1// SPDX-License-Identifier: GPL-2.0 2/* 3 * GPL HEADER START 4 * 5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License version 2 only, 9 * as published by the Free Software Foundation. 10 * 11 * This program is distributed in the hope that it will be useful, but 12 * WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 * General Public License version 2 for more details (a copy is included 15 * in the LICENSE file that accompanied this code). 16 * 17 * You should have received a copy of the GNU General Public License 18 * version 2 along with this program; If not, see 19 * http://www.gnu.org/licenses/gpl-2.0.html 20 * 21 * GPL HEADER END 22 */ 23/* 24 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. 25 * Use is subject to license terms. 26 * 27 * Copyright (c) 2011, 2015, Intel Corporation. 28 */ 29/* 30 * This file is part of Lustre, http://www.lustre.org/ 31 * Lustre is a trademark of Sun Microsystems, Inc. 32 */ 33#ifndef _LUSTRE_CL_OBJECT_H 34#define _LUSTRE_CL_OBJECT_H 35 36/** \defgroup clio clio 37 * 38 * Client objects implement io operations and cache pages. 39 * 40 * Examples: lov and osc are implementations of cl interface. 41 * 42 * Big Theory Statement. 43 * 44 * Layered objects. 45 * 46 * Client implementation is based on the following data-types: 47 * 48 * - cl_object 49 * 50 * - cl_page 51 * 52 * - cl_lock represents an extent lock on an object. 53 * 54 * - cl_io represents high-level i/o activity such as whole read/write 55 * system call, or write-out of pages from under the lock being 56 * canceled. cl_io has sub-ios that can be stopped and resumed 57 * independently, thus achieving high degree of transfer 58 * parallelism. Single cl_io can be advanced forward by 59 * the multiple threads (although in the most usual case of 60 * read/write system call it is associated with the single user 61 * thread, that issued the system call). 62 * 63 * Terminology 64 * 65 * - to avoid confusion high-level I/O operation like read or write system 66 * call is referred to as "an io", whereas low-level I/O operation, like 67 * RPC, is referred to as "a transfer" 68 * 69 * - "generic code" means generic (not file system specific) code in the 70 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that 71 * is not layer specific. 72 * 73 * Locking. 74 * 75 * - i_mutex 76 * - PG_locked 77 * - cl_object_header::coh_page_guard 78 * - lu_site::ls_guard 79 * 80 * See the top comment in cl_object.c for the description of overall locking and 81 * reference-counting design. 82 * 83 * See comments below for the description of i/o, page, and dlm-locking 84 * design. 85 * 86 * @{ 87 */ 88 89/* 90 * super-class definitions. 91 */ 92#include <lu_object.h> 93#include <lustre_compat.h> 94#include <linux/atomic.h> 95#include <linux/mutex.h> 96#include <linux/radix-tree.h> 97#include <linux/spinlock.h> 98#include <linux/wait.h> 99 100struct inode; 101 102struct cl_device; 103 104struct cl_object; 105 106struct cl_page; 107struct cl_page_slice; 108struct cl_lock; 109struct cl_lock_slice; 110 111struct cl_lock_operations; 112struct cl_page_operations; 113 114struct cl_io; 115struct cl_io_slice; 116 117struct cl_req_attr; 118 119/** 120 * Device in the client stack. 121 * 122 * \see vvp_device, lov_device, lovsub_device, osc_device 123 */ 124struct cl_device { 125 /** Super-class. */ 126 struct lu_device cd_lu_dev; 127}; 128 129/** \addtogroup cl_object cl_object 130 * @{ 131 */ 132/** 133 * "Data attributes" of cl_object. Data attributes can be updated 134 * independently for a sub-object, and top-object's attributes are calculated 135 * from sub-objects' ones. 136 */ 137struct cl_attr { 138 /** Object size, in bytes */ 139 loff_t cat_size; 140 /** 141 * Known minimal size, in bytes. 142 * 143 * This is only valid when at least one DLM lock is held. 144 */ 145 loff_t cat_kms; 146 /** Modification time. Measured in seconds since epoch. */ 147 time64_t cat_mtime; 148 /** Access time. Measured in seconds since epoch. */ 149 time64_t cat_atime; 150 /** Change time. Measured in seconds since epoch. */ 151 time64_t cat_ctime; 152 /** 153 * Blocks allocated to this cl_object on the server file system. 154 * 155 * \todo XXX An interface for block size is needed. 156 */ 157 __u64 cat_blocks; 158 /** 159 * User identifier for quota purposes. 160 */ 161 uid_t cat_uid; 162 /** 163 * Group identifier for quota purposes. 164 */ 165 gid_t cat_gid; 166 167 /* nlink of the directory */ 168 __u64 cat_nlink; 169}; 170 171/** 172 * Fields in cl_attr that are being set. 173 */ 174enum cl_attr_valid { 175 CAT_SIZE = 1 << 0, 176 CAT_KMS = 1 << 1, 177 CAT_MTIME = 1 << 3, 178 CAT_ATIME = 1 << 4, 179 CAT_CTIME = 1 << 5, 180 CAT_BLOCKS = 1 << 6, 181 CAT_UID = 1 << 7, 182 CAT_GID = 1 << 8 183}; 184 185/** 186 * Sub-class of lu_object with methods common for objects on the client 187 * stacks. 188 * 189 * cl_object: represents a regular file system object, both a file and a 190 * stripe. cl_object is based on lu_object: it is identified by a fid, 191 * layered, cached, hashed, and lrued. Important distinction with the server 192 * side, where md_object and dt_object are used, is that cl_object "fans out" 193 * at the lov/sns level: depending on the file layout, single file is 194 * represented as a set of "sub-objects" (stripes). At the implementation 195 * level, struct lov_object contains an array of cl_objects. Each sub-object 196 * is a full-fledged cl_object, having its fid, living in the lru and hash 197 * table. 198 * 199 * This leads to the next important difference with the server side: on the 200 * client, it's quite usual to have objects with the different sequence of 201 * layers. For example, typical top-object is composed of the following 202 * layers: 203 * 204 * - vvp 205 * - lov 206 * 207 * whereas its sub-objects are composed of 208 * 209 * - lovsub 210 * - osc 211 * 212 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep 213 * track of the object-subobject relationship. 214 * 215 * Sub-objects are not cached independently: when top-object is about to 216 * be discarded from the memory, all its sub-objects are torn-down and 217 * destroyed too. 218 * 219 * \see vvp_object, lov_object, lovsub_object, osc_object 220 */ 221struct cl_object { 222 /** super class */ 223 struct lu_object co_lu; 224 /** per-object-layer operations */ 225 const struct cl_object_operations *co_ops; 226 /** offset of page slice in cl_page buffer */ 227 int co_slice_off; 228}; 229 230/** 231 * Description of the client object configuration. This is used for the 232 * creation of a new client object that is identified by a more state than 233 * fid. 234 */ 235struct cl_object_conf { 236 /** Super-class. */ 237 struct lu_object_conf coc_lu; 238 union { 239 /** 240 * Object layout. This is consumed by lov. 241 */ 242 struct lu_buf coc_layout; 243 /** 244 * Description of particular stripe location in the 245 * cluster. This is consumed by osc. 246 */ 247 struct lov_oinfo *coc_oinfo; 248 } u; 249 /** 250 * VFS inode. This is consumed by vvp. 251 */ 252 struct inode *coc_inode; 253 /** 254 * Layout lock handle. 255 */ 256 struct ldlm_lock *coc_lock; 257 /** 258 * Operation to handle layout, OBJECT_CONF_XYZ. 259 */ 260 int coc_opc; 261}; 262 263enum { 264 /** configure layout, set up a new stripe, must be called while 265 * holding layout lock. 266 */ 267 OBJECT_CONF_SET = 0, 268 /** invalidate the current stripe configuration due to losing 269 * layout lock. 270 */ 271 OBJECT_CONF_INVALIDATE = 1, 272 /** wait for old layout to go away so that new layout can be set up. */ 273 OBJECT_CONF_WAIT = 2 274}; 275 276enum { 277 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */ 278 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */ 279}; 280 281struct cl_layout { 282 /** the buffer to return the layout in lov_mds_md format. */ 283 struct lu_buf cl_buf; 284 /** size of layout in lov_mds_md format. */ 285 size_t cl_size; 286 /** Layout generation. */ 287 u32 cl_layout_gen; 288}; 289 290/** 291 * Operations implemented for each cl object layer. 292 * 293 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops 294 */ 295struct cl_object_operations { 296 /** 297 * Initialize page slice for this layer. Called top-to-bottom through 298 * every object layer when a new cl_page is instantiated. Layer 299 * keeping private per-page data, or requiring its own page operations 300 * vector should allocate these data here, and attach then to the page 301 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM 302 * sense). Optional. 303 * 304 * \retval NULL success. 305 * 306 * \retval ERR_PTR(errno) failure code. 307 * 308 * \retval valid-pointer pointer to already existing referenced page 309 * to be used instead of newly created. 310 */ 311 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj, 312 struct cl_page *page, pgoff_t index); 313 /** 314 * Initialize lock slice for this layer. Called top-to-bottom through 315 * every object layer when a new cl_lock is instantiated. Layer 316 * keeping private per-lock data, or requiring its own lock operations 317 * vector should allocate these data here, and attach then to the lock 318 * by calling cl_lock_slice_add(). Mandatory. 319 */ 320 int (*coo_lock_init)(const struct lu_env *env, 321 struct cl_object *obj, struct cl_lock *lock, 322 const struct cl_io *io); 323 /** 324 * Initialize io state for a given layer. 325 * 326 * called top-to-bottom once per io existence to initialize io 327 * state. If layer wants to keep some state for this type of io, it 328 * has to embed struct cl_io_slice in lu_env::le_ses, and register 329 * slice with cl_io_slice_add(). It is guaranteed that all threads 330 * participating in this io share the same session. 331 */ 332 int (*coo_io_init)(const struct lu_env *env, 333 struct cl_object *obj, struct cl_io *io); 334 /** 335 * Fill portion of \a attr that this layer controls. This method is 336 * called top-to-bottom through all object layers. 337 * 338 * \pre cl_object_header::coh_attr_guard of the top-object is locked. 339 * 340 * \return 0: to continue 341 * \return +ve: to stop iterating through layers (but 0 is returned 342 * from enclosing cl_object_attr_get()) 343 * \return -ve: to signal error 344 */ 345 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj, 346 struct cl_attr *attr); 347 /** 348 * Update attributes. 349 * 350 * \a valid is a bitmask composed from enum #cl_attr_valid, and 351 * indicating what attributes are to be set. 352 * 353 * \pre cl_object_header::coh_attr_guard of the top-object is locked. 354 * 355 * \return the same convention as for 356 * cl_object_operations::coo_attr_get() is used. 357 */ 358 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj, 359 const struct cl_attr *attr, unsigned int valid); 360 /** 361 * Update object configuration. Called top-to-bottom to modify object 362 * configuration. 363 * 364 * XXX error conditions and handling. 365 */ 366 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj, 367 const struct cl_object_conf *conf); 368 /** 369 * Glimpse ast. Executed when glimpse ast arrives for a lock on this 370 * object. Layers are supposed to fill parts of \a lvb that will be 371 * shipped to the glimpse originator as a glimpse result. 372 * 373 * \see vvp_object_glimpse(), lovsub_object_glimpse(), 374 * \see osc_object_glimpse() 375 */ 376 int (*coo_glimpse)(const struct lu_env *env, 377 const struct cl_object *obj, struct ost_lvb *lvb); 378 /** 379 * Object prune method. Called when the layout is going to change on 380 * this object, therefore each layer has to clean up their cache, 381 * mainly pages and locks. 382 */ 383 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj); 384 /** 385 * Object getstripe method. 386 */ 387 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj, 388 struct lov_user_md __user *lum); 389 /** 390 * Get FIEMAP mapping from the object. 391 */ 392 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj, 393 struct ll_fiemap_info_key *fmkey, 394 struct fiemap *fiemap, size_t *buflen); 395 /** 396 * Get layout and generation of the object. 397 */ 398 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj, 399 struct cl_layout *layout); 400 /** 401 * Get maximum size of the object. 402 */ 403 loff_t (*coo_maxbytes)(struct cl_object *obj); 404 /** 405 * Set request attributes. 406 */ 407 void (*coo_req_attr_set)(const struct lu_env *env, 408 struct cl_object *obj, 409 struct cl_req_attr *attr); 410}; 411 412/** 413 * Extended header for client object. 414 */ 415struct cl_object_header { 416 /** Standard lu_object_header. cl_object::co_lu::lo_header points 417 * here. 418 */ 419 struct lu_object_header coh_lu; 420 421 /** 422 * Parent object. It is assumed that an object has a well-defined 423 * parent, but not a well-defined child (there may be multiple 424 * sub-objects, for the same top-object). cl_object_header::coh_parent 425 * field allows certain code to be written generically, without 426 * limiting possible cl_object layouts unduly. 427 */ 428 struct cl_object_header *coh_parent; 429 /** 430 * Protects consistency between cl_attr of parent object and 431 * attributes of sub-objects, that the former is calculated ("merged") 432 * from. 433 * 434 * \todo XXX this can be read/write lock if needed. 435 */ 436 spinlock_t coh_attr_guard; 437 /** 438 * Size of cl_page + page slices 439 */ 440 unsigned short coh_page_bufsize; 441 /** 442 * Number of objects above this one: 0 for a top-object, 1 for its 443 * sub-object, etc. 444 */ 445 unsigned char coh_nesting; 446}; 447 448/** 449 * Helper macro: iterate over all layers of the object \a obj, assigning every 450 * layer top-to-bottom to \a slice. 451 */ 452#define cl_object_for_each(slice, obj) \ 453 list_for_each_entry((slice), \ 454 &(obj)->co_lu.lo_header->loh_layers, \ 455 co_lu.lo_linkage) 456/** 457 * Helper macro: iterate over all layers of the object \a obj, assigning every 458 * layer bottom-to-top to \a slice. 459 */ 460#define cl_object_for_each_reverse(slice, obj) \ 461 list_for_each_entry_reverse((slice), \ 462 &(obj)->co_lu.lo_header->loh_layers, \ 463 co_lu.lo_linkage) 464/** @} cl_object */ 465 466#define CL_PAGE_EOF ((pgoff_t)~0ull) 467 468/** \addtogroup cl_page cl_page 469 * @{ 470 */ 471 472/** \struct cl_page 473 * Layered client page. 474 * 475 * cl_page: represents a portion of a file, cached in the memory. All pages 476 * of the given file are of the same size, and are kept in the radix tree 477 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects 478 * of the top-level file object are first class cl_objects, they have their 479 * own radix trees of pages and hence page is implemented as a sequence of 480 * struct cl_pages's, linked into double-linked list through 481 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the 482 * corresponding radix tree at the corresponding logical offset. 483 * 484 * cl_page is associated with VM page of the hosting environment (struct 485 * page in Linux kernel, for example), struct page. It is assumed, that this 486 * association is implemented by one of cl_page layers (top layer in the 487 * current design) that 488 * 489 * - intercepts per-VM-page call-backs made by the environment (e.g., 490 * memory pressure), 491 * 492 * - translates state (page flag bits) and locking between lustre and 493 * environment. 494 * 495 * The association between cl_page and struct page is immutable and 496 * established when cl_page is created. 497 * 498 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing 499 * this io an exclusive access to this page w.r.t. other io attempts and 500 * various events changing page state (such as transfer completion, or 501 * eviction of the page from the memory). Note, that in general cl_io 502 * cannot be identified with a particular thread, and page ownership is not 503 * exactly equal to the current thread holding a lock on the page. Layer 504 * implementing association between cl_page and struct page has to implement 505 * ownership on top of available synchronization mechanisms. 506 * 507 * While lustre client maintains the notion of an page ownership by io, 508 * hosting MM/VM usually has its own page concurrency control 509 * mechanisms. For example, in Linux, page access is synchronized by the 510 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*()) 511 * takes care to acquire and release such locks as necessary around the 512 * calls to the file system methods (->readpage(), ->prepare_write(), 513 * ->commit_write(), etc.). This leads to the situation when there are two 514 * different ways to own a page in the client: 515 * 516 * - client code explicitly and voluntary owns the page (cl_page_own()); 517 * 518 * - VM locks a page and then calls the client, that has "to assume" 519 * the ownership from the VM (cl_page_assume()). 520 * 521 * Dual methods to release ownership are cl_page_disown() and 522 * cl_page_unassume(). 523 * 524 * cl_page is reference counted (cl_page::cp_ref). When reference counter 525 * drops to 0, the page is returned to the cache, unless it is in 526 * cl_page_state::CPS_FREEING state, in which case it is immediately 527 * destroyed. 528 * 529 * The general logic guaranteeing the absence of "existential races" for 530 * pages is the following: 531 * 532 * - there are fixed known ways for a thread to obtain a new reference 533 * to a page: 534 * 535 * - by doing a lookup in the cl_object radix tree, protected by the 536 * spin-lock; 537 * 538 * - by starting from VM-locked struct page and following some 539 * hosting environment method (e.g., following ->private pointer in 540 * the case of Linux kernel), see cl_vmpage_page(); 541 * 542 * - when the page enters cl_page_state::CPS_FREEING state, all these 543 * ways are severed with the proper synchronization 544 * (cl_page_delete()); 545 * 546 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page 547 * lock; 548 * 549 * - no new references to the page in cl_page_state::CPS_FREEING state 550 * are allowed (checked in cl_page_get()). 551 * 552 * Together this guarantees that when last reference to a 553 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the 554 * page, as neither references to it can be acquired at that point, nor 555 * ones exist. 556 * 557 * cl_page is a state machine. States are enumerated in enum 558 * cl_page_state. Possible state transitions are enumerated in 559 * cl_page_state_set(). State transition process (i.e., actual changing of 560 * cl_page::cp_state field) is protected by the lock on the underlying VM 561 * page. 562 * 563 * Linux Kernel implementation. 564 * 565 * Binding between cl_page and struct page (which is a typedef for 566 * struct page) is implemented in the vvp layer. cl_page is attached to the 567 * ->private pointer of the struct page, together with the setting of 568 * PG_private bit in page->flags, and acquiring additional reference on the 569 * struct page (much like struct buffer_head, or any similar file system 570 * private data structures). 571 * 572 * PG_locked lock is used to implement both ownership and transfer 573 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}} 574 * states. No additional references are acquired for the duration of the 575 * transfer. 576 * 577 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where 578 * write-out is "protected" by the special PG_writeback bit. 579 */ 580 581/** 582 * States of cl_page. cl_page.c assumes particular order here. 583 * 584 * The page state machine is rather crude, as it doesn't recognize finer page 585 * states like "dirty" or "up to date". This is because such states are not 586 * always well defined for the whole stack (see, for example, the 587 * implementation of the read-ahead, that hides page up-to-dateness to track 588 * cache hits accurately). Such sub-states are maintained by the layers that 589 * are interested in them. 590 */ 591enum cl_page_state { 592 /** 593 * Page is in the cache, un-owned. Page leaves cached state in the 594 * following cases: 595 * 596 * - [cl_page_state::CPS_OWNED] io comes across the page and 597 * owns it; 598 * 599 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the 600 * req-formation engine decides that it wants to include this page 601 * into an RPC being constructed, and yanks it from the cache; 602 * 603 * - [cl_page_state::CPS_FREEING] VM callback is executed to 604 * evict the page form the memory; 605 * 606 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL 607 */ 608 CPS_CACHED, 609 /** 610 * Page is exclusively owned by some cl_io. Page may end up in this 611 * state as a result of 612 * 613 * - io creating new page and immediately owning it; 614 * 615 * - [cl_page_state::CPS_CACHED] io finding existing cached page 616 * and owning it; 617 * 618 * - [cl_page_state::CPS_OWNED] io finding existing owned page 619 * and waiting for owner to release the page; 620 * 621 * Page leaves owned state in the following cases: 622 * 623 * - [cl_page_state::CPS_CACHED] io decides to leave the page in 624 * the cache, doing nothing; 625 * 626 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for 627 * this page; 628 * 629 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write 630 * transfer for this page; 631 * 632 * - [cl_page_state::CPS_FREEING] io decides to destroy this 633 * page (e.g., as part of truncate or extent lock cancellation). 634 * 635 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL 636 */ 637 CPS_OWNED, 638 /** 639 * Page is being written out, as a part of a transfer. This state is 640 * entered when req-formation logic decided that it wants this page to 641 * be sent through the wire _now_. Specifically, it means that once 642 * this state is achieved, transfer completion handler (with either 643 * success or failure indication) is guaranteed to be executed against 644 * this page independently of any locks and any scheduling decisions 645 * made by the hosting environment (that effectively means that the 646 * page is never put into cl_page_state::CPS_PAGEOUT state "in 647 * advance". This property is mentioned, because it is important when 648 * reasoning about possible dead-locks in the system). The page can 649 * enter this state as a result of 650 * 651 * - [cl_page_state::CPS_OWNED] an io requesting an immediate 652 * write-out of this page, or 653 * 654 * - [cl_page_state::CPS_CACHED] req-forming engine deciding 655 * that it has enough dirty pages cached to issue a "good" 656 * transfer. 657 * 658 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer 659 * is completed---it is moved into cl_page_state::CPS_CACHED state. 660 * 661 * Underlying VM page is locked for the duration of transfer. 662 * 663 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL 664 */ 665 CPS_PAGEOUT, 666 /** 667 * Page is being read in, as a part of a transfer. This is quite 668 * similar to the cl_page_state::CPS_PAGEOUT state, except that 669 * read-in is always "immediate"---there is no such thing a sudden 670 * construction of read request from cached, presumably not up to date, 671 * pages. 672 * 673 * Underlying VM page is locked for the duration of transfer. 674 * 675 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL 676 */ 677 CPS_PAGEIN, 678 /** 679 * Page is being destroyed. This state is entered when client decides 680 * that page has to be deleted from its host object, as, e.g., a part 681 * of truncate. 682 * 683 * Once this state is reached, there is no way to escape it. 684 * 685 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL 686 */ 687 CPS_FREEING, 688 CPS_NR 689}; 690 691enum cl_page_type { 692 /** Host page, the page is from the host inode which the cl_page 693 * belongs to. 694 */ 695 CPT_CACHEABLE = 1, 696 697 /** Transient page, the transient cl_page is used to bind a cl_page 698 * to vmpage which is not belonging to the same object of cl_page. 699 * it is used in DirectIO and lockless IO. 700 */ 701 CPT_TRANSIENT, 702}; 703 704/** 705 * Fields are protected by the lock on struct page, except for atomics and 706 * immutables. 707 * 708 * \invariant Data type invariants are in cl_page_invariant(). Basically: 709 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked 710 * list, consistent with the parent/child pointers in the cl_page::cp_obj and 711 * cl_page::cp_owner (when set). 712 */ 713struct cl_page { 714 /** Reference counter. */ 715 atomic_t cp_ref; 716 /** An object this page is a part of. Immutable after creation. */ 717 struct cl_object *cp_obj; 718 /** vmpage */ 719 struct page *cp_vmpage; 720 /** Linkage of pages within group. Pages must be owned */ 721 struct list_head cp_batch; 722 /** List of slices. Immutable after creation. */ 723 struct list_head cp_layers; 724 /** 725 * Page state. This field is const to avoid accidental update, it is 726 * modified only internally within cl_page.c. Protected by a VM lock. 727 */ 728 const enum cl_page_state cp_state; 729 /** 730 * Page type. Only CPT_TRANSIENT is used so far. Immutable after 731 * creation. 732 */ 733 enum cl_page_type cp_type; 734 735 /** 736 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned 737 * by sub-io. Protected by a VM lock. 738 */ 739 struct cl_io *cp_owner; 740 /** List of references to this page, for debugging. */ 741 struct lu_ref cp_reference; 742 /** Link to an object, for debugging. */ 743 struct lu_ref_link cp_obj_ref; 744 /** Link to a queue, for debugging. */ 745 struct lu_ref_link cp_queue_ref; 746 /** Assigned if doing a sync_io */ 747 struct cl_sync_io *cp_sync_io; 748}; 749 750/** 751 * Per-layer part of cl_page. 752 * 753 * \see vvp_page, lov_page, osc_page 754 */ 755struct cl_page_slice { 756 struct cl_page *cpl_page; 757 pgoff_t cpl_index; 758 /** 759 * Object slice corresponding to this page slice. Immutable after 760 * creation. 761 */ 762 struct cl_object *cpl_obj; 763 const struct cl_page_operations *cpl_ops; 764 /** Linkage into cl_page::cp_layers. Immutable after creation. */ 765 struct list_head cpl_linkage; 766}; 767 768/** 769 * Lock mode. For the client extent locks. 770 * 771 * \ingroup cl_lock 772 */ 773enum cl_lock_mode { 774 CLM_READ, 775 CLM_WRITE, 776 CLM_GROUP 777}; 778 779/** 780 * Requested transfer type. 781 */ 782enum cl_req_type { 783 CRT_READ, 784 CRT_WRITE, 785 CRT_NR 786}; 787 788/** 789 * Per-layer page operations. 790 * 791 * Methods taking an \a io argument are for the activity happening in the 792 * context of given \a io. Page is assumed to be owned by that io, except for 793 * the obvious cases (like cl_page_operations::cpo_own()). 794 * 795 * \see vvp_page_ops, lov_page_ops, osc_page_ops 796 */ 797struct cl_page_operations { 798 /** 799 * cl_page<->struct page methods. Only one layer in the stack has to 800 * implement these. Current code assumes that this functionality is 801 * provided by the topmost layer, see cl_page_disown0() as an example. 802 */ 803 804 /** 805 * Called when \a io acquires this page into the exclusive 806 * ownership. When this method returns, it is guaranteed that the is 807 * not owned by other io, and no transfer is going on against 808 * it. Optional. 809 * 810 * \see cl_page_own() 811 * \see vvp_page_own(), lov_page_own() 812 */ 813 int (*cpo_own)(const struct lu_env *env, 814 const struct cl_page_slice *slice, 815 struct cl_io *io, int nonblock); 816 /** Called when ownership it yielded. Optional. 817 * 818 * \see cl_page_disown() 819 * \see vvp_page_disown() 820 */ 821 void (*cpo_disown)(const struct lu_env *env, 822 const struct cl_page_slice *slice, struct cl_io *io); 823 /** 824 * Called for a page that is already "owned" by \a io from VM point of 825 * view. Optional. 826 * 827 * \see cl_page_assume() 828 * \see vvp_page_assume(), lov_page_assume() 829 */ 830 void (*cpo_assume)(const struct lu_env *env, 831 const struct cl_page_slice *slice, struct cl_io *io); 832 /** Dual to cl_page_operations::cpo_assume(). Optional. Called 833 * bottom-to-top when IO releases a page without actually unlocking 834 * it. 835 * 836 * \see cl_page_unassume() 837 * \see vvp_page_unassume() 838 */ 839 void (*cpo_unassume)(const struct lu_env *env, 840 const struct cl_page_slice *slice, 841 struct cl_io *io); 842 /** 843 * Announces whether the page contains valid data or not by \a uptodate. 844 * 845 * \see cl_page_export() 846 * \see vvp_page_export() 847 */ 848 void (*cpo_export)(const struct lu_env *env, 849 const struct cl_page_slice *slice, int uptodate); 850 /** 851 * Checks whether underlying VM page is locked (in the suitable 852 * sense). Used for assertions. 853 * 854 * \retval -EBUSY: page is protected by a lock of a given mode; 855 * \retval -ENODATA: page is not protected by a lock; 856 * \retval 0: this layer cannot decide. (Should never happen.) 857 */ 858 int (*cpo_is_vmlocked)(const struct lu_env *env, 859 const struct cl_page_slice *slice); 860 /** 861 * Page destruction. 862 */ 863 864 /** 865 * Called when page is truncated from the object. Optional. 866 * 867 * \see cl_page_discard() 868 * \see vvp_page_discard(), osc_page_discard() 869 */ 870 void (*cpo_discard)(const struct lu_env *env, 871 const struct cl_page_slice *slice, 872 struct cl_io *io); 873 /** 874 * Called when page is removed from the cache, and is about to being 875 * destroyed. Optional. 876 * 877 * \see cl_page_delete() 878 * \see vvp_page_delete(), osc_page_delete() 879 */ 880 void (*cpo_delete)(const struct lu_env *env, 881 const struct cl_page_slice *slice); 882 /** Destructor. Frees resources and slice itself. */ 883 void (*cpo_fini)(const struct lu_env *env, 884 struct cl_page_slice *slice); 885 /** 886 * Optional debugging helper. Prints given page slice. 887 * 888 * \see cl_page_print() 889 */ 890 int (*cpo_print)(const struct lu_env *env, 891 const struct cl_page_slice *slice, 892 void *cookie, lu_printer_t p); 893 /** 894 * \name transfer 895 * 896 * Transfer methods. 897 * 898 * @{ 899 */ 900 /** 901 * Request type dependent vector of operations. 902 * 903 * Transfer operations depend on transfer mode (cl_req_type). To avoid 904 * passing transfer mode to each and every of these methods, and to 905 * avoid branching on request type inside of the methods, separate 906 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are 907 * provided. That is, method invocation usually looks like 908 * 909 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...); 910 */ 911 struct { 912 /** 913 * Called when a page is submitted for a transfer as a part of 914 * cl_page_list. 915 * 916 * \return 0 : page is eligible for submission; 917 * \return -EALREADY : skip this page; 918 * \return -ve : error. 919 * 920 * \see cl_page_prep() 921 */ 922 int (*cpo_prep)(const struct lu_env *env, 923 const struct cl_page_slice *slice, 924 struct cl_io *io); 925 /** 926 * Completion handler. This is guaranteed to be eventually 927 * fired after cl_page_operations::cpo_prep() or 928 * cl_page_operations::cpo_make_ready() call. 929 * 930 * This method can be called in a non-blocking context. It is 931 * guaranteed however, that the page involved and its object 932 * are pinned in memory (and, hence, calling cl_page_put() is 933 * safe). 934 * 935 * \see cl_page_completion() 936 */ 937 void (*cpo_completion)(const struct lu_env *env, 938 const struct cl_page_slice *slice, 939 int ioret); 940 /** 941 * Called when cached page is about to be added to the 942 * ptlrpc request as a part of req formation. 943 * 944 * \return 0 : proceed with this page; 945 * \return -EAGAIN : skip this page; 946 * \return -ve : error. 947 * 948 * \see cl_page_make_ready() 949 */ 950 int (*cpo_make_ready)(const struct lu_env *env, 951 const struct cl_page_slice *slice); 952 } io[CRT_NR]; 953 /** 954 * Tell transfer engine that only [to, from] part of a page should be 955 * transmitted. 956 * 957 * This is used for immediate transfers. 958 * 959 * \todo XXX this is not very good interface. It would be much better 960 * if all transfer parameters were supplied as arguments to 961 * cl_io_operations::cio_submit() call, but it is not clear how to do 962 * this for page queues. 963 * 964 * \see cl_page_clip() 965 */ 966 void (*cpo_clip)(const struct lu_env *env, 967 const struct cl_page_slice *slice, 968 int from, int to); 969 /** 970 * \pre the page was queued for transferring. 971 * \post page is removed from client's pending list, or -EBUSY 972 * is returned if it has already been in transferring. 973 * 974 * This is one of seldom page operation which is: 975 * 0. called from top level; 976 * 1. don't have vmpage locked; 977 * 2. every layer should synchronize execution of its ->cpo_cancel() 978 * with completion handlers. Osc uses client obd lock for this 979 * purpose. Based on there is no vvp_page_cancel and 980 * lov_page_cancel(), cpo_cancel is defacto protected by client lock. 981 * 982 * \see osc_page_cancel(). 983 */ 984 int (*cpo_cancel)(const struct lu_env *env, 985 const struct cl_page_slice *slice); 986 /** 987 * Write out a page by kernel. This is only called by ll_writepage 988 * right now. 989 * 990 * \see cl_page_flush() 991 */ 992 int (*cpo_flush)(const struct lu_env *env, 993 const struct cl_page_slice *slice, 994 struct cl_io *io); 995 /** @} transfer */ 996}; 997 998/** 999 * Helper macro, dumping detailed information about \a page into a log. 1000 */
1001#define CL_PAGE_DEBUG(mask, env, page, format, ...) \ 1002do { \ 1003 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \ 1004 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \ 1005 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \ 1006 CDEBUG(mask, format, ## __VA_ARGS__); \ 1007 } \ 1008} while (0) 1009 1010/** 1011 * Helper macro, dumping shorter information about \a page into a log. 1012 */ 1013#define CL_PAGE_HEADER(mask, env, page, format, ...) \ 1014do { \ 1015 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \ 1016 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \ 1017 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \ 1018 CDEBUG(mask, format, ## __VA_ARGS__); \ 1019 } \ 1020} while (0) 1021 1022static inline struct page *cl_page_vmpage(struct cl_page *page) 1023{ 1024 LASSERT(page->cp_vmpage); 1025 return page->cp_vmpage; 1026} 1027 1028/** 1029 * Check if a cl_page is in use. 1030 * 1031 * Client cache holds a refcount, this refcount will be dropped when 1032 * the page is taken out of cache, see vvp_page_delete(). 1033 */ 1034static inline bool __page_in_use(const struct cl_page *page, int refc) 1035{ 1036 return (atomic_read(&page->cp_ref) > refc + 1); 1037} 1038 1039/** 1040 * Caller itself holds a refcount of cl_page. 1041 */ 1042#define cl_page_in_use(pg) __page_in_use(pg, 1) 1043/** 1044 * Caller doesn't hold a refcount. 1045 */ 1046#define cl_page_in_use_noref(pg) __page_in_use(pg, 0) 1047 1048/** @} cl_page */ 1049 1050/** \addtogroup cl_lock cl_lock 1051 * @{ 1052 */ 1053/** \struct cl_lock 1054 * 1055 * Extent locking on the client. 1056 * 1057 * LAYERING 1058 * 1059 * The locking model of the new client code is built around 1060 * 1061 * struct cl_lock 1062 * 1063 * data-type representing an extent lock on a regular file. cl_lock is a 1064 * layered object (much like cl_object and cl_page), it consists of a header 1065 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to 1066 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage. 1067 * 1068 * Typical cl_lock consists of the two layers: 1069 * 1070 * - vvp_lock (vvp specific data), and 1071 * - lov_lock (lov specific data). 1072 * 1073 * lov_lock contains an array of sub-locks. Each of these sub-locks is a 1074 * normal cl_lock: it has a header (struct cl_lock) and a list of layers: 1075 * 1076 * - lovsub_lock, and 1077 * - osc_lock 1078 * 1079 * Each sub-lock is associated with a cl_object (representing stripe 1080 * sub-object or the file to which top-level cl_lock is associated to), and is 1081 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to 1082 * cl_object (that at lov layer also fans out into multiple sub-objects), and 1083 * is different from cl_page, that doesn't fan out (there is usually exactly 1084 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock 1085 * a "top-lock" and its lovsub-osc portion a "sub-lock". 1086 * 1087 * LIFE CYCLE 1088 * 1089 * cl_lock is a cacheless data container for the requirements of locks to 1090 * complete the IO. cl_lock is created before I/O starts and destroyed when the 1091 * I/O is complete. 1092 * 1093 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached 1094 * to cl_lock at OSC layer. LDLM lock is still cacheable. 1095 * 1096 * INTERFACE AND USAGE 1097 * 1098 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A 1099 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue() 1100 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock 1101 * consists of multiple sub cl_locks, each sub locks will be enqueued 1102 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse 1103 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from 1104 * OST side. 1105 * 1106 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel() 1107 * method will be called for each layer to release the resource held by this 1108 * lock. At OSC layer, the reference count of LDLM lock, which is held at 1109 * clo_enqueue time, is released. 1110 * 1111 * LDLM lock can only be canceled if there is no cl_lock using it. 1112 * 1113 * Overall process of the locking during IO operation is as following: 1114 * 1115 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock() 1116 * is called on each layer. Responsibility of this method is to add locks, 1117 * needed by a given layer into cl_io.ci_lockset. 1118 * 1119 * - once locks for all layers were collected, they are sorted to avoid 1120 * dead-locks (cl_io_locks_sort()), and enqueued. 1121 * 1122 * - when all locks are acquired, IO is performed; 1123 * 1124 * - locks are released after IO is complete. 1125 * 1126 * Striping introduces major additional complexity into locking. The 1127 * fundamental problem is that it is generally unsafe to actively use (hold) 1128 * two locks on the different OST servers at the same time, as this introduces 1129 * inter-server dependency and can lead to cascading evictions. 1130 * 1131 * Basic solution is to sub-divide large read/write IOs into smaller pieces so 1132 * that no multi-stripe locks are taken (note that this design abandons POSIX 1133 * read/write semantics). Such pieces ideally can be executed concurrently. At 1134 * the same time, certain types of IO cannot be sub-divived, without 1135 * sacrificing correctness. This includes: 1136 * 1137 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee 1138 * atomicity; 1139 * 1140 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken. 1141 * 1142 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where 1143 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf 1144 * has to be held together with the usual lock on [offset, offset + count]. 1145 * 1146 * Interaction with DLM 1147 * 1148 * In the expected setup, cl_lock is ultimately backed up by a collection of 1149 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is 1150 * implemented in osc layer, that also matches DLM events (ASTs, cancellation, 1151 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed 1152 * description of interaction with DLM. 1153 */ 1154 1155/** 1156 * Lock description. 1157 */ 1158struct cl_lock_descr { 1159 /** Object this lock is granted for. */ 1160 struct cl_object *cld_obj; 1161 /** Index of the first page protected by this lock. */ 1162 pgoff_t cld_start; 1163 /** Index of the last page (inclusive) protected by this lock. */ 1164 pgoff_t cld_end; 1165 /** Group ID, for group lock */ 1166 __u64 cld_gid; 1167 /** Lock mode. */ 1168 enum cl_lock_mode cld_mode; 1169 /** 1170 * flags to enqueue lock. A combination of bit-flags from 1171 * enum cl_enq_flags. 1172 */ 1173 __u32 cld_enq_flags; 1174}; 1175 1176#define DDESCR "%s(%d):[%lu, %lu]:%x" 1177#define PDESCR(descr) \ 1178 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \ 1179 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags 1180 1181const char *cl_lock_mode_name(const enum cl_lock_mode mode); 1182 1183/** 1184 * Layered client lock. 1185 */ 1186struct cl_lock { 1187 /** List of slices. Immutable after creation. */ 1188 struct list_head cll_layers; 1189 /** lock attribute, extent, cl_object, etc. */ 1190 struct cl_lock_descr cll_descr; 1191}; 1192 1193/** 1194 * Per-layer part of cl_lock 1195 * 1196 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock 1197 */ 1198struct cl_lock_slice { 1199 struct cl_lock *cls_lock; 1200 /** Object slice corresponding to this lock slice. Immutable after 1201 * creation. 1202 */ 1203 struct cl_object *cls_obj; 1204 const struct cl_lock_operations *cls_ops; 1205 /** Linkage into cl_lock::cll_layers. Immutable after creation. */ 1206 struct list_head cls_linkage; 1207}; 1208 1209/** 1210 * 1211 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops 1212 */ 1213struct cl_lock_operations { 1214 /** @{ */ 1215 /** 1216 * Attempts to enqueue the lock. Called top-to-bottom. 1217 * 1218 * \retval 0 this layer has enqueued the lock successfully 1219 * \retval >0 this layer has enqueued the lock, but need to wait on 1220 * @anchor for resources 1221 * \retval -ve failure 1222 * 1223 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(), 1224 * \see osc_lock_enqueue() 1225 */ 1226 int (*clo_enqueue)(const struct lu_env *env, 1227 const struct cl_lock_slice *slice, 1228 struct cl_io *io, struct cl_sync_io *anchor); 1229 /** 1230 * Cancel a lock, release its DLM lock ref, while does not cancel the 1231 * DLM lock 1232 */ 1233 void (*clo_cancel)(const struct lu_env *env, 1234 const struct cl_lock_slice *slice); 1235 /** @} */ 1236 /** 1237 * Destructor. Frees resources and the slice. 1238 * 1239 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(), 1240 * \see osc_lock_fini() 1241 */ 1242 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice); 1243 /** 1244 * Optional debugging helper. Prints given lock slice. 1245 */ 1246 int (*clo_print)(const struct lu_env *env, 1247 void *cookie, lu_printer_t p, 1248 const struct cl_lock_slice *slice); 1249}; 1250 1251#define CL_LOCK_DEBUG(mask, env, lock, format, ...) \ 1252do { \ 1253 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \ 1254 \ 1255 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \ 1256 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \ 1257 CDEBUG(mask, format, ## __VA_ARGS__); \ 1258 } \ 1259} while (0) 1260 1261#define CL_LOCK_ASSERT(expr, env, lock) do { \ 1262 if (likely(expr)) \ 1263 break; \ 1264 \ 1265 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \ 1266 LBUG(); \ 1267} while (0) 1268 1269/** @} cl_lock */ 1270 1271/** \addtogroup cl_page_list cl_page_list 1272 * Page list used to perform collective operations on a group of pages. 1273 * 1274 * Pages are added to the list one by one. cl_page_list acquires a reference 1275 * for every page in it. Page list is used to perform collective operations on 1276 * pages: 1277 * 1278 * - submit pages for an immediate transfer, 1279 * 1280 * - own pages on behalf of certain io (waiting for each page in turn), 1281 * 1282 * - discard pages. 1283 * 1284 * When list is finalized, it releases references on all pages it still has. 1285 * 1286 * \todo XXX concurrency control. 1287 * 1288 * @{ 1289 */ 1290struct cl_page_list { 1291 unsigned int pl_nr; 1292 struct list_head pl_pages; 1293 struct task_struct *pl_owner; 1294}; 1295 1296/** 1297 * A 2-queue of pages. A convenience data-type for common use case, 2-queue 1298 * contains an incoming page list and an outgoing page list. 1299 */ 1300struct cl_2queue { 1301 struct cl_page_list c2_qin; 1302 struct cl_page_list c2_qout; 1303}; 1304 1305/** @} cl_page_list */ 1306 1307/** \addtogroup cl_io cl_io 1308 * @{ 1309 */ 1310/** \struct cl_io 1311 * I/O 1312 * 1313 * cl_io represents a high level I/O activity like 1314 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent 1315 * lock. 1316 * 1317 * cl_io is a layered object, much like cl_{object,page,lock} but with one 1318 * important distinction. We want to minimize number of calls to the allocator 1319 * in the fast path, e.g., in the case of read(2) when everything is cached: 1320 * client already owns the lock over region being read, and data are cached 1321 * due to read-ahead. To avoid allocation of cl_io layers in such situations, 1322 * per-layer io state is stored in the session, associated with the io, see 1323 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized 1324 * by using free-lists, see cl_env_get(). 1325 * 1326 * There is a small predefined number of possible io types, enumerated in enum 1327 * cl_io_type. 1328 * 1329 * cl_io is a state machine, that can be advanced concurrently by the multiple 1330 * threads. It is up to these threads to control the concurrency and, 1331 * specifically, to detect when io is done, and its state can be safely 1332 * released. 1333 * 1334 * For read/write io overall execution plan is as following: 1335 * 1336 * (0) initialize io state through all layers; 1337 * 1338 * (1) loop: prepare chunk of work to do 1339 * 1340 * (2) call all layers to collect locks they need to process current chunk 1341 * 1342 * (3) sort all locks to avoid dead-locks, and acquire them 1343 * 1344 * (4) process the chunk: call per-page methods 1345 * cl_io_operations::cio_prepare_write(), 1346 * cl_io_operations::cio_commit_write() for write) 1347 * 1348 * (5) release locks 1349 * 1350 * (6) repeat loop. 1351 * 1352 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to 1353 * address allocation efficiency issues mentioned above), and returns with the 1354 * special error condition from per-page method when current sub-io has to 1355 * block. This causes io loop to be repeated, and lov switches to the next 1356 * sub-io in its cl_io_operations::cio_iter_init() implementation. 1357 */ 1358 1359/** IO types */ 1360enum cl_io_type { 1361 /** read system call */ 1362 CIT_READ = 1, 1363 /** write system call */ 1364 CIT_WRITE, 1365 /** truncate, utime system calls */ 1366 CIT_SETATTR, 1367 /** get data version */ 1368 CIT_DATA_VERSION, 1369 /** 1370 * page fault handling 1371 */ 1372 CIT_FAULT, 1373 /** 1374 * fsync system call handling 1375 * To write out a range of file 1376 */ 1377 CIT_FSYNC, 1378 /** 1379 * Miscellaneous io. This is used for occasional io activity that 1380 * doesn't fit into other types. Currently this is used for: 1381 * 1382 * - cancellation of an extent lock. This io exists as a context 1383 * to write dirty pages from under the lock being canceled back 1384 * to the server; 1385 * 1386 * - VM induced page write-out. An io context for writing page out 1387 * for memory cleansing; 1388 * 1389 * - glimpse. An io context to acquire glimpse lock. 1390 * 1391 * - grouplock. An io context to acquire group lock. 1392 * 1393 * CIT_MISC io is used simply as a context in which locks and pages 1394 * are manipulated. Such io has no internal "process", that is, 1395 * cl_io_loop() is never called for it. 1396 */ 1397 CIT_MISC, 1398 CIT_OP_NR 1399}; 1400 1401/** 1402 * States of cl_io state machine 1403 */ 1404enum cl_io_state { 1405 /** Not initialized. */ 1406 CIS_ZERO, 1407 /** Initialized. */ 1408 CIS_INIT, 1409 /** IO iteration started. */ 1410 CIS_IT_STARTED, 1411 /** Locks taken. */ 1412 CIS_LOCKED, 1413 /** Actual IO is in progress. */ 1414 CIS_IO_GOING, 1415 /** IO for the current iteration finished. */ 1416 CIS_IO_FINISHED, 1417 /** Locks released. */ 1418 CIS_UNLOCKED, 1419 /** Iteration completed. */ 1420 CIS_IT_ENDED, 1421 /** cl_io finalized. */ 1422 CIS_FINI 1423}; 1424 1425/** 1426 * IO state private for a layer. 1427 * 1428 * This is usually embedded into layer session data, rather than allocated 1429 * dynamically. 1430 * 1431 * \see vvp_io, lov_io, osc_io 1432 */ 1433struct cl_io_slice { 1434 struct cl_io *cis_io; 1435 /** corresponding object slice. Immutable after creation. */ 1436 struct cl_object *cis_obj; 1437 /** io operations. Immutable after creation. */ 1438 const struct cl_io_operations *cis_iop; 1439 /** 1440 * linkage into a list of all slices for a given cl_io, hanging off 1441 * cl_io::ci_layers. Immutable after creation. 1442 */ 1443 struct list_head cis_linkage; 1444}; 1445 1446typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *, 1447 struct cl_page *); 1448 1449struct cl_read_ahead { 1450 /* 1451 * Maximum page index the readahead window will end. 1452 * This is determined DLM lock coverage, RPC and stripe boundary. 1453 * cra_end is included. 1454 */ 1455 pgoff_t cra_end; 1456 /* optimal RPC size for this read, by pages */ 1457 unsigned long cra_rpc_size; 1458 /* 1459 * Release callback. If readahead holds resources underneath, this 1460 * function should be called to release it. 1461 */ 1462 void (*cra_release)(const struct lu_env *env, void *cbdata); 1463 /* Callback data for cra_release routine */ 1464 void *cra_cbdata; 1465}; 1466 1467static inline void cl_read_ahead_release(const struct lu_env *env, 1468 struct cl_read_ahead *ra) 1469{ 1470 if (ra->cra_release) 1471 ra->cra_release(env, ra->cra_cbdata); 1472 memset(ra, 0, sizeof(*ra)); 1473} 1474 1475/** 1476 * Per-layer io operations. 1477 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops 1478 */ 1479struct cl_io_operations { 1480 /** 1481 * Vector of io state transition methods for every io type. 1482 * 1483 * \see cl_page_operations::io 1484 */ 1485 struct { 1486 /** 1487 * Prepare io iteration at a given layer. 1488 * 1489 * Called top-to-bottom at the beginning of each iteration of 1490 * "io loop" (if it makes sense for this type of io). Here 1491 * layer selects what work it will do during this iteration. 1492 * 1493 * \see cl_io_operations::cio_iter_fini() 1494 */ 1495 int (*cio_iter_init)(const struct lu_env *env, 1496 const struct cl_io_slice *slice); 1497 /** 1498 * Finalize io iteration. 1499 * 1500 * Called bottom-to-top at the end of each iteration of "io 1501 * loop". Here layers can decide whether IO has to be 1502 * continued. 1503 * 1504 * \see cl_io_operations::cio_iter_init() 1505 */ 1506 void (*cio_iter_fini)(const struct lu_env *env, 1507 const struct cl_io_slice *slice); 1508 /** 1509 * Collect locks for the current iteration of io. 1510 * 1511 * Called top-to-bottom to collect all locks necessary for 1512 * this iteration. This methods shouldn't actually enqueue 1513 * anything, instead it should post a lock through 1514 * cl_io_lock_add(). Once all locks are collected, they are 1515 * sorted and enqueued in the proper order. 1516 */ 1517 int (*cio_lock)(const struct lu_env *env, 1518 const struct cl_io_slice *slice); 1519 /** 1520 * Finalize unlocking. 1521 * 1522 * Called bottom-to-top to finish layer specific unlocking 1523 * functionality, after generic code released all locks 1524 * acquired by cl_io_operations::cio_lock(). 1525 */ 1526 void (*cio_unlock)(const struct lu_env *env, 1527 const struct cl_io_slice *slice); 1528 /** 1529 * Start io iteration. 1530 * 1531 * Once all locks are acquired, called top-to-bottom to 1532 * commence actual IO. In the current implementation, 1533 * top-level vvp_io_{read,write}_start() does all the work 1534 * synchronously by calling generic_file_*(), so other layers 1535 * are called when everything is done. 1536 */ 1537 int (*cio_start)(const struct lu_env *env, 1538 const struct cl_io_slice *slice); 1539 /** 1540 * Called top-to-bottom at the end of io loop. Here layer 1541 * might wait for an unfinished asynchronous io. 1542 */ 1543 void (*cio_end)(const struct lu_env *env, 1544 const struct cl_io_slice *slice); 1545 /** 1546 * Called bottom-to-top to notify layers that read/write IO 1547 * iteration finished, with \a nob bytes transferred. 1548 */ 1549 void (*cio_advance)(const struct lu_env *env, 1550 const struct cl_io_slice *slice, 1551 size_t nob); 1552 /** 1553 * Called once per io, bottom-to-top to release io resources. 1554 */ 1555 void (*cio_fini)(const struct lu_env *env, 1556 const struct cl_io_slice *slice); 1557 } op[CIT_OP_NR]; 1558 1559 /** 1560 * Submit pages from \a queue->c2_qin for IO, and move 1561 * successfully submitted pages into \a queue->c2_qout. Return 1562 * non-zero if failed to submit even the single page. If 1563 * submission failed after some pages were moved into \a 1564 * queue->c2_qout, completion callback with non-zero ioret is 1565 * executed on them. 1566 */ 1567 int (*cio_submit)(const struct lu_env *env, 1568 const struct cl_io_slice *slice, 1569 enum cl_req_type crt, 1570 struct cl_2queue *queue); 1571 /** 1572 * Queue async page for write. 1573 * The difference between cio_submit and cio_queue is that 1574 * cio_submit is for urgent request. 1575 */ 1576 int (*cio_commit_async)(const struct lu_env *env, 1577 const struct cl_io_slice *slice, 1578 struct cl_page_list *queue, int from, int to, 1579 cl_commit_cbt cb); 1580 /** 1581 * Decide maximum read ahead extent 1582 * 1583 * \pre io->ci_type == CIT_READ 1584 */ 1585 int (*cio_read_ahead)(const struct lu_env *env, 1586 const struct cl_io_slice *slice, 1587 pgoff_t start, struct cl_read_ahead *ra); 1588 /** 1589 * Optional debugging helper. Print given io slice. 1590 */ 1591 int (*cio_print)(const struct lu_env *env, void *cookie, 1592 lu_printer_t p, const struct cl_io_slice *slice); 1593}; 1594 1595/** 1596 * Flags to lock enqueue procedure. 1597 * \ingroup cl_lock 1598 */ 1599enum cl_enq_flags { 1600 /** 1601 * instruct server to not block, if conflicting lock is found. Instead 1602 * -EWOULDBLOCK is returned immediately. 1603 */ 1604 CEF_NONBLOCK = 0x00000001, 1605 /** 1606 * take lock asynchronously (out of order), as it cannot 1607 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing. 1608 */ 1609 CEF_ASYNC = 0x00000002, 1610 /** 1611 * tell the server to instruct (though a flag in the blocking ast) an 1612 * owner of the conflicting lock, that it can drop dirty pages 1613 * protected by this lock, without sending them to the server. 1614 */ 1615 CEF_DISCARD_DATA = 0x00000004, 1616 /** 1617 * tell the sub layers that it must be a `real' lock. This is used for 1618 * mmapped-buffer locks and glimpse locks that must be never converted 1619 * into lockless mode. 1620 * 1621 * \see vvp_mmap_locks(), cl_glimpse_lock(). 1622 */ 1623 CEF_MUST = 0x00000008, 1624 /** 1625 * tell the sub layers that never request a `real' lock. This flag is 1626 * not used currently. 1627 * 1628 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless 1629 * conversion policy: ci_lockreq describes generic information of lock 1630 * requirement for this IO, especially for locks which belong to the 1631 * object doing IO; however, lock itself may have precise requirements 1632 * that are described by the enqueue flags. 1633 */ 1634 CEF_NEVER = 0x00000010, 1635 /** 1636 * for async glimpse lock. 1637 */ 1638 CEF_AGL = 0x00000020, 1639 /** 1640 * enqueue a lock to test DLM lock existence. 1641 */ 1642 CEF_PEEK = 0x00000040, 1643 /** 1644 * Lock match only. Used by group lock in I/O as group lock 1645 * is known to exist. 1646 */ 1647 CEF_LOCK_MATCH = BIT(7), 1648 /** 1649 * mask of enq_flags. 1650 */ 1651 CEF_MASK = 0x000000ff, 1652}; 1653 1654/** 1655 * Link between lock and io. Intermediate structure is needed, because the 1656 * same lock can be part of multiple io's simultaneously. 1657 */ 1658struct cl_io_lock_link { 1659 /** linkage into one of cl_lockset lists. */ 1660 struct list_head cill_linkage; 1661 struct cl_lock cill_lock; 1662 /** optional destructor */ 1663 void (*cill_fini)(const struct lu_env *env, 1664 struct cl_io_lock_link *link); 1665}; 1666#define cill_descr cill_lock.cll_descr 1667 1668/** 1669 * Lock-set represents a collection of locks, that io needs at a 1670 * time. Generally speaking, client tries to avoid holding multiple locks when 1671 * possible, because 1672 * 1673 * - holding extent locks over multiple ost's introduces the danger of 1674 * "cascading timeouts"; 1675 * 1676 * - holding multiple locks over the same ost is still dead-lock prone, 1677 * see comment in osc_lock_enqueue(), 1678 * 1679 * but there are certain situations where this is unavoidable: 1680 * 1681 * - O_APPEND writes have to take [0, EOF] lock for correctness; 1682 * 1683 * - truncate has to take [new-size, EOF] lock for correctness; 1684 * 1685 * - SNS has to take locks across full stripe for correctness; 1686 * 1687 * - in the case when user level buffer, supplied to {read,write}(file0), 1688 * is a part of a memory mapped lustre file, client has to take a dlm 1689 * locks on file0, and all files that back up the buffer (or a part of 1690 * the buffer, that is being processed in the current chunk, in any 1691 * case, there are situations where at least 2 locks are necessary). 1692 * 1693 * In such cases we at least try to take locks in the same consistent 1694 * order. To this end, all locks are first collected, then sorted, and then 1695 * enqueued. 1696 */ 1697struct cl_lockset { 1698 /** locks to be acquired. */ 1699 struct list_head cls_todo; 1700 /** locks acquired. */ 1701 struct list_head cls_done; 1702}; 1703 1704/** 1705 * Lock requirements(demand) for IO. It should be cl_io_lock_req, 1706 * but 'req' is always to be thought as 'request' :-) 1707 */ 1708enum cl_io_lock_dmd { 1709 /** Always lock data (e.g., O_APPEND). */ 1710 CILR_MANDATORY = 0, 1711 /** Layers are free to decide between local and global locking. */ 1712 CILR_MAYBE, 1713 /** Never lock: there is no cache (e.g., lockless IO). */ 1714 CILR_NEVER 1715}; 1716 1717enum cl_fsync_mode { 1718 /** start writeback, do not wait for them to finish */ 1719 CL_FSYNC_NONE = 0, 1720 /** start writeback and wait for them to finish */ 1721 CL_FSYNC_LOCAL = 1, 1722 /** discard all of dirty pages in a specific file range */ 1723 CL_FSYNC_DISCARD = 2, 1724 /** start writeback and make sure they have reached storage before 1725 * return. OST_SYNC RPC must be issued and finished 1726 */ 1727 CL_FSYNC_ALL = 3 1728}; 1729 1730struct cl_io_rw_common { 1731 loff_t crw_pos; 1732 size_t crw_count; 1733 int crw_nonblock; 1734}; 1735 1736/** 1737 * State for io. 1738 * 1739 * cl_io is shared by all threads participating in this IO (in current 1740 * implementation only one thread advances IO, but parallel IO design and 1741 * concurrent copy_*_user() require multiple threads acting on the same IO. It 1742 * is up to these threads to serialize their activities, including updates to 1743 * mutable cl_io fields. 1744 */ 1745struct cl_io { 1746 /** type of this IO. Immutable after creation. */ 1747 enum cl_io_type ci_type; 1748 /** current state of cl_io state machine. */ 1749 enum cl_io_state ci_state; 1750 /** main object this io is against. Immutable after creation. */ 1751 struct cl_object *ci_obj; 1752 /** 1753 * Upper layer io, of which this io is a part of. Immutable after 1754 * creation. 1755 */ 1756 struct cl_io *ci_parent; 1757 /** List of slices. Immutable after creation. */ 1758 struct list_head ci_layers; 1759 /** list of locks (to be) acquired by this io. */ 1760 struct cl_lockset ci_lockset; 1761 /** lock requirements, this is just a help info for sublayers. */ 1762 enum cl_io_lock_dmd ci_lockreq; 1763 union { 1764 struct cl_rd_io { 1765 struct cl_io_rw_common rd; 1766 } ci_rd; 1767 struct cl_wr_io { 1768 struct cl_io_rw_common wr; 1769 int wr_append; 1770 int wr_sync; 1771 } ci_wr; 1772 struct cl_io_rw_common ci_rw; 1773 struct cl_setattr_io { 1774 struct ost_lvb sa_attr; 1775 unsigned int sa_attr_flags; 1776 unsigned int sa_valid; 1777 int sa_stripe_index; 1778 const struct lu_fid *sa_parent_fid; 1779 } ci_setattr; 1780 struct cl_data_version_io { 1781 u64 dv_data_version; 1782 int dv_flags; 1783 } ci_data_version; 1784 struct cl_fault_io { 1785 /** page index within file. */ 1786 pgoff_t ft_index; 1787 /** bytes valid byte on a faulted page. */ 1788 size_t ft_nob; 1789 /** writable page? for nopage() only */ 1790 int ft_writable; 1791 /** page of an executable? */ 1792 int ft_executable; 1793 /** page_mkwrite() */ 1794 int ft_mkwrite; 1795 /** resulting page */ 1796 struct cl_page *ft_page; 1797 } ci_fault; 1798 struct cl_fsync_io { 1799 loff_t fi_start; 1800 loff_t fi_end; 1801 /** file system level fid */ 1802 struct lu_fid *fi_fid; 1803 enum cl_fsync_mode fi_mode; 1804 /* how many pages were written/discarded */ 1805 unsigned int fi_nr_written; 1806 } ci_fsync; 1807 } u; 1808 struct cl_2queue ci_queue; 1809 size_t ci_nob; 1810 int ci_result; 1811 unsigned int ci_continue:1, 1812 /** 1813 * This io has held grouplock, to inform sublayers that 1814 * don't do lockless i/o. 1815 */ 1816 ci_no_srvlock:1, 1817 /** 1818 * The whole IO need to be restarted because layout has been changed 1819 */ 1820 ci_need_restart:1, 1821 /** 1822 * to not refresh layout - the IO issuer knows that the layout won't 1823 * change(page operations, layout change causes all page to be 1824 * discarded), or it doesn't matter if it changes(sync). 1825 */ 1826 ci_ignore_layout:1, 1827 /** 1828 * Check if layout changed after the IO finishes. Mainly for HSM 1829 * requirement. If IO occurs to openning files, it doesn't need to 1830 * verify layout because HSM won't release openning files. 1831 * Right now, only two operations need to verify layout: glimpse 1832 * and setattr. 1833 */ 1834 ci_verify_layout:1, 1835 /** 1836 * file is released, restore has to to be triggered by vvp layer 1837 */ 1838 ci_restore_needed:1, 1839 /** 1840 * O_NOATIME 1841 */ 1842 ci_noatime:1; 1843 /** 1844 * Number of pages owned by this IO. For invariant checking. 1845 */ 1846 unsigned int ci_owned_nr; 1847}; 1848 1849/** @} cl_io */ 1850 1851/** 1852 * Per-transfer attributes. 1853 */ 1854struct cl_req_attr { 1855 enum cl_req_type cra_type; 1856 u64 cra_flags; 1857 struct cl_page *cra_page; 1858 1859 /** Generic attributes for the server consumption. */ 1860 struct obdo *cra_oa; 1861 /** Jobid */ 1862 char cra_jobid[LUSTRE_JOBID_SIZE]; 1863}; 1864 1865enum cache_stats_item { 1866 /** how many cache lookups were performed */ 1867 CS_lookup = 0, 1868 /** how many times cache lookup resulted in a hit */ 1869 CS_hit, 1870 /** how many entities are in the cache right now */ 1871 CS_total, 1872 /** how many entities in the cache are actively used (and cannot be 1873 * evicted) right now 1874 */ 1875 CS_busy, 1876 /** how many entities were created at all */ 1877 CS_create, 1878 CS_NR 1879}; 1880 1881#define CS_NAMES { "lookup", "hit", "total", "busy", "create" } 1882 1883/** 1884 * Stats for a generic cache (similar to inode, lu_object, etc. caches). 1885 */ 1886struct cache_stats { 1887 const char *cs_name; 1888 atomic_t cs_stats[CS_NR]; 1889}; 1890 1891/** These are not exported so far */ 1892void cache_stats_init(struct cache_stats *cs, const char *name); 1893 1894/** 1895 * Client-side site. This represents particular client stack. "Global" 1896 * variables should (directly or indirectly) be added here to allow multiple 1897 * clients to co-exist in the single address space. 1898 */ 1899struct cl_site { 1900 struct lu_site cs_lu; 1901 /** 1902 * Statistical counters. Atomics do not scale, something better like 1903 * per-cpu counters is needed. 1904 * 1905 * These are exported as /sys/kernel/debug/lustre/llite/.../site 1906 * 1907 * When interpreting keep in mind that both sub-locks (and sub-pages) 1908 * and top-locks (and top-pages) are accounted here. 1909 */ 1910 struct cache_stats cs_pages; 1911 atomic_t cs_pages_state[CPS_NR]; 1912}; 1913 1914int cl_site_init(struct cl_site *s, struct cl_device *top); 1915void cl_site_fini(struct cl_site *s); 1916void cl_stack_fini(const struct lu_env *env, struct cl_device *cl); 1917 1918/** 1919 * Output client site statistical counters into a buffer. Suitable for 1920 * ll_rd_*()-style functions. 1921 */ 1922int cl_site_stats_print(const struct cl_site *site, struct seq_file *m); 1923 1924/** 1925 * \name helpers 1926 * 1927 * Type conversion and accessory functions. 1928 */ 1929/** @{ */ 1930 1931static inline struct cl_site *lu2cl_site(const struct lu_site *site) 1932{ 1933 return container_of(site, struct cl_site, cs_lu); 1934} 1935 1936static inline int lu_device_is_cl(const struct lu_device *d) 1937{ 1938 return d->ld_type->ldt_tags & LU_DEVICE_CL; 1939} 1940 1941static inline struct cl_device *lu2cl_dev(const struct lu_device *d) 1942{ 1943 LASSERT(!d || IS_ERR(d) || lu_device_is_cl(d)); 1944 return container_of0(d, struct cl_device, cd_lu_dev); 1945} 1946 1947static inline struct lu_device *cl2lu_dev(struct cl_device *d) 1948{ 1949 return &d->cd_lu_dev; 1950} 1951 1952static inline struct cl_object *lu2cl(const struct lu_object *o) 1953{ 1954 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->lo_dev)); 1955 return container_of0(o, struct cl_object, co_lu); 1956} 1957 1958static inline const struct cl_object_conf * 1959lu2cl_conf(const struct lu_object_conf *conf) 1960{ 1961 return container_of0(conf, struct cl_object_conf, coc_lu); 1962} 1963 1964static inline struct cl_object *cl_object_next(const struct cl_object *obj) 1965{ 1966 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL; 1967} 1968 1969static inline struct cl_device *cl_object_device(const struct cl_object *o) 1970{ 1971 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev)); 1972 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev); 1973} 1974 1975static inline struct cl_object_header *luh2coh(const struct lu_object_header *h) 1976{ 1977 return container_of0(h, struct cl_object_header, coh_lu); 1978} 1979 1980static inline struct cl_site *cl_object_site(const struct cl_object *obj) 1981{ 1982 return lu2cl_site(obj->co_lu.lo_dev->ld_site); 1983} 1984 1985static inline 1986struct cl_object_header *cl_object_header(const struct cl_object *obj) 1987{ 1988 return luh2coh(obj->co_lu.lo_header); 1989} 1990 1991static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t) 1992{ 1993 return lu_device_init(&d->cd_lu_dev, t); 1994} 1995 1996static inline void cl_device_fini(struct cl_device *d) 1997{ 1998 lu_device_fini(&d->cd_lu_dev); 1999} 2000
2001void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice, 2002 struct cl_object *obj, pgoff_t index, 2003 const struct cl_page_operations *ops); 2004void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice, 2005 struct cl_object *obj, 2006 const struct cl_lock_operations *ops); 2007void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice, 2008 struct cl_object *obj, const struct cl_io_operations *ops); 2009/** @} helpers */ 2010 2011/** \defgroup cl_object cl_object 2012 * @{ 2013 */ 2014struct cl_object *cl_object_top(struct cl_object *o); 2015struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd, 2016 const struct lu_fid *fid, 2017 const struct cl_object_conf *c); 2018 2019int cl_object_header_init(struct cl_object_header *h); 2020void cl_object_put(const struct lu_env *env, struct cl_object *o); 2021void cl_object_get(struct cl_object *o); 2022void cl_object_attr_lock(struct cl_object *o); 2023void cl_object_attr_unlock(struct cl_object *o); 2024int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj, 2025 struct cl_attr *attr); 2026int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj, 2027 const struct cl_attr *attr, unsigned int valid); 2028int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj, 2029 struct ost_lvb *lvb); 2030int cl_conf_set(const struct lu_env *env, struct cl_object *obj, 2031 const struct cl_object_conf *conf); 2032int cl_object_prune(const struct lu_env *env, struct cl_object *obj); 2033void cl_object_kill(const struct lu_env *env, struct cl_object *obj); 2034int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj, 2035 struct lov_user_md __user *lum); 2036int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj, 2037 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap, 2038 size_t *buflen); 2039int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj, 2040 struct cl_layout *cl); 2041loff_t cl_object_maxbytes(struct cl_object *obj); 2042 2043/** 2044 * Returns true, iff \a o0 and \a o1 are slices of the same object. 2045 */ 2046static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1) 2047{ 2048 return cl_object_header(o0) == cl_object_header(o1); 2049} 2050 2051static inline void cl_object_page_init(struct cl_object *clob, int size) 2052{ 2053 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize; 2054 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size); 2055 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512); 2056} 2057 2058static inline void *cl_object_page_slice(struct cl_object *clob, 2059 struct cl_page *page) 2060{ 2061 return (void *)((char *)page + clob->co_slice_off); 2062} 2063 2064/** 2065 * Return refcount of cl_object. 2066 */ 2067static inline int cl_object_refc(struct cl_object *clob) 2068{ 2069 struct lu_object_header *header = clob->co_lu.lo_header; 2070 2071 return atomic_read(&header->loh_ref); 2072} 2073 2074/** @} cl_object */ 2075 2076/** \defgroup cl_page cl_page 2077 * @{ 2078 */ 2079enum { 2080 CLP_GANG_OKAY = 0, 2081 CLP_GANG_RESCHED, 2082 CLP_GANG_AGAIN, 2083 CLP_GANG_ABORT 2084}; 2085 2086/* callback of cl_page_gang_lookup() */ 2087struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *obj, 2088 pgoff_t idx, struct page *vmpage, 2089 enum cl_page_type type); 2090struct cl_page *cl_page_alloc(const struct lu_env *env, 2091 struct cl_object *o, pgoff_t ind, 2092 struct page *vmpage, 2093 enum cl_page_type type); 2094void cl_page_get(struct cl_page *page); 2095void cl_page_put(const struct lu_env *env, struct cl_page *page); 2096void cl_page_print(const struct lu_env *env, void *cookie, lu_printer_t printer, 2097 const struct cl_page *pg); 2098void cl_page_header_print(const struct lu_env *env, void *cookie, 2099 lu_printer_t printer, const struct cl_page *pg); 2100struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj); 2101 2102const struct cl_page_slice *cl_page_at(const struct cl_page *page, 2103 const struct lu_device_type *dtype); 2104 2105/** 2106 * \name ownership 2107 * 2108 * Functions dealing with the ownership of page by io. 2109 */ 2110/** @{ */ 2111 2112int cl_page_own(const struct lu_env *env, 2113 struct cl_io *io, struct cl_page *page); 2114int cl_page_own_try(const struct lu_env *env, 2115 struct cl_io *io, struct cl_page *page); 2116void cl_page_assume(const struct lu_env *env, 2117 struct cl_io *io, struct cl_page *page); 2118void cl_page_unassume(const struct lu_env *env, 2119 struct cl_io *io, struct cl_page *pg); 2120void cl_page_disown(const struct lu_env *env, 2121 struct cl_io *io, struct cl_page *page); 2122void cl_page_disown0(const struct lu_env *env, 2123 struct cl_io *io, struct cl_page *pg); 2124int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io); 2125 2126/** @} ownership */ 2127 2128/** 2129 * \name transfer 2130 * 2131 * Functions dealing with the preparation of a page for a transfer, and 2132 * tracking transfer state. 2133 */ 2134/** @{ */ 2135int cl_page_prep(const struct lu_env *env, struct cl_io *io, 2136 struct cl_page *pg, enum cl_req_type crt); 2137void cl_page_completion(const struct lu_env *env, 2138 struct cl_page *pg, enum cl_req_type crt, int ioret); 2139int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg, 2140 enum cl_req_type crt); 2141int cl_page_cache_add(const struct lu_env *env, struct cl_io *io, 2142 struct cl_page *pg, enum cl_req_type crt); 2143void cl_page_clip(const struct lu_env *env, struct cl_page *pg, 2144 int from, int to); 2145int cl_page_cancel(const struct lu_env *env, struct cl_page *page); 2146int cl_page_flush(const struct lu_env *env, struct cl_io *io, 2147 struct cl_page *pg); 2148 2149/** @} transfer */ 2150 2151/** 2152 * \name helper routines 2153 * Functions to discard, delete and export a cl_page. 2154 */ 2155/** @{ */ 2156void cl_page_discard(const struct lu_env *env, struct cl_io *io, 2157 struct cl_page *pg); 2158void cl_page_delete(const struct lu_env *env, struct cl_page *pg); 2159int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg); 2160void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate); 2161loff_t cl_offset(const struct cl_object *obj, pgoff_t idx); 2162pgoff_t cl_index(const struct cl_object *obj, loff_t offset); 2163size_t cl_page_size(const struct cl_object *obj); 2164int cl_pages_prune(const struct lu_env *env, struct cl_object *obj); 2165 2166void cl_lock_print(const struct lu_env *env, void *cookie, 2167 lu_printer_t printer, const struct cl_lock *lock); 2168void cl_lock_descr_print(const struct lu_env *env, void *cookie, 2169 lu_printer_t printer, 2170 const struct cl_lock_descr *descr); 2171/* @} helper */ 2172 2173/** 2174 * Data structure managing a client's cached pages. A count of 2175 * "unstable" pages is maintained, and an LRU of clean pages is 2176 * maintained. "unstable" pages are pages pinned by the ptlrpc 2177 * layer for recovery purposes. 2178 */ 2179struct cl_client_cache { 2180 /** 2181 * # of client cache refcount 2182 * # of users (OSCs) + 2 (held by llite and lov) 2183 */ 2184 atomic_t ccc_users; 2185 /** 2186 * # of threads are doing shrinking 2187 */ 2188 unsigned int ccc_lru_shrinkers; 2189 /** 2190 * # of LRU entries available 2191 */ 2192 atomic_long_t ccc_lru_left; 2193 /** 2194 * List of entities(OSCs) for this LRU cache 2195 */ 2196 struct list_head ccc_lru; 2197 /** 2198 * Max # of LRU entries 2199 */ 2200 unsigned long ccc_lru_max; 2201 /** 2202 * Lock to protect ccc_lru list 2203 */ 2204 spinlock_t ccc_lru_lock; 2205 /** 2206 * Set if unstable check is enabled 2207 */ 2208 unsigned int ccc_unstable_check:1; 2209 /** 2210 * # of unstable pages for this mount point 2211 */ 2212 atomic_long_t ccc_unstable_nr; 2213 /** 2214 * Waitq for awaiting unstable pages to reach zero. 2215 * Used at umounting time and signaled on BRW commit 2216 */ 2217 wait_queue_head_t ccc_unstable_waitq; 2218 2219}; 2220 2221/** 2222 * cl_cache functions 2223 */ 2224struct cl_client_cache *cl_cache_init(unsigned long lru_page_max); 2225void cl_cache_incref(struct cl_client_cache *cache); 2226void cl_cache_decref(struct cl_client_cache *cache); 2227 2228/** @} cl_page */ 2229 2230/** \defgroup cl_lock cl_lock 2231 * @{ 2232 */ 2233 2234int cl_lock_request(const struct lu_env *env, struct cl_io *io, 2235 struct cl_lock *lock); 2236int cl_lock_init(const struct lu_env *env, struct cl_lock *lock, 2237 const struct cl_io *io); 2238void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock); 2239const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock, 2240 const struct lu_device_type *dtype); 2241void cl_lock_release(const struct lu_env *env, struct cl_lock *lock); 2242int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io, 2243 struct cl_lock *lock, struct cl_sync_io *anchor); 2244void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock); 2245 2246/** @} cl_lock */ 2247 2248/** \defgroup cl_io cl_io 2249 * @{ 2250 */ 2251 2252int cl_io_init(const struct lu_env *env, struct cl_io *io, 2253 enum cl_io_type iot, struct cl_object *obj); 2254int cl_io_sub_init(const struct lu_env *env, struct cl_io *io, 2255 enum cl_io_type iot, struct cl_object *obj); 2256int cl_io_rw_init(const struct lu_env *env, struct cl_io *io, 2257 enum cl_io_type iot, loff_t pos, size_t count); 2258int cl_io_loop(const struct lu_env *env, struct cl_io *io); 2259 2260void cl_io_fini(const struct lu_env *env, struct cl_io *io); 2261int cl_io_iter_init(const struct lu_env *env, struct cl_io *io); 2262void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io); 2263int cl_io_lock(const struct lu_env *env, struct cl_io *io); 2264void cl_io_unlock(const struct lu_env *env, struct cl_io *io); 2265int cl_io_start(const struct lu_env *env, struct cl_io *io); 2266void cl_io_end(const struct lu_env *env, struct cl_io *io); 2267int cl_io_lock_add(const struct lu_env *env, struct cl_io *io, 2268 struct cl_io_lock_link *link); 2269int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io, 2270 struct cl_lock_descr *descr); 2271int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io, 2272 enum cl_req_type iot, struct cl_2queue *queue); 2273int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io, 2274 enum cl_req_type iot, struct cl_2queue *queue, 2275 long timeout); 2276int cl_io_commit_async(const struct lu_env *env, struct cl_io *io, 2277 struct cl_page_list *queue, int from, int to, 2278 cl_commit_cbt cb); 2279int cl_io_read_ahead(const struct lu_env *env, struct cl_io *io, 2280 pgoff_t start, struct cl_read_ahead *ra); 2281int cl_io_is_going(const struct lu_env *env); 2282 2283/** 2284 * True, iff \a io is an O_APPEND write(2). 2285 */ 2286static inline int cl_io_is_append(const struct cl_io *io) 2287{ 2288 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append; 2289} 2290 2291static inline int cl_io_is_sync_write(const struct cl_io *io) 2292{ 2293 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync; 2294} 2295 2296static inline int cl_io_is_mkwrite(const struct cl_io *io) 2297{ 2298 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite; 2299} 2300 2301/** 2302 * True, iff \a io is a truncate(2). 2303 */ 2304static inline int cl_io_is_trunc(const struct cl_io *io) 2305{ 2306 return io->ci_type == CIT_SETATTR && 2307 (io->u.ci_setattr.sa_valid & ATTR_SIZE); 2308} 2309 2310struct cl_io *cl_io_top(struct cl_io *io); 2311 2312#define CL_IO_SLICE_CLEAN(foo_io, base) \ 2313do { \ 2314 typeof(foo_io) __foo_io = (foo_io); \ 2315 \ 2316 BUILD_BUG_ON(offsetof(typeof(*__foo_io), base) != 0); \ 2317 memset(&__foo_io->base + 1, 0, \ 2318 sizeof(*__foo_io) - sizeof(__foo_io->base)); \ 2319} while (0) 2320 2321/** @} cl_io */ 2322 2323/** \defgroup cl_page_list cl_page_list 2324 * @{ 2325 */ 2326 2327/** 2328 * Last page in the page list. 2329 */ 2330static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist) 2331{ 2332 LASSERT(plist->pl_nr > 0); 2333 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch); 2334} 2335 2336static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist) 2337{ 2338 LASSERT(plist->pl_nr > 0); 2339 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch); 2340} 2341 2342/** 2343 * Iterate over pages in a page list. 2344 */ 2345#define cl_page_list_for_each(page, list) \ 2346 list_for_each_entry((page), &(list)->pl_pages, cp_batch) 2347 2348/** 2349 * Iterate over pages in a page list, taking possible removals into account. 2350 */ 2351#define cl_page_list_for_each_safe(page, temp, list) \ 2352 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch) 2353 2354void cl_page_list_init(struct cl_page_list *plist); 2355void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page); 2356void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src, 2357 struct cl_page *page); 2358void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src, 2359 struct cl_page *page); 2360void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head); 2361void cl_page_list_del(const struct lu_env *env, struct cl_page_list *plist, 2362 struct cl_page *page); 2363void cl_page_list_disown(const struct lu_env *env, 2364 struct cl_io *io, struct cl_page_list *plist); 2365void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist); 2366 2367void cl_2queue_init(struct cl_2queue *queue); 2368void cl_2queue_disown(const struct lu_env *env, 2369 struct cl_io *io, struct cl_2queue *queue); 2370void cl_2queue_discard(const struct lu_env *env, 2371 struct cl_io *io, struct cl_2queue *queue); 2372void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue); 2373void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page); 2374 2375/** @} cl_page_list */ 2376 2377void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj, 2378 struct cl_req_attr *attr); 2379 2380/** \defgroup cl_sync_io cl_sync_io 2381 * @{ 2382 */ 2383 2384/** 2385 * Anchor for synchronous transfer. This is allocated on a stack by thread 2386 * doing synchronous transfer, and a pointer to this structure is set up in 2387 * every page submitted for transfer. Transfer completion routine updates 2388 * anchor and wakes up waiting thread when transfer is complete. 2389 */ 2390struct cl_sync_io { 2391 /** number of pages yet to be transferred. */ 2392 atomic_t csi_sync_nr; 2393 /** error code. */ 2394 int csi_sync_rc; 2395 /** barrier of destroy this structure */ 2396 atomic_t csi_barrier; 2397 /** completion to be signaled when transfer is complete. */ 2398 wait_queue_head_t csi_waitq; 2399 /** callback to invoke when this IO is finished */ 2400 void (*csi_end_io)(const struct lu_env *, 2401 struct cl_sync_io *); 2402}; 2403 2404void cl_sync_io_init(struct cl_sync_io *anchor, int nr, 2405 void (*end)(const struct lu_env *, struct cl_sync_io *)); 2406int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor, 2407 long timeout); 2408void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor, 2409 int ioret); 2410void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor); 2411 2412/** @} cl_sync_io */ 2413 2414/** \defgroup cl_env cl_env 2415 * 2416 * lu_env handling for a client. 2417 * 2418 * lu_env is an environment within which lustre code executes. Its major part 2419 * is lu_context---a fast memory allocation mechanism that is used to conserve 2420 * precious kernel stack space. Originally lu_env was designed for a server, 2421 * where 2422 * 2423 * - there is a (mostly) fixed number of threads, and 2424 * 2425 * - call chains have no non-lustre portions inserted between lustre code. 2426 * 2427 * On a client both these assumption fails, because every user thread can 2428 * potentially execute lustre code as part of a system call, and lustre calls 2429 * into VFS or MM that call back into lustre. 2430 * 2431 * To deal with that, cl_env wrapper functions implement the following 2432 * optimizations: 2433 * 2434 * - allocation and destruction of environment is amortized by caching no 2435 * longer used environments instead of destroying them; 2436 * 2437 * \see lu_env, lu_context, lu_context_key 2438 * @{ 2439 */ 2440 2441struct lu_env *cl_env_get(u16 *refcheck); 2442struct lu_env *cl_env_alloc(u16 *refcheck, __u32 tags); 2443void cl_env_put(struct lu_env *env, u16 *refcheck); 2444unsigned int cl_env_cache_purge(unsigned int nr); 2445struct lu_env *cl_env_percpu_get(void); 2446void cl_env_percpu_put(struct lu_env *env); 2447 2448/** @} cl_env */ 2449 2450/* 2451 * Misc 2452 */ 2453void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb); 2454 2455struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site, 2456 struct lu_device_type *ldt, 2457 struct lu_device *next); 2458/** @} clio */ 2459 2460int cl_global_init(void); 2461void cl_global_fini(void); 2462 2463#endif /* _LINUX_CL_OBJECT_H */ 2464