4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 #ifndef _LUSTRE_CL_OBJECT_H
37 #define _LUSTRE_CL_OBJECT_H
39 /** \defgroup clio clio
41 * Client objects implement io operations and cache pages.
43 * Examples: lov and osc are implementations of cl interface.
45 * Big Theory Statement.
49 * Client implementation is based on the following data-types:
55 * - cl_lock represents an extent lock on an object.
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).
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.
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"
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.
84 * - cl_object_header::coh_page_guard
85 * - cl_object_header::coh_lock_guard
88 * See the top comment in cl_object.c for the description of overall locking and
89 * reference-counting design.
91 * See comments below for the description of i/o, page, and dlm-locking
98 * super-class definitions.
100 #include <lu_object.h>
102 # include <linux/mutex.h>
103 # include <linux/radix-tree.h>
109 struct cl_device_operations;
112 struct cl_object_page_operations;
113 struct cl_object_lock_operations;
116 struct cl_page_slice;
118 struct cl_lock_slice;
120 struct cl_lock_operations;
121 struct cl_page_operations;
130 * Operations for each data device in the client stack.
132 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
134 struct cl_device_operations {
136 * Initialize cl_req. This method is called top-to-bottom on all
137 * devices in the stack to get them a chance to allocate layer-private
138 * data, and to attach them to the cl_req by calling
139 * cl_req_slice_add().
141 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
142 * \see ccc_req_init()
144 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
149 * Device in the client stack.
151 * \see ccc_device, lov_device, lovsub_device, osc_device
155 struct lu_device cd_lu_dev;
156 /** Per-layer operation vector. */
157 const struct cl_device_operations *cd_ops;
160 /** \addtogroup cl_object cl_object
163 * "Data attributes" of cl_object. Data attributes can be updated
164 * independently for a sub-object, and top-object's attributes are calculated
165 * from sub-objects' ones.
168 /** Object size, in bytes */
171 * Known minimal size, in bytes.
173 * This is only valid when at least one DLM lock is held.
176 /** Modification time. Measured in seconds since epoch. */
178 /** Access time. Measured in seconds since epoch. */
180 /** Change time. Measured in seconds since epoch. */
183 * Blocks allocated to this cl_object on the server file system.
185 * \todo XXX An interface for block size is needed.
189 * User identifier for quota purposes.
193 * Group identifier for quota purposes.
199 * Fields in cl_attr that are being set.
213 * Sub-class of lu_object with methods common for objects on the client
216 * cl_object: represents a regular file system object, both a file and a
217 * stripe. cl_object is based on lu_object: it is identified by a fid,
218 * layered, cached, hashed, and lrued. Important distinction with the server
219 * side, where md_object and dt_object are used, is that cl_object "fans out"
220 * at the lov/sns level: depending on the file layout, single file is
221 * represented as a set of "sub-objects" (stripes). At the implementation
222 * level, struct lov_object contains an array of cl_objects. Each sub-object
223 * is a full-fledged cl_object, having its fid, living in the lru and hash
226 * This leads to the next important difference with the server side: on the
227 * client, it's quite usual to have objects with the different sequence of
228 * layers. For example, typical top-object is composed of the following
234 * whereas its sub-objects are composed of
239 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
240 * track of the object-subobject relationship.
242 * Sub-objects are not cached independently: when top-object is about to
243 * be discarded from the memory, all its sub-objects are torn-down and
246 * \see ccc_object, lov_object, lovsub_object, osc_object
250 struct lu_object co_lu;
251 /** per-object-layer operations */
252 const struct cl_object_operations *co_ops;
253 /** offset of page slice in cl_page buffer */
258 * Description of the client object configuration. This is used for the
259 * creation of a new client object that is identified by a more state than
262 struct cl_object_conf {
264 struct lu_object_conf coc_lu;
267 * Object layout. This is consumed by lov.
269 struct lustre_md *coc_md;
271 * Description of particular stripe location in the
272 * cluster. This is consumed by osc.
274 struct lov_oinfo *coc_oinfo;
277 * VFS inode. This is consumed by vvp.
279 struct inode *coc_inode;
281 * Layout lock handle.
283 struct ldlm_lock *coc_lock;
285 * Operation to handle layout, OBJECT_CONF_XYZ.
291 /** configure layout, set up a new stripe, must be called while
292 * holding layout lock. */
294 /** invalidate the current stripe configuration due to losing
296 OBJECT_CONF_INVALIDATE = 1,
297 /** wait for old layout to go away so that new layout can be
303 * Operations implemented for each cl object layer.
305 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
307 struct cl_object_operations {
309 * Initialize page slice for this layer. Called top-to-bottom through
310 * every object layer when a new cl_page is instantiated. Layer
311 * keeping private per-page data, or requiring its own page operations
312 * vector should allocate these data here, and attach then to the page
313 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
316 * \retval NULL success.
318 * \retval ERR_PTR(errno) failure code.
320 * \retval valid-pointer pointer to already existing referenced page
321 * to be used instead of newly created.
323 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
324 struct cl_page *page, pgoff_t index);
326 * Initialize lock slice for this layer. Called top-to-bottom through
327 * every object layer when a new cl_lock is instantiated. Layer
328 * keeping private per-lock data, or requiring its own lock operations
329 * vector should allocate these data here, and attach then to the lock
330 * by calling cl_lock_slice_add(). Mandatory.
332 int (*coo_lock_init)(const struct lu_env *env,
333 struct cl_object *obj, struct cl_lock *lock,
334 const struct cl_io *io);
336 * Initialize io state for a given layer.
338 * called top-to-bottom once per io existence to initialize io
339 * state. If layer wants to keep some state for this type of io, it
340 * has to embed struct cl_io_slice in lu_env::le_ses, and register
341 * slice with cl_io_slice_add(). It is guaranteed that all threads
342 * participating in this io share the same session.
344 int (*coo_io_init)(const struct lu_env *env,
345 struct cl_object *obj, struct cl_io *io);
347 * Fill portion of \a attr that this layer controls. This method is
348 * called top-to-bottom through all object layers.
350 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
352 * \return 0: to continue
353 * \return +ve: to stop iterating through layers (but 0 is returned
354 * from enclosing cl_object_attr_get())
355 * \return -ve: to signal error
357 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
358 struct cl_attr *attr);
362 * \a valid is a bitmask composed from enum #cl_attr_valid, and
363 * indicating what attributes are to be set.
365 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
367 * \return the same convention as for
368 * cl_object_operations::coo_attr_get() is used.
370 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj,
371 const struct cl_attr *attr, unsigned valid);
373 * Update object configuration. Called top-to-bottom to modify object
376 * XXX error conditions and handling.
378 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
379 const struct cl_object_conf *conf);
381 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
382 * object. Layers are supposed to fill parts of \a lvb that will be
383 * shipped to the glimpse originator as a glimpse result.
385 * \see ccc_object_glimpse(), lovsub_object_glimpse(),
386 * \see osc_object_glimpse()
388 int (*coo_glimpse)(const struct lu_env *env,
389 const struct cl_object *obj, struct ost_lvb *lvb);
391 * Object prune method. Called when the layout is going to change on
392 * this object, therefore each layer has to clean up their cache,
393 * mainly pages and locks.
395 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
399 * Extended header for client object.
401 struct cl_object_header {
402 /** Standard lu_object_header. cl_object::co_lu::lo_header points
404 struct lu_object_header coh_lu;
406 * \todo XXX move locks below to the separate cache-lines, they are
407 * mostly useless otherwise.
410 /** Lock protecting lock list. */
411 spinlock_t coh_lock_guard;
413 /** List of cl_lock's granted for this object. */
414 cfs_list_t coh_locks;
417 * Parent object. It is assumed that an object has a well-defined
418 * parent, but not a well-defined child (there may be multiple
419 * sub-objects, for the same top-object). cl_object_header::coh_parent
420 * field allows certain code to be written generically, without
421 * limiting possible cl_object layouts unduly.
423 struct cl_object_header *coh_parent;
425 * Protects consistency between cl_attr of parent object and
426 * attributes of sub-objects, that the former is calculated ("merged")
429 * \todo XXX this can be read/write lock if needed.
431 spinlock_t coh_attr_guard;
433 * Size of cl_page + page slices
435 unsigned short coh_page_bufsize;
437 * Number of objects above this one: 0 for a top-object, 1 for its
440 unsigned char coh_nesting;
444 * Helper macro: iterate over all layers of the object \a obj, assigning every
445 * layer top-to-bottom to \a slice.
447 #define cl_object_for_each(slice, obj) \
448 cfs_list_for_each_entry((slice), \
449 &(obj)->co_lu.lo_header->loh_layers, \
452 * Helper macro: iterate over all layers of the object \a obj, assigning every
453 * layer bottom-to-top to \a slice.
455 #define cl_object_for_each_reverse(slice, obj) \
456 cfs_list_for_each_entry_reverse((slice), \
457 &(obj)->co_lu.lo_header->loh_layers, \
461 #define CL_PAGE_EOF ((pgoff_t)~0ull)
463 /** \addtogroup cl_page cl_page
467 * Layered client page.
469 * cl_page: represents a portion of a file, cached in the memory. All pages
470 * of the given file are of the same size, and are kept in the radix tree
471 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
472 * of the top-level file object are first class cl_objects, they have their
473 * own radix trees of pages and hence page is implemented as a sequence of
474 * struct cl_pages's, linked into double-linked list through
475 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
476 * corresponding radix tree at the corresponding logical offset.
478 * cl_page is associated with VM page of the hosting environment (struct
479 * page in Linux kernel, for example), struct page. It is assumed, that this
480 * association is implemented by one of cl_page layers (top layer in the
481 * current design) that
483 * - intercepts per-VM-page call-backs made by the environment (e.g.,
486 * - translates state (page flag bits) and locking between lustre and
489 * The association between cl_page and struct page is immutable and
490 * established when cl_page is created.
492 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
493 * this io an exclusive access to this page w.r.t. other io attempts and
494 * various events changing page state (such as transfer completion, or
495 * eviction of the page from the memory). Note, that in general cl_io
496 * cannot be identified with a particular thread, and page ownership is not
497 * exactly equal to the current thread holding a lock on the page. Layer
498 * implementing association between cl_page and struct page has to implement
499 * ownership on top of available synchronization mechanisms.
501 * While lustre client maintains the notion of an page ownership by io,
502 * hosting MM/VM usually has its own page concurrency control
503 * mechanisms. For example, in Linux, page access is synchronized by the
504 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
505 * takes care to acquire and release such locks as necessary around the
506 * calls to the file system methods (->readpage(), ->prepare_write(),
507 * ->commit_write(), etc.). This leads to the situation when there are two
508 * different ways to own a page in the client:
510 * - client code explicitly and voluntary owns the page (cl_page_own());
512 * - VM locks a page and then calls the client, that has "to assume"
513 * the ownership from the VM (cl_page_assume()).
515 * Dual methods to release ownership are cl_page_disown() and
516 * cl_page_unassume().
518 * cl_page is reference counted (cl_page::cp_ref). When reference counter
519 * drops to 0, the page is returned to the cache, unless it is in
520 * cl_page_state::CPS_FREEING state, in which case it is immediately
523 * The general logic guaranteeing the absence of "existential races" for
524 * pages is the following:
526 * - there are fixed known ways for a thread to obtain a new reference
529 * - by doing a lookup in the cl_object radix tree, protected by the
532 * - by starting from VM-locked struct page and following some
533 * hosting environment method (e.g., following ->private pointer in
534 * the case of Linux kernel), see cl_vmpage_page();
536 * - when the page enters cl_page_state::CPS_FREEING state, all these
537 * ways are severed with the proper synchronization
538 * (cl_page_delete());
540 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
543 * - no new references to the page in cl_page_state::CPS_FREEING state
544 * are allowed (checked in cl_page_get()).
546 * Together this guarantees that when last reference to a
547 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
548 * page, as neither references to it can be acquired at that point, nor
551 * cl_page is a state machine. States are enumerated in enum
552 * cl_page_state. Possible state transitions are enumerated in
553 * cl_page_state_set(). State transition process (i.e., actual changing of
554 * cl_page::cp_state field) is protected by the lock on the underlying VM
557 * Linux Kernel implementation.
559 * Binding between cl_page and struct page (which is a typedef for
560 * struct page) is implemented in the vvp layer. cl_page is attached to the
561 * ->private pointer of the struct page, together with the setting of
562 * PG_private bit in page->flags, and acquiring additional reference on the
563 * struct page (much like struct buffer_head, or any similar file system
564 * private data structures).
566 * PG_locked lock is used to implement both ownership and transfer
567 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
568 * states. No additional references are acquired for the duration of the
571 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
572 * write-out is "protected" by the special PG_writeback bit.
576 * States of cl_page. cl_page.c assumes particular order here.
578 * The page state machine is rather crude, as it doesn't recognize finer page
579 * states like "dirty" or "up to date". This is because such states are not
580 * always well defined for the whole stack (see, for example, the
581 * implementation of the read-ahead, that hides page up-to-dateness to track
582 * cache hits accurately). Such sub-states are maintained by the layers that
583 * are interested in them.
587 * Page is in the cache, un-owned. Page leaves cached state in the
590 * - [cl_page_state::CPS_OWNED] io comes across the page and
593 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
594 * req-formation engine decides that it wants to include this page
595 * into an cl_req being constructed, and yanks it from the cache;
597 * - [cl_page_state::CPS_FREEING] VM callback is executed to
598 * evict the page form the memory;
600 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
604 * Page is exclusively owned by some cl_io. Page may end up in this
605 * state as a result of
607 * - io creating new page and immediately owning it;
609 * - [cl_page_state::CPS_CACHED] io finding existing cached page
612 * - [cl_page_state::CPS_OWNED] io finding existing owned page
613 * and waiting for owner to release the page;
615 * Page leaves owned state in the following cases:
617 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
618 * the cache, doing nothing;
620 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
623 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
624 * transfer for this page;
626 * - [cl_page_state::CPS_FREEING] io decides to destroy this
627 * page (e.g., as part of truncate or extent lock cancellation).
629 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
633 * Page is being written out, as a part of a transfer. This state is
634 * entered when req-formation logic decided that it wants this page to
635 * be sent through the wire _now_. Specifically, it means that once
636 * this state is achieved, transfer completion handler (with either
637 * success or failure indication) is guaranteed to be executed against
638 * this page independently of any locks and any scheduling decisions
639 * made by the hosting environment (that effectively means that the
640 * page is never put into cl_page_state::CPS_PAGEOUT state "in
641 * advance". This property is mentioned, because it is important when
642 * reasoning about possible dead-locks in the system). The page can
643 * enter this state as a result of
645 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
646 * write-out of this page, or
648 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
649 * that it has enough dirty pages cached to issue a "good"
652 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
653 * is completed---it is moved into cl_page_state::CPS_CACHED state.
655 * Underlying VM page is locked for the duration of transfer.
657 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
661 * Page is being read in, as a part of a transfer. This is quite
662 * similar to the cl_page_state::CPS_PAGEOUT state, except that
663 * read-in is always "immediate"---there is no such thing a sudden
664 * construction of read cl_req from cached, presumably not up to date,
667 * Underlying VM page is locked for the duration of transfer.
669 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
673 * Page is being destroyed. This state is entered when client decides
674 * that page has to be deleted from its host object, as, e.g., a part
677 * Once this state is reached, there is no way to escape it.
679 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
686 /** Host page, the page is from the host inode which the cl_page
690 /** Transient page, the transient cl_page is used to bind a cl_page
691 * to vmpage which is not belonging to the same object of cl_page.
692 * it is used in DirectIO, lockless IO and liblustre. */
697 * Flags maintained for every cl_page.
701 * Set when pagein completes. Used for debugging (read completes at
702 * most once for a page).
704 CPF_READ_COMPLETED = 1 << 0
708 * Fields are protected by the lock on struct page, except for atomics and
711 * \invariant Data type invariants are in cl_page_invariant(). Basically:
712 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
713 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
714 * cl_page::cp_owner (when set).
717 /** Reference counter. */
719 /** An object this page is a part of. Immutable after creation. */
720 struct cl_object *cp_obj;
721 /** List of slices. Immutable after creation. */
722 cfs_list_t cp_layers;
723 struct page *cp_vmpage;
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;
729 /** Linkage of pages within group. Protected by cl_page::cp_mutex. */
731 /** Mutex serializing membership of a page in a batch. */
732 struct mutex cp_mutex;
733 /** Linkage of pages within cl_req. */
734 cfs_list_t cp_flight;
735 /** Transfer error. */
739 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
742 enum cl_page_type cp_type;
745 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
746 * by sub-io. Protected by a VM lock.
748 struct cl_io *cp_owner;
750 * Debug information, the task is owning the page.
752 struct task_struct *cp_task;
754 * Owning IO request in cl_page_state::CPS_PAGEOUT and
755 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
756 * the top-level pages. Protected by a VM lock.
758 struct cl_req *cp_req;
759 /** List of references to this page, for debugging. */
760 struct lu_ref cp_reference;
761 /** Link to an object, for debugging. */
762 struct lu_ref_link cp_obj_ref;
763 /** Link to a queue, for debugging. */
764 struct lu_ref_link cp_queue_ref;
765 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */
767 /** Assigned if doing a sync_io */
768 struct cl_sync_io *cp_sync_io;
772 * Per-layer part of cl_page.
774 * \see ccc_page, lov_page, osc_page
776 struct cl_page_slice {
777 struct cl_page *cpl_page;
780 * Object slice corresponding to this page slice. Immutable after
783 struct cl_object *cpl_obj;
784 const struct cl_page_operations *cpl_ops;
785 /** Linkage into cl_page::cp_layers. Immutable after creation. */
786 cfs_list_t cpl_linkage;
790 * Lock mode. For the client extent locks.
792 * \warning: cl_lock_mode_match() assumes particular ordering here.
797 * Mode of a lock that protects no data, and exists only as a
798 * placeholder. This is used for `glimpse' requests. A phantom lock
799 * might get promoted to real lock at some point.
808 * Requested transfer type.
818 * Per-layer page operations.
820 * Methods taking an \a io argument are for the activity happening in the
821 * context of given \a io. Page is assumed to be owned by that io, except for
822 * the obvious cases (like cl_page_operations::cpo_own()).
824 * \see vvp_page_ops, lov_page_ops, osc_page_ops
826 struct cl_page_operations {
828 * cl_page<->struct page methods. Only one layer in the stack has to
829 * implement these. Current code assumes that this functionality is
830 * provided by the topmost layer, see cl_page_disown0() as an example.
834 * Called when \a io acquires this page into the exclusive
835 * ownership. When this method returns, it is guaranteed that the is
836 * not owned by other io, and no transfer is going on against
840 * \see vvp_page_own(), lov_page_own()
842 int (*cpo_own)(const struct lu_env *env,
843 const struct cl_page_slice *slice,
844 struct cl_io *io, int nonblock);
845 /** Called when ownership it yielded. Optional.
847 * \see cl_page_disown()
848 * \see vvp_page_disown()
850 void (*cpo_disown)(const struct lu_env *env,
851 const struct cl_page_slice *slice, struct cl_io *io);
853 * Called for a page that is already "owned" by \a io from VM point of
856 * \see cl_page_assume()
857 * \see vvp_page_assume(), lov_page_assume()
859 void (*cpo_assume)(const struct lu_env *env,
860 const struct cl_page_slice *slice, struct cl_io *io);
861 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
862 * bottom-to-top when IO releases a page without actually unlocking
865 * \see cl_page_unassume()
866 * \see vvp_page_unassume()
868 void (*cpo_unassume)(const struct lu_env *env,
869 const struct cl_page_slice *slice,
872 * Announces whether the page contains valid data or not by \a uptodate.
874 * \see cl_page_export()
875 * \see vvp_page_export()
877 void (*cpo_export)(const struct lu_env *env,
878 const struct cl_page_slice *slice, int uptodate);
880 * Checks whether underlying VM page is locked (in the suitable
881 * sense). Used for assertions.
883 * \retval -EBUSY: page is protected by a lock of a given mode;
884 * \retval -ENODATA: page is not protected by a lock;
885 * \retval 0: this layer cannot decide. (Should never happen.)
887 int (*cpo_is_vmlocked)(const struct lu_env *env,
888 const struct cl_page_slice *slice);
894 * Called when page is truncated from the object. Optional.
896 * \see cl_page_discard()
897 * \see vvp_page_discard(), osc_page_discard()
899 void (*cpo_discard)(const struct lu_env *env,
900 const struct cl_page_slice *slice,
903 * Called when page is removed from the cache, and is about to being
904 * destroyed. Optional.
906 * \see cl_page_delete()
907 * \see vvp_page_delete(), osc_page_delete()
909 void (*cpo_delete)(const struct lu_env *env,
910 const struct cl_page_slice *slice);
911 /** Destructor. Frees resources and slice itself. */
912 void (*cpo_fini)(const struct lu_env *env,
913 struct cl_page_slice *slice);
916 * Checks whether the page is protected by a cl_lock. This is a
917 * per-layer method, because certain layers have ways to check for the
918 * lock much more efficiently than through the generic locks scan, or
919 * implement locking mechanisms separate from cl_lock, e.g.,
920 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
921 * being canceled, or scheduled for cancellation as soon as the last
922 * user goes away, too.
924 * \retval -EBUSY: page is protected by a lock of a given mode;
925 * \retval -ENODATA: page is not protected by a lock;
926 * \retval 0: this layer cannot decide.
928 * \see cl_page_is_under_lock()
930 int (*cpo_is_under_lock)(const struct lu_env *env,
931 const struct cl_page_slice *slice,
935 * Optional debugging helper. Prints given page slice.
937 * \see cl_page_print()
939 int (*cpo_print)(const struct lu_env *env,
940 const struct cl_page_slice *slice,
941 void *cookie, lu_printer_t p);
945 * Transfer methods. See comment on cl_req for a description of
946 * transfer formation and life-cycle.
951 * Request type dependent vector of operations.
953 * Transfer operations depend on transfer mode (cl_req_type). To avoid
954 * passing transfer mode to each and every of these methods, and to
955 * avoid branching on request type inside of the methods, separate
956 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
957 * provided. That is, method invocation usually looks like
959 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
963 * Called when a page is submitted for a transfer as a part of
966 * \return 0 : page is eligible for submission;
967 * \return -EALREADY : skip this page;
968 * \return -ve : error.
970 * \see cl_page_prep()
972 int (*cpo_prep)(const struct lu_env *env,
973 const struct cl_page_slice *slice,
976 * Completion handler. This is guaranteed to be eventually
977 * fired after cl_page_operations::cpo_prep() or
978 * cl_page_operations::cpo_make_ready() call.
980 * This method can be called in a non-blocking context. It is
981 * guaranteed however, that the page involved and its object
982 * are pinned in memory (and, hence, calling cl_page_put() is
985 * \see cl_page_completion()
987 void (*cpo_completion)(const struct lu_env *env,
988 const struct cl_page_slice *slice,
991 * Called when cached page is about to be added to the
992 * cl_req as a part of req formation.
994 * \return 0 : proceed with this page;
995 * \return -EAGAIN : skip this page;
996 * \return -ve : error.
998 * \see cl_page_make_ready()
1000 int (*cpo_make_ready)(const struct lu_env *env,
1001 const struct cl_page_slice *slice);
1004 * Tell transfer engine that only [to, from] part of a page should be
1007 * This is used for immediate transfers.
1009 * \todo XXX this is not very good interface. It would be much better
1010 * if all transfer parameters were supplied as arguments to
1011 * cl_io_operations::cio_submit() call, but it is not clear how to do
1012 * this for page queues.
1014 * \see cl_page_clip()
1016 void (*cpo_clip)(const struct lu_env *env,
1017 const struct cl_page_slice *slice,
1020 * \pre the page was queued for transferring.
1021 * \post page is removed from client's pending list, or -EBUSY
1022 * is returned if it has already been in transferring.
1024 * This is one of seldom page operation which is:
1025 * 0. called from top level;
1026 * 1. don't have vmpage locked;
1027 * 2. every layer should synchronize execution of its ->cpo_cancel()
1028 * with completion handlers. Osc uses client obd lock for this
1029 * purpose. Based on there is no vvp_page_cancel and
1030 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1032 * \see osc_page_cancel().
1034 int (*cpo_cancel)(const struct lu_env *env,
1035 const struct cl_page_slice *slice);
1037 * Write out a page by kernel. This is only called by ll_writepage
1040 * \see cl_page_flush()
1042 int (*cpo_flush)(const struct lu_env *env,
1043 const struct cl_page_slice *slice,
1049 * Helper macro, dumping detailed information about \a page into a log.
1051 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1053 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1054 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1055 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1056 CDEBUG(mask, format , ## __VA_ARGS__); \
1061 * Helper macro, dumping shorter information about \a page into a log.
1063 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1065 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1066 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1067 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1068 CDEBUG(mask, format , ## __VA_ARGS__); \
1072 static inline int __page_in_use(const struct cl_page *page, int refc)
1074 if (page->cp_type == CPT_CACHEABLE)
1076 LASSERT(cfs_atomic_read(&page->cp_ref) > 0);
1077 return (cfs_atomic_read(&page->cp_ref) > refc);
1079 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1080 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1082 static inline struct page *cl_page_vmpage(struct cl_page *page)
1084 LASSERT(page->cp_vmpage != NULL);
1085 return page->cp_vmpage;
1090 /** \addtogroup cl_lock cl_lock
1094 * Extent locking on the client.
1098 * The locking model of the new client code is built around
1102 * data-type representing an extent lock on a regular file. cl_lock is a
1103 * layered object (much like cl_object and cl_page), it consists of a header
1104 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1105 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1107 * All locks for a given object are linked into cl_object_header::coh_locks
1108 * list (protected by cl_object_header::coh_lock_guard spin-lock) through
1109 * cl_lock::cll_linkage. Currently this list is not sorted in any way. We can
1110 * sort it in starting lock offset, or use altogether different data structure
1113 * Typical cl_lock consists of the two layers:
1115 * - vvp_lock (vvp specific data), and
1116 * - lov_lock (lov specific data).
1118 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1119 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1121 * - lovsub_lock, and
1124 * Each sub-lock is associated with a cl_object (representing stripe
1125 * sub-object or the file to which top-level cl_lock is associated to), and is
1126 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1127 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1128 * is different from cl_page, that doesn't fan out (there is usually exactly
1129 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1130 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1134 * cl_lock is reference counted. When reference counter drops to 0, lock is
1135 * placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING
1136 * lock is destroyed when last reference is released. Referencing between
1137 * top-lock and its sub-locks is described in the lov documentation module.
1141 * Also, cl_lock is a state machine. This requires some clarification. One of
1142 * the goals of client IO re-write was to make IO path non-blocking, or at
1143 * least to make it easier to make it non-blocking in the future. Here
1144 * `non-blocking' means that when a system call (read, write, truncate)
1145 * reaches a situation where it has to wait for a communication with the
1146 * server, it should --instead of waiting-- remember its current state and
1147 * switch to some other work. E.g,. instead of waiting for a lock enqueue,
1148 * client should proceed doing IO on the next stripe, etc. Obviously this is
1149 * rather radical redesign, and it is not planned to be fully implemented at
1150 * this time, instead we are putting some infrastructure in place, that would
1151 * make it easier to do asynchronous non-blocking IO easier in the
1152 * future. Specifically, where old locking code goes to sleep (waiting for
1153 * enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When
1154 * enqueue reply comes, its completion handler signals that lock state-machine
1155 * is ready to transit to the next state. There is some generic code in
1156 * cl_lock.c that sleeps, waiting for these signals. As a result, for users of
1157 * this cl_lock.c code, it looks like locking is done in normal blocking
1158 * fashion, and it the same time it is possible to switch to the non-blocking
1159 * locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c
1162 * For a description of state machine states and transitions see enum
1165 * There are two ways to restrict a set of states which lock might move to:
1167 * - placing a "hold" on a lock guarantees that lock will not be moved
1168 * into cl_lock_state::CLS_FREEING state until hold is released. Hold
1169 * can be only acquired on a lock that is not in
1170 * cl_lock_state::CLS_FREEING. All holds on a lock are counted in
1171 * cl_lock::cll_holds. Hold protects lock from cancellation and
1172 * destruction. Requests to cancel and destroy a lock on hold will be
1173 * recorded, but only honored when last hold on a lock is released;
1175 * - placing a "user" on a lock guarantees that lock will not leave
1176 * cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING,
1177 * cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of
1178 * states, once it enters this set. That is, if a user is added onto a
1179 * lock in a state not from this set, it doesn't immediately enforce
1180 * lock to move to this set, but once lock enters this set it will
1181 * remain there until all users are removed. Lock users are counted in
1182 * cl_lock::cll_users.
1184 * User is used to assure that lock is not canceled or destroyed while
1185 * it is being enqueued, or actively used by some IO.
1187 * Currently, a user always comes with a hold (cl_lock_invariant()
1188 * checks that a number of holds is not less than a number of users).
1192 * This is how lock state-machine operates. struct cl_lock contains a mutex
1193 * cl_lock::cll_guard that protects struct fields.
1195 * - mutex is taken, and cl_lock::cll_state is examined.
1197 * - for every state there are possible target states where lock can move
1198 * into. They are tried in order. Attempts to move into next state are
1199 * done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try().
1201 * - if the transition can be performed immediately, state is changed,
1202 * and mutex is released.
1204 * - if the transition requires blocking, _try() function returns
1205 * cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to
1206 * sleep, waiting for possibility of lock state change. It is woken
1207 * up when some event occurs, that makes lock state change possible
1208 * (e.g., the reception of the reply from the server), and repeats
1211 * Top-lock and sub-lock has separate mutexes and the latter has to be taken
1212 * first to avoid dead-lock.
1214 * To see an example of interaction of all these issues, take a look at the
1215 * lov_cl.c:lov_lock_enqueue() function. It is called as a part of
1216 * cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by
1217 * advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note
1218 * also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It
1219 * also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be
1220 * done in parallel, rather than one after another (this is used for glimpse
1221 * locks, that cannot dead-lock).
1223 * INTERFACE AND USAGE
1225 * struct cl_lock_operations provide a number of call-backs that are invoked
1226 * when events of interest occurs. Layers can intercept and handle glimpse,
1227 * blocking, cancel ASTs and a reception of the reply from the server.
1229 * One important difference with the old client locking model is that new
1230 * client has a representation for the top-lock, whereas in the old code only
1231 * sub-locks existed as real data structures and file-level locks are
1232 * represented by "request sets" that are created and destroyed on each and
1233 * every lock creation.
1235 * Top-locks are cached, and can be found in the cache by the system calls. It
1236 * is possible that top-lock is in cache, but some of its sub-locks were
1237 * canceled and destroyed. In that case top-lock has to be enqueued again
1238 * before it can be used.
1240 * Overall process of the locking during IO operation is as following:
1242 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1243 * is called on each layer. Responsibility of this method is to add locks,
1244 * needed by a given layer into cl_io.ci_lockset.
1246 * - once locks for all layers were collected, they are sorted to avoid
1247 * dead-locks (cl_io_locks_sort()), and enqueued.
1249 * - when all locks are acquired, IO is performed;
1251 * - locks are released into cache.
1253 * Striping introduces major additional complexity into locking. The
1254 * fundamental problem is that it is generally unsafe to actively use (hold)
1255 * two locks on the different OST servers at the same time, as this introduces
1256 * inter-server dependency and can lead to cascading evictions.
1258 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1259 * that no multi-stripe locks are taken (note that this design abandons POSIX
1260 * read/write semantics). Such pieces ideally can be executed concurrently. At
1261 * the same time, certain types of IO cannot be sub-divived, without
1262 * sacrificing correctness. This includes:
1264 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1267 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1269 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1270 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1271 * has to be held together with the usual lock on [offset, offset + count].
1273 * As multi-stripe locks have to be allowed, it makes sense to cache them, so
1274 * that, for example, a sequence of O_APPEND writes can proceed quickly
1275 * without going down to the individual stripes to do lock matching. On the
1276 * other hand, multi-stripe locks shouldn't be used by normal read/write
1277 * calls. To achieve this, every layer can implement ->clo_fits_into() method,
1278 * that is called by lock matching code (cl_lock_lookup()), and that can be
1279 * used to selectively disable matching of certain locks for certain IOs. For
1280 * exmaple, lov layer implements lov_lock_fits_into() that allow multi-stripe
1281 * locks to be matched only for truncates and O_APPEND writes.
1283 * Interaction with DLM
1285 * In the expected setup, cl_lock is ultimately backed up by a collection of
1286 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1287 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1288 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1289 * description of interaction with DLM.
1295 struct cl_lock_descr {
1296 /** Object this lock is granted for. */
1297 struct cl_object *cld_obj;
1298 /** Index of the first page protected by this lock. */
1300 /** Index of the last page (inclusive) protected by this lock. */
1302 /** Group ID, for group lock */
1305 enum cl_lock_mode cld_mode;
1307 * flags to enqueue lock. A combination of bit-flags from
1308 * enum cl_enq_flags.
1310 __u32 cld_enq_flags;
1313 #define DDESCR "%s(%d):[%lu, %lu]"
1314 #define PDESCR(descr) \
1315 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1316 (descr)->cld_start, (descr)->cld_end
1318 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1321 * Lock state-machine states.
1326 * Possible state transitions:
1328 * +------------------>NEW
1330 * | | cl_enqueue_try()
1332 * | cl_unuse_try() V
1333 * | +--------------QUEUING (*)
1335 * | | | cl_enqueue_try()
1337 * | | cl_unuse_try() V
1338 * sub-lock | +-------------ENQUEUED (*)
1340 * | | | cl_wait_try()
1345 * | | HELD<---------+
1347 * | | | | cl_use_try()
1348 * | | cl_unuse_try() | |
1351 * | +------------>INTRANSIT (D) <--+
1353 * | cl_unuse_try() | | cached lock found
1354 * | | | cl_use_try()
1357 * +------------------CACHED---------+
1366 * In states marked with (*) transition to the same state (i.e., a loop
1367 * in the diagram) is possible.
1369 * (R) is the point where Receive call-back is invoked: it allows layers
1370 * to handle arrival of lock reply.
1372 * (C) is the point where Cancellation call-back is invoked.
1374 * (D) is the transit state which means the lock is changing.
1376 * Transition to FREEING state is possible from any other state in the
1377 * diagram in case of unrecoverable error.
1381 * These states are for individual cl_lock object. Top-lock and its sub-locks
1382 * can be in the different states. Another way to say this is that we have
1383 * nested state-machines.
1385 * Separate QUEUING and ENQUEUED states are needed to support non-blocking
1386 * operation for locks with multiple sub-locks. Imagine lock on a file F, that
1387 * intersects 3 stripes S0, S1, and S2. To enqueue F client has to send
1388 * enqueue to S0, wait for its completion, then send enqueue for S1, wait for
1389 * its completion and at last enqueue lock for S2, and wait for its
1390 * completion. In that case, top-lock is in QUEUING state while S0, S1 are
1391 * handled, and is in ENQUEUED state after enqueue to S2 has been sent (note
1392 * that in this case, sub-locks move from state to state, and top-lock remains
1393 * in the same state).
1395 enum cl_lock_state {
1397 * Lock that wasn't yet enqueued
1401 * Enqueue is in progress, blocking for some intermediate interaction
1402 * with the other side.
1406 * Lock is fully enqueued, waiting for server to reply when it is
1411 * Lock granted, actively used by some IO.
1415 * This state is used to mark the lock is being used, or unused.
1416 * We need this state because the lock may have several sublocks,
1417 * so it's impossible to have an atomic way to bring all sublocks
1418 * into CLS_HELD state at use case, or all sublocks to CLS_CACHED
1420 * If a thread is referring to a lock, and it sees the lock is in this
1421 * state, it must wait for the lock.
1422 * See state diagram for details.
1426 * Lock granted, not used.
1430 * Lock is being destroyed.
1436 enum cl_lock_flags {
1438 * lock has been cancelled. This flag is never cleared once set (by
1439 * cl_lock_cancel0()).
1441 CLF_CANCELLED = 1 << 0,
1442 /** cancellation is pending for this lock. */
1443 CLF_CANCELPEND = 1 << 1,
1444 /** destruction is pending for this lock. */
1445 CLF_DOOMED = 1 << 2,
1446 /** from enqueue RPC reply upcall. */
1447 CLF_FROM_UPCALL= 1 << 3,
1453 * Lock closure is a collection of locks (both top-locks and sub-locks) that
1454 * might be updated in a result of an operation on a certain lock (which lock
1455 * this is a closure of).
1457 * Closures are needed to guarantee dead-lock freedom in the presence of
1459 * - nested state-machines (top-lock state-machine composed of sub-lock
1460 * state-machines), and
1462 * - shared sub-locks.
1464 * Specifically, many operations, such as lock enqueue, wait, unlock,
1465 * etc. start from a top-lock, and then operate on a sub-locks of this
1466 * top-lock, holding a top-lock mutex. When sub-lock state changes as a result
1467 * of such operation, this change has to be propagated to all top-locks that
1468 * share this sub-lock. Obviously, no natural lock ordering (e.g.,
1469 * top-to-bottom or bottom-to-top) captures this scenario, so try-locking has
1470 * to be used. Lock closure systematizes this try-and-repeat logic.
1472 struct cl_lock_closure {
1474 * Lock that is mutexed when closure construction is started. When
1475 * closure in is `wait' mode (cl_lock_closure::clc_wait), mutex on
1476 * origin is released before waiting.
1478 struct cl_lock *clc_origin;
1480 * List of enclosed locks, so far. Locks are linked here through
1481 * cl_lock::cll_inclosure.
1483 cfs_list_t clc_list;
1485 * True iff closure is in a `wait' mode. This determines what
1486 * cl_lock_enclosure() does when a lock L to be added to the closure
1487 * is currently mutexed by some other thread.
1489 * If cl_lock_closure::clc_wait is not set, then closure construction
1490 * fails with CLO_REPEAT immediately.
1492 * In wait mode, cl_lock_enclosure() waits until next attempt to build
1493 * a closure might succeed. To this end it releases an origin mutex
1494 * (cl_lock_closure::clc_origin), that has to be the only lock mutex
1495 * owned by the current thread, and then waits on L mutex (by grabbing
1496 * it and immediately releasing), before returning CLO_REPEAT to the
1500 /** Number of locks in the closure. */
1505 * Layered client lock.
1508 /** Reference counter. */
1509 cfs_atomic_t cll_ref;
1510 /** List of slices. Immutable after creation. */
1511 cfs_list_t cll_layers;
1513 * Linkage into cl_lock::cll_descr::cld_obj::coh_locks list. Protected
1514 * by cl_lock::cll_descr::cld_obj::coh_lock_guard.
1516 cfs_list_t cll_linkage;
1518 * Parameters of this lock. Protected by
1519 * cl_lock::cll_descr::cld_obj::coh_lock_guard nested within
1520 * cl_lock::cll_guard. Modified only on lock creation and in
1523 struct cl_lock_descr cll_descr;
1524 /** Protected by cl_lock::cll_guard. */
1525 enum cl_lock_state cll_state;
1526 /** signals state changes. */
1527 wait_queue_head_t cll_wq;
1529 * Recursive lock, most fields in cl_lock{} are protected by this.
1531 * Locking rules: this mutex is never held across network
1532 * communication, except when lock is being canceled.
1534 * Lock ordering: a mutex of a sub-lock is taken first, then a mutex
1535 * on a top-lock. Other direction is implemented through a
1536 * try-lock-repeat loop. Mutices of unrelated locks can be taken only
1539 * \see osc_lock_enqueue_wait(), lov_lock_cancel(), lov_sublock_wait().
1541 struct mutex cll_guard;
1542 struct task_struct *cll_guarder;
1546 * the owner for INTRANSIT state
1548 struct task_struct *cll_intransit_owner;
1551 * Number of holds on a lock. A hold prevents a lock from being
1552 * canceled and destroyed. Protected by cl_lock::cll_guard.
1554 * \see cl_lock_hold(), cl_lock_unhold(), cl_lock_release()
1558 * Number of lock users. Valid in cl_lock_state::CLS_HELD state
1559 * only. Lock user pins lock in CLS_HELD state. Protected by
1560 * cl_lock::cll_guard.
1562 * \see cl_wait(), cl_unuse().
1566 * Flag bit-mask. Values from enum cl_lock_flags. Updates are
1567 * protected by cl_lock::cll_guard.
1569 unsigned long cll_flags;
1571 * A linkage into a list of locks in a closure.
1573 * \see cl_lock_closure
1575 cfs_list_t cll_inclosure;
1577 * Confict lock at queuing time.
1579 struct cl_lock *cll_conflict;
1581 * A list of references to this lock, for debugging.
1583 struct lu_ref cll_reference;
1585 * A list of holds on this lock, for debugging.
1587 struct lu_ref cll_holders;
1589 * A reference for cl_lock::cll_descr::cld_obj. For debugging.
1591 struct lu_ref_link cll_obj_ref;
1592 #ifdef CONFIG_LOCKDEP
1593 /* "dep_map" name is assumed by lockdep.h macros. */
1594 struct lockdep_map dep_map;
1599 * Per-layer part of cl_lock
1601 * \see ccc_lock, lov_lock, lovsub_lock, osc_lock
1603 struct cl_lock_slice {
1604 struct cl_lock *cls_lock;
1605 /** Object slice corresponding to this lock slice. Immutable after
1607 struct cl_object *cls_obj;
1608 const struct cl_lock_operations *cls_ops;
1609 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1610 cfs_list_t cls_linkage;
1614 * Possible (non-error) return values of ->clo_{enqueue,wait,unlock}().
1616 * NOTE: lov_subresult() depends on ordering here.
1618 enum cl_lock_transition {
1619 /** operation cannot be completed immediately. Wait for state change. */
1621 /** operation had to release lock mutex, restart. */
1623 /** lower layer re-enqueued. */
1629 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1631 struct cl_lock_operations {
1633 * \name statemachine
1635 * State machine transitions. These 3 methods are called to transfer
1636 * lock from one state to another, as described in the commentary
1637 * above enum #cl_lock_state.
1639 * \retval 0 this layer has nothing more to do to before
1640 * transition to the target state happens;
1642 * \retval CLO_REPEAT method had to release and re-acquire cl_lock
1643 * mutex, repeat invocation of transition method
1644 * across all layers;
1646 * \retval CLO_WAIT this layer cannot move to the target state
1647 * immediately, as it has to wait for certain event
1648 * (e.g., the communication with the server). It
1649 * is guaranteed, that when the state transfer
1650 * becomes possible, cl_lock::cll_wq wait-queue
1651 * is signaled. Caller can wait for this event by
1652 * calling cl_lock_state_wait();
1654 * \retval -ve failure, abort state transition, move the lock
1655 * into cl_lock_state::CLS_FREEING state, and set
1656 * cl_lock::cll_error.
1658 * Once all layers voted to agree to transition (by returning 0), lock
1659 * is moved into corresponding target state. All state transition
1660 * methods are optional.
1664 * Attempts to enqueue the lock. Called top-to-bottom.
1666 * \see ccc_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1667 * \see osc_lock_enqueue()
1669 int (*clo_enqueue)(const struct lu_env *env,
1670 const struct cl_lock_slice *slice,
1671 struct cl_io *io, __u32 enqflags);
1673 * Attempts to wait for enqueue result. Called top-to-bottom.
1675 * \see ccc_lock_wait(), lov_lock_wait(), osc_lock_wait()
1677 int (*clo_wait)(const struct lu_env *env,
1678 const struct cl_lock_slice *slice);
1680 * Attempts to unlock the lock. Called bottom-to-top. In addition to
1681 * usual return values of lock state-machine methods, this can return
1682 * -ESTALE to indicate that lock cannot be returned to the cache, and
1683 * has to be re-initialized.
1684 * unuse is a one-shot operation, so it must NOT return CLO_WAIT.
1686 * \see ccc_lock_unuse(), lov_lock_unuse(), osc_lock_unuse()
1688 int (*clo_unuse)(const struct lu_env *env,
1689 const struct cl_lock_slice *slice);
1691 * Notifies layer that cached lock is started being used.
1693 * \pre lock->cll_state == CLS_CACHED
1695 * \see lov_lock_use(), osc_lock_use()
1697 int (*clo_use)(const struct lu_env *env,
1698 const struct cl_lock_slice *slice);
1699 /** @} statemachine */
1701 * A method invoked when lock state is changed (as a result of state
1702 * transition). This is used, for example, to track when the state of
1703 * a sub-lock changes, to propagate this change to the corresponding
1704 * top-lock. Optional
1706 * \see lovsub_lock_state()
1708 void (*clo_state)(const struct lu_env *env,
1709 const struct cl_lock_slice *slice,
1710 enum cl_lock_state st);
1712 * Returns true, iff given lock is suitable for the given io, idea
1713 * being, that there are certain "unsafe" locks, e.g., ones acquired
1714 * for O_APPEND writes, that we don't want to re-use for a normal
1715 * write, to avoid the danger of cascading evictions. Optional. Runs
1716 * under cl_object_header::coh_lock_guard.
1718 * XXX this should take more information about lock needed by
1719 * io. Probably lock description or something similar.
1721 * \see lov_fits_into()
1723 int (*clo_fits_into)(const struct lu_env *env,
1724 const struct cl_lock_slice *slice,
1725 const struct cl_lock_descr *need,
1726 const struct cl_io *io);
1729 * Asynchronous System Traps. All of then are optional, all are
1730 * executed bottom-to-top.
1735 * Cancellation callback. Cancel a lock voluntarily, or under
1736 * the request of server.
1738 void (*clo_cancel)(const struct lu_env *env,
1739 const struct cl_lock_slice *slice);
1741 * Lock weighting ast. Executed to estimate how precious this lock
1742 * is. The sum of results across all layers is used to determine
1743 * whether lock worth keeping in cache given present memory usage.
1745 * \see osc_lock_weigh(), vvp_lock_weigh(), lovsub_lock_weigh().
1747 unsigned long (*clo_weigh)(const struct lu_env *env,
1748 const struct cl_lock_slice *slice);
1752 * \see lovsub_lock_closure()
1754 int (*clo_closure)(const struct lu_env *env,
1755 const struct cl_lock_slice *slice,
1756 struct cl_lock_closure *closure);
1758 * Executed bottom-to-top when lock description changes (e.g., as a
1759 * result of server granting more generous lock than was requested).
1761 * \see lovsub_lock_modify()
1763 int (*clo_modify)(const struct lu_env *env,
1764 const struct cl_lock_slice *slice,
1765 const struct cl_lock_descr *updated);
1767 * Notifies layers (bottom-to-top) that lock is going to be
1768 * destroyed. Responsibility of layers is to prevent new references on
1769 * this lock from being acquired once this method returns.
1771 * This can be called multiple times due to the races.
1773 * \see cl_lock_delete()
1774 * \see osc_lock_delete(), lovsub_lock_delete()
1776 void (*clo_delete)(const struct lu_env *env,
1777 const struct cl_lock_slice *slice);
1779 * Destructor. Frees resources and the slice.
1781 * \see ccc_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1782 * \see osc_lock_fini()
1784 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1786 * Optional debugging helper. Prints given lock slice.
1788 int (*clo_print)(const struct lu_env *env,
1789 void *cookie, lu_printer_t p,
1790 const struct cl_lock_slice *slice);
1793 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1795 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1796 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1797 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1798 CDEBUG(mask, format , ## __VA_ARGS__); \
1802 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1806 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1812 /** \addtogroup cl_page_list cl_page_list
1813 * Page list used to perform collective operations on a group of pages.
1815 * Pages are added to the list one by one. cl_page_list acquires a reference
1816 * for every page in it. Page list is used to perform collective operations on
1819 * - submit pages for an immediate transfer,
1821 * - own pages on behalf of certain io (waiting for each page in turn),
1825 * When list is finalized, it releases references on all pages it still has.
1827 * \todo XXX concurrency control.
1831 struct cl_page_list {
1833 cfs_list_t pl_pages;
1834 struct task_struct *pl_owner;
1838 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1839 * contains an incoming page list and an outgoing page list.
1842 struct cl_page_list c2_qin;
1843 struct cl_page_list c2_qout;
1846 /** @} cl_page_list */
1848 /** \addtogroup cl_io cl_io
1853 * cl_io represents a high level I/O activity like
1854 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1857 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1858 * important distinction. We want to minimize number of calls to the allocator
1859 * in the fast path, e.g., in the case of read(2) when everything is cached:
1860 * client already owns the lock over region being read, and data are cached
1861 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1862 * per-layer io state is stored in the session, associated with the io, see
1863 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1864 * by using free-lists, see cl_env_get().
1866 * There is a small predefined number of possible io types, enumerated in enum
1869 * cl_io is a state machine, that can be advanced concurrently by the multiple
1870 * threads. It is up to these threads to control the concurrency and,
1871 * specifically, to detect when io is done, and its state can be safely
1874 * For read/write io overall execution plan is as following:
1876 * (0) initialize io state through all layers;
1878 * (1) loop: prepare chunk of work to do
1880 * (2) call all layers to collect locks they need to process current chunk
1882 * (3) sort all locks to avoid dead-locks, and acquire them
1884 * (4) process the chunk: call per-page methods
1885 * (cl_io_operations::cio_read_page() for read,
1886 * cl_io_operations::cio_prepare_write(),
1887 * cl_io_operations::cio_commit_write() for write)
1893 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1894 * address allocation efficiency issues mentioned above), and returns with the
1895 * special error condition from per-page method when current sub-io has to
1896 * block. This causes io loop to be repeated, and lov switches to the next
1897 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1902 /** read system call */
1904 /** write system call */
1906 /** truncate, utime system calls */
1909 * page fault handling
1913 * fsync system call handling
1914 * To write out a range of file
1918 * Miscellaneous io. This is used for occasional io activity that
1919 * doesn't fit into other types. Currently this is used for:
1921 * - cancellation of an extent lock. This io exists as a context
1922 * to write dirty pages from under the lock being canceled back
1925 * - VM induced page write-out. An io context for writing page out
1926 * for memory cleansing;
1928 * - glimpse. An io context to acquire glimpse lock.
1930 * - grouplock. An io context to acquire group lock.
1932 * CIT_MISC io is used simply as a context in which locks and pages
1933 * are manipulated. Such io has no internal "process", that is,
1934 * cl_io_loop() is never called for it.
1941 * States of cl_io state machine
1944 /** Not initialized. */
1948 /** IO iteration started. */
1952 /** Actual IO is in progress. */
1954 /** IO for the current iteration finished. */
1956 /** Locks released. */
1958 /** Iteration completed. */
1960 /** cl_io finalized. */
1965 * IO state private for a layer.
1967 * This is usually embedded into layer session data, rather than allocated
1970 * \see vvp_io, lov_io, osc_io, ccc_io
1972 struct cl_io_slice {
1973 struct cl_io *cis_io;
1974 /** corresponding object slice. Immutable after creation. */
1975 struct cl_object *cis_obj;
1976 /** io operations. Immutable after creation. */
1977 const struct cl_io_operations *cis_iop;
1979 * linkage into a list of all slices for a given cl_io, hanging off
1980 * cl_io::ci_layers. Immutable after creation.
1982 cfs_list_t cis_linkage;
1985 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1989 * Per-layer io operations.
1990 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1992 struct cl_io_operations {
1994 * Vector of io state transition methods for every io type.
1996 * \see cl_page_operations::io
2000 * Prepare io iteration at a given layer.
2002 * Called top-to-bottom at the beginning of each iteration of
2003 * "io loop" (if it makes sense for this type of io). Here
2004 * layer selects what work it will do during this iteration.
2006 * \see cl_io_operations::cio_iter_fini()
2008 int (*cio_iter_init) (const struct lu_env *env,
2009 const struct cl_io_slice *slice);
2011 * Finalize io iteration.
2013 * Called bottom-to-top at the end of each iteration of "io
2014 * loop". Here layers can decide whether IO has to be
2017 * \see cl_io_operations::cio_iter_init()
2019 void (*cio_iter_fini) (const struct lu_env *env,
2020 const struct cl_io_slice *slice);
2022 * Collect locks for the current iteration of io.
2024 * Called top-to-bottom to collect all locks necessary for
2025 * this iteration. This methods shouldn't actually enqueue
2026 * anything, instead it should post a lock through
2027 * cl_io_lock_add(). Once all locks are collected, they are
2028 * sorted and enqueued in the proper order.
2030 int (*cio_lock) (const struct lu_env *env,
2031 const struct cl_io_slice *slice);
2033 * Finalize unlocking.
2035 * Called bottom-to-top to finish layer specific unlocking
2036 * functionality, after generic code released all locks
2037 * acquired by cl_io_operations::cio_lock().
2039 void (*cio_unlock)(const struct lu_env *env,
2040 const struct cl_io_slice *slice);
2042 * Start io iteration.
2044 * Once all locks are acquired, called top-to-bottom to
2045 * commence actual IO. In the current implementation,
2046 * top-level vvp_io_{read,write}_start() does all the work
2047 * synchronously by calling generic_file_*(), so other layers
2048 * are called when everything is done.
2050 int (*cio_start)(const struct lu_env *env,
2051 const struct cl_io_slice *slice);
2053 * Called top-to-bottom at the end of io loop. Here layer
2054 * might wait for an unfinished asynchronous io.
2056 void (*cio_end) (const struct lu_env *env,
2057 const struct cl_io_slice *slice);
2059 * Called bottom-to-top to notify layers that read/write IO
2060 * iteration finished, with \a nob bytes transferred.
2062 void (*cio_advance)(const struct lu_env *env,
2063 const struct cl_io_slice *slice,
2066 * Called once per io, bottom-to-top to release io resources.
2068 void (*cio_fini) (const struct lu_env *env,
2069 const struct cl_io_slice *slice);
2073 * Submit pages from \a queue->c2_qin for IO, and move
2074 * successfully submitted pages into \a queue->c2_qout. Return
2075 * non-zero if failed to submit even the single page. If
2076 * submission failed after some pages were moved into \a
2077 * queue->c2_qout, completion callback with non-zero ioret is
2080 int (*cio_submit)(const struct lu_env *env,
2081 const struct cl_io_slice *slice,
2082 enum cl_req_type crt,
2083 struct cl_2queue *queue);
2085 * Queue async page for write.
2086 * The difference between cio_submit and cio_queue is that
2087 * cio_submit is for urgent request.
2089 int (*cio_commit_async)(const struct lu_env *env,
2090 const struct cl_io_slice *slice,
2091 struct cl_page_list *queue, int from, int to,
2094 * Read missing page.
2096 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
2097 * method, when it hits not-up-to-date page in the range. Optional.
2099 * \pre io->ci_type == CIT_READ
2101 int (*cio_read_page)(const struct lu_env *env,
2102 const struct cl_io_slice *slice,
2103 const struct cl_page_slice *page);
2105 * Optional debugging helper. Print given io slice.
2107 int (*cio_print)(const struct lu_env *env, void *cookie,
2108 lu_printer_t p, const struct cl_io_slice *slice);
2112 * Flags to lock enqueue procedure.
2117 * instruct server to not block, if conflicting lock is found. Instead
2118 * -EWOULDBLOCK is returned immediately.
2120 CEF_NONBLOCK = 0x00000001,
2122 * take lock asynchronously (out of order), as it cannot
2123 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
2125 CEF_ASYNC = 0x00000002,
2127 * tell the server to instruct (though a flag in the blocking ast) an
2128 * owner of the conflicting lock, that it can drop dirty pages
2129 * protected by this lock, without sending them to the server.
2131 CEF_DISCARD_DATA = 0x00000004,
2133 * tell the sub layers that it must be a `real' lock. This is used for
2134 * mmapped-buffer locks and glimpse locks that must be never converted
2135 * into lockless mode.
2137 * \see vvp_mmap_locks(), cl_glimpse_lock().
2139 CEF_MUST = 0x00000008,
2141 * tell the sub layers that never request a `real' lock. This flag is
2142 * not used currently.
2144 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
2145 * conversion policy: ci_lockreq describes generic information of lock
2146 * requirement for this IO, especially for locks which belong to the
2147 * object doing IO; however, lock itself may have precise requirements
2148 * that are described by the enqueue flags.
2150 CEF_NEVER = 0x00000010,
2152 * for async glimpse lock.
2154 CEF_AGL = 0x00000020,
2156 * mask of enq_flags.
2158 CEF_MASK = 0x0000003f,
2162 * Link between lock and io. Intermediate structure is needed, because the
2163 * same lock can be part of multiple io's simultaneously.
2165 struct cl_io_lock_link {
2166 /** linkage into one of cl_lockset lists. */
2167 cfs_list_t cill_linkage;
2168 struct cl_lock_descr cill_descr;
2169 struct cl_lock *cill_lock;
2170 /** optional destructor */
2171 void (*cill_fini)(const struct lu_env *env,
2172 struct cl_io_lock_link *link);
2176 * Lock-set represents a collection of locks, that io needs at a
2177 * time. Generally speaking, client tries to avoid holding multiple locks when
2180 * - holding extent locks over multiple ost's introduces the danger of
2181 * "cascading timeouts";
2183 * - holding multiple locks over the same ost is still dead-lock prone,
2184 * see comment in osc_lock_enqueue(),
2186 * but there are certain situations where this is unavoidable:
2188 * - O_APPEND writes have to take [0, EOF] lock for correctness;
2190 * - truncate has to take [new-size, EOF] lock for correctness;
2192 * - SNS has to take locks across full stripe for correctness;
2194 * - in the case when user level buffer, supplied to {read,write}(file0),
2195 * is a part of a memory mapped lustre file, client has to take a dlm
2196 * locks on file0, and all files that back up the buffer (or a part of
2197 * the buffer, that is being processed in the current chunk, in any
2198 * case, there are situations where at least 2 locks are necessary).
2200 * In such cases we at least try to take locks in the same consistent
2201 * order. To this end, all locks are first collected, then sorted, and then
2205 /** locks to be acquired. */
2206 cfs_list_t cls_todo;
2207 /** locks currently being processed. */
2208 cfs_list_t cls_curr;
2209 /** locks acquired. */
2210 cfs_list_t cls_done;
2214 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
2215 * but 'req' is always to be thought as 'request' :-)
2217 enum cl_io_lock_dmd {
2218 /** Always lock data (e.g., O_APPEND). */
2220 /** Layers are free to decide between local and global locking. */
2222 /** Never lock: there is no cache (e.g., liblustre). */
2226 enum cl_fsync_mode {
2227 /** start writeback, do not wait for them to finish */
2229 /** start writeback and wait for them to finish */
2231 /** discard all of dirty pages in a specific file range */
2232 CL_FSYNC_DISCARD = 2,
2233 /** start writeback and make sure they have reached storage before
2234 * return. OST_SYNC RPC must be issued and finished */
2238 struct cl_io_rw_common {
2248 * cl_io is shared by all threads participating in this IO (in current
2249 * implementation only one thread advances IO, but parallel IO design and
2250 * concurrent copy_*_user() require multiple threads acting on the same IO. It
2251 * is up to these threads to serialize their activities, including updates to
2252 * mutable cl_io fields.
2255 /** type of this IO. Immutable after creation. */
2256 enum cl_io_type ci_type;
2257 /** current state of cl_io state machine. */
2258 enum cl_io_state ci_state;
2259 /** main object this io is against. Immutable after creation. */
2260 struct cl_object *ci_obj;
2262 * Upper layer io, of which this io is a part of. Immutable after
2265 struct cl_io *ci_parent;
2266 /** List of slices. Immutable after creation. */
2267 cfs_list_t ci_layers;
2268 /** list of locks (to be) acquired by this io. */
2269 struct cl_lockset ci_lockset;
2270 /** lock requirements, this is just a help info for sublayers. */
2271 enum cl_io_lock_dmd ci_lockreq;
2274 struct cl_io_rw_common rd;
2277 struct cl_io_rw_common wr;
2281 struct cl_io_rw_common ci_rw;
2282 struct cl_setattr_io {
2283 struct ost_lvb sa_attr;
2284 unsigned int sa_valid;
2285 struct obd_capa *sa_capa;
2287 struct cl_fault_io {
2288 /** page index within file. */
2290 /** bytes valid byte on a faulted page. */
2292 /** writable page? for nopage() only */
2294 /** page of an executable? */
2296 /** page_mkwrite() */
2298 /** resulting page */
2299 struct cl_page *ft_page;
2301 struct cl_fsync_io {
2304 struct obd_capa *fi_capa;
2305 /** file system level fid */
2306 struct lu_fid *fi_fid;
2307 enum cl_fsync_mode fi_mode;
2308 /* how many pages were written/discarded */
2309 unsigned int fi_nr_written;
2312 struct cl_2queue ci_queue;
2315 unsigned int ci_continue:1,
2317 * This io has held grouplock, to inform sublayers that
2318 * don't do lockless i/o.
2322 * The whole IO need to be restarted because layout has been changed
2326 * to not refresh layout - the IO issuer knows that the layout won't
2327 * change(page operations, layout change causes all page to be
2328 * discarded), or it doesn't matter if it changes(sync).
2332 * Check if layout changed after the IO finishes. Mainly for HSM
2333 * requirement. If IO occurs to openning files, it doesn't need to
2334 * verify layout because HSM won't release openning files.
2335 * Right now, only two opertaions need to verify layout: glimpse
2340 * file is released, restore has to to be triggered by vvp layer
2342 ci_restore_needed:1,
2348 * Number of pages owned by this IO. For invariant checking.
2350 unsigned ci_owned_nr;
2355 /** \addtogroup cl_req cl_req
2360 * There are two possible modes of transfer initiation on the client:
2362 * - immediate transfer: this is started when a high level io wants a page
2363 * or a collection of pages to be transferred right away. Examples:
2364 * read-ahead, synchronous read in the case of non-page aligned write,
2365 * page write-out as a part of extent lock cancellation, page write-out
2366 * as a part of memory cleansing. Immediate transfer can be both
2367 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
2369 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
2370 * when io wants to transfer a page to the server some time later, when
2371 * it can be done efficiently. Example: pages dirtied by the write(2)
2374 * In any case, transfer takes place in the form of a cl_req, which is a
2375 * representation for a network RPC.
2377 * Pages queued for an opportunistic transfer are cached until it is decided
2378 * that efficient RPC can be composed of them. This decision is made by "a
2379 * req-formation engine", currently implemented as a part of osc
2380 * layer. Req-formation depends on many factors: the size of the resulting
2381 * RPC, whether or not multi-object RPCs are supported by the server,
2382 * max-rpc-in-flight limitations, size of the dirty cache, etc.
2384 * For the immediate transfer io submits a cl_page_list, that req-formation
2385 * engine slices into cl_req's, possibly adding cached pages to some of
2386 * the resulting req's.
2388 * Whenever a page from cl_page_list is added to a newly constructed req, its
2389 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
2390 * page state is atomically changed from cl_page_state::CPS_OWNED to
2391 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
2392 * is zeroed, and cl_page::cp_req is set to the
2393 * req. cl_page_operations::cpo_prep() method at the particular layer might
2394 * return -EALREADY to indicate that it does not need to submit this page
2395 * at all. This is possible, for example, if page, submitted for read,
2396 * became up-to-date in the meantime; and for write, the page don't have
2397 * dirty bit marked. \see cl_io_submit_rw()
2399 * Whenever a cached page is added to a newly constructed req, its
2400 * cl_page_operations::cpo_make_ready() layer methods are called. At that
2401 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
2402 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
2403 * req. cl_page_operations::cpo_make_ready() method at the particular layer
2404 * might return -EAGAIN to indicate that this page is not eligible for the
2405 * transfer right now.
2409 * Plan is to divide transfers into "priority bands" (indicated when
2410 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
2411 * and allow glueing of cached pages to immediate transfers only within single
2412 * band. This would make high priority transfers (like lock cancellation or
2413 * memory pressure induced write-out) really high priority.
2418 * Per-transfer attributes.
2420 struct cl_req_attr {
2421 /** Generic attributes for the server consumption. */
2422 struct obdo *cra_oa;
2424 struct obd_capa *cra_capa;
2426 char cra_jobid[JOBSTATS_JOBID_SIZE];
2430 * Transfer request operations definable at every layer.
2432 * Concurrency: transfer formation engine synchronizes calls to all transfer
2435 struct cl_req_operations {
2437 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
2438 * complete (all pages are added).
2440 * \see osc_req_prep()
2442 int (*cro_prep)(const struct lu_env *env,
2443 const struct cl_req_slice *slice);
2445 * Called top-to-bottom to fill in \a oa fields. This is called twice
2446 * with different flags, see bug 10150 and osc_build_req().
2448 * \param obj an object from cl_req which attributes are to be set in
2451 * \param oa struct obdo where attributes are placed
2453 * \param flags \a oa fields to be filled.
2455 void (*cro_attr_set)(const struct lu_env *env,
2456 const struct cl_req_slice *slice,
2457 const struct cl_object *obj,
2458 struct cl_req_attr *attr, obd_valid flags);
2460 * Called top-to-bottom from cl_req_completion() to notify layers that
2461 * transfer completed. Has to free all state allocated by
2462 * cl_device_operations::cdo_req_init().
2464 void (*cro_completion)(const struct lu_env *env,
2465 const struct cl_req_slice *slice, int ioret);
2469 * A per-object state that (potentially multi-object) transfer request keeps.
2472 /** object itself */
2473 struct cl_object *ro_obj;
2474 /** reference to cl_req_obj::ro_obj. For debugging. */
2475 struct lu_ref_link ro_obj_ref;
2476 /* something else? Number of pages for a given object? */
2482 * Transfer requests are not reference counted, because IO sub-system owns
2483 * them exclusively and knows when to free them.
2487 * cl_req is created by cl_req_alloc() that calls
2488 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2489 * state in every layer.
2491 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2492 * contains pages for.
2494 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2495 * called top-to-bottom. At that point layers can modify req, let it pass, or
2496 * deny it completely. This is to support things like SNS that have transfer
2497 * ordering requirements invisible to the individual req-formation engine.
2499 * On transfer completion (or transfer timeout, or failure to initiate the
2500 * transfer of an allocated req), cl_req_operations::cro_completion() method
2501 * is called, after execution of cl_page_operations::cpo_completion() of all
2505 enum cl_req_type crq_type;
2506 /** A list of pages being transfered */
2507 cfs_list_t crq_pages;
2508 /** Number of pages in cl_req::crq_pages */
2509 unsigned crq_nrpages;
2510 /** An array of objects which pages are in ->crq_pages */
2511 struct cl_req_obj *crq_o;
2512 /** Number of elements in cl_req::crq_objs[] */
2513 unsigned crq_nrobjs;
2514 cfs_list_t crq_layers;
2518 * Per-layer state for request.
2520 struct cl_req_slice {
2521 struct cl_req *crs_req;
2522 struct cl_device *crs_dev;
2523 cfs_list_t crs_linkage;
2524 const struct cl_req_operations *crs_ops;
2529 enum cache_stats_item {
2530 /** how many cache lookups were performed */
2532 /** how many times cache lookup resulted in a hit */
2534 /** how many entities are in the cache right now */
2536 /** how many entities in the cache are actively used (and cannot be
2537 * evicted) right now */
2539 /** how many entities were created at all */
2544 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2547 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2549 struct cache_stats {
2550 const char *cs_name;
2551 cfs_atomic_t cs_stats[CS_NR];
2554 /** These are not exported so far */
2555 void cache_stats_init (struct cache_stats *cs, const char *name);
2556 int cache_stats_print(const struct cache_stats *cs,
2557 char *page, int count, int header);
2560 * Client-side site. This represents particular client stack. "Global"
2561 * variables should (directly or indirectly) be added here to allow multiple
2562 * clients to co-exist in the single address space.
2565 struct lu_site cs_lu;
2567 * Statistical counters. Atomics do not scale, something better like
2568 * per-cpu counters is needed.
2570 * These are exported as /proc/fs/lustre/llite/.../site
2572 * When interpreting keep in mind that both sub-locks (and sub-pages)
2573 * and top-locks (and top-pages) are accounted here.
2575 struct cache_stats cs_pages;
2576 struct cache_stats cs_locks;
2577 cfs_atomic_t cs_pages_state[CPS_NR];
2578 cfs_atomic_t cs_locks_state[CLS_NR];
2581 int cl_site_init (struct cl_site *s, struct cl_device *top);
2582 void cl_site_fini (struct cl_site *s);
2583 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2586 * Output client site statistical counters into a buffer. Suitable for
2587 * ll_rd_*()-style functions.
2589 int cl_site_stats_print(const struct cl_site *s, char *page, int count);
2594 * Type conversion and accessory functions.
2598 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2600 return container_of(site, struct cl_site, cs_lu);
2603 static inline int lu_device_is_cl(const struct lu_device *d)
2605 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2608 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2610 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2611 return container_of0(d, struct cl_device, cd_lu_dev);
2614 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2616 return &d->cd_lu_dev;
2619 static inline struct cl_object *lu2cl(const struct lu_object *o)
2621 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2622 return container_of0(o, struct cl_object, co_lu);
2625 static inline const struct cl_object_conf *
2626 lu2cl_conf(const struct lu_object_conf *conf)
2628 return container_of0(conf, struct cl_object_conf, coc_lu);
2631 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2633 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2636 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2638 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2639 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2642 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2644 return container_of0(h, struct cl_object_header, coh_lu);
2647 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2649 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2653 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2655 return luh2coh(obj->co_lu.lo_header);
2658 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2660 return lu_device_init(&d->cd_lu_dev, t);
2663 static inline void cl_device_fini(struct cl_device *d)
2665 lu_device_fini(&d->cd_lu_dev);
2668 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2669 struct cl_object *obj,
2670 const struct cl_page_operations *ops);
2671 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2672 struct cl_object *obj,
2673 const struct cl_lock_operations *ops);
2674 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2675 struct cl_object *obj, const struct cl_io_operations *ops);
2676 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2677 struct cl_device *dev,
2678 const struct cl_req_operations *ops);
2681 /** \defgroup cl_object cl_object
2683 struct cl_object *cl_object_top (struct cl_object *o);
2684 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2685 const struct lu_fid *fid,
2686 const struct cl_object_conf *c);
2688 int cl_object_header_init(struct cl_object_header *h);
2689 void cl_object_header_fini(struct cl_object_header *h);
2690 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2691 void cl_object_get (struct cl_object *o);
2692 void cl_object_attr_lock (struct cl_object *o);
2693 void cl_object_attr_unlock(struct cl_object *o);
2694 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2695 struct cl_attr *attr);
2696 int cl_object_attr_set (const struct lu_env *env, struct cl_object *obj,
2697 const struct cl_attr *attr, unsigned valid);
2698 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2699 struct ost_lvb *lvb);
2700 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2701 const struct cl_object_conf *conf);
2702 void cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2703 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2704 int cl_object_has_locks (struct cl_object *obj);
2707 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2709 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2711 return cl_object_header(o0) == cl_object_header(o1);
2714 static inline void cl_object_page_init(struct cl_object *clob, int size)
2716 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2717 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2720 static inline void *cl_object_page_slice(struct cl_object *clob,
2721 struct cl_page *page)
2723 return (void *)((char *)page + clob->co_slice_off);
2727 * Return refcount of cl_object.
2729 static inline int cl_object_refc(struct cl_object *clob)
2731 struct lu_object_header *header = clob->co_lu.lo_header;
2732 return cfs_atomic_read(&header->loh_ref);
2737 /** \defgroup cl_page cl_page
2745 /* callback of cl_page_gang_lookup() */
2747 struct cl_page *cl_page_find (const struct lu_env *env,
2748 struct cl_object *obj,
2749 pgoff_t idx, struct page *vmpage,
2750 enum cl_page_type type);
2751 struct cl_page *cl_page_alloc (const struct lu_env *env,
2752 struct cl_object *o, pgoff_t ind,
2753 struct page *vmpage,
2754 enum cl_page_type type);
2755 void cl_page_get (struct cl_page *page);
2756 void cl_page_put (const struct lu_env *env,
2757 struct cl_page *page);
2758 void cl_page_print (const struct lu_env *env, void *cookie,
2759 lu_printer_t printer,
2760 const struct cl_page *pg);
2761 void cl_page_header_print(const struct lu_env *env, void *cookie,
2762 lu_printer_t printer,
2763 const struct cl_page *pg);
2764 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2765 struct cl_page *cl_page_top (struct cl_page *page);
2767 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2768 const struct lu_device_type *dtype);
2773 * Functions dealing with the ownership of page by io.
2777 int cl_page_own (const struct lu_env *env,
2778 struct cl_io *io, struct cl_page *page);
2779 int cl_page_own_try (const struct lu_env *env,
2780 struct cl_io *io, struct cl_page *page);
2781 void cl_page_assume (const struct lu_env *env,
2782 struct cl_io *io, struct cl_page *page);
2783 void cl_page_unassume (const struct lu_env *env,
2784 struct cl_io *io, struct cl_page *pg);
2785 void cl_page_disown (const struct lu_env *env,
2786 struct cl_io *io, struct cl_page *page);
2787 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2794 * Functions dealing with the preparation of a page for a transfer, and
2795 * tracking transfer state.
2798 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2799 struct cl_page *pg, enum cl_req_type crt);
2800 void cl_page_completion (const struct lu_env *env,
2801 struct cl_page *pg, enum cl_req_type crt, int ioret);
2802 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2803 enum cl_req_type crt);
2804 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2805 struct cl_page *pg, enum cl_req_type crt);
2806 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2808 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2809 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2810 struct cl_page *pg);
2816 * \name helper routines
2817 * Functions to discard, delete and export a cl_page.
2820 void cl_page_discard (const struct lu_env *env, struct cl_io *io,
2821 struct cl_page *pg);
2822 void cl_page_delete (const struct lu_env *env, struct cl_page *pg);
2823 int cl_page_is_vmlocked (const struct lu_env *env,
2824 const struct cl_page *pg);
2825 void cl_page_export (const struct lu_env *env,
2826 struct cl_page *pg, int uptodate);
2827 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2828 struct cl_page *page);
2829 loff_t cl_offset (const struct cl_object *obj, pgoff_t idx);
2830 pgoff_t cl_index (const struct cl_object *obj, loff_t offset);
2831 int cl_page_size (const struct cl_object *obj);
2832 int cl_pages_prune (const struct lu_env *env, struct cl_object *obj);
2834 void cl_lock_print (const struct lu_env *env, void *cookie,
2835 lu_printer_t printer, const struct cl_lock *lock);
2836 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2837 lu_printer_t printer,
2838 const struct cl_lock_descr *descr);
2843 /** \defgroup cl_lock cl_lock
2846 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2847 const struct cl_lock_descr *need,
2848 const char *scope, const void *source);
2849 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
2850 const struct cl_lock_descr *need,
2851 const char *scope, const void *source);
2852 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2853 const struct cl_lock_descr *need,
2854 const char *scope, const void *source);
2855 struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env,
2856 struct cl_object *obj, pgoff_t index,
2857 struct cl_lock *except, int pending,
2859 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2860 const struct lu_device_type *dtype);
2862 void cl_lock_get (struct cl_lock *lock);
2863 void cl_lock_get_trust (struct cl_lock *lock);
2864 void cl_lock_put (const struct lu_env *env, struct cl_lock *lock);
2865 void cl_lock_hold_add (const struct lu_env *env, struct cl_lock *lock,
2866 const char *scope, const void *source);
2867 void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
2868 const char *scope, const void *source);
2869 void cl_lock_unhold (const struct lu_env *env, struct cl_lock *lock,
2870 const char *scope, const void *source);
2871 void cl_lock_release (const struct lu_env *env, struct cl_lock *lock,
2872 const char *scope, const void *source);
2873 void cl_lock_user_add (const struct lu_env *env, struct cl_lock *lock);
2874 void cl_lock_user_del (const struct lu_env *env, struct cl_lock *lock);
2876 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
2877 struct cl_lock *lock);
2878 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
2879 enum cl_lock_state state);
2880 int cl_lock_is_intransit(struct cl_lock *lock);
2882 int cl_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock,
2885 /** \name statemachine statemachine
2886 * Interface to lock state machine consists of 3 parts:
2888 * - "try" functions that attempt to effect a state transition. If state
2889 * transition is not possible right now (e.g., if it has to wait for some
2890 * asynchronous event to occur), these functions return
2891 * cl_lock_transition::CLO_WAIT.
2893 * - "non-try" functions that implement synchronous blocking interface on
2894 * top of non-blocking "try" functions. These functions repeatedly call
2895 * corresponding "try" versions, and if state transition is not possible
2896 * immediately, wait for lock state change.
2898 * - methods from cl_lock_operations, called by "try" functions. Lock can
2899 * be advanced to the target state only when all layers voted that they
2900 * are ready for this transition. "Try" functions call methods under lock
2901 * mutex. If a layer had to release a mutex, it re-acquires it and returns
2902 * cl_lock_transition::CLO_REPEAT, causing "try" function to call all
2905 * TRY NON-TRY METHOD FINAL STATE
2907 * cl_enqueue_try() cl_enqueue() cl_lock_operations::clo_enqueue() CLS_ENQUEUED
2909 * cl_wait_try() cl_wait() cl_lock_operations::clo_wait() CLS_HELD
2911 * cl_unuse_try() cl_unuse() cl_lock_operations::clo_unuse() CLS_CACHED
2913 * cl_use_try() NONE cl_lock_operations::clo_use() CLS_HELD
2917 int cl_enqueue (const struct lu_env *env, struct cl_lock *lock,
2918 struct cl_io *io, __u32 flags);
2919 int cl_wait (const struct lu_env *env, struct cl_lock *lock);
2920 void cl_unuse (const struct lu_env *env, struct cl_lock *lock);
2921 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
2922 struct cl_io *io, __u32 flags);
2923 int cl_unuse_try (const struct lu_env *env, struct cl_lock *lock);
2924 int cl_wait_try (const struct lu_env *env, struct cl_lock *lock);
2925 int cl_use_try (const struct lu_env *env, struct cl_lock *lock, int atomic);
2927 /** @} statemachine */
2929 void cl_lock_signal (const struct lu_env *env, struct cl_lock *lock);
2930 int cl_lock_state_wait (const struct lu_env *env, struct cl_lock *lock);
2931 void cl_lock_state_set (const struct lu_env *env, struct cl_lock *lock,
2932 enum cl_lock_state state);
2933 int cl_queue_match (const cfs_list_t *queue,
2934 const struct cl_lock_descr *need);
2936 void cl_lock_mutex_get (const struct lu_env *env, struct cl_lock *lock);
2937 int cl_lock_mutex_try (const struct lu_env *env, struct cl_lock *lock);
2938 void cl_lock_mutex_put (const struct lu_env *env, struct cl_lock *lock);
2939 int cl_lock_is_mutexed (struct cl_lock *lock);
2940 int cl_lock_nr_mutexed (const struct lu_env *env);
2941 int cl_lock_discard_pages(const struct lu_env *env, struct cl_lock *lock);
2942 int cl_lock_ext_match (const struct cl_lock_descr *has,
2943 const struct cl_lock_descr *need);
2944 int cl_lock_descr_match(const struct cl_lock_descr *has,
2945 const struct cl_lock_descr *need);
2946 int cl_lock_mode_match (enum cl_lock_mode has, enum cl_lock_mode need);
2947 int cl_lock_modify (const struct lu_env *env, struct cl_lock *lock,
2948 const struct cl_lock_descr *desc);
2950 void cl_lock_closure_init (const struct lu_env *env,
2951 struct cl_lock_closure *closure,
2952 struct cl_lock *origin, int wait);
2953 void cl_lock_closure_fini (struct cl_lock_closure *closure);
2954 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
2955 struct cl_lock_closure *closure);
2956 void cl_lock_disclosure (const struct lu_env *env,
2957 struct cl_lock_closure *closure);
2958 int cl_lock_enclosure (const struct lu_env *env, struct cl_lock *lock,
2959 struct cl_lock_closure *closure);
2961 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2962 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock);
2963 void cl_lock_error (const struct lu_env *env, struct cl_lock *lock, int error);
2964 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int wait);
2966 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock);
2970 /** \defgroup cl_io cl_io
2973 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2974 enum cl_io_type iot, struct cl_object *obj);
2975 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2976 enum cl_io_type iot, struct cl_object *obj);
2977 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2978 enum cl_io_type iot, loff_t pos, size_t count);
2979 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2981 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2982 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2983 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2984 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2985 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2986 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2987 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2988 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2989 struct cl_io_lock_link *link);
2990 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2991 struct cl_lock_descr *descr);
2992 int cl_io_read_page (const struct lu_env *env, struct cl_io *io,
2993 struct cl_page *page);
2994 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2995 enum cl_req_type iot, struct cl_2queue *queue);
2996 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2997 enum cl_req_type iot, struct cl_2queue *queue,
2999 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
3000 struct cl_page_list *queue, int from, int to,
3002 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
3004 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
3005 struct cl_page_list *queue);
3006 int cl_io_is_going (const struct lu_env *env);
3009 * True, iff \a io is an O_APPEND write(2).
3011 static inline int cl_io_is_append(const struct cl_io *io)
3013 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
3016 static inline int cl_io_is_sync_write(const struct cl_io *io)
3018 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
3021 static inline int cl_io_is_mkwrite(const struct cl_io *io)
3023 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
3027 * True, iff \a io is a truncate(2).
3029 static inline int cl_io_is_trunc(const struct cl_io *io)
3031 return io->ci_type == CIT_SETATTR &&
3032 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
3035 struct cl_io *cl_io_top(struct cl_io *io);
3037 void cl_io_print(const struct lu_env *env, void *cookie,
3038 lu_printer_t printer, const struct cl_io *io);
3040 #define CL_IO_SLICE_CLEAN(foo_io, base) \
3042 typeof(foo_io) __foo_io = (foo_io); \
3044 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
3045 memset(&__foo_io->base + 1, 0, \
3046 (sizeof *__foo_io) - sizeof __foo_io->base); \
3051 /** \defgroup cl_page_list cl_page_list
3055 * Last page in the page list.
3057 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
3059 LASSERT(plist->pl_nr > 0);
3060 return cfs_list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
3063 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
3065 LASSERT(plist->pl_nr > 0);
3066 return cfs_list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
3070 * Iterate over pages in a page list.
3072 #define cl_page_list_for_each(page, list) \
3073 cfs_list_for_each_entry((page), &(list)->pl_pages, cp_batch)
3076 * Iterate over pages in a page list, taking possible removals into account.
3078 #define cl_page_list_for_each_safe(page, temp, list) \
3079 cfs_list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
3081 void cl_page_list_init (struct cl_page_list *plist);
3082 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
3083 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
3084 struct cl_page *page);
3085 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
3086 struct cl_page *page);
3087 void cl_page_list_splice (struct cl_page_list *list,
3088 struct cl_page_list *head);
3089 void cl_page_list_del (const struct lu_env *env,
3090 struct cl_page_list *plist, struct cl_page *page);
3091 void cl_page_list_disown (const struct lu_env *env,
3092 struct cl_io *io, struct cl_page_list *plist);
3093 int cl_page_list_own (const struct lu_env *env,
3094 struct cl_io *io, struct cl_page_list *plist);
3095 void cl_page_list_assume (const struct lu_env *env,
3096 struct cl_io *io, struct cl_page_list *plist);
3097 void cl_page_list_discard(const struct lu_env *env,
3098 struct cl_io *io, struct cl_page_list *plist);
3099 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
3101 void cl_2queue_init (struct cl_2queue *queue);
3102 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
3103 void cl_2queue_disown (const struct lu_env *env,
3104 struct cl_io *io, struct cl_2queue *queue);
3105 void cl_2queue_assume (const struct lu_env *env,
3106 struct cl_io *io, struct cl_2queue *queue);
3107 void cl_2queue_discard (const struct lu_env *env,
3108 struct cl_io *io, struct cl_2queue *queue);
3109 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
3110 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
3112 /** @} cl_page_list */
3114 /** \defgroup cl_req cl_req
3116 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
3117 enum cl_req_type crt, int nr_objects);
3119 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
3120 struct cl_page *page);
3121 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
3122 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
3123 void cl_req_attr_set (const struct lu_env *env, struct cl_req *req,
3124 struct cl_req_attr *attr, obd_valid flags);
3125 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
3127 /** \defgroup cl_sync_io cl_sync_io
3131 * Anchor for synchronous transfer. This is allocated on a stack by thread
3132 * doing synchronous transfer, and a pointer to this structure is set up in
3133 * every page submitted for transfer. Transfer completion routine updates
3134 * anchor and wakes up waiting thread when transfer is complete.
3137 /** number of pages yet to be transferred. */
3138 cfs_atomic_t csi_sync_nr;
3141 /** barrier of destroy this structure */
3142 cfs_atomic_t csi_barrier;
3143 /** completion to be signaled when transfer is complete. */
3144 wait_queue_head_t csi_waitq;
3147 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages);
3148 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
3149 struct cl_page_list *queue, struct cl_sync_io *anchor,
3151 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret);
3153 /** @} cl_sync_io */
3157 /** \defgroup cl_env cl_env
3159 * lu_env handling for a client.
3161 * lu_env is an environment within which lustre code executes. Its major part
3162 * is lu_context---a fast memory allocation mechanism that is used to conserve
3163 * precious kernel stack space. Originally lu_env was designed for a server,
3166 * - there is a (mostly) fixed number of threads, and
3168 * - call chains have no non-lustre portions inserted between lustre code.
3170 * On a client both these assumtpion fails, because every user thread can
3171 * potentially execute lustre code as part of a system call, and lustre calls
3172 * into VFS or MM that call back into lustre.
3174 * To deal with that, cl_env wrapper functions implement the following
3177 * - allocation and destruction of environment is amortized by caching no
3178 * longer used environments instead of destroying them;
3180 * - there is a notion of "current" environment, attached to the kernel
3181 * data structure representing current thread Top-level lustre code
3182 * allocates an environment and makes it current, then calls into
3183 * non-lustre code, that in turn calls lustre back. Low-level lustre
3184 * code thus called can fetch environment created by the top-level code
3185 * and reuse it, avoiding additional environment allocation.
3186 * Right now, three interfaces can attach the cl_env to running thread:
3189 * - cl_env_reexit(cl_env_reenter had to be called priorly)
3191 * \see lu_env, lu_context, lu_context_key
3194 struct cl_env_nest {
3199 struct lu_env *cl_env_peek (int *refcheck);
3200 struct lu_env *cl_env_get (int *refcheck);
3201 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
3202 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
3203 void cl_env_put (struct lu_env *env, int *refcheck);
3204 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
3205 void *cl_env_reenter (void);
3206 void cl_env_reexit (void *cookie);
3207 void cl_env_implant (struct lu_env *env, int *refcheck);
3208 void cl_env_unplant (struct lu_env *env, int *refcheck);
3209 unsigned cl_env_cache_purge(unsigned nr);
3210 struct lu_env *cl_env_percpu_get (void);
3211 void cl_env_percpu_put (struct lu_env *env);
3218 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
3219 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
3221 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
3222 struct lu_device_type *ldt,
3223 struct lu_device *next);
3226 int cl_global_init(void);
3227 void cl_global_fini(void);
3229 #endif /* _LINUX_CL_OBJECT_H */