1/* 2 * GPL HEADER START 3 * 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This program is free software; you can redistribute it and/or modify 7 * it under the terms of the GNU General Public License version 2 only, 8 * as published by the Free Software Foundation. 9 * 10 * This program is distributed in the hope that it will be useful, but 11 * WITHOUT ANY WARRANTY; without even the implied warranty of 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 13 * General Public License version 2 for more details (a copy is included 14 * in the LICENSE file that accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License 17 * version 2 along with this program; If not, see 18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf 19 * 20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 21 * CA 95054 USA or visit www.sun.com if you need additional information or 22 * have any questions. 23 * 24 * GPL HEADER END 25 */ 26/* 27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. 28 * Use is subject to license terms. 29 * 30 * Copyright (c) 2011, 2012, Intel Corporation. 31 */ 32/* 33 * This file is part of Lustre, http://www.lustre.org/ 34 * Lustre is a trademark of Sun Microsystems, Inc. 35 */ 36#ifndef _LUSTRE_CL_OBJECT_H 37#define _LUSTRE_CL_OBJECT_H 38 39/** \defgroup clio clio 40 * 41 * Client objects implement io operations and cache pages. 42 * 43 * Examples: lov and osc are implementations of cl interface. 44 * 45 * Big Theory Statement. 46 * 47 * Layered objects. 48 * 49 * Client implementation is based on the following data-types: 50 * 51 * - cl_object 52 * 53 * - cl_page 54 * 55 * - cl_lock represents an extent lock on an object. 56 * 57 * - cl_io represents high-level i/o activity such as whole read/write 58 * system call, or write-out of pages from under the lock being 59 * canceled. cl_io has sub-ios that can be stopped and resumed 60 * independently, thus achieving high degree of transfer 61 * parallelism. Single cl_io can be advanced forward by 62 * the multiple threads (although in the most usual case of 63 * read/write system call it is associated with the single user 64 * thread, that issued the system call). 65 * 66 * - cl_req represents a collection of pages for a transfer. cl_req is 67 * constructed by req-forming engine that tries to saturate 68 * transport with large and continuous transfers. 69 * 70 * Terminology 71 * 72 * - to avoid confusion high-level I/O operation like read or write system 73 * call is referred to as "an io", whereas low-level I/O operation, like 74 * RPC, is referred to as "a transfer" 75 * 76 * - "generic code" means generic (not file system specific) code in the 77 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that 78 * is not layer specific. 79 * 80 * Locking. 81 * 82 * - i_mutex 83 * - PG_locked 84 * - cl_object_header::coh_page_guard 85 * - cl_object_header::coh_lock_guard 86 * - lu_site::ls_guard 87 * 88 * See the top comment in cl_object.c for the description of overall locking and 89 * reference-counting design. 90 * 91 * See comments below for the description of i/o, page, and dlm-locking 92 * design. 93 * 94 * @{ 95 */ 96 97/* 98 * super-class definitions. 99 */ 100#include <lu_object.h> 101#include <lvfs.h> 102# include <linux/mutex.h> 103# include <linux/radix-tree.h> 104 105struct inode; 106 107struct cl_device; 108struct cl_device_operations; 109 110struct cl_object; 111struct cl_object_page_operations; 112struct cl_object_lock_operations; 113 114struct cl_page; 115struct cl_page_slice; 116struct cl_lock; 117struct cl_lock_slice; 118 119struct cl_lock_operations; 120struct cl_page_operations; 121 122struct cl_io; 123struct cl_io_slice; 124 125struct cl_req; 126struct cl_req_slice; 127 128/** 129 * Operations for each data device in the client stack. 130 * 131 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops 132 */ 133struct cl_device_operations { 134 /** 135 * Initialize cl_req. This method is called top-to-bottom on all 136 * devices in the stack to get them a chance to allocate layer-private 137 * data, and to attach them to the cl_req by calling 138 * cl_req_slice_add(). 139 * 140 * \see osc_req_init(), lov_req_init(), lovsub_req_init() 141 * \see ccc_req_init() 142 */ 143 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev, 144 struct cl_req *req); 145}; 146 147/** 148 * Device in the client stack. 149 * 150 * \see ccc_device, lov_device, lovsub_device, osc_device 151 */ 152struct cl_device { 153 /** Super-class. */ 154 struct lu_device cd_lu_dev; 155 /** Per-layer operation vector. */ 156 const struct cl_device_operations *cd_ops; 157}; 158 159/** \addtogroup cl_object cl_object 160 * @{ */ 161/** 162 * "Data attributes" of cl_object. Data attributes can be updated 163 * independently for a sub-object, and top-object's attributes are calculated 164 * from sub-objects' ones. 165 */ 166struct cl_attr { 167 /** Object size, in bytes */ 168 loff_t cat_size; 169 /** 170 * Known minimal size, in bytes. 171 * 172 * This is only valid when at least one DLM lock is held. 173 */ 174 loff_t cat_kms; 175 /** Modification time. Measured in seconds since epoch. */ 176 time_t cat_mtime; 177 /** Access time. Measured in seconds since epoch. */ 178 time_t cat_atime; 179 /** Change time. Measured in seconds since epoch. */ 180 time_t cat_ctime; 181 /** 182 * Blocks allocated to this cl_object on the server file system. 183 * 184 * \todo XXX An interface for block size is needed. 185 */ 186 __u64 cat_blocks; 187 /** 188 * User identifier for quota purposes. 189 */ 190 uid_t cat_uid; 191 /** 192 * Group identifier for quota purposes. 193 */ 194 gid_t cat_gid; 195}; 196 197/** 198 * Fields in cl_attr that are being set. 199 */ 200enum cl_attr_valid { 201 CAT_SIZE = 1 << 0, 202 CAT_KMS = 1 << 1, 203 CAT_MTIME = 1 << 3, 204 CAT_ATIME = 1 << 4, 205 CAT_CTIME = 1 << 5, 206 CAT_BLOCKS = 1 << 6, 207 CAT_UID = 1 << 7, 208 CAT_GID = 1 << 8 209}; 210 211/** 212 * Sub-class of lu_object with methods common for objects on the client 213 * stacks. 214 * 215 * cl_object: represents a regular file system object, both a file and a 216 * stripe. cl_object is based on lu_object: it is identified by a fid, 217 * layered, cached, hashed, and lrued. Important distinction with the server 218 * side, where md_object and dt_object are used, is that cl_object "fans out" 219 * at the lov/sns level: depending on the file layout, single file is 220 * represented as a set of "sub-objects" (stripes). At the implementation 221 * level, struct lov_object contains an array of cl_objects. Each sub-object 222 * is a full-fledged cl_object, having its fid, living in the lru and hash 223 * table. 224 * 225 * This leads to the next important difference with the server side: on the 226 * client, it's quite usual to have objects with the different sequence of 227 * layers. For example, typical top-object is composed of the following 228 * layers: 229 * 230 * - vvp 231 * - lov 232 * 233 * whereas its sub-objects are composed of 234 * 235 * - lovsub 236 * - osc 237 * 238 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep 239 * track of the object-subobject relationship. 240 * 241 * Sub-objects are not cached independently: when top-object is about to 242 * be discarded from the memory, all its sub-objects are torn-down and 243 * destroyed too. 244 * 245 * \see ccc_object, lov_object, lovsub_object, osc_object 246 */ 247struct cl_object { 248 /** super class */ 249 struct lu_object co_lu; 250 /** per-object-layer operations */ 251 const struct cl_object_operations *co_ops; 252 /** offset of page slice in cl_page buffer */ 253 int co_slice_off; 254}; 255 256/** 257 * Description of the client object configuration. This is used for the 258 * creation of a new client object that is identified by a more state than 259 * fid. 260 */ 261struct cl_object_conf { 262 /** Super-class. */ 263 struct lu_object_conf coc_lu; 264 union { 265 /** 266 * Object layout. This is consumed by lov. 267 */ 268 struct lustre_md *coc_md; 269 /** 270 * Description of particular stripe location in the 271 * cluster. This is consumed by osc. 272 */ 273 struct lov_oinfo *coc_oinfo; 274 } u; 275 /** 276 * VFS inode. This is consumed by vvp. 277 */ 278 struct inode *coc_inode; 279 /** 280 * Layout lock handle. 281 */ 282 struct ldlm_lock *coc_lock; 283 /** 284 * Operation to handle layout, OBJECT_CONF_XYZ. 285 */ 286 int coc_opc; 287}; 288 289enum { 290 /** configure layout, set up a new stripe, must be called while 291 * holding layout lock. */ 292 OBJECT_CONF_SET = 0, 293 /** invalidate the current stripe configuration due to losing 294 * layout lock. */ 295 OBJECT_CONF_INVALIDATE = 1, 296 /** wait for old layout to go away so that new layout can be 297 * set up. */ 298 OBJECT_CONF_WAIT = 2 299}; 300 301/** 302 * Operations implemented for each cl object layer. 303 * 304 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops 305 */ 306struct cl_object_operations { 307 /** 308 * Initialize page slice for this layer. Called top-to-bottom through 309 * every object layer when a new cl_page is instantiated. Layer 310 * keeping private per-page data, or requiring its own page operations 311 * vector should allocate these data here, and attach then to the page 312 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM 313 * sense). Optional. 314 * 315 * \retval NULL success. 316 * 317 * \retval ERR_PTR(errno) failure code. 318 * 319 * \retval valid-pointer pointer to already existing referenced page 320 * to be used instead of newly created. 321 */ 322 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj, 323 struct cl_page *page, struct page *vmpage); 324 /** 325 * Initialize lock slice for this layer. Called top-to-bottom through 326 * every object layer when a new cl_lock is instantiated. Layer 327 * keeping private per-lock data, or requiring its own lock operations 328 * vector should allocate these data here, and attach then to the lock 329 * by calling cl_lock_slice_add(). Mandatory. 330 */ 331 int (*coo_lock_init)(const struct lu_env *env, 332 struct cl_object *obj, struct cl_lock *lock, 333 const struct cl_io *io); 334 /** 335 * Initialize io state for a given layer. 336 * 337 * called top-to-bottom once per io existence to initialize io 338 * state. If layer wants to keep some state for this type of io, it 339 * has to embed struct cl_io_slice in lu_env::le_ses, and register 340 * slice with cl_io_slice_add(). It is guaranteed that all threads 341 * participating in this io share the same session. 342 */ 343 int (*coo_io_init)(const struct lu_env *env, 344 struct cl_object *obj, struct cl_io *io); 345 /** 346 * Fill portion of \a attr that this layer controls. This method is 347 * called top-to-bottom through all object layers. 348 * 349 * \pre cl_object_header::coh_attr_guard of the top-object is locked. 350 * 351 * \return 0: to continue 352 * \return +ve: to stop iterating through layers (but 0 is returned 353 * from enclosing cl_object_attr_get()) 354 * \return -ve: to signal error 355 */ 356 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj, 357 struct cl_attr *attr); 358 /** 359 * Update attributes. 360 * 361 * \a valid is a bitmask composed from enum #cl_attr_valid, and 362 * indicating what attributes are to be set. 363 * 364 * \pre cl_object_header::coh_attr_guard of the top-object is locked. 365 * 366 * \return the same convention as for 367 * cl_object_operations::coo_attr_get() is used. 368 */ 369 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj, 370 const struct cl_attr *attr, unsigned valid); 371 /** 372 * Update object configuration. Called top-to-bottom to modify object 373 * configuration. 374 * 375 * XXX error conditions and handling. 376 */ 377 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj, 378 const struct cl_object_conf *conf); 379 /** 380 * Glimpse ast. Executed when glimpse ast arrives for a lock on this 381 * object. Layers are supposed to fill parts of \a lvb that will be 382 * shipped to the glimpse originator as a glimpse result. 383 * 384 * \see ccc_object_glimpse(), lovsub_object_glimpse(), 385 * \see osc_object_glimpse() 386 */ 387 int (*coo_glimpse)(const struct lu_env *env, 388 const struct cl_object *obj, struct ost_lvb *lvb); 389}; 390 391/** 392 * Extended header for client object. 393 */ 394struct cl_object_header { 395 /** Standard lu_object_header. cl_object::co_lu::lo_header points 396 * here. */ 397 struct lu_object_header coh_lu; 398 /** \name locks 399 * \todo XXX move locks below to the separate cache-lines, they are 400 * mostly useless otherwise. 401 */ 402 /** @{ */ 403 /** Lock protecting page tree. */ 404 spinlock_t coh_page_guard; 405 /** Lock protecting lock list. */ 406 spinlock_t coh_lock_guard; 407 /** @} locks */ 408 /** Radix tree of cl_page's, cached for this object. */ 409 struct radix_tree_root coh_tree; 410 /** # of pages in radix tree. */ 411 unsigned long coh_pages; 412 /** List of cl_lock's granted for this object. */ 413 struct list_head coh_locks; 414 415 /** 416 * Parent object. It is assumed that an object has a well-defined 417 * parent, but not a well-defined child (there may be multiple 418 * sub-objects, for the same top-object). cl_object_header::coh_parent 419 * field allows certain code to be written generically, without 420 * limiting possible cl_object layouts unduly. 421 */ 422 struct cl_object_header *coh_parent; 423 /** 424 * Protects consistency between cl_attr of parent object and 425 * attributes of sub-objects, that the former is calculated ("merged") 426 * from. 427 * 428 * \todo XXX this can be read/write lock if needed. 429 */ 430 spinlock_t coh_attr_guard; 431 /** 432 * Size of cl_page + page slices 433 */ 434 unsigned short coh_page_bufsize; 435 /** 436 * Number of objects above this one: 0 for a top-object, 1 for its 437 * sub-object, etc. 438 */ 439 unsigned char coh_nesting; 440}; 441 442/** 443 * Helper macro: iterate over all layers of the object \a obj, assigning every 444 * layer top-to-bottom to \a slice. 445 */ 446#define cl_object_for_each(slice, obj) \ 447 list_for_each_entry((slice), \ 448 &(obj)->co_lu.lo_header->loh_layers, \ 449 co_lu.lo_linkage) 450/** 451 * Helper macro: iterate over all layers of the object \a obj, assigning every 452 * layer bottom-to-top to \a slice. 453 */ 454#define cl_object_for_each_reverse(slice, obj) \ 455 list_for_each_entry_reverse((slice), \ 456 &(obj)->co_lu.lo_header->loh_layers, \ 457 co_lu.lo_linkage) 458/** @} cl_object */ 459 460#ifndef pgoff_t 461#define pgoff_t unsigned long 462#endif 463 464#define CL_PAGE_EOF ((pgoff_t)~0ull) 465 466/** \addtogroup cl_page cl_page 467 * @{ */ 468 469/** \struct cl_page 470 * Layered client page. 471 * 472 * cl_page: represents a portion of a file, cached in the memory. All pages 473 * of the given file are of the same size, and are kept in the radix tree 474 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects 475 * of the top-level file object are first class cl_objects, they have their 476 * own radix trees of pages and hence page is implemented as a sequence of 477 * struct cl_pages's, linked into double-linked list through 478 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the 479 * corresponding radix tree at the corresponding logical offset. 480 * 481 * cl_page is associated with VM page of the hosting environment (struct 482 * page in Linux kernel, for example), struct page. It is assumed, that this 483 * association is implemented by one of cl_page layers (top layer in the 484 * current design) that 485 * 486 * - intercepts per-VM-page call-backs made by the environment (e.g., 487 * memory pressure), 488 * 489 * - translates state (page flag bits) and locking between lustre and 490 * environment. 491 * 492 * The association between cl_page and struct page is immutable and 493 * established when cl_page is created. 494 * 495 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing 496 * this io an exclusive access to this page w.r.t. other io attempts and 497 * various events changing page state (such as transfer completion, or 498 * eviction of the page from the memory). Note, that in general cl_io 499 * cannot be identified with a particular thread, and page ownership is not 500 * exactly equal to the current thread holding a lock on the page. Layer 501 * implementing association between cl_page and struct page has to implement 502 * ownership on top of available synchronization mechanisms. 503 * 504 * While lustre client maintains the notion of an page ownership by io, 505 * hosting MM/VM usually has its own page concurrency control 506 * mechanisms. For example, in Linux, page access is synchronized by the 507 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*()) 508 * takes care to acquire and release such locks as necessary around the 509 * calls to the file system methods (->readpage(), ->prepare_write(), 510 * ->commit_write(), etc.). This leads to the situation when there are two 511 * different ways to own a page in the client: 512 * 513 * - client code explicitly and voluntary owns the page (cl_page_own()); 514 * 515 * - VM locks a page and then calls the client, that has "to assume" 516 * the ownership from the VM (cl_page_assume()). 517 * 518 * Dual methods to release ownership are cl_page_disown() and 519 * cl_page_unassume(). 520 * 521 * cl_page is reference counted (cl_page::cp_ref). When reference counter 522 * drops to 0, the page is returned to the cache, unless it is in 523 * cl_page_state::CPS_FREEING state, in which case it is immediately 524 * destroyed. 525 * 526 * The general logic guaranteeing the absence of "existential races" for 527 * pages is the following: 528 * 529 * - there are fixed known ways for a thread to obtain a new reference 530 * to a page: 531 * 532 * - by doing a lookup in the cl_object radix tree, protected by the 533 * spin-lock; 534 * 535 * - by starting from VM-locked struct page and following some 536 * hosting environment method (e.g., following ->private pointer in 537 * the case of Linux kernel), see cl_vmpage_page(); 538 * 539 * - when the page enters cl_page_state::CPS_FREEING state, all these 540 * ways are severed with the proper synchronization 541 * (cl_page_delete()); 542 * 543 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page 544 * lock; 545 * 546 * - no new references to the page in cl_page_state::CPS_FREEING state 547 * are allowed (checked in cl_page_get()). 548 * 549 * Together this guarantees that when last reference to a 550 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the 551 * page, as neither references to it can be acquired at that point, nor 552 * ones exist. 553 * 554 * cl_page is a state machine. States are enumerated in enum 555 * cl_page_state. Possible state transitions are enumerated in 556 * cl_page_state_set(). State transition process (i.e., actual changing of 557 * cl_page::cp_state field) is protected by the lock on the underlying VM 558 * page. 559 * 560 * Linux Kernel implementation. 561 * 562 * Binding between cl_page and struct page (which is a typedef for 563 * struct page) is implemented in the vvp layer. cl_page is attached to the 564 * ->private pointer of the struct page, together with the setting of 565 * PG_private bit in page->flags, and acquiring additional reference on the 566 * struct page (much like struct buffer_head, or any similar file system 567 * private data structures). 568 * 569 * PG_locked lock is used to implement both ownership and transfer 570 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}} 571 * states. No additional references are acquired for the duration of the 572 * transfer. 573 * 574 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where 575 * write-out is "protected" by the special PG_writeback bit. 576 */ 577 578/** 579 * States of cl_page. cl_page.c assumes particular order here. 580 * 581 * The page state machine is rather crude, as it doesn't recognize finer page 582 * states like "dirty" or "up to date". This is because such states are not 583 * always well defined for the whole stack (see, for example, the 584 * implementation of the read-ahead, that hides page up-to-dateness to track 585 * cache hits accurately). Such sub-states are maintained by the layers that 586 * are interested in them. 587 */ 588enum cl_page_state { 589 /** 590 * Page is in the cache, un-owned. Page leaves cached state in the 591 * following cases: 592 * 593 * - [cl_page_state::CPS_OWNED] io comes across the page and 594 * owns it; 595 * 596 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the 597 * req-formation engine decides that it wants to include this page 598 * into an cl_req being constructed, and yanks it from the cache; 599 * 600 * - [cl_page_state::CPS_FREEING] VM callback is executed to 601 * evict the page form the memory; 602 * 603 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL 604 */ 605 CPS_CACHED, 606 /** 607 * Page is exclusively owned by some cl_io. Page may end up in this 608 * state as a result of 609 * 610 * - io creating new page and immediately owning it; 611 * 612 * - [cl_page_state::CPS_CACHED] io finding existing cached page 613 * and owning it; 614 * 615 * - [cl_page_state::CPS_OWNED] io finding existing owned page 616 * and waiting for owner to release the page; 617 * 618 * Page leaves owned state in the following cases: 619 * 620 * - [cl_page_state::CPS_CACHED] io decides to leave the page in 621 * the cache, doing nothing; 622 * 623 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for 624 * this page; 625 * 626 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write 627 * transfer for this page; 628 * 629 * - [cl_page_state::CPS_FREEING] io decides to destroy this 630 * page (e.g., as part of truncate or extent lock cancellation). 631 * 632 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL 633 */ 634 CPS_OWNED, 635 /** 636 * Page is being written out, as a part of a transfer. This state is 637 * entered when req-formation logic decided that it wants this page to 638 * be sent through the wire _now_. Specifically, it means that once 639 * this state is achieved, transfer completion handler (with either 640 * success or failure indication) is guaranteed to be executed against 641 * this page independently of any locks and any scheduling decisions 642 * made by the hosting environment (that effectively means that the 643 * page is never put into cl_page_state::CPS_PAGEOUT state "in 644 * advance". This property is mentioned, because it is important when 645 * reasoning about possible dead-locks in the system). The page can 646 * enter this state as a result of 647 * 648 * - [cl_page_state::CPS_OWNED] an io requesting an immediate 649 * write-out of this page, or 650 * 651 * - [cl_page_state::CPS_CACHED] req-forming engine deciding 652 * that it has enough dirty pages cached to issue a "good" 653 * transfer. 654 * 655 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer 656 * is completed---it is moved into cl_page_state::CPS_CACHED state. 657 * 658 * Underlying VM page is locked for the duration of transfer. 659 * 660 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL 661 */ 662 CPS_PAGEOUT, 663 /** 664 * Page is being read in, as a part of a transfer. This is quite 665 * similar to the cl_page_state::CPS_PAGEOUT state, except that 666 * read-in is always "immediate"---there is no such thing a sudden 667 * construction of read cl_req from cached, presumably not up to date, 668 * pages. 669 * 670 * Underlying VM page is locked for the duration of transfer. 671 * 672 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL 673 */ 674 CPS_PAGEIN, 675 /** 676 * Page is being destroyed. This state is entered when client decides 677 * that page has to be deleted from its host object, as, e.g., a part 678 * of truncate. 679 * 680 * Once this state is reached, there is no way to escape it. 681 * 682 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL 683 */ 684 CPS_FREEING, 685 CPS_NR 686}; 687 688enum cl_page_type { 689 /** Host page, the page is from the host inode which the cl_page 690 * belongs to. */ 691 CPT_CACHEABLE = 1, 692 693 /** Transient page, the transient cl_page is used to bind a cl_page 694 * to vmpage which is not belonging to the same object of cl_page. 695 * it is used in DirectIO, lockless IO and liblustre. */ 696 CPT_TRANSIENT, 697}; 698 699/** 700 * Flags maintained for every cl_page. 701 */ 702enum cl_page_flags { 703 /** 704 * Set when pagein completes. Used for debugging (read completes at 705 * most once for a page). 706 */ 707 CPF_READ_COMPLETED = 1 << 0 708}; 709 710/** 711 * Fields are protected by the lock on struct page, except for atomics and 712 * immutables. 713 * 714 * \invariant Data type invariants are in cl_page_invariant(). Basically: 715 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked 716 * list, consistent with the parent/child pointers in the cl_page::cp_obj and 717 * cl_page::cp_owner (when set). 718 */ 719struct cl_page { 720 /** Reference counter. */ 721 atomic_t cp_ref; 722 /** An object this page is a part of. Immutable after creation. */ 723 struct cl_object *cp_obj; 724 /** Logical page index within the object. Immutable after creation. */ 725 pgoff_t cp_index; 726 /** List of slices. Immutable after creation. */ 727 struct list_head cp_layers; 728 /** Parent page, NULL for top-level page. Immutable after creation. */ 729 struct cl_page *cp_parent; 730 /** Lower-layer page. NULL for bottommost page. Immutable after 731 * creation. */ 732 struct cl_page *cp_child; 733 /** 734 * Page state. This field is const to avoid accidental update, it is 735 * modified only internally within cl_page.c. Protected by a VM lock. 736 */ 737 const enum cl_page_state cp_state; 738 /** Linkage of pages within group. Protected by cl_page::cp_mutex. */ 739 struct list_head cp_batch; 740 /** Mutex serializing membership of a page in a batch. */ 741 struct mutex cp_mutex; 742 /** Linkage of pages within cl_req. */ 743 struct list_head cp_flight; 744 /** Transfer error. */ 745 int cp_error; 746 747 /** 748 * Page type. Only CPT_TRANSIENT is used so far. Immutable after 749 * creation. 750 */ 751 enum cl_page_type cp_type; 752 753 /** 754 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned 755 * by sub-io. Protected by a VM lock. 756 */ 757 struct cl_io *cp_owner; 758 /** 759 * Debug information, the task is owning the page. 760 */ 761 task_t *cp_task; 762 /** 763 * Owning IO request in cl_page_state::CPS_PAGEOUT and 764 * cl_page_state::CPS_PAGEIN states. This field is maintained only in 765 * the top-level pages. Protected by a VM lock. 766 */ 767 struct cl_req *cp_req; 768 /** List of references to this page, for debugging. */ 769 struct lu_ref cp_reference; 770 /** Link to an object, for debugging. */ 771 struct lu_ref_link *cp_obj_ref; 772 /** Link to a queue, for debugging. */ 773 struct lu_ref_link *cp_queue_ref; 774 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */ 775 unsigned cp_flags; 776 /** Assigned if doing a sync_io */ 777 struct cl_sync_io *cp_sync_io; 778}; 779 780/** 781 * Per-layer part of cl_page. 782 * 783 * \see ccc_page, lov_page, osc_page 784 */ 785struct cl_page_slice { 786 struct cl_page *cpl_page; 787 /** 788 * Object slice corresponding to this page slice. Immutable after 789 * creation. 790 */ 791 struct cl_object *cpl_obj; 792 const struct cl_page_operations *cpl_ops; 793 /** Linkage into cl_page::cp_layers. Immutable after creation. */ 794 struct list_head cpl_linkage; 795}; 796 797/** 798 * Lock mode. For the client extent locks. 799 * 800 * \warning: cl_lock_mode_match() assumes particular ordering here. 801 * \ingroup cl_lock 802 */ 803enum cl_lock_mode { 804 /** 805 * Mode of a lock that protects no data, and exists only as a 806 * placeholder. This is used for `glimpse' requests. A phantom lock 807 * might get promoted to real lock at some point. 808 */ 809 CLM_PHANTOM, 810 CLM_READ, 811 CLM_WRITE, 812 CLM_GROUP 813}; 814 815/** 816 * Requested transfer type. 817 * \ingroup cl_req 818 */ 819enum cl_req_type { 820 CRT_READ, 821 CRT_WRITE, 822 CRT_NR 823}; 824 825/** 826 * Per-layer page operations. 827 * 828 * Methods taking an \a io argument are for the activity happening in the 829 * context of given \a io. Page is assumed to be owned by that io, except for 830 * the obvious cases (like cl_page_operations::cpo_own()). 831 * 832 * \see vvp_page_ops, lov_page_ops, osc_page_ops 833 */ 834struct cl_page_operations { 835 /** 836 * cl_page<->struct page methods. Only one layer in the stack has to 837 * implement these. Current code assumes that this functionality is 838 * provided by the topmost layer, see cl_page_disown0() as an example. 839 */ 840 841 /** 842 * \return the underlying VM page. Optional. 843 */ 844 struct page *(*cpo_vmpage)(const struct lu_env *env, 845 const struct cl_page_slice *slice); 846 /** 847 * Called when \a io acquires this page into the exclusive 848 * ownership. When this method returns, it is guaranteed that the is 849 * not owned by other io, and no transfer is going on against 850 * it. Optional. 851 * 852 * \see cl_page_own() 853 * \see vvp_page_own(), lov_page_own() 854 */ 855 int (*cpo_own)(const struct lu_env *env, 856 const struct cl_page_slice *slice, 857 struct cl_io *io, int nonblock); 858 /** Called when ownership it yielded. Optional. 859 * 860 * \see cl_page_disown() 861 * \see vvp_page_disown() 862 */ 863 void (*cpo_disown)(const struct lu_env *env, 864 const struct cl_page_slice *slice, struct cl_io *io); 865 /** 866 * Called for a page that is already "owned" by \a io from VM point of 867 * view. Optional. 868 * 869 * \see cl_page_assume() 870 * \see vvp_page_assume(), lov_page_assume() 871 */ 872 void (*cpo_assume)(const struct lu_env *env, 873 const struct cl_page_slice *slice, struct cl_io *io); 874 /** Dual to cl_page_operations::cpo_assume(). Optional. Called 875 * bottom-to-top when IO releases a page without actually unlocking 876 * it. 877 * 878 * \see cl_page_unassume() 879 * \see vvp_page_unassume() 880 */ 881 void (*cpo_unassume)(const struct lu_env *env, 882 const struct cl_page_slice *slice, 883 struct cl_io *io); 884 /** 885 * Announces whether the page contains valid data or not by \a uptodate. 886 * 887 * \see cl_page_export() 888 * \see vvp_page_export() 889 */ 890 void (*cpo_export)(const struct lu_env *env, 891 const struct cl_page_slice *slice, int uptodate); 892 /** 893 * Unmaps page from the user space (if it is mapped). 894 * 895 * \see cl_page_unmap() 896 * \see vvp_page_unmap() 897 */ 898 int (*cpo_unmap)(const struct lu_env *env, 899 const struct cl_page_slice *slice, struct cl_io *io); 900 /** 901 * Checks whether underlying VM page is locked (in the suitable 902 * sense). Used for assertions. 903 * 904 * \retval -EBUSY: page is protected by a lock of a given mode; 905 * \retval -ENODATA: page is not protected by a lock; 906 * \retval 0: this layer cannot decide. (Should never happen.) 907 */ 908 int (*cpo_is_vmlocked)(const struct lu_env *env, 909 const struct cl_page_slice *slice); 910 /** 911 * Page destruction. 912 */ 913 914 /** 915 * Called when page is truncated from the object. Optional. 916 * 917 * \see cl_page_discard() 918 * \see vvp_page_discard(), osc_page_discard() 919 */ 920 void (*cpo_discard)(const struct lu_env *env, 921 const struct cl_page_slice *slice, 922 struct cl_io *io); 923 /** 924 * Called when page is removed from the cache, and is about to being 925 * destroyed. Optional. 926 * 927 * \see cl_page_delete() 928 * \see vvp_page_delete(), osc_page_delete() 929 */ 930 void (*cpo_delete)(const struct lu_env *env, 931 const struct cl_page_slice *slice); 932 /** Destructor. Frees resources and slice itself. */ 933 void (*cpo_fini)(const struct lu_env *env, 934 struct cl_page_slice *slice); 935 936 /** 937 * Checks whether the page is protected by a cl_lock. This is a 938 * per-layer method, because certain layers have ways to check for the 939 * lock much more efficiently than through the generic locks scan, or 940 * implement locking mechanisms separate from cl_lock, e.g., 941 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks 942 * being canceled, or scheduled for cancellation as soon as the last 943 * user goes away, too. 944 * 945 * \retval -EBUSY: page is protected by a lock of a given mode; 946 * \retval -ENODATA: page is not protected by a lock; 947 * \retval 0: this layer cannot decide. 948 * 949 * \see cl_page_is_under_lock() 950 */ 951 int (*cpo_is_under_lock)(const struct lu_env *env, 952 const struct cl_page_slice *slice, 953 struct cl_io *io); 954 955 /** 956 * Optional debugging helper. Prints given page slice. 957 * 958 * \see cl_page_print() 959 */ 960 int (*cpo_print)(const struct lu_env *env, 961 const struct cl_page_slice *slice, 962 void *cookie, lu_printer_t p); 963 /** 964 * \name transfer 965 * 966 * Transfer methods. See comment on cl_req for a description of 967 * transfer formation and life-cycle. 968 * 969 * @{ 970 */ 971 /** 972 * Request type dependent vector of operations. 973 * 974 * Transfer operations depend on transfer mode (cl_req_type). To avoid 975 * passing transfer mode to each and every of these methods, and to 976 * avoid branching on request type inside of the methods, separate 977 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are 978 * provided. That is, method invocation usually looks like 979 * 980 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...); 981 */ 982 struct { 983 /** 984 * Called when a page is submitted for a transfer as a part of 985 * cl_page_list. 986 * 987 * \return 0 : page is eligible for submission; 988 * \return -EALREADY : skip this page; 989 * \return -ve : error. 990 * 991 * \see cl_page_prep() 992 */ 993 int (*cpo_prep)(const struct lu_env *env, 994 const struct cl_page_slice *slice, 995 struct cl_io *io); 996 /** 997 * Completion handler. This is guaranteed to be eventually 998 * fired after cl_page_operations::cpo_prep() or 999 * cl_page_operations::cpo_make_ready() call. 1000 *
1001 * This method can be called in a non-blocking context. It is 1002 * guaranteed however, that the page involved and its object 1003 * are pinned in memory (and, hence, calling cl_page_put() is 1004 * safe). 1005 * 1006 * \see cl_page_completion() 1007 */ 1008 void (*cpo_completion)(const struct lu_env *env, 1009 const struct cl_page_slice *slice, 1010 int ioret); 1011 /** 1012 * Called when cached page is about to be added to the 1013 * cl_req as a part of req formation. 1014 * 1015 * \return 0 : proceed with this page; 1016 * \return -EAGAIN : skip this page; 1017 * \return -ve : error. 1018 * 1019 * \see cl_page_make_ready() 1020 */ 1021 int (*cpo_make_ready)(const struct lu_env *env, 1022 const struct cl_page_slice *slice); 1023 /** 1024 * Announce that this page is to be written out 1025 * opportunistically, that is, page is dirty, it is not 1026 * necessary to start write-out transfer right now, but 1027 * eventually page has to be written out. 1028 * 1029 * Main caller of this is the write path (see 1030 * vvp_io_commit_write()), using this method to build a 1031 * "transfer cache" from which large transfers are then 1032 * constructed by the req-formation engine. 1033 * 1034 * \todo XXX it would make sense to add page-age tracking 1035 * semantics here, and to oblige the req-formation engine to 1036 * send the page out not later than it is too old. 1037 * 1038 * \see cl_page_cache_add() 1039 */ 1040 int (*cpo_cache_add)(const struct lu_env *env, 1041 const struct cl_page_slice *slice, 1042 struct cl_io *io); 1043 } io[CRT_NR]; 1044 /** 1045 * Tell transfer engine that only [to, from] part of a page should be 1046 * transmitted. 1047 * 1048 * This is used for immediate transfers. 1049 * 1050 * \todo XXX this is not very good interface. It would be much better 1051 * if all transfer parameters were supplied as arguments to 1052 * cl_io_operations::cio_submit() call, but it is not clear how to do 1053 * this for page queues. 1054 * 1055 * \see cl_page_clip() 1056 */ 1057 void (*cpo_clip)(const struct lu_env *env, 1058 const struct cl_page_slice *slice, 1059 int from, int to); 1060 /** 1061 * \pre the page was queued for transferring. 1062 * \post page is removed from client's pending list, or -EBUSY 1063 * is returned if it has already been in transferring. 1064 * 1065 * This is one of seldom page operation which is: 1066 * 0. called from top level; 1067 * 1. don't have vmpage locked; 1068 * 2. every layer should synchronize execution of its ->cpo_cancel() 1069 * with completion handlers. Osc uses client obd lock for this 1070 * purpose. Based on there is no vvp_page_cancel and 1071 * lov_page_cancel(), cpo_cancel is defacto protected by client lock. 1072 * 1073 * \see osc_page_cancel(). 1074 */ 1075 int (*cpo_cancel)(const struct lu_env *env, 1076 const struct cl_page_slice *slice); 1077 /** 1078 * Write out a page by kernel. This is only called by ll_writepage 1079 * right now. 1080 * 1081 * \see cl_page_flush() 1082 */ 1083 int (*cpo_flush)(const struct lu_env *env, 1084 const struct cl_page_slice *slice, 1085 struct cl_io *io); 1086 /** @} transfer */ 1087}; 1088 1089/** 1090 * Helper macro, dumping detailed information about \a page into a log. 1091 */ 1092#define CL_PAGE_DEBUG(mask, env, page, format, ...) \ 1093do { \ 1094 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \ 1095 \ 1096 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \ 1097 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \ 1098 CDEBUG(mask, format , ## __VA_ARGS__); \ 1099 } \ 1100} while (0) 1101 1102/** 1103 * Helper macro, dumping shorter information about \a page into a log. 1104 */ 1105#define CL_PAGE_HEADER(mask, env, page, format, ...) \ 1106do { \ 1107 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \ 1108 \ 1109 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \ 1110 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \ 1111 CDEBUG(mask, format , ## __VA_ARGS__); \ 1112 } \ 1113} while (0) 1114 1115static inline int __page_in_use(const struct cl_page *page, int refc) 1116{ 1117 if (page->cp_type == CPT_CACHEABLE) 1118 ++refc; 1119 LASSERT(atomic_read(&page->cp_ref) > 0); 1120 return (atomic_read(&page->cp_ref) > refc); 1121} 1122#define cl_page_in_use(pg) __page_in_use(pg, 1) 1123#define cl_page_in_use_noref(pg) __page_in_use(pg, 0) 1124 1125/** @} cl_page */ 1126 1127/** \addtogroup cl_lock cl_lock 1128 * @{ */ 1129/** \struct cl_lock 1130 * 1131 * Extent locking on the client. 1132 * 1133 * LAYERING 1134 * 1135 * The locking model of the new client code is built around 1136 * 1137 * struct cl_lock 1138 * 1139 * data-type representing an extent lock on a regular file. cl_lock is a 1140 * layered object (much like cl_object and cl_page), it consists of a header 1141 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to 1142 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage. 1143 * 1144 * All locks for a given object are linked into cl_object_header::coh_locks 1145 * list (protected by cl_object_header::coh_lock_guard spin-lock) through 1146 * cl_lock::cll_linkage. Currently this list is not sorted in any way. We can 1147 * sort it in starting lock offset, or use altogether different data structure 1148 * like a tree. 1149 * 1150 * Typical cl_lock consists of the two layers: 1151 * 1152 * - vvp_lock (vvp specific data), and 1153 * - lov_lock (lov specific data). 1154 * 1155 * lov_lock contains an array of sub-locks. Each of these sub-locks is a 1156 * normal cl_lock: it has a header (struct cl_lock) and a list of layers: 1157 * 1158 * - lovsub_lock, and 1159 * - osc_lock 1160 * 1161 * Each sub-lock is associated with a cl_object (representing stripe 1162 * sub-object or the file to which top-level cl_lock is associated to), and is 1163 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to 1164 * cl_object (that at lov layer also fans out into multiple sub-objects), and 1165 * is different from cl_page, that doesn't fan out (there is usually exactly 1166 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock 1167 * a "top-lock" and its lovsub-osc portion a "sub-lock". 1168 * 1169 * LIFE CYCLE 1170 * 1171 * cl_lock is reference counted. When reference counter drops to 0, lock is 1172 * placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING 1173 * lock is destroyed when last reference is released. Referencing between 1174 * top-lock and its sub-locks is described in the lov documentation module. 1175 * 1176 * STATE MACHINE 1177 * 1178 * Also, cl_lock is a state machine. This requires some clarification. One of 1179 * the goals of client IO re-write was to make IO path non-blocking, or at 1180 * least to make it easier to make it non-blocking in the future. Here 1181 * `non-blocking' means that when a system call (read, write, truncate) 1182 * reaches a situation where it has to wait for a communication with the 1183 * server, it should --instead of waiting-- remember its current state and 1184 * switch to some other work. E.g,. instead of waiting for a lock enqueue, 1185 * client should proceed doing IO on the next stripe, etc. Obviously this is 1186 * rather radical redesign, and it is not planned to be fully implemented at 1187 * this time, instead we are putting some infrastructure in place, that would 1188 * make it easier to do asynchronous non-blocking IO easier in the 1189 * future. Specifically, where old locking code goes to sleep (waiting for 1190 * enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When 1191 * enqueue reply comes, its completion handler signals that lock state-machine 1192 * is ready to transit to the next state. There is some generic code in 1193 * cl_lock.c that sleeps, waiting for these signals. As a result, for users of 1194 * this cl_lock.c code, it looks like locking is done in normal blocking 1195 * fashion, and it the same time it is possible to switch to the non-blocking 1196 * locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c 1197 * functions). 1198 * 1199 * For a description of state machine states and transitions see enum 1200 * cl_lock_state. 1201 * 1202 * There are two ways to restrict a set of states which lock might move to: 1203 * 1204 * - placing a "hold" on a lock guarantees that lock will not be moved 1205 * into cl_lock_state::CLS_FREEING state until hold is released. Hold 1206 * can be only acquired on a lock that is not in 1207 * cl_lock_state::CLS_FREEING. All holds on a lock are counted in 1208 * cl_lock::cll_holds. Hold protects lock from cancellation and 1209 * destruction. Requests to cancel and destroy a lock on hold will be 1210 * recorded, but only honored when last hold on a lock is released; 1211 * 1212 * - placing a "user" on a lock guarantees that lock will not leave 1213 * cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING, 1214 * cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of 1215 * states, once it enters this set. That is, if a user is added onto a 1216 * lock in a state not from this set, it doesn't immediately enforce 1217 * lock to move to this set, but once lock enters this set it will 1218 * remain there until all users are removed. Lock users are counted in 1219 * cl_lock::cll_users. 1220 * 1221 * User is used to assure that lock is not canceled or destroyed while 1222 * it is being enqueued, or actively used by some IO. 1223 * 1224 * Currently, a user always comes with a hold (cl_lock_invariant() 1225 * checks that a number of holds is not less than a number of users). 1226 * 1227 * CONCURRENCY 1228 * 1229 * This is how lock state-machine operates. struct cl_lock contains a mutex 1230 * cl_lock::cll_guard that protects struct fields. 1231 * 1232 * - mutex is taken, and cl_lock::cll_state is examined. 1233 * 1234 * - for every state there are possible target states where lock can move 1235 * into. They are tried in order. Attempts to move into next state are 1236 * done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try(). 1237 * 1238 * - if the transition can be performed immediately, state is changed, 1239 * and mutex is released. 1240 * 1241 * - if the transition requires blocking, _try() function returns 1242 * cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to 1243 * sleep, waiting for possibility of lock state change. It is woken 1244 * up when some event occurs, that makes lock state change possible 1245 * (e.g., the reception of the reply from the server), and repeats 1246 * the loop. 1247 * 1248 * Top-lock and sub-lock has separate mutexes and the latter has to be taken 1249 * first to avoid dead-lock. 1250 * 1251 * To see an example of interaction of all these issues, take a look at the 1252 * lov_cl.c:lov_lock_enqueue() function. It is called as a part of 1253 * cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by 1254 * advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note 1255 * also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It 1256 * also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be 1257 * done in parallel, rather than one after another (this is used for glimpse 1258 * locks, that cannot dead-lock). 1259 * 1260 * INTERFACE AND USAGE 1261 * 1262 * struct cl_lock_operations provide a number of call-backs that are invoked 1263 * when events of interest occurs. Layers can intercept and handle glimpse, 1264 * blocking, cancel ASTs and a reception of the reply from the server. 1265 * 1266 * One important difference with the old client locking model is that new 1267 * client has a representation for the top-lock, whereas in the old code only 1268 * sub-locks existed as real data structures and file-level locks are 1269 * represented by "request sets" that are created and destroyed on each and 1270 * every lock creation. 1271 * 1272 * Top-locks are cached, and can be found in the cache by the system calls. It 1273 * is possible that top-lock is in cache, but some of its sub-locks were 1274 * canceled and destroyed. In that case top-lock has to be enqueued again 1275 * before it can be used. 1276 * 1277 * Overall process of the locking during IO operation is as following: 1278 * 1279 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock() 1280 * is called on each layer. Responsibility of this method is to add locks, 1281 * needed by a given layer into cl_io.ci_lockset. 1282 * 1283 * - once locks for all layers were collected, they are sorted to avoid 1284 * dead-locks (cl_io_locks_sort()), and enqueued. 1285 * 1286 * - when all locks are acquired, IO is performed; 1287 * 1288 * - locks are released into cache. 1289 * 1290 * Striping introduces major additional complexity into locking. The 1291 * fundamental problem is that it is generally unsafe to actively use (hold) 1292 * two locks on the different OST servers at the same time, as this introduces 1293 * inter-server dependency and can lead to cascading evictions. 1294 * 1295 * Basic solution is to sub-divide large read/write IOs into smaller pieces so 1296 * that no multi-stripe locks are taken (note that this design abandons POSIX 1297 * read/write semantics). Such pieces ideally can be executed concurrently. At 1298 * the same time, certain types of IO cannot be sub-divived, without 1299 * sacrificing correctness. This includes: 1300 * 1301 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee 1302 * atomicity; 1303 * 1304 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken. 1305 * 1306 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where 1307 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf 1308 * has to be held together with the usual lock on [offset, offset + count]. 1309 * 1310 * As multi-stripe locks have to be allowed, it makes sense to cache them, so 1311 * that, for example, a sequence of O_APPEND writes can proceed quickly 1312 * without going down to the individual stripes to do lock matching. On the 1313 * other hand, multi-stripe locks shouldn't be used by normal read/write 1314 * calls. To achieve this, every layer can implement ->clo_fits_into() method, 1315 * that is called by lock matching code (cl_lock_lookup()), and that can be 1316 * used to selectively disable matching of certain locks for certain IOs. For 1317 * exmaple, lov layer implements lov_lock_fits_into() that allow multi-stripe 1318 * locks to be matched only for truncates and O_APPEND writes. 1319 * 1320 * Interaction with DLM 1321 * 1322 * In the expected setup, cl_lock is ultimately backed up by a collection of 1323 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is 1324 * implemented in osc layer, that also matches DLM events (ASTs, cancellation, 1325 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed 1326 * description of interaction with DLM. 1327 */ 1328 1329/** 1330 * Lock description. 1331 */ 1332struct cl_lock_descr { 1333 /** Object this lock is granted for. */ 1334 struct cl_object *cld_obj; 1335 /** Index of the first page protected by this lock. */ 1336 pgoff_t cld_start; 1337 /** Index of the last page (inclusive) protected by this lock. */ 1338 pgoff_t cld_end; 1339 /** Group ID, for group lock */ 1340 __u64 cld_gid; 1341 /** Lock mode. */ 1342 enum cl_lock_mode cld_mode; 1343 /** 1344 * flags to enqueue lock. A combination of bit-flags from 1345 * enum cl_enq_flags. 1346 */ 1347 __u32 cld_enq_flags; 1348}; 1349 1350#define DDESCR "%s(%d):[%lu, %lu]" 1351#define PDESCR(descr) \ 1352 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \ 1353 (descr)->cld_start, (descr)->cld_end 1354 1355const char *cl_lock_mode_name(const enum cl_lock_mode mode); 1356 1357/** 1358 * Lock state-machine states. 1359 * 1360 * \htmlonly 1361 * <pre> 1362 * 1363 * Possible state transitions: 1364 * 1365 * +------------------>NEW 1366 * | | 1367 * | | cl_enqueue_try() 1368 * | | 1369 * | cl_unuse_try() V 1370 * | +--------------QUEUING (*) 1371 * | | | 1372 * | | | cl_enqueue_try() 1373 * | | | 1374 * | | cl_unuse_try() V 1375 * sub-lock | +-------------ENQUEUED (*) 1376 * canceled | | | 1377 * | | | cl_wait_try() 1378 * | | | 1379 * | | (R) 1380 * | | | 1381 * | | V 1382 * | | HELD<---------+ 1383 * | | | | 1384 * | | | | cl_use_try() 1385 * | | cl_unuse_try() | | 1386 * | | | | 1387 * | | V ---+ 1388 * | +------------>INTRANSIT (D) <--+ 1389 * | | | 1390 * | cl_unuse_try() | | cached lock found 1391 * | | | cl_use_try() 1392 * | | | 1393 * | V | 1394 * +------------------CACHED---------+ 1395 * | 1396 * (C) 1397 * | 1398 * V 1399 * FREEING 1400 * 1401 * Legend: 1402 * 1403 * In states marked with (*) transition to the same state (i.e., a loop 1404 * in the diagram) is possible. 1405 * 1406 * (R) is the point where Receive call-back is invoked: it allows layers 1407 * to handle arrival of lock reply. 1408 * 1409 * (C) is the point where Cancellation call-back is invoked. 1410 * 1411 * (D) is the transit state which means the lock is changing. 1412 * 1413 * Transition to FREEING state is possible from any other state in the 1414 * diagram in case of unrecoverable error. 1415 * </pre> 1416 * \endhtmlonly 1417 * 1418 * These states are for individual cl_lock object. Top-lock and its sub-locks 1419 * can be in the different states. Another way to say this is that we have 1420 * nested state-machines. 1421 * 1422 * Separate QUEUING and ENQUEUED states are needed to support non-blocking 1423 * operation for locks with multiple sub-locks. Imagine lock on a file F, that 1424 * intersects 3 stripes S0, S1, and S2. To enqueue F client has to send 1425 * enqueue to S0, wait for its completion, then send enqueue for S1, wait for 1426 * its completion and at last enqueue lock for S2, and wait for its 1427 * completion. In that case, top-lock is in QUEUING state while S0, S1 are 1428 * handled, and is in ENQUEUED state after enqueue to S2 has been sent (note 1429 * that in this case, sub-locks move from state to state, and top-lock remains 1430 * in the same state). 1431 */ 1432enum cl_lock_state { 1433 /** 1434 * Lock that wasn't yet enqueued 1435 */ 1436 CLS_NEW, 1437 /** 1438 * Enqueue is in progress, blocking for some intermediate interaction 1439 * with the other side. 1440 */ 1441 CLS_QUEUING, 1442 /** 1443 * Lock is fully enqueued, waiting for server to reply when it is 1444 * granted. 1445 */ 1446 CLS_ENQUEUED, 1447 /** 1448 * Lock granted, actively used by some IO. 1449 */ 1450 CLS_HELD, 1451 /** 1452 * This state is used to mark the lock is being used, or unused. 1453 * We need this state because the lock may have several sublocks, 1454 * so it's impossible to have an atomic way to bring all sublocks 1455 * into CLS_HELD state at use case, or all sublocks to CLS_CACHED 1456 * at unuse case. 1457 * If a thread is referring to a lock, and it sees the lock is in this 1458 * state, it must wait for the lock. 1459 * See state diagram for details. 1460 */ 1461 CLS_INTRANSIT, 1462 /** 1463 * Lock granted, not used. 1464 */ 1465 CLS_CACHED, 1466 /** 1467 * Lock is being destroyed. 1468 */ 1469 CLS_FREEING, 1470 CLS_NR 1471}; 1472 1473enum cl_lock_flags { 1474 /** 1475 * lock has been cancelled. This flag is never cleared once set (by 1476 * cl_lock_cancel0()). 1477 */ 1478 CLF_CANCELLED = 1 << 0, 1479 /** cancellation is pending for this lock. */ 1480 CLF_CANCELPEND = 1 << 1, 1481 /** destruction is pending for this lock. */ 1482 CLF_DOOMED = 1 << 2, 1483 /** from enqueue RPC reply upcall. */ 1484 CLF_FROM_UPCALL= 1 << 3, 1485}; 1486 1487/** 1488 * Lock closure. 1489 * 1490 * Lock closure is a collection of locks (both top-locks and sub-locks) that 1491 * might be updated in a result of an operation on a certain lock (which lock 1492 * this is a closure of). 1493 * 1494 * Closures are needed to guarantee dead-lock freedom in the presence of 1495 * 1496 * - nested state-machines (top-lock state-machine composed of sub-lock 1497 * state-machines), and 1498 * 1499 * - shared sub-locks. 1500 * 1501 * Specifically, many operations, such as lock enqueue, wait, unlock, 1502 * etc. start from a top-lock, and then operate on a sub-locks of this 1503 * top-lock, holding a top-lock mutex. When sub-lock state changes as a result 1504 * of such operation, this change has to be propagated to all top-locks that 1505 * share this sub-lock. Obviously, no natural lock ordering (e.g., 1506 * top-to-bottom or bottom-to-top) captures this scenario, so try-locking has 1507 * to be used. Lock closure systematizes this try-and-repeat logic. 1508 */ 1509struct cl_lock_closure { 1510 /** 1511 * Lock that is mutexed when closure construction is started. When 1512 * closure in is `wait' mode (cl_lock_closure::clc_wait), mutex on 1513 * origin is released before waiting. 1514 */ 1515 struct cl_lock *clc_origin; 1516 /** 1517 * List of enclosed locks, so far. Locks are linked here through 1518 * cl_lock::cll_inclosure. 1519 */ 1520 struct list_head clc_list; 1521 /** 1522 * True iff closure is in a `wait' mode. This determines what 1523 * cl_lock_enclosure() does when a lock L to be added to the closure 1524 * is currently mutexed by some other thread. 1525 * 1526 * If cl_lock_closure::clc_wait is not set, then closure construction 1527 * fails with CLO_REPEAT immediately. 1528 * 1529 * In wait mode, cl_lock_enclosure() waits until next attempt to build 1530 * a closure might succeed. To this end it releases an origin mutex 1531 * (cl_lock_closure::clc_origin), that has to be the only lock mutex 1532 * owned by the current thread, and then waits on L mutex (by grabbing 1533 * it and immediately releasing), before returning CLO_REPEAT to the 1534 * caller. 1535 */ 1536 int clc_wait; 1537 /** Number of locks in the closure. */ 1538 int clc_nr; 1539}; 1540 1541/** 1542 * Layered client lock. 1543 */ 1544struct cl_lock { 1545 /** Reference counter. */ 1546 atomic_t cll_ref; 1547 /** List of slices. Immutable after creation. */ 1548 struct list_head cll_layers; 1549 /** 1550 * Linkage into cl_lock::cll_descr::cld_obj::coh_locks list. Protected 1551 * by cl_lock::cll_descr::cld_obj::coh_lock_guard. 1552 */ 1553 struct list_head cll_linkage; 1554 /** 1555 * Parameters of this lock. Protected by 1556 * cl_lock::cll_descr::cld_obj::coh_lock_guard nested within 1557 * cl_lock::cll_guard. Modified only on lock creation and in 1558 * cl_lock_modify(). 1559 */ 1560 struct cl_lock_descr cll_descr; 1561 /** Protected by cl_lock::cll_guard. */ 1562 enum cl_lock_state cll_state; 1563 /** signals state changes. */ 1564 wait_queue_head_t cll_wq; 1565 /** 1566 * Recursive lock, most fields in cl_lock{} are protected by this. 1567 * 1568 * Locking rules: this mutex is never held across network 1569 * communication, except when lock is being canceled. 1570 * 1571 * Lock ordering: a mutex of a sub-lock is taken first, then a mutex 1572 * on a top-lock. Other direction is implemented through a 1573 * try-lock-repeat loop. Mutices of unrelated locks can be taken only 1574 * by try-locking. 1575 * 1576 * \see osc_lock_enqueue_wait(), lov_lock_cancel(), lov_sublock_wait(). 1577 */ 1578 struct mutex cll_guard; 1579 task_t *cll_guarder; 1580 int cll_depth; 1581 1582 /** 1583 * the owner for INTRANSIT state 1584 */ 1585 task_t *cll_intransit_owner; 1586 int cll_error; 1587 /** 1588 * Number of holds on a lock. A hold prevents a lock from being 1589 * canceled and destroyed. Protected by cl_lock::cll_guard. 1590 * 1591 * \see cl_lock_hold(), cl_lock_unhold(), cl_lock_release() 1592 */ 1593 int cll_holds; 1594 /** 1595 * Number of lock users. Valid in cl_lock_state::CLS_HELD state 1596 * only. Lock user pins lock in CLS_HELD state. Protected by 1597 * cl_lock::cll_guard. 1598 * 1599 * \see cl_wait(), cl_unuse(). 1600 */ 1601 int cll_users; 1602 /** 1603 * Flag bit-mask. Values from enum cl_lock_flags. Updates are 1604 * protected by cl_lock::cll_guard. 1605 */ 1606 unsigned long cll_flags; 1607 /** 1608 * A linkage into a list of locks in a closure. 1609 * 1610 * \see cl_lock_closure 1611 */ 1612 struct list_head cll_inclosure; 1613 /** 1614 * Confict lock at queuing time. 1615 */ 1616 struct cl_lock *cll_conflict; 1617 /** 1618 * A list of references to this lock, for debugging. 1619 */ 1620 struct lu_ref cll_reference; 1621 /** 1622 * A list of holds on this lock, for debugging. 1623 */ 1624 struct lu_ref cll_holders; 1625 /** 1626 * A reference for cl_lock::cll_descr::cld_obj. For debugging. 1627 */ 1628 struct lu_ref_link *cll_obj_ref; 1629#ifdef CONFIG_LOCKDEP 1630 /* "dep_map" name is assumed by lockdep.h macros. */ 1631 struct lockdep_map dep_map; 1632#endif 1633}; 1634 1635/** 1636 * Per-layer part of cl_lock 1637 * 1638 * \see ccc_lock, lov_lock, lovsub_lock, osc_lock 1639 */ 1640struct cl_lock_slice { 1641 struct cl_lock *cls_lock; 1642 /** Object slice corresponding to this lock slice. Immutable after 1643 * creation. */ 1644 struct cl_object *cls_obj; 1645 const struct cl_lock_operations *cls_ops; 1646 /** Linkage into cl_lock::cll_layers. Immutable after creation. */ 1647 struct list_head cls_linkage; 1648}; 1649 1650/** 1651 * Possible (non-error) return values of ->clo_{enqueue,wait,unlock}(). 1652 * 1653 * NOTE: lov_subresult() depends on ordering here. 1654 */ 1655enum cl_lock_transition { 1656 /** operation cannot be completed immediately. Wait for state change. */ 1657 CLO_WAIT = 1, 1658 /** operation had to release lock mutex, restart. */ 1659 CLO_REPEAT = 2, 1660 /** lower layer re-enqueued. */ 1661 CLO_REENQUEUED = 3, 1662}; 1663 1664/** 1665 * 1666 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops 1667 */ 1668struct cl_lock_operations { 1669 /** 1670 * \name statemachine 1671 * 1672 * State machine transitions. These 3 methods are called to transfer 1673 * lock from one state to another, as described in the commentary 1674 * above enum #cl_lock_state. 1675 * 1676 * \retval 0 this layer has nothing more to do to before 1677 * transition to the target state happens; 1678 * 1679 * \retval CLO_REPEAT method had to release and re-acquire cl_lock 1680 * mutex, repeat invocation of transition method 1681 * across all layers; 1682 * 1683 * \retval CLO_WAIT this layer cannot move to the target state 1684 * immediately, as it has to wait for certain event 1685 * (e.g., the communication with the server). It 1686 * is guaranteed, that when the state transfer 1687 * becomes possible, cl_lock::cll_wq wait-queue 1688 * is signaled. Caller can wait for this event by 1689 * calling cl_lock_state_wait(); 1690 * 1691 * \retval -ve failure, abort state transition, move the lock 1692 * into cl_lock_state::CLS_FREEING state, and set 1693 * cl_lock::cll_error. 1694 * 1695 * Once all layers voted to agree to transition (by returning 0), lock 1696 * is moved into corresponding target state. All state transition 1697 * methods are optional. 1698 */ 1699 /** @{ */ 1700 /** 1701 * Attempts to enqueue the lock. Called top-to-bottom. 1702 * 1703 * \see ccc_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(), 1704 * \see osc_lock_enqueue() 1705 */ 1706 int (*clo_enqueue)(const struct lu_env *env, 1707 const struct cl_lock_slice *slice, 1708 struct cl_io *io, __u32 enqflags); 1709 /** 1710 * Attempts to wait for enqueue result. Called top-to-bottom. 1711 * 1712 * \see ccc_lock_wait(), lov_lock_wait(), osc_lock_wait() 1713 */ 1714 int (*clo_wait)(const struct lu_env *env, 1715 const struct cl_lock_slice *slice); 1716 /** 1717 * Attempts to unlock the lock. Called bottom-to-top. In addition to 1718 * usual return values of lock state-machine methods, this can return 1719 * -ESTALE to indicate that lock cannot be returned to the cache, and 1720 * has to be re-initialized. 1721 * unuse is a one-shot operation, so it must NOT return CLO_WAIT. 1722 * 1723 * \see ccc_lock_unuse(), lov_lock_unuse(), osc_lock_unuse() 1724 */ 1725 int (*clo_unuse)(const struct lu_env *env, 1726 const struct cl_lock_slice *slice); 1727 /** 1728 * Notifies layer that cached lock is started being used. 1729 * 1730 * \pre lock->cll_state == CLS_CACHED 1731 * 1732 * \see lov_lock_use(), osc_lock_use() 1733 */ 1734 int (*clo_use)(const struct lu_env *env, 1735 const struct cl_lock_slice *slice); 1736 /** @} statemachine */ 1737 /** 1738 * A method invoked when lock state is changed (as a result of state 1739 * transition). This is used, for example, to track when the state of 1740 * a sub-lock changes, to propagate this change to the corresponding 1741 * top-lock. Optional 1742 * 1743 * \see lovsub_lock_state() 1744 */ 1745 void (*clo_state)(const struct lu_env *env, 1746 const struct cl_lock_slice *slice, 1747 enum cl_lock_state st); 1748 /** 1749 * Returns true, iff given lock is suitable for the given io, idea 1750 * being, that there are certain "unsafe" locks, e.g., ones acquired 1751 * for O_APPEND writes, that we don't want to re-use for a normal 1752 * write, to avoid the danger of cascading evictions. Optional. Runs 1753 * under cl_object_header::coh_lock_guard. 1754 * 1755 * XXX this should take more information about lock needed by 1756 * io. Probably lock description or something similar. 1757 * 1758 * \see lov_fits_into() 1759 */ 1760 int (*clo_fits_into)(const struct lu_env *env, 1761 const struct cl_lock_slice *slice, 1762 const struct cl_lock_descr *need, 1763 const struct cl_io *io); 1764 /** 1765 * \name ast 1766 * Asynchronous System Traps. All of then are optional, all are 1767 * executed bottom-to-top. 1768 */ 1769 /** @{ */ 1770 1771 /** 1772 * Cancellation callback. Cancel a lock voluntarily, or under 1773 * the request of server. 1774 */ 1775 void (*clo_cancel)(const struct lu_env *env, 1776 const struct cl_lock_slice *slice); 1777 /** 1778 * Lock weighting ast. Executed to estimate how precious this lock 1779 * is. The sum of results across all layers is used to determine 1780 * whether lock worth keeping in cache given present memory usage. 1781 * 1782 * \see osc_lock_weigh(), vvp_lock_weigh(), lovsub_lock_weigh(). 1783 */ 1784 unsigned long (*clo_weigh)(const struct lu_env *env, 1785 const struct cl_lock_slice *slice); 1786 /** @} ast */ 1787 1788 /** 1789 * \see lovsub_lock_closure() 1790 */ 1791 int (*clo_closure)(const struct lu_env *env, 1792 const struct cl_lock_slice *slice, 1793 struct cl_lock_closure *closure); 1794 /** 1795 * Executed bottom-to-top when lock description changes (e.g., as a 1796 * result of server granting more generous lock than was requested). 1797 * 1798 * \see lovsub_lock_modify() 1799 */ 1800 int (*clo_modify)(const struct lu_env *env, 1801 const struct cl_lock_slice *slice, 1802 const struct cl_lock_descr *updated); 1803 /** 1804 * Notifies layers (bottom-to-top) that lock is going to be 1805 * destroyed. Responsibility of layers is to prevent new references on 1806 * this lock from being acquired once this method returns. 1807 * 1808 * This can be called multiple times due to the races. 1809 * 1810 * \see cl_lock_delete() 1811 * \see osc_lock_delete(), lovsub_lock_delete() 1812 */ 1813 void (*clo_delete)(const struct lu_env *env, 1814 const struct cl_lock_slice *slice); 1815 /** 1816 * Destructor. Frees resources and the slice. 1817 * 1818 * \see ccc_lock_fini(), lov_lock_fini(), lovsub_lock_fini(), 1819 * \see osc_lock_fini() 1820 */ 1821 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice); 1822 /** 1823 * Optional debugging helper. Prints given lock slice. 1824 */ 1825 int (*clo_print)(const struct lu_env *env, 1826 void *cookie, lu_printer_t p, 1827 const struct cl_lock_slice *slice); 1828}; 1829 1830#define CL_LOCK_DEBUG(mask, env, lock, format, ...) \ 1831do { \ 1832 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \ 1833 \ 1834 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \ 1835 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \ 1836 CDEBUG(mask, format , ## __VA_ARGS__); \ 1837 } \ 1838} while (0) 1839 1840#define CL_LOCK_ASSERT(expr, env, lock) do { \ 1841 if (likely(expr)) \ 1842 break; \ 1843 \ 1844 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \ 1845 LBUG(); \ 1846} while (0) 1847 1848/** @} cl_lock */ 1849 1850/** \addtogroup cl_page_list cl_page_list 1851 * Page list used to perform collective operations on a group of pages. 1852 * 1853 * Pages are added to the list one by one. cl_page_list acquires a reference 1854 * for every page in it. Page list is used to perform collective operations on 1855 * pages: 1856 * 1857 * - submit pages for an immediate transfer, 1858 * 1859 * - own pages on behalf of certain io (waiting for each page in turn), 1860 * 1861 * - discard pages. 1862 * 1863 * When list is finalized, it releases references on all pages it still has. 1864 * 1865 * \todo XXX concurrency control. 1866 * 1867 * @{ 1868 */ 1869struct cl_page_list { 1870 unsigned pl_nr; 1871 struct list_head pl_pages; 1872 task_t *pl_owner; 1873}; 1874 1875/** 1876 * A 2-queue of pages. A convenience data-type for common use case, 2-queue 1877 * contains an incoming page list and an outgoing page list. 1878 */ 1879struct cl_2queue { 1880 struct cl_page_list c2_qin; 1881 struct cl_page_list c2_qout; 1882}; 1883 1884/** @} cl_page_list */ 1885 1886/** \addtogroup cl_io cl_io 1887 * @{ */ 1888/** \struct cl_io 1889 * I/O 1890 * 1891 * cl_io represents a high level I/O activity like 1892 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent 1893 * lock. 1894 * 1895 * cl_io is a layered object, much like cl_{object,page,lock} but with one 1896 * important distinction. We want to minimize number of calls to the allocator 1897 * in the fast path, e.g., in the case of read(2) when everything is cached: 1898 * client already owns the lock over region being read, and data are cached 1899 * due to read-ahead. To avoid allocation of cl_io layers in such situations, 1900 * per-layer io state is stored in the session, associated with the io, see 1901 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized 1902 * by using free-lists, see cl_env_get(). 1903 * 1904 * There is a small predefined number of possible io types, enumerated in enum 1905 * cl_io_type. 1906 * 1907 * cl_io is a state machine, that can be advanced concurrently by the multiple 1908 * threads. It is up to these threads to control the concurrency and, 1909 * specifically, to detect when io is done, and its state can be safely 1910 * released. 1911 * 1912 * For read/write io overall execution plan is as following: 1913 * 1914 * (0) initialize io state through all layers; 1915 * 1916 * (1) loop: prepare chunk of work to do 1917 * 1918 * (2) call all layers to collect locks they need to process current chunk 1919 * 1920 * (3) sort all locks to avoid dead-locks, and acquire them 1921 * 1922 * (4) process the chunk: call per-page methods 1923 * (cl_io_operations::cio_read_page() for read, 1924 * cl_io_operations::cio_prepare_write(), 1925 * cl_io_operations::cio_commit_write() for write) 1926 * 1927 * (5) release locks 1928 * 1929 * (6) repeat loop. 1930 * 1931 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to 1932 * address allocation efficiency issues mentioned above), and returns with the 1933 * special error condition from per-page method when current sub-io has to 1934 * block. This causes io loop to be repeated, and lov switches to the next 1935 * sub-io in its cl_io_operations::cio_iter_init() implementation. 1936 */ 1937 1938/** IO types */ 1939enum cl_io_type { 1940 /** read system call */ 1941 CIT_READ, 1942 /** write system call */ 1943 CIT_WRITE, 1944 /** truncate, utime system calls */ 1945 CIT_SETATTR, 1946 /** 1947 * page fault handling 1948 */ 1949 CIT_FAULT, 1950 /** 1951 * fsync system call handling 1952 * To write out a range of file 1953 */ 1954 CIT_FSYNC, 1955 /** 1956 * Miscellaneous io. This is used for occasional io activity that 1957 * doesn't fit into other types. Currently this is used for: 1958 * 1959 * - cancellation of an extent lock. This io exists as a context 1960 * to write dirty pages from under the lock being canceled back 1961 * to the server; 1962 * 1963 * - VM induced page write-out. An io context for writing page out 1964 * for memory cleansing; 1965 * 1966 * - glimpse. An io context to acquire glimpse lock. 1967 * 1968 * - grouplock. An io context to acquire group lock. 1969 * 1970 * CIT_MISC io is used simply as a context in which locks and pages 1971 * are manipulated. Such io has no internal "process", that is, 1972 * cl_io_loop() is never called for it. 1973 */ 1974 CIT_MISC, 1975 CIT_OP_NR 1976}; 1977 1978/** 1979 * States of cl_io state machine 1980 */ 1981enum cl_io_state { 1982 /** Not initialized. */ 1983 CIS_ZERO, 1984 /** Initialized. */ 1985 CIS_INIT, 1986 /** IO iteration started. */ 1987 CIS_IT_STARTED, 1988 /** Locks taken. */ 1989 CIS_LOCKED, 1990 /** Actual IO is in progress. */ 1991 CIS_IO_GOING, 1992 /** IO for the current iteration finished. */ 1993 CIS_IO_FINISHED, 1994 /** Locks released. */ 1995 CIS_UNLOCKED, 1996 /** Iteration completed. */ 1997 CIS_IT_ENDED, 1998 /** cl_io finalized. */ 1999 CIS_FINI 2000};
2001 2002/** 2003 * IO state private for a layer. 2004 * 2005 * This is usually embedded into layer session data, rather than allocated 2006 * dynamically. 2007 * 2008 * \see vvp_io, lov_io, osc_io, ccc_io 2009 */ 2010struct cl_io_slice { 2011 struct cl_io *cis_io; 2012 /** corresponding object slice. Immutable after creation. */ 2013 struct cl_object *cis_obj; 2014 /** io operations. Immutable after creation. */ 2015 const struct cl_io_operations *cis_iop; 2016 /** 2017 * linkage into a list of all slices for a given cl_io, hanging off 2018 * cl_io::ci_layers. Immutable after creation. 2019 */ 2020 struct list_head cis_linkage; 2021}; 2022 2023 2024/** 2025 * Per-layer io operations. 2026 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops 2027 */ 2028struct cl_io_operations { 2029 /** 2030 * Vector of io state transition methods for every io type. 2031 * 2032 * \see cl_page_operations::io 2033 */ 2034 struct { 2035 /** 2036 * Prepare io iteration at a given layer. 2037 * 2038 * Called top-to-bottom at the beginning of each iteration of 2039 * "io loop" (if it makes sense for this type of io). Here 2040 * layer selects what work it will do during this iteration. 2041 * 2042 * \see cl_io_operations::cio_iter_fini() 2043 */ 2044 int (*cio_iter_init) (const struct lu_env *env, 2045 const struct cl_io_slice *slice); 2046 /** 2047 * Finalize io iteration. 2048 * 2049 * Called bottom-to-top at the end of each iteration of "io 2050 * loop". Here layers can decide whether IO has to be 2051 * continued. 2052 * 2053 * \see cl_io_operations::cio_iter_init() 2054 */ 2055 void (*cio_iter_fini) (const struct lu_env *env, 2056 const struct cl_io_slice *slice); 2057 /** 2058 * Collect locks for the current iteration of io. 2059 * 2060 * Called top-to-bottom to collect all locks necessary for 2061 * this iteration. This methods shouldn't actually enqueue 2062 * anything, instead it should post a lock through 2063 * cl_io_lock_add(). Once all locks are collected, they are 2064 * sorted and enqueued in the proper order. 2065 */ 2066 int (*cio_lock) (const struct lu_env *env, 2067 const struct cl_io_slice *slice); 2068 /** 2069 * Finalize unlocking. 2070 * 2071 * Called bottom-to-top to finish layer specific unlocking 2072 * functionality, after generic code released all locks 2073 * acquired by cl_io_operations::cio_lock(). 2074 */ 2075 void (*cio_unlock)(const struct lu_env *env, 2076 const struct cl_io_slice *slice); 2077 /** 2078 * Start io iteration. 2079 * 2080 * Once all locks are acquired, called top-to-bottom to 2081 * commence actual IO. In the current implementation, 2082 * top-level vvp_io_{read,write}_start() does all the work 2083 * synchronously by calling generic_file_*(), so other layers 2084 * are called when everything is done. 2085 */ 2086 int (*cio_start)(const struct lu_env *env, 2087 const struct cl_io_slice *slice); 2088 /** 2089 * Called top-to-bottom at the end of io loop. Here layer 2090 * might wait for an unfinished asynchronous io. 2091 */ 2092 void (*cio_end) (const struct lu_env *env, 2093 const struct cl_io_slice *slice); 2094 /** 2095 * Called bottom-to-top to notify layers that read/write IO 2096 * iteration finished, with \a nob bytes transferred. 2097 */ 2098 void (*cio_advance)(const struct lu_env *env, 2099 const struct cl_io_slice *slice, 2100 size_t nob); 2101 /** 2102 * Called once per io, bottom-to-top to release io resources. 2103 */ 2104 void (*cio_fini) (const struct lu_env *env, 2105 const struct cl_io_slice *slice); 2106 } op[CIT_OP_NR]; 2107 struct { 2108 /** 2109 * Submit pages from \a queue->c2_qin for IO, and move 2110 * successfully submitted pages into \a queue->c2_qout. Return 2111 * non-zero if failed to submit even the single page. If 2112 * submission failed after some pages were moved into \a 2113 * queue->c2_qout, completion callback with non-zero ioret is 2114 * executed on them. 2115 */ 2116 int (*cio_submit)(const struct lu_env *env, 2117 const struct cl_io_slice *slice, 2118 enum cl_req_type crt, 2119 struct cl_2queue *queue); 2120 } req_op[CRT_NR]; 2121 /** 2122 * Read missing page. 2123 * 2124 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start() 2125 * method, when it hits not-up-to-date page in the range. Optional. 2126 * 2127 * \pre io->ci_type == CIT_READ 2128 */ 2129 int (*cio_read_page)(const struct lu_env *env, 2130 const struct cl_io_slice *slice, 2131 const struct cl_page_slice *page); 2132 /** 2133 * Prepare write of a \a page. Called bottom-to-top by a top-level 2134 * cl_io_operations::op[CIT_WRITE]::cio_start() to prepare page for 2135 * get data from user-level buffer. 2136 * 2137 * \pre io->ci_type == CIT_WRITE 2138 * 2139 * \see vvp_io_prepare_write(), lov_io_prepare_write(), 2140 * osc_io_prepare_write(). 2141 */ 2142 int (*cio_prepare_write)(const struct lu_env *env, 2143 const struct cl_io_slice *slice, 2144 const struct cl_page_slice *page, 2145 unsigned from, unsigned to); 2146 /** 2147 * 2148 * \pre io->ci_type == CIT_WRITE 2149 * 2150 * \see vvp_io_commit_write(), lov_io_commit_write(), 2151 * osc_io_commit_write(). 2152 */ 2153 int (*cio_commit_write)(const struct lu_env *env, 2154 const struct cl_io_slice *slice, 2155 const struct cl_page_slice *page, 2156 unsigned from, unsigned to); 2157 /** 2158 * Optional debugging helper. Print given io slice. 2159 */ 2160 int (*cio_print)(const struct lu_env *env, void *cookie, 2161 lu_printer_t p, const struct cl_io_slice *slice); 2162}; 2163 2164/** 2165 * Flags to lock enqueue procedure. 2166 * \ingroup cl_lock 2167 */ 2168enum cl_enq_flags { 2169 /** 2170 * instruct server to not block, if conflicting lock is found. Instead 2171 * -EWOULDBLOCK is returned immediately. 2172 */ 2173 CEF_NONBLOCK = 0x00000001, 2174 /** 2175 * take lock asynchronously (out of order), as it cannot 2176 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing. 2177 */ 2178 CEF_ASYNC = 0x00000002, 2179 /** 2180 * tell the server to instruct (though a flag in the blocking ast) an 2181 * owner of the conflicting lock, that it can drop dirty pages 2182 * protected by this lock, without sending them to the server. 2183 */ 2184 CEF_DISCARD_DATA = 0x00000004, 2185 /** 2186 * tell the sub layers that it must be a `real' lock. This is used for 2187 * mmapped-buffer locks and glimpse locks that must be never converted 2188 * into lockless mode. 2189 * 2190 * \see vvp_mmap_locks(), cl_glimpse_lock(). 2191 */ 2192 CEF_MUST = 0x00000008, 2193 /** 2194 * tell the sub layers that never request a `real' lock. This flag is 2195 * not used currently. 2196 * 2197 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless 2198 * conversion policy: ci_lockreq describes generic information of lock 2199 * requirement for this IO, especially for locks which belong to the 2200 * object doing IO; however, lock itself may have precise requirements 2201 * that are described by the enqueue flags. 2202 */ 2203 CEF_NEVER = 0x00000010, 2204 /** 2205 * for async glimpse lock. 2206 */ 2207 CEF_AGL = 0x00000020, 2208 /** 2209 * mask of enq_flags. 2210 */ 2211 CEF_MASK = 0x0000003f, 2212}; 2213 2214/** 2215 * Link between lock and io. Intermediate structure is needed, because the 2216 * same lock can be part of multiple io's simultaneously. 2217 */ 2218struct cl_io_lock_link { 2219 /** linkage into one of cl_lockset lists. */ 2220 struct list_head cill_linkage; 2221 struct cl_lock_descr cill_descr; 2222 struct cl_lock *cill_lock; 2223 /** optional destructor */ 2224 void (*cill_fini)(const struct lu_env *env, 2225 struct cl_io_lock_link *link); 2226}; 2227 2228/** 2229 * Lock-set represents a collection of locks, that io needs at a 2230 * time. Generally speaking, client tries to avoid holding multiple locks when 2231 * possible, because 2232 * 2233 * - holding extent locks over multiple ost's introduces the danger of 2234 * "cascading timeouts"; 2235 * 2236 * - holding multiple locks over the same ost is still dead-lock prone, 2237 * see comment in osc_lock_enqueue(), 2238 * 2239 * but there are certain situations where this is unavoidable: 2240 * 2241 * - O_APPEND writes have to take [0, EOF] lock for correctness; 2242 * 2243 * - truncate has to take [new-size, EOF] lock for correctness; 2244 * 2245 * - SNS has to take locks across full stripe for correctness; 2246 * 2247 * - in the case when user level buffer, supplied to {read,write}(file0), 2248 * is a part of a memory mapped lustre file, client has to take a dlm 2249 * locks on file0, and all files that back up the buffer (or a part of 2250 * the buffer, that is being processed in the current chunk, in any 2251 * case, there are situations where at least 2 locks are necessary). 2252 * 2253 * In such cases we at least try to take locks in the same consistent 2254 * order. To this end, all locks are first collected, then sorted, and then 2255 * enqueued. 2256 */ 2257struct cl_lockset { 2258 /** locks to be acquired. */ 2259 struct list_head cls_todo; 2260 /** locks currently being processed. */ 2261 struct list_head cls_curr; 2262 /** locks acquired. */ 2263 struct list_head cls_done; 2264}; 2265 2266/** 2267 * Lock requirements(demand) for IO. It should be cl_io_lock_req, 2268 * but 'req' is always to be thought as 'request' :-) 2269 */ 2270enum cl_io_lock_dmd { 2271 /** Always lock data (e.g., O_APPEND). */ 2272 CILR_MANDATORY = 0, 2273 /** Layers are free to decide between local and global locking. */ 2274 CILR_MAYBE, 2275 /** Never lock: there is no cache (e.g., liblustre). */ 2276 CILR_NEVER 2277}; 2278 2279enum cl_fsync_mode { 2280 /** start writeback, do not wait for them to finish */ 2281 CL_FSYNC_NONE = 0, 2282 /** start writeback and wait for them to finish */ 2283 CL_FSYNC_LOCAL = 1, 2284 /** discard all of dirty pages in a specific file range */ 2285 CL_FSYNC_DISCARD = 2, 2286 /** start writeback and make sure they have reached storage before 2287 * return. OST_SYNC RPC must be issued and finished */ 2288 CL_FSYNC_ALL = 3 2289}; 2290 2291struct cl_io_rw_common { 2292 loff_t crw_pos; 2293 size_t crw_count; 2294 int crw_nonblock; 2295}; 2296 2297 2298/** 2299 * State for io. 2300 * 2301 * cl_io is shared by all threads participating in this IO (in current 2302 * implementation only one thread advances IO, but parallel IO design and 2303 * concurrent copy_*_user() require multiple threads acting on the same IO. It 2304 * is up to these threads to serialize their activities, including updates to 2305 * mutable cl_io fields. 2306 */ 2307struct cl_io { 2308 /** type of this IO. Immutable after creation. */ 2309 enum cl_io_type ci_type; 2310 /** current state of cl_io state machine. */ 2311 enum cl_io_state ci_state; 2312 /** main object this io is against. Immutable after creation. */ 2313 struct cl_object *ci_obj; 2314 /** 2315 * Upper layer io, of which this io is a part of. Immutable after 2316 * creation. 2317 */ 2318 struct cl_io *ci_parent; 2319 /** List of slices. Immutable after creation. */ 2320 struct list_head ci_layers; 2321 /** list of locks (to be) acquired by this io. */ 2322 struct cl_lockset ci_lockset; 2323 /** lock requirements, this is just a help info for sublayers. */ 2324 enum cl_io_lock_dmd ci_lockreq; 2325 union { 2326 struct cl_rd_io { 2327 struct cl_io_rw_common rd; 2328 } ci_rd; 2329 struct cl_wr_io { 2330 struct cl_io_rw_common wr; 2331 int wr_append; 2332 int wr_sync; 2333 } ci_wr; 2334 struct cl_io_rw_common ci_rw; 2335 struct cl_setattr_io { 2336 struct ost_lvb sa_attr; 2337 unsigned int sa_valid; 2338 struct obd_capa *sa_capa; 2339 } ci_setattr; 2340 struct cl_fault_io { 2341 /** page index within file. */ 2342 pgoff_t ft_index; 2343 /** bytes valid byte on a faulted page. */ 2344 int ft_nob; 2345 /** writable page? for nopage() only */ 2346 int ft_writable; 2347 /** page of an executable? */ 2348 int ft_executable; 2349 /** page_mkwrite() */ 2350 int ft_mkwrite; 2351 /** resulting page */ 2352 struct cl_page *ft_page; 2353 } ci_fault; 2354 struct cl_fsync_io { 2355 loff_t fi_start; 2356 loff_t fi_end; 2357 struct obd_capa *fi_capa; 2358 /** file system level fid */ 2359 struct lu_fid *fi_fid; 2360 enum cl_fsync_mode fi_mode; 2361 /* how many pages were written/discarded */ 2362 unsigned int fi_nr_written; 2363 } ci_fsync; 2364 } u; 2365 struct cl_2queue ci_queue; 2366 size_t ci_nob; 2367 int ci_result; 2368 unsigned int ci_continue:1, 2369 /** 2370 * This io has held grouplock, to inform sublayers that 2371 * don't do lockless i/o. 2372 */ 2373 ci_no_srvlock:1, 2374 /** 2375 * The whole IO need to be restarted because layout has been changed 2376 */ 2377 ci_need_restart:1, 2378 /** 2379 * to not refresh layout - the IO issuer knows that the layout won't 2380 * change(page operations, layout change causes all page to be 2381 * discarded), or it doesn't matter if it changes(sync). 2382 */ 2383 ci_ignore_layout:1, 2384 /** 2385 * Check if layout changed after the IO finishes. Mainly for HSM 2386 * requirement. If IO occurs to openning files, it doesn't need to 2387 * verify layout because HSM won't release openning files. 2388 * Right now, only two opertaions need to verify layout: glimpse 2389 * and setattr. 2390 */ 2391 ci_verify_layout:1; 2392 /** 2393 * Number of pages owned by this IO. For invariant checking. 2394 */ 2395 unsigned ci_owned_nr; 2396}; 2397 2398/** @} cl_io */ 2399 2400/** \addtogroup cl_req cl_req 2401 * @{ */ 2402/** \struct cl_req 2403 * Transfer. 2404 * 2405 * There are two possible modes of transfer initiation on the client: 2406 * 2407 * - immediate transfer: this is started when a high level io wants a page 2408 * or a collection of pages to be transferred right away. Examples: 2409 * read-ahead, synchronous read in the case of non-page aligned write, 2410 * page write-out as a part of extent lock cancellation, page write-out 2411 * as a part of memory cleansing. Immediate transfer can be both 2412 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE; 2413 * 2414 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens 2415 * when io wants to transfer a page to the server some time later, when 2416 * it can be done efficiently. Example: pages dirtied by the write(2) 2417 * path. 2418 * 2419 * In any case, transfer takes place in the form of a cl_req, which is a 2420 * representation for a network RPC. 2421 * 2422 * Pages queued for an opportunistic transfer are cached until it is decided 2423 * that efficient RPC can be composed of them. This decision is made by "a 2424 * req-formation engine", currently implemented as a part of osc 2425 * layer. Req-formation depends on many factors: the size of the resulting 2426 * RPC, whether or not multi-object RPCs are supported by the server, 2427 * max-rpc-in-flight limitations, size of the dirty cache, etc. 2428 * 2429 * For the immediate transfer io submits a cl_page_list, that req-formation 2430 * engine slices into cl_req's, possibly adding cached pages to some of 2431 * the resulting req's. 2432 * 2433 * Whenever a page from cl_page_list is added to a newly constructed req, its 2434 * cl_page_operations::cpo_prep() layer methods are called. At that moment, 2435 * page state is atomically changed from cl_page_state::CPS_OWNED to 2436 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner 2437 * is zeroed, and cl_page::cp_req is set to the 2438 * req. cl_page_operations::cpo_prep() method at the particular layer might 2439 * return -EALREADY to indicate that it does not need to submit this page 2440 * at all. This is possible, for example, if page, submitted for read, 2441 * became up-to-date in the meantime; and for write, the page don't have 2442 * dirty bit marked. \see cl_io_submit_rw() 2443 * 2444 * Whenever a cached page is added to a newly constructed req, its 2445 * cl_page_operations::cpo_make_ready() layer methods are called. At that 2446 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to 2447 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to 2448 * req. cl_page_operations::cpo_make_ready() method at the particular layer 2449 * might return -EAGAIN to indicate that this page is not eligible for the 2450 * transfer right now. 2451 * 2452 * FUTURE 2453 * 2454 * Plan is to divide transfers into "priority bands" (indicated when 2455 * submitting cl_page_list, and queuing a page for the opportunistic transfer) 2456 * and allow glueing of cached pages to immediate transfers only within single 2457 * band. This would make high priority transfers (like lock cancellation or 2458 * memory pressure induced write-out) really high priority. 2459 * 2460 */ 2461 2462/** 2463 * Per-transfer attributes. 2464 */ 2465struct cl_req_attr { 2466 /** Generic attributes for the server consumption. */ 2467 struct obdo *cra_oa; 2468 /** Capability. */ 2469 struct obd_capa *cra_capa; 2470 /** Jobid */ 2471 char cra_jobid[JOBSTATS_JOBID_SIZE]; 2472}; 2473 2474/** 2475 * Transfer request operations definable at every layer. 2476 * 2477 * Concurrency: transfer formation engine synchronizes calls to all transfer 2478 * methods. 2479 */ 2480struct cl_req_operations { 2481 /** 2482 * Invoked top-to-bottom by cl_req_prep() when transfer formation is 2483 * complete (all pages are added). 2484 * 2485 * \see osc_req_prep() 2486 */ 2487 int (*cro_prep)(const struct lu_env *env, 2488 const struct cl_req_slice *slice); 2489 /** 2490 * Called top-to-bottom to fill in \a oa fields. This is called twice 2491 * with different flags, see bug 10150 and osc_build_req(). 2492 * 2493 * \param obj an object from cl_req which attributes are to be set in 2494 * \a oa. 2495 * 2496 * \param oa struct obdo where attributes are placed 2497 * 2498 * \param flags \a oa fields to be filled. 2499 */ 2500 void (*cro_attr_set)(const struct lu_env *env, 2501 const struct cl_req_slice *slice, 2502 const struct cl_object *obj, 2503 struct cl_req_attr *attr, obd_valid flags); 2504 /** 2505 * Called top-to-bottom from cl_req_completion() to notify layers that 2506 * transfer completed. Has to free all state allocated by 2507 * cl_device_operations::cdo_req_init(). 2508 */ 2509 void (*cro_completion)(const struct lu_env *env, 2510 const struct cl_req_slice *slice, int ioret); 2511}; 2512 2513/** 2514 * A per-object state that (potentially multi-object) transfer request keeps. 2515 */ 2516struct cl_req_obj { 2517 /** object itself */ 2518 struct cl_object *ro_obj; 2519 /** reference to cl_req_obj::ro_obj. For debugging. */ 2520 struct lu_ref_link *ro_obj_ref; 2521 /* something else? Number of pages for a given object? */ 2522}; 2523 2524/** 2525 * Transfer request. 2526 * 2527 * Transfer requests are not reference counted, because IO sub-system owns 2528 * them exclusively and knows when to free them. 2529 * 2530 * Life cycle. 2531 * 2532 * cl_req is created by cl_req_alloc() that calls 2533 * cl_device_operations::cdo_req_init() device methods to allocate per-req 2534 * state in every layer. 2535 * 2536 * Then pages are added (cl_req_page_add()), req keeps track of all objects it 2537 * contains pages for. 2538 * 2539 * Once all pages were collected, cl_page_operations::cpo_prep() method is 2540 * called top-to-bottom. At that point layers can modify req, let it pass, or 2541 * deny it completely. This is to support things like SNS that have transfer 2542 * ordering requirements invisible to the individual req-formation engine. 2543 * 2544 * On transfer completion (or transfer timeout, or failure to initiate the 2545 * transfer of an allocated req), cl_req_operations::cro_completion() method 2546 * is called, after execution of cl_page_operations::cpo_completion() of all 2547 * req's pages. 2548 */ 2549struct cl_req { 2550 enum cl_req_type crq_type; 2551 /** A list of pages being transfered */ 2552 struct list_head crq_pages; 2553 /** Number of pages in cl_req::crq_pages */ 2554 unsigned crq_nrpages; 2555 /** An array of objects which pages are in ->crq_pages */ 2556 struct cl_req_obj *crq_o; 2557 /** Number of elements in cl_req::crq_objs[] */ 2558 unsigned crq_nrobjs; 2559 struct list_head crq_layers; 2560}; 2561 2562/** 2563 * Per-layer state for request. 2564 */ 2565struct cl_req_slice { 2566 struct cl_req *crs_req; 2567 struct cl_device *crs_dev; 2568 struct list_head crs_linkage; 2569 const struct cl_req_operations *crs_ops; 2570}; 2571 2572/* @} cl_req */ 2573 2574enum cache_stats_item { 2575 /** how many cache lookups were performed */ 2576 CS_lookup = 0, 2577 /** how many times cache lookup resulted in a hit */ 2578 CS_hit, 2579 /** how many entities are in the cache right now */ 2580 CS_total, 2581 /** how many entities in the cache are actively used (and cannot be 2582 * evicted) right now */ 2583 CS_busy, 2584 /** how many entities were created at all */ 2585 CS_create, 2586 CS_NR 2587}; 2588 2589#define CS_NAMES { "lookup", "hit", "total", "busy", "create" } 2590 2591/** 2592 * Stats for a generic cache (similar to inode, lu_object, etc. caches). 2593 */ 2594struct cache_stats { 2595 const char *cs_name; 2596 atomic_t cs_stats[CS_NR]; 2597}; 2598 2599/** These are not exported so far */ 2600void cache_stats_init (struct cache_stats *cs, const char *name); 2601 2602/** 2603 * Client-side site. This represents particular client stack. "Global" 2604 * variables should (directly or indirectly) be added here to allow multiple 2605 * clients to co-exist in the single address space. 2606 */ 2607struct cl_site { 2608 struct lu_site cs_lu; 2609 /** 2610 * Statistical counters. Atomics do not scale, something better like 2611 * per-cpu counters is needed. 2612 * 2613 * These are exported as /proc/fs/lustre/llite/.../site 2614 * 2615 * When interpreting keep in mind that both sub-locks (and sub-pages) 2616 * and top-locks (and top-pages) are accounted here. 2617 */ 2618 struct cache_stats cs_pages; 2619 struct cache_stats cs_locks; 2620 atomic_t cs_pages_state[CPS_NR]; 2621 atomic_t cs_locks_state[CLS_NR]; 2622}; 2623 2624int cl_site_init (struct cl_site *s, struct cl_device *top); 2625void cl_site_fini (struct cl_site *s); 2626void cl_stack_fini(const struct lu_env *env, struct cl_device *cl); 2627 2628/** 2629 * Output client site statistical counters into a buffer. Suitable for 2630 * ll_rd_*()-style functions. 2631 */ 2632int cl_site_stats_print(const struct cl_site *site, struct seq_file *m); 2633 2634/** 2635 * \name helpers 2636 * 2637 * Type conversion and accessory functions. 2638 */ 2639/** @{ */ 2640 2641static inline struct cl_site *lu2cl_site(const struct lu_site *site) 2642{ 2643 return container_of(site, struct cl_site, cs_lu); 2644} 2645 2646static inline int lu_device_is_cl(const struct lu_device *d) 2647{ 2648 return d->ld_type->ldt_tags & LU_DEVICE_CL; 2649} 2650 2651static inline struct cl_device *lu2cl_dev(const struct lu_device *d) 2652{ 2653 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d)); 2654 return container_of0(d, struct cl_device, cd_lu_dev); 2655} 2656 2657static inline struct lu_device *cl2lu_dev(struct cl_device *d) 2658{ 2659 return &d->cd_lu_dev; 2660} 2661 2662static inline struct cl_object *lu2cl(const struct lu_object *o) 2663{ 2664 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev)); 2665 return container_of0(o, struct cl_object, co_lu); 2666} 2667 2668static inline const struct cl_object_conf * 2669lu2cl_conf(const struct lu_object_conf *conf) 2670{ 2671 return container_of0(conf, struct cl_object_conf, coc_lu); 2672} 2673 2674static inline struct cl_object *cl_object_next(const struct cl_object *obj) 2675{ 2676 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL; 2677} 2678 2679static inline struct cl_device *cl_object_device(const struct cl_object *o) 2680{ 2681 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev)); 2682 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev); 2683} 2684 2685static inline struct cl_object_header *luh2coh(const struct lu_object_header *h) 2686{ 2687 return container_of0(h, struct cl_object_header, coh_lu); 2688} 2689 2690static inline struct cl_site *cl_object_site(const struct cl_object *obj) 2691{ 2692 return lu2cl_site(obj->co_lu.lo_dev->ld_site); 2693} 2694 2695static inline 2696struct cl_object_header *cl_object_header(const struct cl_object *obj) 2697{ 2698 return luh2coh(obj->co_lu.lo_header); 2699} 2700 2701static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t) 2702{ 2703 return lu_device_init(&d->cd_lu_dev, t); 2704} 2705 2706static inline void cl_device_fini(struct cl_device *d) 2707{ 2708 lu_device_fini(&d->cd_lu_dev); 2709} 2710 2711void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice, 2712 struct cl_object *obj, 2713 const struct cl_page_operations *ops); 2714void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice, 2715 struct cl_object *obj, 2716 const struct cl_lock_operations *ops); 2717void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice, 2718 struct cl_object *obj, const struct cl_io_operations *ops); 2719void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice, 2720 struct cl_device *dev, 2721 const struct cl_req_operations *ops); 2722/** @} helpers */ 2723 2724/** \defgroup cl_object cl_object 2725 * @{ */ 2726struct cl_object *cl_object_top (struct cl_object *o); 2727struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd, 2728 const struct lu_fid *fid, 2729 const struct cl_object_conf *c); 2730 2731int cl_object_header_init(struct cl_object_header *h); 2732void cl_object_header_fini(struct cl_object_header *h); 2733void cl_object_put (const struct lu_env *env, struct cl_object *o); 2734void cl_object_get (struct cl_object *o); 2735void cl_object_attr_lock (struct cl_object *o); 2736void cl_object_attr_unlock(struct cl_object *o); 2737int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj, 2738 struct cl_attr *attr); 2739int cl_object_attr_set (const struct lu_env *env, struct cl_object *obj, 2740 const struct cl_attr *attr, unsigned valid); 2741int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj, 2742 struct ost_lvb *lvb); 2743int cl_conf_set (const struct lu_env *env, struct cl_object *obj, 2744 const struct cl_object_conf *conf); 2745void cl_object_prune (const struct lu_env *env, struct cl_object *obj); 2746void cl_object_kill (const struct lu_env *env, struct cl_object *obj); 2747int cl_object_has_locks (struct cl_object *obj); 2748 2749/** 2750 * Returns true, iff \a o0 and \a o1 are slices of the same object. 2751 */ 2752static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1) 2753{ 2754 return cl_object_header(o0) == cl_object_header(o1); 2755} 2756 2757static inline void cl_object_page_init(struct cl_object *clob, int size) 2758{ 2759 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize; 2760 cl_object_header(clob)->coh_page_bufsize += ALIGN(size, 8); 2761} 2762 2763static inline void *cl_object_page_slice(struct cl_object *clob, 2764 struct cl_page *page) 2765{ 2766 return (void *)((char *)page + clob->co_slice_off); 2767} 2768 2769/** @} cl_object */ 2770 2771/** \defgroup cl_page cl_page 2772 * @{ */ 2773enum { 2774 CLP_GANG_OKAY = 0, 2775 CLP_GANG_RESCHED, 2776 CLP_GANG_AGAIN, 2777 CLP_GANG_ABORT 2778}; 2779 2780/* callback of cl_page_gang_lookup() */ 2781typedef int (*cl_page_gang_cb_t) (const struct lu_env *, struct cl_io *, 2782 struct cl_page *, void *); 2783int cl_page_gang_lookup (const struct lu_env *env, 2784 struct cl_object *obj, 2785 struct cl_io *io, 2786 pgoff_t start, pgoff_t end, 2787 cl_page_gang_cb_t cb, void *cbdata); 2788struct cl_page *cl_page_lookup (struct cl_object_header *hdr, 2789 pgoff_t index); 2790struct cl_page *cl_page_find (const struct lu_env *env, 2791 struct cl_object *obj, 2792 pgoff_t idx, struct page *vmpage, 2793 enum cl_page_type type); 2794struct cl_page *cl_page_find_sub (const struct lu_env *env, 2795 struct cl_object *obj, 2796 pgoff_t idx, struct page *vmpage, 2797 struct cl_page *parent); 2798void cl_page_get (struct cl_page *page); 2799void cl_page_put (const struct lu_env *env, 2800 struct cl_page *page); 2801void cl_page_print (const struct lu_env *env, void *cookie, 2802 lu_printer_t printer, 2803 const struct cl_page *pg); 2804void cl_page_header_print(const struct lu_env *env, void *cookie, 2805 lu_printer_t printer, 2806 const struct cl_page *pg); 2807struct page *cl_page_vmpage (const struct lu_env *env, 2808 struct cl_page *page); 2809struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj); 2810struct cl_page *cl_page_top (struct cl_page *page); 2811 2812const struct cl_page_slice *cl_page_at(const struct cl_page *page, 2813 const struct lu_device_type *dtype); 2814 2815/** 2816 * \name ownership 2817 * 2818 * Functions dealing with the ownership of page by io. 2819 */ 2820/** @{ */ 2821 2822int cl_page_own (const struct lu_env *env, 2823 struct cl_io *io, struct cl_page *page); 2824int cl_page_own_try (const struct lu_env *env, 2825 struct cl_io *io, struct cl_page *page); 2826void cl_page_assume (const struct lu_env *env, 2827 struct cl_io *io, struct cl_page *page); 2828void cl_page_unassume (const struct lu_env *env, 2829 struct cl_io *io, struct cl_page *pg); 2830void cl_page_disown (const struct lu_env *env, 2831 struct cl_io *io, struct cl_page *page); 2832int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io); 2833 2834/** @} ownership */ 2835 2836/** 2837 * \name transfer 2838 * 2839 * Functions dealing with the preparation of a page for a transfer, and 2840 * tracking transfer state. 2841 */ 2842/** @{ */ 2843int cl_page_prep (const struct lu_env *env, struct cl_io *io, 2844 struct cl_page *pg, enum cl_req_type crt); 2845void cl_page_completion (const struct lu_env *env, 2846 struct cl_page *pg, enum cl_req_type crt, int ioret); 2847int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg, 2848 enum cl_req_type crt); 2849int cl_page_cache_add (const struct lu_env *env, struct cl_io *io, 2850 struct cl_page *pg, enum cl_req_type crt); 2851void cl_page_clip (const struct lu_env *env, struct cl_page *pg, 2852 int from, int to); 2853int cl_page_cancel (const struct lu_env *env, struct cl_page *page); 2854int cl_page_flush (const struct lu_env *env, struct cl_io *io, 2855 struct cl_page *pg); 2856 2857/** @} transfer */ 2858 2859 2860/** 2861 * \name helper routines 2862 * Functions to discard, delete and export a cl_page. 2863 */ 2864/** @{ */ 2865void cl_page_discard (const struct lu_env *env, struct cl_io *io, 2866 struct cl_page *pg); 2867void cl_page_delete (const struct lu_env *env, struct cl_page *pg); 2868int cl_page_unmap (const struct lu_env *env, struct cl_io *io, 2869 struct cl_page *pg); 2870int cl_page_is_vmlocked (const struct lu_env *env, 2871 const struct cl_page *pg); 2872void cl_page_export (const struct lu_env *env, 2873 struct cl_page *pg, int uptodate); 2874int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io, 2875 struct cl_page *page); 2876loff_t cl_offset (const struct cl_object *obj, pgoff_t idx); 2877pgoff_t cl_index (const struct cl_object *obj, loff_t offset); 2878int cl_page_size (const struct cl_object *obj); 2879int cl_pages_prune (const struct lu_env *env, struct cl_object *obj); 2880 2881void cl_lock_print (const struct lu_env *env, void *cookie, 2882 lu_printer_t printer, const struct cl_lock *lock); 2883void cl_lock_descr_print(const struct lu_env *env, void *cookie, 2884 lu_printer_t printer, 2885 const struct cl_lock_descr *descr); 2886/* @} helper */ 2887 2888/** @} cl_page */ 2889 2890/** \defgroup cl_lock cl_lock 2891 * @{ */ 2892 2893struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io, 2894 const struct cl_lock_descr *need, 2895 const char *scope, const void *source); 2896struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io, 2897 const struct cl_lock_descr *need, 2898 const char *scope, const void *source); 2899struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io, 2900 const struct cl_lock_descr *need, 2901 const char *scope, const void *source); 2902struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env, 2903 struct cl_object *obj, pgoff_t index, 2904 struct cl_lock *except, int pending, 2905 int canceld); 2906static inline struct cl_lock *cl_lock_at_page(const struct lu_env *env, 2907 struct cl_object *obj, 2908 struct cl_page *page, 2909 struct cl_lock *except, 2910 int pending, int canceld) 2911{ 2912 LASSERT(cl_object_header(obj) == cl_object_header(page->cp_obj)); 2913 return cl_lock_at_pgoff(env, obj, page->cp_index, except, 2914 pending, canceld); 2915} 2916 2917const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock, 2918 const struct lu_device_type *dtype); 2919 2920void cl_lock_get (struct cl_lock *lock); 2921void cl_lock_get_trust (struct cl_lock *lock); 2922void cl_lock_put (const struct lu_env *env, struct cl_lock *lock); 2923void cl_lock_hold_add (const struct lu_env *env, struct cl_lock *lock, 2924 const char *scope, const void *source); 2925void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock, 2926 const char *scope, const void *source); 2927void cl_lock_unhold (const struct lu_env *env, struct cl_lock *lock, 2928 const char *scope, const void *source); 2929void cl_lock_release (const struct lu_env *env, struct cl_lock *lock, 2930 const char *scope, const void *source); 2931void cl_lock_user_add (const struct lu_env *env, struct cl_lock *lock); 2932void cl_lock_user_del (const struct lu_env *env, struct cl_lock *lock); 2933 2934enum cl_lock_state cl_lock_intransit(const struct lu_env *env, 2935 struct cl_lock *lock); 2936void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock, 2937 enum cl_lock_state state); 2938int cl_lock_is_intransit(struct cl_lock *lock); 2939 2940int cl_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock, 2941 int keep_mutex); 2942 2943/** \name statemachine statemachine 2944 * Interface to lock state machine consists of 3 parts: 2945 * 2946 * - "try" functions that attempt to effect a state transition. If state 2947 * transition is not possible right now (e.g., if it has to wait for some 2948 * asynchronous event to occur), these functions return 2949 * cl_lock_transition::CLO_WAIT. 2950 * 2951 * - "non-try" functions that implement synchronous blocking interface on 2952 * top of non-blocking "try" functions. These functions repeatedly call 2953 * corresponding "try" versions, and if state transition is not possible 2954 * immediately, wait for lock state change. 2955 * 2956 * - methods from cl_lock_operations, called by "try" functions. Lock can 2957 * be advanced to the target state only when all layers voted that they 2958 * are ready for this transition. "Try" functions call methods under lock 2959 * mutex. If a layer had to release a mutex, it re-acquires it and returns 2960 * cl_lock_transition::CLO_REPEAT, causing "try" function to call all 2961 * layers again. 2962 * 2963 * TRY NON-TRY METHOD FINAL STATE 2964 * 2965 * cl_enqueue_try() cl_enqueue() cl_lock_operations::clo_enqueue() CLS_ENQUEUED 2966 * 2967 * cl_wait_try() cl_wait() cl_lock_operations::clo_wait() CLS_HELD 2968 * 2969 * cl_unuse_try() cl_unuse() cl_lock_operations::clo_unuse() CLS_CACHED 2970 * 2971 * cl_use_try() NONE cl_lock_operations::clo_use() CLS_HELD 2972 * 2973 * @{ */ 2974 2975int cl_enqueue (const struct lu_env *env, struct cl_lock *lock, 2976 struct cl_io *io, __u32 flags); 2977int cl_wait (const struct lu_env *env, struct cl_lock *lock); 2978void cl_unuse (const struct lu_env *env, struct cl_lock *lock); 2979int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock, 2980 struct cl_io *io, __u32 flags); 2981int cl_unuse_try (const struct lu_env *env, struct cl_lock *lock); 2982int cl_wait_try (const struct lu_env *env, struct cl_lock *lock); 2983int cl_use_try (const struct lu_env *env, struct cl_lock *lock, int atomic); 2984 2985/** @} statemachine */ 2986 2987void cl_lock_signal (const struct lu_env *env, struct cl_lock *lock); 2988int cl_lock_state_wait (const struct lu_env *env, struct cl_lock *lock); 2989void cl_lock_state_set (const struct lu_env *env, struct cl_lock *lock, 2990 enum cl_lock_state state); 2991int cl_queue_match (const struct list_head *queue, 2992 const struct cl_lock_descr *need); 2993 2994void cl_lock_mutex_get (const struct lu_env *env, struct cl_lock *lock); 2995int cl_lock_mutex_try (const struct lu_env *env, struct cl_lock *lock); 2996void cl_lock_mutex_put (const struct lu_env *env, struct cl_lock *lock); 2997int cl_lock_is_mutexed (struct cl_lock *lock); 2998int cl_lock_nr_mutexed (const struct lu_env *env); 2999int cl_lock_discard_pages(const struct lu_env *env, struct cl_lock *lock); 3000int cl_lock_ext_match (const struct cl_lock_descr *has,
3001 const struct cl_lock_descr *need); 3002int cl_lock_descr_match(const struct cl_lock_descr *has, 3003 const struct cl_lock_descr *need); 3004int cl_lock_mode_match (enum cl_lock_mode has, enum cl_lock_mode need); 3005int cl_lock_modify (const struct lu_env *env, struct cl_lock *lock, 3006 const struct cl_lock_descr *desc); 3007 3008void cl_lock_closure_init (const struct lu_env *env, 3009 struct cl_lock_closure *closure, 3010 struct cl_lock *origin, int wait); 3011void cl_lock_closure_fini (struct cl_lock_closure *closure); 3012int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock, 3013 struct cl_lock_closure *closure); 3014void cl_lock_disclosure (const struct lu_env *env, 3015 struct cl_lock_closure *closure); 3016int cl_lock_enclosure (const struct lu_env *env, struct cl_lock *lock, 3017 struct cl_lock_closure *closure); 3018 3019void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock); 3020void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock); 3021void cl_lock_error (const struct lu_env *env, struct cl_lock *lock, int error); 3022void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int wait); 3023 3024unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock); 3025 3026/** @} cl_lock */ 3027 3028/** \defgroup cl_io cl_io 3029 * @{ */ 3030 3031int cl_io_init (const struct lu_env *env, struct cl_io *io, 3032 enum cl_io_type iot, struct cl_object *obj); 3033int cl_io_sub_init (const struct lu_env *env, struct cl_io *io, 3034 enum cl_io_type iot, struct cl_object *obj); 3035int cl_io_rw_init (const struct lu_env *env, struct cl_io *io, 3036 enum cl_io_type iot, loff_t pos, size_t count); 3037int cl_io_loop (const struct lu_env *env, struct cl_io *io); 3038 3039void cl_io_fini (const struct lu_env *env, struct cl_io *io); 3040int cl_io_iter_init (const struct lu_env *env, struct cl_io *io); 3041void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io); 3042int cl_io_lock (const struct lu_env *env, struct cl_io *io); 3043void cl_io_unlock (const struct lu_env *env, struct cl_io *io); 3044int cl_io_start (const struct lu_env *env, struct cl_io *io); 3045void cl_io_end (const struct lu_env *env, struct cl_io *io); 3046int cl_io_lock_add (const struct lu_env *env, struct cl_io *io, 3047 struct cl_io_lock_link *link); 3048int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io, 3049 struct cl_lock_descr *descr); 3050int cl_io_read_page (const struct lu_env *env, struct cl_io *io, 3051 struct cl_page *page); 3052int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io, 3053 struct cl_page *page, unsigned from, unsigned to); 3054int cl_io_commit_write (const struct lu_env *env, struct cl_io *io, 3055 struct cl_page *page, unsigned from, unsigned to); 3056int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io, 3057 enum cl_req_type iot, struct cl_2queue *queue); 3058int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io, 3059 enum cl_req_type iot, struct cl_2queue *queue, 3060 long timeout); 3061void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io, 3062 size_t nob); 3063int cl_io_cancel (const struct lu_env *env, struct cl_io *io, 3064 struct cl_page_list *queue); 3065int cl_io_is_going (const struct lu_env *env); 3066 3067/** 3068 * True, iff \a io is an O_APPEND write(2). 3069 */ 3070static inline int cl_io_is_append(const struct cl_io *io) 3071{ 3072 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append; 3073} 3074 3075static inline int cl_io_is_sync_write(const struct cl_io *io) 3076{ 3077 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync; 3078} 3079 3080static inline int cl_io_is_mkwrite(const struct cl_io *io) 3081{ 3082 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite; 3083} 3084 3085/** 3086 * True, iff \a io is a truncate(2). 3087 */ 3088static inline int cl_io_is_trunc(const struct cl_io *io) 3089{ 3090 return io->ci_type == CIT_SETATTR && 3091 (io->u.ci_setattr.sa_valid & ATTR_SIZE); 3092} 3093 3094struct cl_io *cl_io_top(struct cl_io *io); 3095 3096void cl_io_print(const struct lu_env *env, void *cookie, 3097 lu_printer_t printer, const struct cl_io *io); 3098 3099#define CL_IO_SLICE_CLEAN(foo_io, base) \ 3100do { \ 3101 typeof(foo_io) __foo_io = (foo_io); \ 3102 \ 3103 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \ 3104 memset(&__foo_io->base + 1, 0, \ 3105 (sizeof *__foo_io) - sizeof __foo_io->base); \ 3106} while (0) 3107 3108/** @} cl_io */ 3109 3110/** \defgroup cl_page_list cl_page_list 3111 * @{ */ 3112 3113/** 3114 * Last page in the page list. 3115 */ 3116static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist) 3117{ 3118 LASSERT(plist->pl_nr > 0); 3119 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch); 3120} 3121 3122/** 3123 * Iterate over pages in a page list. 3124 */ 3125#define cl_page_list_for_each(page, list) \ 3126 list_for_each_entry((page), &(list)->pl_pages, cp_batch) 3127 3128/** 3129 * Iterate over pages in a page list, taking possible removals into account. 3130 */ 3131#define cl_page_list_for_each_safe(page, temp, list) \ 3132 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch) 3133 3134void cl_page_list_init (struct cl_page_list *plist); 3135void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page); 3136void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src, 3137 struct cl_page *page); 3138void cl_page_list_splice (struct cl_page_list *list, 3139 struct cl_page_list *head); 3140void cl_page_list_del (const struct lu_env *env, 3141 struct cl_page_list *plist, struct cl_page *page); 3142void cl_page_list_disown (const struct lu_env *env, 3143 struct cl_io *io, struct cl_page_list *plist); 3144int cl_page_list_own (const struct lu_env *env, 3145 struct cl_io *io, struct cl_page_list *plist); 3146void cl_page_list_assume (const struct lu_env *env, 3147 struct cl_io *io, struct cl_page_list *plist); 3148void cl_page_list_discard(const struct lu_env *env, 3149 struct cl_io *io, struct cl_page_list *plist); 3150int cl_page_list_unmap (const struct lu_env *env, 3151 struct cl_io *io, struct cl_page_list *plist); 3152void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist); 3153 3154void cl_2queue_init (struct cl_2queue *queue); 3155void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page); 3156void cl_2queue_disown (const struct lu_env *env, 3157 struct cl_io *io, struct cl_2queue *queue); 3158void cl_2queue_assume (const struct lu_env *env, 3159 struct cl_io *io, struct cl_2queue *queue); 3160void cl_2queue_discard (const struct lu_env *env, 3161 struct cl_io *io, struct cl_2queue *queue); 3162void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue); 3163void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page); 3164 3165/** @} cl_page_list */ 3166 3167/** \defgroup cl_req cl_req 3168 * @{ */ 3169struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page, 3170 enum cl_req_type crt, int nr_objects); 3171 3172void cl_req_page_add (const struct lu_env *env, struct cl_req *req, 3173 struct cl_page *page); 3174void cl_req_page_done (const struct lu_env *env, struct cl_page *page); 3175int cl_req_prep (const struct lu_env *env, struct cl_req *req); 3176void cl_req_attr_set (const struct lu_env *env, struct cl_req *req, 3177 struct cl_req_attr *attr, obd_valid flags); 3178void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret); 3179 3180/** \defgroup cl_sync_io cl_sync_io 3181 * @{ */ 3182 3183/** 3184 * Anchor for synchronous transfer. This is allocated on a stack by thread 3185 * doing synchronous transfer, and a pointer to this structure is set up in 3186 * every page submitted for transfer. Transfer completion routine updates 3187 * anchor and wakes up waiting thread when transfer is complete. 3188 */ 3189struct cl_sync_io { 3190 /** number of pages yet to be transferred. */ 3191 atomic_t csi_sync_nr; 3192 /** error code. */ 3193 int csi_sync_rc; 3194 /** barrier of destroy this structure */ 3195 atomic_t csi_barrier; 3196 /** completion to be signaled when transfer is complete. */ 3197 wait_queue_head_t csi_waitq; 3198}; 3199 3200void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages); 3201int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io, 3202 struct cl_page_list *queue, struct cl_sync_io *anchor, 3203 long timeout); 3204void cl_sync_io_note(struct cl_sync_io *anchor, int ioret); 3205 3206/** @} cl_sync_io */ 3207 3208/** @} cl_req */ 3209 3210/** \defgroup cl_env cl_env 3211 * 3212 * lu_env handling for a client. 3213 * 3214 * lu_env is an environment within which lustre code executes. Its major part 3215 * is lu_context---a fast memory allocation mechanism that is used to conserve 3216 * precious kernel stack space. Originally lu_env was designed for a server, 3217 * where 3218 * 3219 * - there is a (mostly) fixed number of threads, and 3220 * 3221 * - call chains have no non-lustre portions inserted between lustre code. 3222 * 3223 * On a client both these assumtpion fails, because every user thread can 3224 * potentially execute lustre code as part of a system call, and lustre calls 3225 * into VFS or MM that call back into lustre. 3226 * 3227 * To deal with that, cl_env wrapper functions implement the following 3228 * optimizations: 3229 * 3230 * - allocation and destruction of environment is amortized by caching no 3231 * longer used environments instead of destroying them; 3232 * 3233 * - there is a notion of "current" environment, attached to the kernel 3234 * data structure representing current thread Top-level lustre code 3235 * allocates an environment and makes it current, then calls into 3236 * non-lustre code, that in turn calls lustre back. Low-level lustre 3237 * code thus called can fetch environment created by the top-level code 3238 * and reuse it, avoiding additional environment allocation. 3239 * Right now, three interfaces can attach the cl_env to running thread: 3240 * - cl_env_get 3241 * - cl_env_implant 3242 * - cl_env_reexit(cl_env_reenter had to be called priorly) 3243 * 3244 * \see lu_env, lu_context, lu_context_key 3245 * @{ */ 3246 3247struct cl_env_nest { 3248 int cen_refcheck; 3249 void *cen_cookie; 3250}; 3251 3252struct lu_env *cl_env_peek (int *refcheck); 3253struct lu_env *cl_env_get (int *refcheck); 3254struct lu_env *cl_env_alloc (int *refcheck, __u32 tags); 3255struct lu_env *cl_env_nested_get (struct cl_env_nest *nest); 3256void cl_env_put (struct lu_env *env, int *refcheck); 3257void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env); 3258void *cl_env_reenter (void); 3259void cl_env_reexit (void *cookie); 3260void cl_env_implant (struct lu_env *env, int *refcheck); 3261void cl_env_unplant (struct lu_env *env, int *refcheck); 3262 3263/** @} cl_env */ 3264 3265/* 3266 * Misc 3267 */ 3268void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr); 3269void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb); 3270 3271struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site, 3272 struct lu_device_type *ldt, 3273 struct lu_device *next); 3274/** @} clio */ 3275 3276int cl_global_init(void); 3277void cl_global_fini(void); 3278 3279#endif /* _LINUX_CL_OBJECT_H */ 3280