1 // SPDX-License-Identifier: GPL-2.0
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 only,
9 * as published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful, but
12 * WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * General Public License version 2 for more details (a copy is included
15 * in the LICENSE file that accompanied this code).
17 * You should have received a copy of the GNU General Public License
18 * version 2 along with this program; If not, see
19 * http://www.gnu.org/licenses/gpl-2.0.html
24 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
25 * Use is subject to license terms.
27 * Copyright (c) 2011, 2015, Intel Corporation.
30 * This file is part of Lustre, http://www.lustre.org/
31 * Lustre is a trademark of Sun Microsystems, Inc.
33 #ifndef _LUSTRE_CL_OBJECT_H
34 #define _LUSTRE_CL_OBJECT_H
36 /** \defgroup clio clio
38 * Client objects implement io operations and cache pages.
40 * Examples: lov and osc are implementations of cl interface.
42 * Big Theory Statement.
46 * Client implementation is based on the following data-types:
52 * - cl_lock represents an extent lock on an object.
54 * - cl_io represents high-level i/o activity such as whole read/write
55 * system call, or write-out of pages from under the lock being
56 * canceled. cl_io has sub-ios that can be stopped and resumed
57 * independently, thus achieving high degree of transfer
58 * parallelism. Single cl_io can be advanced forward by
59 * the multiple threads (although in the most usual case of
60 * read/write system call it is associated with the single user
61 * thread, that issued the system call).
65 * - to avoid confusion high-level I/O operation like read or write system
66 * call is referred to as "an io", whereas low-level I/O operation, like
67 * RPC, is referred to as "a transfer"
69 * - "generic code" means generic (not file system specific) code in the
70 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
71 * is not layer specific.
77 * - cl_object_header::coh_page_guard
80 * See the top comment in cl_object.c for the description of overall locking and
81 * reference-counting design.
83 * See comments below for the description of i/o, page, and dlm-locking
90 * super-class definitions.
92 #include <lu_object.h>
93 #include <lustre_compat.h>
94 #include <linux/atomic.h>
95 #include <linux/mutex.h>
96 #include <linux/radix-tree.h>
97 #include <linux/spinlock.h>
98 #include <linux/wait.h>
107 struct cl_page_slice;
109 struct cl_lock_slice;
111 struct cl_lock_operations;
112 struct cl_page_operations;
120 * Device in the client stack.
122 * \see vvp_device, lov_device, lovsub_device, osc_device
126 struct lu_device cd_lu_dev;
129 /** \addtogroup cl_object cl_object
133 * "Data attributes" of cl_object. Data attributes can be updated
134 * independently for a sub-object, and top-object's attributes are calculated
135 * from sub-objects' ones.
138 /** Object size, in bytes */
141 * Known minimal size, in bytes.
143 * This is only valid when at least one DLM lock is held.
146 /** Modification time. Measured in seconds since epoch. */
148 /** Access time. Measured in seconds since epoch. */
150 /** Change time. Measured in seconds since epoch. */
153 * Blocks allocated to this cl_object on the server file system.
155 * \todo XXX An interface for block size is needed.
159 * User identifier for quota purposes.
163 * Group identifier for quota purposes.
167 /* nlink of the directory */
172 * Fields in cl_attr that are being set.
186 * Sub-class of lu_object with methods common for objects on the client
189 * cl_object: represents a regular file system object, both a file and a
190 * stripe. cl_object is based on lu_object: it is identified by a fid,
191 * layered, cached, hashed, and lrued. Important distinction with the server
192 * side, where md_object and dt_object are used, is that cl_object "fans out"
193 * at the lov/sns level: depending on the file layout, single file is
194 * represented as a set of "sub-objects" (stripes). At the implementation
195 * level, struct lov_object contains an array of cl_objects. Each sub-object
196 * is a full-fledged cl_object, having its fid, living in the lru and hash
199 * This leads to the next important difference with the server side: on the
200 * client, it's quite usual to have objects with the different sequence of
201 * layers. For example, typical top-object is composed of the following
207 * whereas its sub-objects are composed of
212 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
213 * track of the object-subobject relationship.
215 * Sub-objects are not cached independently: when top-object is about to
216 * be discarded from the memory, all its sub-objects are torn-down and
219 * \see vvp_object, lov_object, lovsub_object, osc_object
223 struct lu_object co_lu;
224 /** per-object-layer operations */
225 const struct cl_object_operations *co_ops;
226 /** offset of page slice in cl_page buffer */
231 * Description of the client object configuration. This is used for the
232 * creation of a new client object that is identified by a more state than
235 struct cl_object_conf {
237 struct lu_object_conf coc_lu;
240 * Object layout. This is consumed by lov.
242 struct lu_buf coc_layout;
244 * Description of particular stripe location in the
245 * cluster. This is consumed by osc.
247 struct lov_oinfo *coc_oinfo;
250 * VFS inode. This is consumed by vvp.
252 struct inode *coc_inode;
254 * Layout lock handle.
256 struct ldlm_lock *coc_lock;
258 * Operation to handle layout, OBJECT_CONF_XYZ.
264 /** configure layout, set up a new stripe, must be called while
265 * holding layout lock.
268 /** invalidate the current stripe configuration due to losing
271 OBJECT_CONF_INVALIDATE = 1,
272 /** wait for old layout to go away so that new layout can be set up. */
277 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
278 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
282 /** the buffer to return the layout in lov_mds_md format. */
283 struct lu_buf cl_buf;
284 /** size of layout in lov_mds_md format. */
286 /** Layout generation. */
291 * Operations implemented for each cl object layer.
293 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
295 struct cl_object_operations {
297 * Initialize page slice for this layer. Called top-to-bottom through
298 * every object layer when a new cl_page is instantiated. Layer
299 * keeping private per-page data, or requiring its own page operations
300 * vector should allocate these data here, and attach then to the page
301 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
304 * \retval NULL success.
306 * \retval ERR_PTR(errno) failure code.
308 * \retval valid-pointer pointer to already existing referenced page
309 * to be used instead of newly created.
311 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
312 struct cl_page *page, pgoff_t index);
314 * Initialize lock slice for this layer. Called top-to-bottom through
315 * every object layer when a new cl_lock is instantiated. Layer
316 * keeping private per-lock data, or requiring its own lock operations
317 * vector should allocate these data here, and attach then to the lock
318 * by calling cl_lock_slice_add(). Mandatory.
320 int (*coo_lock_init)(const struct lu_env *env,
321 struct cl_object *obj, struct cl_lock *lock,
322 const struct cl_io *io);
324 * Initialize io state for a given layer.
326 * called top-to-bottom once per io existence to initialize io
327 * state. If layer wants to keep some state for this type of io, it
328 * has to embed struct cl_io_slice in lu_env::le_ses, and register
329 * slice with cl_io_slice_add(). It is guaranteed that all threads
330 * participating in this io share the same session.
332 int (*coo_io_init)(const struct lu_env *env,
333 struct cl_object *obj, struct cl_io *io);
335 * Fill portion of \a attr that this layer controls. This method is
336 * called top-to-bottom through all object layers.
338 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
340 * \return 0: to continue
341 * \return +ve: to stop iterating through layers (but 0 is returned
342 * from enclosing cl_object_attr_get())
343 * \return -ve: to signal error
345 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
346 struct cl_attr *attr);
350 * \a valid is a bitmask composed from enum #cl_attr_valid, and
351 * indicating what attributes are to be set.
353 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
355 * \return the same convention as for
356 * cl_object_operations::coo_attr_get() is used.
358 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
359 const struct cl_attr *attr, unsigned int valid);
361 * Update object configuration. Called top-to-bottom to modify object
364 * XXX error conditions and handling.
366 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
367 const struct cl_object_conf *conf);
369 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
370 * object. Layers are supposed to fill parts of \a lvb that will be
371 * shipped to the glimpse originator as a glimpse result.
373 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
374 * \see osc_object_glimpse()
376 int (*coo_glimpse)(const struct lu_env *env,
377 const struct cl_object *obj, struct ost_lvb *lvb);
379 * Object prune method. Called when the layout is going to change on
380 * this object, therefore each layer has to clean up their cache,
381 * mainly pages and locks.
383 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
385 * Object getstripe method.
387 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
388 struct lov_user_md __user *lum);
390 * Get FIEMAP mapping from the object.
392 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
393 struct ll_fiemap_info_key *fmkey,
394 struct fiemap *fiemap, size_t *buflen);
396 * Get layout and generation of the object.
398 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
399 struct cl_layout *layout);
401 * Get maximum size of the object.
403 loff_t (*coo_maxbytes)(struct cl_object *obj);
405 * Set request attributes.
407 void (*coo_req_attr_set)(const struct lu_env *env,
408 struct cl_object *obj,
409 struct cl_req_attr *attr);
413 * Extended header for client object.
415 struct cl_object_header {
416 /** Standard lu_object_header. cl_object::co_lu::lo_header points
419 struct lu_object_header coh_lu;
422 * Parent object. It is assumed that an object has a well-defined
423 * parent, but not a well-defined child (there may be multiple
424 * sub-objects, for the same top-object). cl_object_header::coh_parent
425 * field allows certain code to be written generically, without
426 * limiting possible cl_object layouts unduly.
428 struct cl_object_header *coh_parent;
430 * Protects consistency between cl_attr of parent object and
431 * attributes of sub-objects, that the former is calculated ("merged")
434 * \todo XXX this can be read/write lock if needed.
436 spinlock_t coh_attr_guard;
438 * Size of cl_page + page slices
440 unsigned short coh_page_bufsize;
442 * Number of objects above this one: 0 for a top-object, 1 for its
445 unsigned char coh_nesting;
449 * Helper macro: iterate over all layers of the object \a obj, assigning every
450 * layer top-to-bottom to \a slice.
452 #define cl_object_for_each(slice, obj) \
453 list_for_each_entry((slice), \
454 &(obj)->co_lu.lo_header->loh_layers, \
457 * Helper macro: iterate over all layers of the object \a obj, assigning every
458 * layer bottom-to-top to \a slice.
460 #define cl_object_for_each_reverse(slice, obj) \
461 list_for_each_entry_reverse((slice), \
462 &(obj)->co_lu.lo_header->loh_layers, \
466 #define CL_PAGE_EOF ((pgoff_t)~0ull)
468 /** \addtogroup cl_page cl_page
473 * Layered client page.
475 * cl_page: represents a portion of a file, cached in the memory. All pages
476 * of the given file are of the same size, and are kept in the radix tree
477 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
478 * of the top-level file object are first class cl_objects, they have their
479 * own radix trees of pages and hence page is implemented as a sequence of
480 * struct cl_pages's, linked into double-linked list through
481 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
482 * corresponding radix tree at the corresponding logical offset.
484 * cl_page is associated with VM page of the hosting environment (struct
485 * page in Linux kernel, for example), struct page. It is assumed, that this
486 * association is implemented by one of cl_page layers (top layer in the
487 * current design) that
489 * - intercepts per-VM-page call-backs made by the environment (e.g.,
492 * - translates state (page flag bits) and locking between lustre and
495 * The association between cl_page and struct page is immutable and
496 * established when cl_page is created.
498 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
499 * this io an exclusive access to this page w.r.t. other io attempts and
500 * various events changing page state (such as transfer completion, or
501 * eviction of the page from the memory). Note, that in general cl_io
502 * cannot be identified with a particular thread, and page ownership is not
503 * exactly equal to the current thread holding a lock on the page. Layer
504 * implementing association between cl_page and struct page has to implement
505 * ownership on top of available synchronization mechanisms.
507 * While lustre client maintains the notion of an page ownership by io,
508 * hosting MM/VM usually has its own page concurrency control
509 * mechanisms. For example, in Linux, page access is synchronized by the
510 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
511 * takes care to acquire and release such locks as necessary around the
512 * calls to the file system methods (->readpage(), ->prepare_write(),
513 * ->commit_write(), etc.). This leads to the situation when there are two
514 * different ways to own a page in the client:
516 * - client code explicitly and voluntary owns the page (cl_page_own());
518 * - VM locks a page and then calls the client, that has "to assume"
519 * the ownership from the VM (cl_page_assume()).
521 * Dual methods to release ownership are cl_page_disown() and
522 * cl_page_unassume().
524 * cl_page is reference counted (cl_page::cp_ref). When reference counter
525 * drops to 0, the page is returned to the cache, unless it is in
526 * cl_page_state::CPS_FREEING state, in which case it is immediately
529 * The general logic guaranteeing the absence of "existential races" for
530 * pages is the following:
532 * - there are fixed known ways for a thread to obtain a new reference
535 * - by doing a lookup in the cl_object radix tree, protected by the
538 * - by starting from VM-locked struct page and following some
539 * hosting environment method (e.g., following ->private pointer in
540 * the case of Linux kernel), see cl_vmpage_page();
542 * - when the page enters cl_page_state::CPS_FREEING state, all these
543 * ways are severed with the proper synchronization
544 * (cl_page_delete());
546 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
549 * - no new references to the page in cl_page_state::CPS_FREEING state
550 * are allowed (checked in cl_page_get()).
552 * Together this guarantees that when last reference to a
553 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
554 * page, as neither references to it can be acquired at that point, nor
557 * cl_page is a state machine. States are enumerated in enum
558 * cl_page_state. Possible state transitions are enumerated in
559 * cl_page_state_set(). State transition process (i.e., actual changing of
560 * cl_page::cp_state field) is protected by the lock on the underlying VM
563 * Linux Kernel implementation.
565 * Binding between cl_page and struct page (which is a typedef for
566 * struct page) is implemented in the vvp layer. cl_page is attached to the
567 * ->private pointer of the struct page, together with the setting of
568 * PG_private bit in page->flags, and acquiring additional reference on the
569 * struct page (much like struct buffer_head, or any similar file system
570 * private data structures).
572 * PG_locked lock is used to implement both ownership and transfer
573 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
574 * states. No additional references are acquired for the duration of the
577 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
578 * write-out is "protected" by the special PG_writeback bit.
582 * States of cl_page. cl_page.c assumes particular order here.
584 * The page state machine is rather crude, as it doesn't recognize finer page
585 * states like "dirty" or "up to date". This is because such states are not
586 * always well defined for the whole stack (see, for example, the
587 * implementation of the read-ahead, that hides page up-to-dateness to track
588 * cache hits accurately). Such sub-states are maintained by the layers that
589 * are interested in them.
593 * Page is in the cache, un-owned. Page leaves cached state in the
596 * - [cl_page_state::CPS_OWNED] io comes across the page and
599 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
600 * req-formation engine decides that it wants to include this page
601 * into an RPC being constructed, and yanks it from the cache;
603 * - [cl_page_state::CPS_FREEING] VM callback is executed to
604 * evict the page form the memory;
606 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
610 * Page is exclusively owned by some cl_io. Page may end up in this
611 * state as a result of
613 * - io creating new page and immediately owning it;
615 * - [cl_page_state::CPS_CACHED] io finding existing cached page
618 * - [cl_page_state::CPS_OWNED] io finding existing owned page
619 * and waiting for owner to release the page;
621 * Page leaves owned state in the following cases:
623 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
624 * the cache, doing nothing;
626 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
629 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
630 * transfer for this page;
632 * - [cl_page_state::CPS_FREEING] io decides to destroy this
633 * page (e.g., as part of truncate or extent lock cancellation).
635 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
639 * Page is being written out, as a part of a transfer. This state is
640 * entered when req-formation logic decided that it wants this page to
641 * be sent through the wire _now_. Specifically, it means that once
642 * this state is achieved, transfer completion handler (with either
643 * success or failure indication) is guaranteed to be executed against
644 * this page independently of any locks and any scheduling decisions
645 * made by the hosting environment (that effectively means that the
646 * page is never put into cl_page_state::CPS_PAGEOUT state "in
647 * advance". This property is mentioned, because it is important when
648 * reasoning about possible dead-locks in the system). The page can
649 * enter this state as a result of
651 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
652 * write-out of this page, or
654 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
655 * that it has enough dirty pages cached to issue a "good"
658 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
659 * is completed---it is moved into cl_page_state::CPS_CACHED state.
661 * Underlying VM page is locked for the duration of transfer.
663 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
667 * Page is being read in, as a part of a transfer. This is quite
668 * similar to the cl_page_state::CPS_PAGEOUT state, except that
669 * read-in is always "immediate"---there is no such thing a sudden
670 * construction of read request from cached, presumably not up to date,
673 * Underlying VM page is locked for the duration of transfer.
675 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
679 * Page is being destroyed. This state is entered when client decides
680 * that page has to be deleted from its host object, as, e.g., a part
683 * Once this state is reached, there is no way to escape it.
685 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
692 /** Host page, the page is from the host inode which the cl_page
697 /** Transient page, the transient cl_page is used to bind a cl_page
698 * to vmpage which is not belonging to the same object of cl_page.
699 * it is used in DirectIO and lockless IO.
705 * Fields are protected by the lock on struct page, except for atomics and
708 * \invariant Data type invariants are in cl_page_invariant(). Basically:
709 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
710 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
711 * cl_page::cp_owner (when set).
714 /** Reference counter. */
716 /** An object this page is a part of. Immutable after creation. */
717 struct cl_object *cp_obj;
719 struct page *cp_vmpage;
720 /** Linkage of pages within group. Pages must be owned */
721 struct list_head cp_batch;
722 /** List of slices. Immutable after creation. */
723 struct list_head cp_layers;
725 * Page state. This field is const to avoid accidental update, it is
726 * modified only internally within cl_page.c. Protected by a VM lock.
728 const enum cl_page_state cp_state;
730 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
733 enum cl_page_type cp_type;
736 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
737 * by sub-io. Protected by a VM lock.
739 struct cl_io *cp_owner;
740 /** List of references to this page, for debugging. */
741 struct lu_ref cp_reference;
742 /** Link to an object, for debugging. */
743 struct lu_ref_link cp_obj_ref;
744 /** Link to a queue, for debugging. */
745 struct lu_ref_link cp_queue_ref;
746 /** Assigned if doing a sync_io */
747 struct cl_sync_io *cp_sync_io;
751 * Per-layer part of cl_page.
753 * \see vvp_page, lov_page, osc_page
755 struct cl_page_slice {
756 struct cl_page *cpl_page;
759 * Object slice corresponding to this page slice. Immutable after
762 struct cl_object *cpl_obj;
763 const struct cl_page_operations *cpl_ops;
764 /** Linkage into cl_page::cp_layers. Immutable after creation. */
765 struct list_head cpl_linkage;
769 * Lock mode. For the client extent locks.
780 * Requested transfer type.
789 * Per-layer page operations.
791 * Methods taking an \a io argument are for the activity happening in the
792 * context of given \a io. Page is assumed to be owned by that io, except for
793 * the obvious cases (like cl_page_operations::cpo_own()).
795 * \see vvp_page_ops, lov_page_ops, osc_page_ops
797 struct cl_page_operations {
799 * cl_page<->struct page methods. Only one layer in the stack has to
800 * implement these. Current code assumes that this functionality is
801 * provided by the topmost layer, see cl_page_disown0() as an example.
805 * Called when \a io acquires this page into the exclusive
806 * ownership. When this method returns, it is guaranteed that the is
807 * not owned by other io, and no transfer is going on against
811 * \see vvp_page_own(), lov_page_own()
813 int (*cpo_own)(const struct lu_env *env,
814 const struct cl_page_slice *slice,
815 struct cl_io *io, int nonblock);
816 /** Called when ownership it yielded. Optional.
818 * \see cl_page_disown()
819 * \see vvp_page_disown()
821 void (*cpo_disown)(const struct lu_env *env,
822 const struct cl_page_slice *slice, struct cl_io *io);
824 * Called for a page that is already "owned" by \a io from VM point of
827 * \see cl_page_assume()
828 * \see vvp_page_assume(), lov_page_assume()
830 void (*cpo_assume)(const struct lu_env *env,
831 const struct cl_page_slice *slice, struct cl_io *io);
832 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
833 * bottom-to-top when IO releases a page without actually unlocking
836 * \see cl_page_unassume()
837 * \see vvp_page_unassume()
839 void (*cpo_unassume)(const struct lu_env *env,
840 const struct cl_page_slice *slice,
843 * Announces whether the page contains valid data or not by \a uptodate.
845 * \see cl_page_export()
846 * \see vvp_page_export()
848 void (*cpo_export)(const struct lu_env *env,
849 const struct cl_page_slice *slice, int uptodate);
851 * Checks whether underlying VM page is locked (in the suitable
852 * sense). Used for assertions.
854 * \retval -EBUSY: page is protected by a lock of a given mode;
855 * \retval -ENODATA: page is not protected by a lock;
856 * \retval 0: this layer cannot decide. (Should never happen.)
858 int (*cpo_is_vmlocked)(const struct lu_env *env,
859 const struct cl_page_slice *slice);
865 * Called when page is truncated from the object. Optional.
867 * \see cl_page_discard()
868 * \see vvp_page_discard(), osc_page_discard()
870 void (*cpo_discard)(const struct lu_env *env,
871 const struct cl_page_slice *slice,
874 * Called when page is removed from the cache, and is about to being
875 * destroyed. Optional.
877 * \see cl_page_delete()
878 * \see vvp_page_delete(), osc_page_delete()
880 void (*cpo_delete)(const struct lu_env *env,
881 const struct cl_page_slice *slice);
882 /** Destructor. Frees resources and slice itself. */
883 void (*cpo_fini)(const struct lu_env *env,
884 struct cl_page_slice *slice);
886 * Optional debugging helper. Prints given page slice.
888 * \see cl_page_print()
890 int (*cpo_print)(const struct lu_env *env,
891 const struct cl_page_slice *slice,
892 void *cookie, lu_printer_t p);
901 * Request type dependent vector of operations.
903 * Transfer operations depend on transfer mode (cl_req_type). To avoid
904 * passing transfer mode to each and every of these methods, and to
905 * avoid branching on request type inside of the methods, separate
906 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
907 * provided. That is, method invocation usually looks like
909 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
913 * Called when a page is submitted for a transfer as a part of
916 * \return 0 : page is eligible for submission;
917 * \return -EALREADY : skip this page;
918 * \return -ve : error.
920 * \see cl_page_prep()
922 int (*cpo_prep)(const struct lu_env *env,
923 const struct cl_page_slice *slice,
926 * Completion handler. This is guaranteed to be eventually
927 * fired after cl_page_operations::cpo_prep() or
928 * cl_page_operations::cpo_make_ready() call.
930 * This method can be called in a non-blocking context. It is
931 * guaranteed however, that the page involved and its object
932 * are pinned in memory (and, hence, calling cl_page_put() is
935 * \see cl_page_completion()
937 void (*cpo_completion)(const struct lu_env *env,
938 const struct cl_page_slice *slice,
941 * Called when cached page is about to be added to the
942 * ptlrpc request as a part of req formation.
944 * \return 0 : proceed with this page;
945 * \return -EAGAIN : skip this page;
946 * \return -ve : error.
948 * \see cl_page_make_ready()
950 int (*cpo_make_ready)(const struct lu_env *env,
951 const struct cl_page_slice *slice);
954 * Tell transfer engine that only [to, from] part of a page should be
957 * This is used for immediate transfers.
959 * \todo XXX this is not very good interface. It would be much better
960 * if all transfer parameters were supplied as arguments to
961 * cl_io_operations::cio_submit() call, but it is not clear how to do
962 * this for page queues.
964 * \see cl_page_clip()
966 void (*cpo_clip)(const struct lu_env *env,
967 const struct cl_page_slice *slice,
970 * \pre the page was queued for transferring.
971 * \post page is removed from client's pending list, or -EBUSY
972 * is returned if it has already been in transferring.
974 * This is one of seldom page operation which is:
975 * 0. called from top level;
976 * 1. don't have vmpage locked;
977 * 2. every layer should synchronize execution of its ->cpo_cancel()
978 * with completion handlers. Osc uses client obd lock for this
979 * purpose. Based on there is no vvp_page_cancel and
980 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
982 * \see osc_page_cancel().
984 int (*cpo_cancel)(const struct lu_env *env,
985 const struct cl_page_slice *slice);
987 * Write out a page by kernel. This is only called by ll_writepage
990 * \see cl_page_flush()
992 int (*cpo_flush)(const struct lu_env *env,
993 const struct cl_page_slice *slice,
999 * Helper macro, dumping detailed information about \a page into a log.
1001 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1003 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1004 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1005 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1006 CDEBUG(mask, format, ## __VA_ARGS__); \
1011 * Helper macro, dumping shorter information about \a page into a log.
1013 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1015 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1016 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1017 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1018 CDEBUG(mask, format, ## __VA_ARGS__); \
1022 static inline struct page *cl_page_vmpage(struct cl_page *page)
1024 LASSERT(page->cp_vmpage);
1025 return page->cp_vmpage;
1029 * Check if a cl_page is in use.
1031 * Client cache holds a refcount, this refcount will be dropped when
1032 * the page is taken out of cache, see vvp_page_delete().
1034 static inline bool __page_in_use(const struct cl_page *page, int refc)
1036 return (atomic_read(&page->cp_ref) > refc + 1);
1040 * Caller itself holds a refcount of cl_page.
1042 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1044 * Caller doesn't hold a refcount.
1046 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1050 /** \addtogroup cl_lock cl_lock
1055 * Extent locking on the client.
1059 * The locking model of the new client code is built around
1063 * data-type representing an extent lock on a regular file. cl_lock is a
1064 * layered object (much like cl_object and cl_page), it consists of a header
1065 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1066 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1068 * Typical cl_lock consists of the two layers:
1070 * - vvp_lock (vvp specific data), and
1071 * - lov_lock (lov specific data).
1073 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1074 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1076 * - lovsub_lock, and
1079 * Each sub-lock is associated with a cl_object (representing stripe
1080 * sub-object or the file to which top-level cl_lock is associated to), and is
1081 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1082 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1083 * is different from cl_page, that doesn't fan out (there is usually exactly
1084 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1085 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1089 * cl_lock is a cacheless data container for the requirements of locks to
1090 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1093 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1094 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1096 * INTERFACE AND USAGE
1098 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1099 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1100 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1101 * consists of multiple sub cl_locks, each sub locks will be enqueued
1102 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1103 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1106 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1107 * method will be called for each layer to release the resource held by this
1108 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1109 * clo_enqueue time, is released.
1111 * LDLM lock can only be canceled if there is no cl_lock using it.
1113 * Overall process of the locking during IO operation is as following:
1115 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1116 * is called on each layer. Responsibility of this method is to add locks,
1117 * needed by a given layer into cl_io.ci_lockset.
1119 * - once locks for all layers were collected, they are sorted to avoid
1120 * dead-locks (cl_io_locks_sort()), and enqueued.
1122 * - when all locks are acquired, IO is performed;
1124 * - locks are released after IO is complete.
1126 * Striping introduces major additional complexity into locking. The
1127 * fundamental problem is that it is generally unsafe to actively use (hold)
1128 * two locks on the different OST servers at the same time, as this introduces
1129 * inter-server dependency and can lead to cascading evictions.
1131 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1132 * that no multi-stripe locks are taken (note that this design abandons POSIX
1133 * read/write semantics). Such pieces ideally can be executed concurrently. At
1134 * the same time, certain types of IO cannot be sub-divived, without
1135 * sacrificing correctness. This includes:
1137 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1140 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1142 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1143 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1144 * has to be held together with the usual lock on [offset, offset + count].
1146 * Interaction with DLM
1148 * In the expected setup, cl_lock is ultimately backed up by a collection of
1149 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1150 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1151 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1152 * description of interaction with DLM.
1158 struct cl_lock_descr {
1159 /** Object this lock is granted for. */
1160 struct cl_object *cld_obj;
1161 /** Index of the first page protected by this lock. */
1163 /** Index of the last page (inclusive) protected by this lock. */
1165 /** Group ID, for group lock */
1168 enum cl_lock_mode cld_mode;
1170 * flags to enqueue lock. A combination of bit-flags from
1171 * enum cl_enq_flags.
1173 __u32 cld_enq_flags;
1176 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1177 #define PDESCR(descr) \
1178 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1179 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1181 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1184 * Layered client lock.
1187 /** List of slices. Immutable after creation. */
1188 struct list_head cll_layers;
1189 /** lock attribute, extent, cl_object, etc. */
1190 struct cl_lock_descr cll_descr;
1194 * Per-layer part of cl_lock
1196 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1198 struct cl_lock_slice {
1199 struct cl_lock *cls_lock;
1200 /** Object slice corresponding to this lock slice. Immutable after
1203 struct cl_object *cls_obj;
1204 const struct cl_lock_operations *cls_ops;
1205 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1206 struct list_head cls_linkage;
1211 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1213 struct cl_lock_operations {
1216 * Attempts to enqueue the lock. Called top-to-bottom.
1218 * \retval 0 this layer has enqueued the lock successfully
1219 * \retval >0 this layer has enqueued the lock, but need to wait on
1220 * @anchor for resources
1221 * \retval -ve failure
1223 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1224 * \see osc_lock_enqueue()
1226 int (*clo_enqueue)(const struct lu_env *env,
1227 const struct cl_lock_slice *slice,
1228 struct cl_io *io, struct cl_sync_io *anchor);
1230 * Cancel a lock, release its DLM lock ref, while does not cancel the
1233 void (*clo_cancel)(const struct lu_env *env,
1234 const struct cl_lock_slice *slice);
1237 * Destructor. Frees resources and the slice.
1239 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1240 * \see osc_lock_fini()
1242 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1244 * Optional debugging helper. Prints given lock slice.
1246 int (*clo_print)(const struct lu_env *env,
1247 void *cookie, lu_printer_t p,
1248 const struct cl_lock_slice *slice);
1251 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1253 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1255 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1256 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1257 CDEBUG(mask, format, ## __VA_ARGS__); \
1261 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1265 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1271 /** \addtogroup cl_page_list cl_page_list
1272 * Page list used to perform collective operations on a group of pages.
1274 * Pages are added to the list one by one. cl_page_list acquires a reference
1275 * for every page in it. Page list is used to perform collective operations on
1278 * - submit pages for an immediate transfer,
1280 * - own pages on behalf of certain io (waiting for each page in turn),
1284 * When list is finalized, it releases references on all pages it still has.
1286 * \todo XXX concurrency control.
1290 struct cl_page_list {
1292 struct list_head pl_pages;
1293 struct task_struct *pl_owner;
1297 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1298 * contains an incoming page list and an outgoing page list.
1301 struct cl_page_list c2_qin;
1302 struct cl_page_list c2_qout;
1305 /** @} cl_page_list */
1307 /** \addtogroup cl_io cl_io
1313 * cl_io represents a high level I/O activity like
1314 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1317 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1318 * important distinction. We want to minimize number of calls to the allocator
1319 * in the fast path, e.g., in the case of read(2) when everything is cached:
1320 * client already owns the lock over region being read, and data are cached
1321 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1322 * per-layer io state is stored in the session, associated with the io, see
1323 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1324 * by using free-lists, see cl_env_get().
1326 * There is a small predefined number of possible io types, enumerated in enum
1329 * cl_io is a state machine, that can be advanced concurrently by the multiple
1330 * threads. It is up to these threads to control the concurrency and,
1331 * specifically, to detect when io is done, and its state can be safely
1334 * For read/write io overall execution plan is as following:
1336 * (0) initialize io state through all layers;
1338 * (1) loop: prepare chunk of work to do
1340 * (2) call all layers to collect locks they need to process current chunk
1342 * (3) sort all locks to avoid dead-locks, and acquire them
1344 * (4) process the chunk: call per-page methods
1345 * cl_io_operations::cio_prepare_write(),
1346 * cl_io_operations::cio_commit_write() for write)
1352 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1353 * address allocation efficiency issues mentioned above), and returns with the
1354 * special error condition from per-page method when current sub-io has to
1355 * block. This causes io loop to be repeated, and lov switches to the next
1356 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1361 /** read system call */
1363 /** write system call */
1365 /** truncate, utime system calls */
1367 /** get data version */
1370 * page fault handling
1374 * fsync system call handling
1375 * To write out a range of file
1379 * Miscellaneous io. This is used for occasional io activity that
1380 * doesn't fit into other types. Currently this is used for:
1382 * - cancellation of an extent lock. This io exists as a context
1383 * to write dirty pages from under the lock being canceled back
1386 * - VM induced page write-out. An io context for writing page out
1387 * for memory cleansing;
1389 * - glimpse. An io context to acquire glimpse lock.
1391 * - grouplock. An io context to acquire group lock.
1393 * CIT_MISC io is used simply as a context in which locks and pages
1394 * are manipulated. Such io has no internal "process", that is,
1395 * cl_io_loop() is never called for it.
1402 * States of cl_io state machine
1405 /** Not initialized. */
1409 /** IO iteration started. */
1413 /** Actual IO is in progress. */
1415 /** IO for the current iteration finished. */
1417 /** Locks released. */
1419 /** Iteration completed. */
1421 /** cl_io finalized. */
1426 * IO state private for a layer.
1428 * This is usually embedded into layer session data, rather than allocated
1431 * \see vvp_io, lov_io, osc_io
1433 struct cl_io_slice {
1434 struct cl_io *cis_io;
1435 /** corresponding object slice. Immutable after creation. */
1436 struct cl_object *cis_obj;
1437 /** io operations. Immutable after creation. */
1438 const struct cl_io_operations *cis_iop;
1440 * linkage into a list of all slices for a given cl_io, hanging off
1441 * cl_io::ci_layers. Immutable after creation.
1443 struct list_head cis_linkage;
1446 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1449 struct cl_read_ahead {
1451 * Maximum page index the readahead window will end.
1452 * This is determined DLM lock coverage, RPC and stripe boundary.
1453 * cra_end is included.
1456 /* optimal RPC size for this read, by pages */
1457 unsigned long cra_rpc_size;
1459 * Release callback. If readahead holds resources underneath, this
1460 * function should be called to release it.
1462 void (*cra_release)(const struct lu_env *env, void *cbdata);
1463 /* Callback data for cra_release routine */
1467 static inline void cl_read_ahead_release(const struct lu_env *env,
1468 struct cl_read_ahead *ra)
1470 if (ra->cra_release)
1471 ra->cra_release(env, ra->cra_cbdata);
1472 memset(ra, 0, sizeof(*ra));
1476 * Per-layer io operations.
1477 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1479 struct cl_io_operations {
1481 * Vector of io state transition methods for every io type.
1483 * \see cl_page_operations::io
1487 * Prepare io iteration at a given layer.
1489 * Called top-to-bottom at the beginning of each iteration of
1490 * "io loop" (if it makes sense for this type of io). Here
1491 * layer selects what work it will do during this iteration.
1493 * \see cl_io_operations::cio_iter_fini()
1495 int (*cio_iter_init)(const struct lu_env *env,
1496 const struct cl_io_slice *slice);
1498 * Finalize io iteration.
1500 * Called bottom-to-top at the end of each iteration of "io
1501 * loop". Here layers can decide whether IO has to be
1504 * \see cl_io_operations::cio_iter_init()
1506 void (*cio_iter_fini)(const struct lu_env *env,
1507 const struct cl_io_slice *slice);
1509 * Collect locks for the current iteration of io.
1511 * Called top-to-bottom to collect all locks necessary for
1512 * this iteration. This methods shouldn't actually enqueue
1513 * anything, instead it should post a lock through
1514 * cl_io_lock_add(). Once all locks are collected, they are
1515 * sorted and enqueued in the proper order.
1517 int (*cio_lock)(const struct lu_env *env,
1518 const struct cl_io_slice *slice);
1520 * Finalize unlocking.
1522 * Called bottom-to-top to finish layer specific unlocking
1523 * functionality, after generic code released all locks
1524 * acquired by cl_io_operations::cio_lock().
1526 void (*cio_unlock)(const struct lu_env *env,
1527 const struct cl_io_slice *slice);
1529 * Start io iteration.
1531 * Once all locks are acquired, called top-to-bottom to
1532 * commence actual IO. In the current implementation,
1533 * top-level vvp_io_{read,write}_start() does all the work
1534 * synchronously by calling generic_file_*(), so other layers
1535 * are called when everything is done.
1537 int (*cio_start)(const struct lu_env *env,
1538 const struct cl_io_slice *slice);
1540 * Called top-to-bottom at the end of io loop. Here layer
1541 * might wait for an unfinished asynchronous io.
1543 void (*cio_end)(const struct lu_env *env,
1544 const struct cl_io_slice *slice);
1546 * Called bottom-to-top to notify layers that read/write IO
1547 * iteration finished, with \a nob bytes transferred.
1549 void (*cio_advance)(const struct lu_env *env,
1550 const struct cl_io_slice *slice,
1553 * Called once per io, bottom-to-top to release io resources.
1555 void (*cio_fini)(const struct lu_env *env,
1556 const struct cl_io_slice *slice);
1560 * Submit pages from \a queue->c2_qin for IO, and move
1561 * successfully submitted pages into \a queue->c2_qout. Return
1562 * non-zero if failed to submit even the single page. If
1563 * submission failed after some pages were moved into \a
1564 * queue->c2_qout, completion callback with non-zero ioret is
1567 int (*cio_submit)(const struct lu_env *env,
1568 const struct cl_io_slice *slice,
1569 enum cl_req_type crt,
1570 struct cl_2queue *queue);
1572 * Queue async page for write.
1573 * The difference between cio_submit and cio_queue is that
1574 * cio_submit is for urgent request.
1576 int (*cio_commit_async)(const struct lu_env *env,
1577 const struct cl_io_slice *slice,
1578 struct cl_page_list *queue, int from, int to,
1581 * Decide maximum read ahead extent
1583 * \pre io->ci_type == CIT_READ
1585 int (*cio_read_ahead)(const struct lu_env *env,
1586 const struct cl_io_slice *slice,
1587 pgoff_t start, struct cl_read_ahead *ra);
1589 * Optional debugging helper. Print given io slice.
1591 int (*cio_print)(const struct lu_env *env, void *cookie,
1592 lu_printer_t p, const struct cl_io_slice *slice);
1596 * Flags to lock enqueue procedure.
1601 * instruct server to not block, if conflicting lock is found. Instead
1602 * -EWOULDBLOCK is returned immediately.
1604 CEF_NONBLOCK = 0x00000001,
1606 * take lock asynchronously (out of order), as it cannot
1607 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1609 CEF_ASYNC = 0x00000002,
1611 * tell the server to instruct (though a flag in the blocking ast) an
1612 * owner of the conflicting lock, that it can drop dirty pages
1613 * protected by this lock, without sending them to the server.
1615 CEF_DISCARD_DATA = 0x00000004,
1617 * tell the sub layers that it must be a `real' lock. This is used for
1618 * mmapped-buffer locks and glimpse locks that must be never converted
1619 * into lockless mode.
1621 * \see vvp_mmap_locks(), cl_glimpse_lock().
1623 CEF_MUST = 0x00000008,
1625 * tell the sub layers that never request a `real' lock. This flag is
1626 * not used currently.
1628 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1629 * conversion policy: ci_lockreq describes generic information of lock
1630 * requirement for this IO, especially for locks which belong to the
1631 * object doing IO; however, lock itself may have precise requirements
1632 * that are described by the enqueue flags.
1634 CEF_NEVER = 0x00000010,
1636 * for async glimpse lock.
1638 CEF_AGL = 0x00000020,
1640 * enqueue a lock to test DLM lock existence.
1642 CEF_PEEK = 0x00000040,
1644 * Lock match only. Used by group lock in I/O as group lock
1645 * is known to exist.
1647 CEF_LOCK_MATCH = BIT(7),
1649 * mask of enq_flags.
1651 CEF_MASK = 0x000000ff,
1655 * Link between lock and io. Intermediate structure is needed, because the
1656 * same lock can be part of multiple io's simultaneously.
1658 struct cl_io_lock_link {
1659 /** linkage into one of cl_lockset lists. */
1660 struct list_head cill_linkage;
1661 struct cl_lock cill_lock;
1662 /** optional destructor */
1663 void (*cill_fini)(const struct lu_env *env,
1664 struct cl_io_lock_link *link);
1666 #define cill_descr cill_lock.cll_descr
1669 * Lock-set represents a collection of locks, that io needs at a
1670 * time. Generally speaking, client tries to avoid holding multiple locks when
1673 * - holding extent locks over multiple ost's introduces the danger of
1674 * "cascading timeouts";
1676 * - holding multiple locks over the same ost is still dead-lock prone,
1677 * see comment in osc_lock_enqueue(),
1679 * but there are certain situations where this is unavoidable:
1681 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1683 * - truncate has to take [new-size, EOF] lock for correctness;
1685 * - SNS has to take locks across full stripe for correctness;
1687 * - in the case when user level buffer, supplied to {read,write}(file0),
1688 * is a part of a memory mapped lustre file, client has to take a dlm
1689 * locks on file0, and all files that back up the buffer (or a part of
1690 * the buffer, that is being processed in the current chunk, in any
1691 * case, there are situations where at least 2 locks are necessary).
1693 * In such cases we at least try to take locks in the same consistent
1694 * order. To this end, all locks are first collected, then sorted, and then
1698 /** locks to be acquired. */
1699 struct list_head cls_todo;
1700 /** locks acquired. */
1701 struct list_head cls_done;
1705 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1706 * but 'req' is always to be thought as 'request' :-)
1708 enum cl_io_lock_dmd {
1709 /** Always lock data (e.g., O_APPEND). */
1711 /** Layers are free to decide between local and global locking. */
1713 /** Never lock: there is no cache (e.g., lockless IO). */
1717 enum cl_fsync_mode {
1718 /** start writeback, do not wait for them to finish */
1720 /** start writeback and wait for them to finish */
1722 /** discard all of dirty pages in a specific file range */
1723 CL_FSYNC_DISCARD = 2,
1724 /** start writeback and make sure they have reached storage before
1725 * return. OST_SYNC RPC must be issued and finished
1730 struct cl_io_rw_common {
1739 * cl_io is shared by all threads participating in this IO (in current
1740 * implementation only one thread advances IO, but parallel IO design and
1741 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1742 * is up to these threads to serialize their activities, including updates to
1743 * mutable cl_io fields.
1746 /** type of this IO. Immutable after creation. */
1747 enum cl_io_type ci_type;
1748 /** current state of cl_io state machine. */
1749 enum cl_io_state ci_state;
1750 /** main object this io is against. Immutable after creation. */
1751 struct cl_object *ci_obj;
1753 * Upper layer io, of which this io is a part of. Immutable after
1756 struct cl_io *ci_parent;
1757 /** List of slices. Immutable after creation. */
1758 struct list_head ci_layers;
1759 /** list of locks (to be) acquired by this io. */
1760 struct cl_lockset ci_lockset;
1761 /** lock requirements, this is just a help info for sublayers. */
1762 enum cl_io_lock_dmd ci_lockreq;
1765 struct cl_io_rw_common rd;
1768 struct cl_io_rw_common wr;
1772 struct cl_io_rw_common ci_rw;
1773 struct cl_setattr_io {
1774 struct ost_lvb sa_attr;
1775 unsigned int sa_attr_flags;
1776 unsigned int sa_valid;
1777 int sa_stripe_index;
1778 const struct lu_fid *sa_parent_fid;
1780 struct cl_data_version_io {
1781 u64 dv_data_version;
1784 struct cl_fault_io {
1785 /** page index within file. */
1787 /** bytes valid byte on a faulted page. */
1789 /** writable page? for nopage() only */
1791 /** page of an executable? */
1793 /** page_mkwrite() */
1795 /** resulting page */
1796 struct cl_page *ft_page;
1798 struct cl_fsync_io {
1801 /** file system level fid */
1802 struct lu_fid *fi_fid;
1803 enum cl_fsync_mode fi_mode;
1804 /* how many pages were written/discarded */
1805 unsigned int fi_nr_written;
1808 struct cl_2queue ci_queue;
1811 unsigned int ci_continue:1,
1813 * This io has held grouplock, to inform sublayers that
1814 * don't do lockless i/o.
1818 * The whole IO need to be restarted because layout has been changed
1822 * to not refresh layout - the IO issuer knows that the layout won't
1823 * change(page operations, layout change causes all page to be
1824 * discarded), or it doesn't matter if it changes(sync).
1828 * Check if layout changed after the IO finishes. Mainly for HSM
1829 * requirement. If IO occurs to openning files, it doesn't need to
1830 * verify layout because HSM won't release openning files.
1831 * Right now, only two operations need to verify layout: glimpse
1836 * file is released, restore has to to be triggered by vvp layer
1838 ci_restore_needed:1,
1844 * Number of pages owned by this IO. For invariant checking.
1846 unsigned int ci_owned_nr;
1852 * Per-transfer attributes.
1854 struct cl_req_attr {
1855 enum cl_req_type cra_type;
1857 struct cl_page *cra_page;
1859 /** Generic attributes for the server consumption. */
1860 struct obdo *cra_oa;
1862 char cra_jobid[LUSTRE_JOBID_SIZE];
1865 enum cache_stats_item {
1866 /** how many cache lookups were performed */
1868 /** how many times cache lookup resulted in a hit */
1870 /** how many entities are in the cache right now */
1872 /** how many entities in the cache are actively used (and cannot be
1873 * evicted) right now
1876 /** how many entities were created at all */
1881 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
1884 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
1886 struct cache_stats {
1887 const char *cs_name;
1888 atomic_t cs_stats[CS_NR];
1891 /** These are not exported so far */
1892 void cache_stats_init(struct cache_stats *cs, const char *name);
1895 * Client-side site. This represents particular client stack. "Global"
1896 * variables should (directly or indirectly) be added here to allow multiple
1897 * clients to co-exist in the single address space.
1900 struct lu_site cs_lu;
1902 * Statistical counters. Atomics do not scale, something better like
1903 * per-cpu counters is needed.
1905 * These are exported as /sys/kernel/debug/lustre/llite/.../site
1907 * When interpreting keep in mind that both sub-locks (and sub-pages)
1908 * and top-locks (and top-pages) are accounted here.
1910 struct cache_stats cs_pages;
1911 atomic_t cs_pages_state[CPS_NR];
1914 int cl_site_init(struct cl_site *s, struct cl_device *top);
1915 void cl_site_fini(struct cl_site *s);
1916 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
1919 * Output client site statistical counters into a buffer. Suitable for
1920 * ll_rd_*()-style functions.
1922 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
1927 * Type conversion and accessory functions.
1931 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
1933 return container_of(site, struct cl_site, cs_lu);
1936 static inline int lu_device_is_cl(const struct lu_device *d)
1938 return d->ld_type->ldt_tags & LU_DEVICE_CL;
1941 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
1943 LASSERT(!d || IS_ERR(d) || lu_device_is_cl(d));
1944 return container_of0(d, struct cl_device, cd_lu_dev);
1947 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
1949 return &d->cd_lu_dev;
1952 static inline struct cl_object *lu2cl(const struct lu_object *o)
1954 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
1955 return container_of0(o, struct cl_object, co_lu);
1958 static inline const struct cl_object_conf *
1959 lu2cl_conf(const struct lu_object_conf *conf)
1961 return container_of0(conf, struct cl_object_conf, coc_lu);
1964 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
1966 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
1969 static inline struct cl_device *cl_object_device(const struct cl_object *o)
1971 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
1972 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
1975 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
1977 return container_of0(h, struct cl_object_header, coh_lu);
1980 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
1982 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
1986 struct cl_object_header *cl_object_header(const struct cl_object *obj)
1988 return luh2coh(obj->co_lu.lo_header);
1991 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
1993 return lu_device_init(&d->cd_lu_dev, t);
1996 static inline void cl_device_fini(struct cl_device *d)
1998 lu_device_fini(&d->cd_lu_dev);
2001 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2002 struct cl_object *obj, pgoff_t index,
2003 const struct cl_page_operations *ops);
2004 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2005 struct cl_object *obj,
2006 const struct cl_lock_operations *ops);
2007 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2008 struct cl_object *obj, const struct cl_io_operations *ops);
2011 /** \defgroup cl_object cl_object
2014 struct cl_object *cl_object_top(struct cl_object *o);
2015 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2016 const struct lu_fid *fid,
2017 const struct cl_object_conf *c);
2019 int cl_object_header_init(struct cl_object_header *h);
2020 void cl_object_put(const struct lu_env *env, struct cl_object *o);
2021 void cl_object_get(struct cl_object *o);
2022 void cl_object_attr_lock(struct cl_object *o);
2023 void cl_object_attr_unlock(struct cl_object *o);
2024 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2025 struct cl_attr *attr);
2026 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2027 const struct cl_attr *attr, unsigned int valid);
2028 int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj,
2029 struct ost_lvb *lvb);
2030 int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
2031 const struct cl_object_conf *conf);
2032 int cl_object_prune(const struct lu_env *env, struct cl_object *obj);
2033 void cl_object_kill(const struct lu_env *env, struct cl_object *obj);
2034 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2035 struct lov_user_md __user *lum);
2036 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2037 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2039 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2040 struct cl_layout *cl);
2041 loff_t cl_object_maxbytes(struct cl_object *obj);
2044 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2046 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2048 return cl_object_header(o0) == cl_object_header(o1);
2051 static inline void cl_object_page_init(struct cl_object *clob, int size)
2053 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2054 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2055 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2058 static inline void *cl_object_page_slice(struct cl_object *clob,
2059 struct cl_page *page)
2061 return (void *)((char *)page + clob->co_slice_off);
2065 * Return refcount of cl_object.
2067 static inline int cl_object_refc(struct cl_object *clob)
2069 struct lu_object_header *header = clob->co_lu.lo_header;
2071 return atomic_read(&header->loh_ref);
2076 /** \defgroup cl_page cl_page
2086 /* callback of cl_page_gang_lookup() */
2087 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *obj,
2088 pgoff_t idx, struct page *vmpage,
2089 enum cl_page_type type);
2090 struct cl_page *cl_page_alloc(const struct lu_env *env,
2091 struct cl_object *o, pgoff_t ind,
2092 struct page *vmpage,
2093 enum cl_page_type type);
2094 void cl_page_get(struct cl_page *page);
2095 void cl_page_put(const struct lu_env *env, struct cl_page *page);
2096 void cl_page_print(const struct lu_env *env, void *cookie, lu_printer_t printer,
2097 const struct cl_page *pg);
2098 void cl_page_header_print(const struct lu_env *env, void *cookie,
2099 lu_printer_t printer, const struct cl_page *pg);
2100 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj);
2102 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2103 const struct lu_device_type *dtype);
2108 * Functions dealing with the ownership of page by io.
2112 int cl_page_own(const struct lu_env *env,
2113 struct cl_io *io, struct cl_page *page);
2114 int cl_page_own_try(const struct lu_env *env,
2115 struct cl_io *io, struct cl_page *page);
2116 void cl_page_assume(const struct lu_env *env,
2117 struct cl_io *io, struct cl_page *page);
2118 void cl_page_unassume(const struct lu_env *env,
2119 struct cl_io *io, struct cl_page *pg);
2120 void cl_page_disown(const struct lu_env *env,
2121 struct cl_io *io, struct cl_page *page);
2122 void cl_page_disown0(const struct lu_env *env,
2123 struct cl_io *io, struct cl_page *pg);
2124 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io);
2131 * Functions dealing with the preparation of a page for a transfer, and
2132 * tracking transfer state.
2135 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
2136 struct cl_page *pg, enum cl_req_type crt);
2137 void cl_page_completion(const struct lu_env *env,
2138 struct cl_page *pg, enum cl_req_type crt, int ioret);
2139 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
2140 enum cl_req_type crt);
2141 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
2142 struct cl_page *pg, enum cl_req_type crt);
2143 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
2145 int cl_page_cancel(const struct lu_env *env, struct cl_page *page);
2146 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
2147 struct cl_page *pg);
2152 * \name helper routines
2153 * Functions to discard, delete and export a cl_page.
2156 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2157 struct cl_page *pg);
2158 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2159 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg);
2160 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate);
2161 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2162 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2163 size_t cl_page_size(const struct cl_object *obj);
2164 int cl_pages_prune(const struct lu_env *env, struct cl_object *obj);
2166 void cl_lock_print(const struct lu_env *env, void *cookie,
2167 lu_printer_t printer, const struct cl_lock *lock);
2168 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2169 lu_printer_t printer,
2170 const struct cl_lock_descr *descr);
2174 * Data structure managing a client's cached pages. A count of
2175 * "unstable" pages is maintained, and an LRU of clean pages is
2176 * maintained. "unstable" pages are pages pinned by the ptlrpc
2177 * layer for recovery purposes.
2179 struct cl_client_cache {
2181 * # of client cache refcount
2182 * # of users (OSCs) + 2 (held by llite and lov)
2186 * # of threads are doing shrinking
2188 unsigned int ccc_lru_shrinkers;
2190 * # of LRU entries available
2192 atomic_long_t ccc_lru_left;
2194 * List of entities(OSCs) for this LRU cache
2196 struct list_head ccc_lru;
2198 * Max # of LRU entries
2200 unsigned long ccc_lru_max;
2202 * Lock to protect ccc_lru list
2204 spinlock_t ccc_lru_lock;
2206 * Set if unstable check is enabled
2208 unsigned int ccc_unstable_check:1;
2210 * # of unstable pages for this mount point
2212 atomic_long_t ccc_unstable_nr;
2214 * Waitq for awaiting unstable pages to reach zero.
2215 * Used at umounting time and signaled on BRW commit
2217 wait_queue_head_t ccc_unstable_waitq;
2222 * cl_cache functions
2224 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2225 void cl_cache_incref(struct cl_client_cache *cache);
2226 void cl_cache_decref(struct cl_client_cache *cache);
2230 /** \defgroup cl_lock cl_lock
2234 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2235 struct cl_lock *lock);
2236 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2237 const struct cl_io *io);
2238 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2239 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2240 const struct lu_device_type *dtype);
2241 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2242 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2243 struct cl_lock *lock, struct cl_sync_io *anchor);
2244 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2248 /** \defgroup cl_io cl_io
2252 int cl_io_init(const struct lu_env *env, struct cl_io *io,
2253 enum cl_io_type iot, struct cl_object *obj);
2254 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
2255 enum cl_io_type iot, struct cl_object *obj);
2256 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
2257 enum cl_io_type iot, loff_t pos, size_t count);
2258 int cl_io_loop(const struct lu_env *env, struct cl_io *io);
2260 void cl_io_fini(const struct lu_env *env, struct cl_io *io);
2261 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io);
2262 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io);
2263 int cl_io_lock(const struct lu_env *env, struct cl_io *io);
2264 void cl_io_unlock(const struct lu_env *env, struct cl_io *io);
2265 int cl_io_start(const struct lu_env *env, struct cl_io *io);
2266 void cl_io_end(const struct lu_env *env, struct cl_io *io);
2267 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
2268 struct cl_io_lock_link *link);
2269 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2270 struct cl_lock_descr *descr);
2271 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
2272 enum cl_req_type iot, struct cl_2queue *queue);
2273 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
2274 enum cl_req_type iot, struct cl_2queue *queue,
2276 int cl_io_commit_async(const struct lu_env *env, struct cl_io *io,
2277 struct cl_page_list *queue, int from, int to,
2279 int cl_io_read_ahead(const struct lu_env *env, struct cl_io *io,
2280 pgoff_t start, struct cl_read_ahead *ra);
2281 int cl_io_is_going(const struct lu_env *env);
2284 * True, iff \a io is an O_APPEND write(2).
2286 static inline int cl_io_is_append(const struct cl_io *io)
2288 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2291 static inline int cl_io_is_sync_write(const struct cl_io *io)
2293 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2296 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2298 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2302 * True, iff \a io is a truncate(2).
2304 static inline int cl_io_is_trunc(const struct cl_io *io)
2306 return io->ci_type == CIT_SETATTR &&
2307 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2310 struct cl_io *cl_io_top(struct cl_io *io);
2312 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2314 typeof(foo_io) __foo_io = (foo_io); \
2316 BUILD_BUG_ON(offsetof(typeof(*__foo_io), base) != 0); \
2317 memset(&__foo_io->base + 1, 0, \
2318 sizeof(*__foo_io) - sizeof(__foo_io->base)); \
2323 /** \defgroup cl_page_list cl_page_list
2328 * Last page in the page list.
2330 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2332 LASSERT(plist->pl_nr > 0);
2333 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2336 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2338 LASSERT(plist->pl_nr > 0);
2339 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2343 * Iterate over pages in a page list.
2345 #define cl_page_list_for_each(page, list) \
2346 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2349 * Iterate over pages in a page list, taking possible removals into account.
2351 #define cl_page_list_for_each_safe(page, temp, list) \
2352 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2354 void cl_page_list_init(struct cl_page_list *plist);
2355 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page);
2356 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2357 struct cl_page *page);
2358 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2359 struct cl_page *page);
2360 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head);
2361 void cl_page_list_del(const struct lu_env *env, struct cl_page_list *plist,
2362 struct cl_page *page);
2363 void cl_page_list_disown(const struct lu_env *env,
2364 struct cl_io *io, struct cl_page_list *plist);
2365 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2367 void cl_2queue_init(struct cl_2queue *queue);
2368 void cl_2queue_disown(const struct lu_env *env,
2369 struct cl_io *io, struct cl_2queue *queue);
2370 void cl_2queue_discard(const struct lu_env *env,
2371 struct cl_io *io, struct cl_2queue *queue);
2372 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2373 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2375 /** @} cl_page_list */
2377 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2378 struct cl_req_attr *attr);
2380 /** \defgroup cl_sync_io cl_sync_io
2385 * Anchor for synchronous transfer. This is allocated on a stack by thread
2386 * doing synchronous transfer, and a pointer to this structure is set up in
2387 * every page submitted for transfer. Transfer completion routine updates
2388 * anchor and wakes up waiting thread when transfer is complete.
2391 /** number of pages yet to be transferred. */
2392 atomic_t csi_sync_nr;
2395 /** barrier of destroy this structure */
2396 atomic_t csi_barrier;
2397 /** completion to be signaled when transfer is complete. */
2398 wait_queue_head_t csi_waitq;
2399 /** callback to invoke when this IO is finished */
2400 void (*csi_end_io)(const struct lu_env *,
2401 struct cl_sync_io *);
2404 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2405 void (*end)(const struct lu_env *, struct cl_sync_io *));
2406 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2408 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2410 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2412 /** @} cl_sync_io */
2414 /** \defgroup cl_env cl_env
2416 * lu_env handling for a client.
2418 * lu_env is an environment within which lustre code executes. Its major part
2419 * is lu_context---a fast memory allocation mechanism that is used to conserve
2420 * precious kernel stack space. Originally lu_env was designed for a server,
2423 * - there is a (mostly) fixed number of threads, and
2425 * - call chains have no non-lustre portions inserted between lustre code.
2427 * On a client both these assumption fails, because every user thread can
2428 * potentially execute lustre code as part of a system call, and lustre calls
2429 * into VFS or MM that call back into lustre.
2431 * To deal with that, cl_env wrapper functions implement the following
2434 * - allocation and destruction of environment is amortized by caching no
2435 * longer used environments instead of destroying them;
2437 * \see lu_env, lu_context, lu_context_key
2441 struct lu_env *cl_env_get(u16 *refcheck);
2442 struct lu_env *cl_env_alloc(u16 *refcheck, __u32 tags);
2443 void cl_env_put(struct lu_env *env, u16 *refcheck);
2444 unsigned int cl_env_cache_purge(unsigned int nr);
2445 struct lu_env *cl_env_percpu_get(void);
2446 void cl_env_percpu_put(struct lu_env *env);
2453 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2455 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2456 struct lu_device_type *ldt,
2457 struct lu_device *next);
2460 int cl_global_init(void);
2461 void cl_global_fini(void);
2463 #endif /* _LINUX_CL_OBJECT_H */