1835 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 1836 pgio->pg_lseg = NULL; 1837 return; 1838 } 1839 } 1840 /* If no lseg, fall back to read through mds */ 1841 if (pgio->pg_lseg == NULL) 1842 nfs_pageio_reset_read_mds(pgio); 1843 1844 } 1845 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read); 1846 1847 void 1848 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, 1849 struct nfs_page *req, u64 wb_size) 1850 { 1851 if (pgio->pg_lseg == NULL) { 1852 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, 1853 req->wb_context, 1854 req_offset(req), 1855 wb_size, 1856 IOMODE_RW, 1857 GFP_NOFS); 1858 if (IS_ERR(pgio->pg_lseg)) { 1859 pgio->pg_error = PTR_ERR(pgio->pg_lseg); 1860 pgio->pg_lseg = NULL; 1861 return; 1862 } 1863 } 1864 /* If no lseg, fall back to write through mds */ 1865 if (pgio->pg_lseg == NULL) 1866 nfs_pageio_reset_write_mds(pgio); 1867 } 1868 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write); 1869 1870 void 1871 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc) 1872 { 1873 if (desc->pg_lseg) { 1874 pnfs_put_lseg(desc->pg_lseg); 1875 desc->pg_lseg = NULL; 1876 } 1877 } 1878 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup); 1879 1880 /* 1881 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number 1882 * of bytes (maximum @req->wb_bytes) that can be coalesced. 1883 */ 1884 size_t 1885 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, 1886 struct nfs_page *prev, struct nfs_page *req) 1887 { 1888 unsigned int size; 1889 u64 seg_end, req_start, seg_left; 1890 1891 size = nfs_generic_pg_test(pgio, prev, req); 1892 if (!size) 1893 return 0; 1894 1895 /* 1896 * 'size' contains the number of bytes left in the current page (up 1897 * to the original size asked for in @req->wb_bytes). 1898 * 1899 * Calculate how many bytes are left in the layout segment 1900 * and if there are less bytes than 'size', return that instead. 1901 * 1902 * Please also note that 'end_offset' is actually the offset of the 1903 * first byte that lies outside the pnfs_layout_range. FIXME? 1904 * 1905 */ 1906 if (pgio->pg_lseg) { 1907 seg_end = end_offset(pgio->pg_lseg->pls_range.offset, 1908 pgio->pg_lseg->pls_range.length); 1909 req_start = req_offset(req); 1910 WARN_ON_ONCE(req_start >= seg_end); 1911 /* start of request is past the last byte of this segment */ 1912 if (req_start >= seg_end) { 1913 /* reference the new lseg */ 1914 if (pgio->pg_ops->pg_cleanup) 1915 pgio->pg_ops->pg_cleanup(pgio); 1916 if (pgio->pg_ops->pg_init) 1917 pgio->pg_ops->pg_init(pgio, req); 1918 return 0; 1919 } 1920 1921 /* adjust 'size' iff there are fewer bytes left in the 1922 * segment than what nfs_generic_pg_test returned */ 1923 seg_left = seg_end - req_start; 1924 if (seg_left < size) 1925 size = (unsigned int)seg_left; 1926 } 1927 1928 return size; 1929 } 1930 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test); 1931 1932 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr) 1933 { 1934 struct nfs_pageio_descriptor pgio; 1935 1936 /* Resend all requests through the MDS */ 1937 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true, 1938 hdr->completion_ops); 1939 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags); 1940 return nfs_pageio_resend(&pgio, hdr); 1941 } 1942 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds); 1943 1944 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr) 1945 { 1946 1947 dprintk("pnfs write error = %d\n", hdr->pnfs_error); 1948 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags & 1949 PNFS_LAYOUTRET_ON_ERROR) { 1950 pnfs_return_layout(hdr->inode); 1951 } 1952 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) 1953 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr); 1954 } 1955 1956 /* 1957 * Called by non rpc-based layout drivers 1958 */ 1959 void pnfs_ld_write_done(struct nfs_pgio_header *hdr) 1960 { 1961 if (likely(!hdr->pnfs_error)) { 1962 pnfs_set_layoutcommit(hdr->inode, hdr->lseg, 1963 hdr->mds_offset + hdr->res.count); 1964 hdr->mds_ops->rpc_call_done(&hdr->task, hdr); 1965 } 1966 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error); 1967 if (unlikely(hdr->pnfs_error)) 1968 pnfs_ld_handle_write_error(hdr); 1969 hdr->mds_ops->rpc_release(hdr); 1970 } 1971 EXPORT_SYMBOL_GPL(pnfs_ld_write_done); 1972 1973 static void 1974 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc, 1975 struct nfs_pgio_header *hdr) 1976 { 1977 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 1978 1979 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 1980 list_splice_tail_init(&hdr->pages, &mirror->pg_list); 1981 nfs_pageio_reset_write_mds(desc); 1982 mirror->pg_recoalesce = 1; 1983 } 1984 nfs_pgio_data_destroy(hdr); 1985 hdr->release(hdr); 1986 } 1987 1988 static enum pnfs_try_status 1989 pnfs_try_to_write_data(struct nfs_pgio_header *hdr, 1990 const struct rpc_call_ops *call_ops, 1991 struct pnfs_layout_segment *lseg, 1992 int how) 1993 { 1994 struct inode *inode = hdr->inode; 1995 enum pnfs_try_status trypnfs; 1996 struct nfs_server *nfss = NFS_SERVER(inode); 1997 1998 hdr->mds_ops = call_ops; 1999 2000 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__, 2001 inode->i_ino, hdr->args.count, hdr->args.offset, how); 2002 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how); 2003 if (trypnfs != PNFS_NOT_ATTEMPTED) 2004 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE); 2005 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs); 2006 return trypnfs; 2007 } 2008 2009 static void 2010 pnfs_do_write(struct nfs_pageio_descriptor *desc, 2011 struct nfs_pgio_header *hdr, int how) 2012 { 2013 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops; 2014 struct pnfs_layout_segment *lseg = desc->pg_lseg; 2015 enum pnfs_try_status trypnfs; 2016 2017 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how); 2018 if (trypnfs == PNFS_NOT_ATTEMPTED) 2019 pnfs_write_through_mds(desc, hdr); 2020 } 2021 2022 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr) 2023 { 2024 pnfs_put_lseg(hdr->lseg); 2025 nfs_pgio_header_free(hdr); 2026 } 2027 2028 int 2029 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc) 2030 { 2031 struct nfs_pgio_header *hdr; 2032 int ret; 2033 2034 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops); 2035 if (!hdr) { 2036 desc->pg_error = -ENOMEM; 2037 return desc->pg_error; 2038 } 2039 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free); 2040 2041 hdr->lseg = pnfs_get_lseg(desc->pg_lseg); 2042 ret = nfs_generic_pgio(desc, hdr); 2043 if (!ret) 2044 pnfs_do_write(desc, hdr, desc->pg_ioflags); 2045 2046 return ret; 2047 } 2048 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages); 2049 2050 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr) 2051 { 2052 struct nfs_pageio_descriptor pgio; 2053 2054 /* Resend all requests through the MDS */ 2055 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops); 2056 return nfs_pageio_resend(&pgio, hdr); 2057 } 2058 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds); 2059 2060 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr) 2061 { 2062 dprintk("pnfs read error = %d\n", hdr->pnfs_error); 2063 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags & 2064 PNFS_LAYOUTRET_ON_ERROR) { 2065 pnfs_return_layout(hdr->inode); 2066 } 2067 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) 2068 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr); 2069 } 2070 2071 /* 2072 * Called by non rpc-based layout drivers 2073 */ 2074 void pnfs_ld_read_done(struct nfs_pgio_header *hdr) 2075 { 2076 if (likely(!hdr->pnfs_error)) { 2077 __nfs4_read_done_cb(hdr); 2078 hdr->mds_ops->rpc_call_done(&hdr->task, hdr); 2079 } 2080 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error); 2081 if (unlikely(hdr->pnfs_error)) 2082 pnfs_ld_handle_read_error(hdr); 2083 hdr->mds_ops->rpc_release(hdr); 2084 } 2085 EXPORT_SYMBOL_GPL(pnfs_ld_read_done); 2086 2087 static void 2088 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc, 2089 struct nfs_pgio_header *hdr) 2090 { 2091 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc); 2092 2093 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { 2094 list_splice_tail_init(&hdr->pages, &mirror->pg_list); 2095 nfs_pageio_reset_read_mds(desc); 2096 mirror->pg_recoalesce = 1; 2097 } 2098 nfs_pgio_data_destroy(hdr); 2099 hdr->release(hdr); 2100 } 2101 2102 /* 2103 * Call the appropriate parallel I/O subsystem read function. 2104 */ 2105 static enum pnfs_try_status 2106 pnfs_try_to_read_data(struct nfs_pgio_header *hdr, 2107 const struct rpc_call_ops *call_ops, 2108 struct pnfs_layout_segment *lseg) 2109 { 2110 struct inode *inode = hdr->inode; 2111 struct nfs_server *nfss = NFS_SERVER(inode); 2112 enum pnfs_try_status trypnfs; 2113 2114 hdr->mds_ops = call_ops; 2115 2116 dprintk("%s: Reading ino:%lu %u@%llu\n", 2117 __func__, inode->i_ino, hdr->args.count, hdr->args.offset); 2118 2119 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr); 2120 if (trypnfs != PNFS_NOT_ATTEMPTED) 2121 nfs_inc_stats(inode, NFSIOS_PNFS_READ); 2122 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs); 2123 return trypnfs; 2124 } 2125 2126 /* Resend all requests through pnfs. */ 2127 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr) 2128 { 2129 struct nfs_pageio_descriptor pgio; 2130 2131 nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops); 2132 return nfs_pageio_resend(&pgio, hdr); 2133 } 2134 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs); 2135 2136 static void 2137 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr) 2138 { 2139 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops; 2140 struct pnfs_layout_segment *lseg = desc->pg_lseg; 2141 enum pnfs_try_status trypnfs; 2142 int err = 0; 2143 2144 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg); 2145 if (trypnfs == PNFS_TRY_AGAIN) 2146 err = pnfs_read_resend_pnfs(hdr); 2147 if (trypnfs == PNFS_NOT_ATTEMPTED || err) 2148 pnfs_read_through_mds(desc, hdr); 2149 } 2150 2151 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr) 2152 { 2153 pnfs_put_lseg(hdr->lseg); 2154 nfs_pgio_header_free(hdr); 2155 } 2156 2157 int 2158 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc) 2159 { 2160 struct nfs_pgio_header *hdr; 2161 int ret; 2162 2163 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops); 2164 if (!hdr) { 2165 desc->pg_error = -ENOMEM; 2166 return desc->pg_error; 2167 } 2168 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free); 2169 hdr->lseg = pnfs_get_lseg(desc->pg_lseg); 2170 ret = nfs_generic_pgio(desc, hdr); 2171 if (!ret) 2172 pnfs_do_read(desc, hdr); 2173 return ret; 2174 } 2175 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages); 2176 2177 static void pnfs_clear_layoutcommitting(struct inode *inode) 2178 { 2179 unsigned long *bitlock = &NFS_I(inode)->flags; 2180 2181 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock); 2182 smp_mb__after_atomic(); 2183 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING); 2184 } 2185 2186 /* 2187 * There can be multiple RW segments. 2188 */ 2189 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp) 2190 { 2191 struct pnfs_layout_segment *lseg; 2192 2193 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) { 2194 if (lseg->pls_range.iomode == IOMODE_RW && 2195 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) 2196 list_add(&lseg->pls_lc_list, listp); 2197 } 2198 } 2199 2200 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp) 2201 { 2202 struct pnfs_layout_segment *lseg, *tmp; 2203 2204 /* Matched by references in pnfs_set_layoutcommit */ 2205 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) { 2206 list_del_init(&lseg->pls_lc_list); 2207 pnfs_put_lseg(lseg); 2208 } 2209 2210 pnfs_clear_layoutcommitting(inode); 2211 } 2212 2213 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg) 2214 { 2215 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode); 2216 } 2217 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail); 2218 2219 void 2220 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg, 2221 loff_t end_pos) 2222 { 2223 struct nfs_inode *nfsi = NFS_I(inode); 2224 bool mark_as_dirty = false; 2225 2226 spin_lock(&inode->i_lock); 2227 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) { 2228 nfsi->layout->plh_lwb = end_pos; 2229 mark_as_dirty = true; 2230 dprintk("%s: Set layoutcommit for inode %lu ", 2231 __func__, inode->i_ino); 2232 } else if (end_pos > nfsi->layout->plh_lwb) 2233 nfsi->layout->plh_lwb = end_pos; 2234 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) { 2235 /* references matched in nfs4_layoutcommit_release */ 2236 pnfs_get_lseg(lseg); 2237 } 2238 spin_unlock(&inode->i_lock); 2239 dprintk("%s: lseg %p end_pos %llu\n", 2240 __func__, lseg, nfsi->layout->plh_lwb); 2241 2242 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one 2243 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */ 2244 if (mark_as_dirty) 2245 mark_inode_dirty_sync(inode); 2246 } 2247 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit); 2248 2249 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data) 2250 { 2251 struct nfs_server *nfss = NFS_SERVER(data->args.inode); 2252 2253 if (nfss->pnfs_curr_ld->cleanup_layoutcommit) 2254 nfss->pnfs_curr_ld->cleanup_layoutcommit(data); 2255 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list); 2256 } 2257 2258 /* 2259 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and 2260 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough 2261 * data to disk to allow the server to recover the data if it crashes. 2262 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag 2263 * is off, and a COMMIT is sent to a data server, or 2264 * if WRITEs to a data server return NFS_DATA_SYNC. 2265 */ 2266 int 2267 pnfs_layoutcommit_inode(struct inode *inode, bool sync) 2268 { 2269 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld; 2270 struct nfs4_layoutcommit_data *data; 2271 struct nfs_inode *nfsi = NFS_I(inode); 2272 loff_t end_pos; 2273 int status; 2274 2275 if (!pnfs_layoutcommit_outstanding(inode)) 2276 return 0; 2277 2278 dprintk("--> %s inode %lu\n", __func__, inode->i_ino); 2279 2280 status = -EAGAIN; 2281 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) { 2282 if (!sync) 2283 goto out; 2284 status = wait_on_bit_lock_action(&nfsi->flags, 2285 NFS_INO_LAYOUTCOMMITTING, 2286 nfs_wait_bit_killable, 2287 TASK_KILLABLE); 2288 if (status) 2289 goto out; 2290 } 2291 2292 status = -ENOMEM; 2293 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */ 2294 data = kzalloc(sizeof(*data), GFP_NOFS); 2295 if (!data) 2296 goto clear_layoutcommitting; 2297 2298 status = 0; 2299 spin_lock(&inode->i_lock); 2300 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) 2301 goto out_unlock; 2302 2303 INIT_LIST_HEAD(&data->lseg_list); 2304 pnfs_list_write_lseg(inode, &data->lseg_list); 2305 2306 end_pos = nfsi->layout->plh_lwb; 2307 2308 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid); 2309 spin_unlock(&inode->i_lock); 2310 2311 data->args.inode = inode; 2312 data->cred = get_rpccred(nfsi->layout->plh_lc_cred); 2313 nfs_fattr_init(&data->fattr); 2314 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask; 2315 data->res.fattr = &data->fattr; 2316 data->args.lastbytewritten = end_pos - 1; 2317 data->res.server = NFS_SERVER(inode); 2318 2319 if (ld->prepare_layoutcommit) { 2320 status = ld->prepare_layoutcommit(&data->args); 2321 if (status) { 2322 put_rpccred(data->cred); 2323 spin_lock(&inode->i_lock); 2324 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags); 2325 if (end_pos > nfsi->layout->plh_lwb) 2326 nfsi->layout->plh_lwb = end_pos; 2327 goto out_unlock; 2328 } 2329 } 2330 2331 2332 status = nfs4_proc_layoutcommit(data, sync); 2333 out: 2334 if (status) 2335 mark_inode_dirty_sync(inode); 2336 dprintk("<-- %s status %d\n", __func__, status); 2337 return status; 2338 out_unlock: 2339 spin_unlock(&inode->i_lock); 2340 kfree(data); 2341 clear_layoutcommitting: 2342 pnfs_clear_layoutcommitting(inode); 2343 goto out; 2344 } 2345 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode); 2346 2347 int 2348 pnfs_generic_sync(struct inode *inode, bool datasync) 2349 { 2350 return pnfs_layoutcommit_inode(inode, true); 2351 } 2352 EXPORT_SYMBOL_GPL(pnfs_generic_sync); 2353 2354 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void) 2355 { 2356 struct nfs4_threshold *thp; 2357 2358 thp = kzalloc(sizeof(*thp), GFP_NOFS); 2359 if (!thp) { 2360 dprintk("%s mdsthreshold allocation failed\n", __func__); 2361 return NULL; 2362 } 2363 return thp; 2364 } 2365 2366 #if IS_ENABLED(CONFIG_NFS_V4_2) 2367 int 2368 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags) 2369 { 2370 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld; 2371 struct nfs_server *server = NFS_SERVER(inode); 2372 struct nfs_inode *nfsi = NFS_I(inode); 2373 struct nfs42_layoutstat_data *data; 2374 struct pnfs_layout_hdr *hdr; 2375 int status = 0; 2376 2377 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats) 2378 goto out; 2379 2380 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS)) 2381 goto out; 2382 2383 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags)) 2384 goto out; 2385 2386 spin_lock(&inode->i_lock); 2387 if (!NFS_I(inode)->layout) { 2388 spin_unlock(&inode->i_lock); 2389 goto out; 2390 } 2391 hdr = NFS_I(inode)->layout; 2392 pnfs_get_layout_hdr(hdr); 2393 spin_unlock(&inode->i_lock); 2394 2395 data = kzalloc(sizeof(*data), gfp_flags); 2396 if (!data) { 2397 status = -ENOMEM; 2398 goto out_put; 2399 } 2400 2401 data->args.fh = NFS_FH(inode); 2402 data->args.inode = inode; 2403 nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid); 2404 status = ld->prepare_layoutstats(&data->args); 2405 if (status) 2406 goto out_free; 2407 2408 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data); 2409 2410 out: 2411 dprintk("%s returns %d\n", __func__, status); 2412 return status; 2413 2414 out_free: 2415 kfree(data); 2416 out_put: 2417 pnfs_put_layout_hdr(hdr); 2418 smp_mb__before_atomic(); 2419 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags); 2420 smp_mb__after_atomic(); 2421 goto out; 2422 } 2423 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat); 2424 #endif 2425 2426 unsigned int layoutstats_timer; 2427 module_param(layoutstats_timer, uint, 0644); 2428 EXPORT_SYMBOL_GPL(layoutstats_timer); 2429