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>
105 # include <liblustre.h>
111 struct cl_device_operations;
114 struct cl_object_page_operations;
115 struct cl_object_lock_operations;
118 struct cl_page_slice;
120 struct cl_lock_slice;
122 struct cl_lock_operations;
123 struct cl_page_operations;
132 * Operations for each data device in the client stack.
134 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
136 struct cl_device_operations {
138 * Initialize cl_req. This method is called top-to-bottom on all
139 * devices in the stack to get them a chance to allocate layer-private
140 * data, and to attach them to the cl_req by calling
141 * cl_req_slice_add().
143 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
144 * \see ccc_req_init()
146 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
151 * Device in the client stack.
153 * \see ccc_device, lov_device, lovsub_device, osc_device
157 struct lu_device cd_lu_dev;
158 /** Per-layer operation vector. */
159 const struct cl_device_operations *cd_ops;
162 /** \addtogroup cl_object cl_object
165 * "Data attributes" of cl_object. Data attributes can be updated
166 * independently for a sub-object, and top-object's attributes are calculated
167 * from sub-objects' ones.
170 /** Object size, in bytes */
173 * Known minimal size, in bytes.
175 * This is only valid when at least one DLM lock is held.
178 /** Modification time. Measured in seconds since epoch. */
180 /** Access time. Measured in seconds since epoch. */
182 /** Change time. Measured in seconds since epoch. */
185 * Blocks allocated to this cl_object on the server file system.
187 * \todo XXX An interface for block size is needed.
191 * User identifier for quota purposes.
195 * Group identifier for quota purposes.
199 /* nlink of the directory */
204 * Fields in cl_attr that are being set.
218 * Sub-class of lu_object with methods common for objects on the client
221 * cl_object: represents a regular file system object, both a file and a
222 * stripe. cl_object is based on lu_object: it is identified by a fid,
223 * layered, cached, hashed, and lrued. Important distinction with the server
224 * side, where md_object and dt_object are used, is that cl_object "fans out"
225 * at the lov/sns level: depending on the file layout, single file is
226 * represented as a set of "sub-objects" (stripes). At the implementation
227 * level, struct lov_object contains an array of cl_objects. Each sub-object
228 * is a full-fledged cl_object, having its fid, living in the lru and hash
231 * This leads to the next important difference with the server side: on the
232 * client, it's quite usual to have objects with the different sequence of
233 * layers. For example, typical top-object is composed of the following
239 * whereas its sub-objects are composed of
244 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
245 * track of the object-subobject relationship.
247 * Sub-objects are not cached independently: when top-object is about to
248 * be discarded from the memory, all its sub-objects are torn-down and
251 * \see ccc_object, lov_object, lovsub_object, osc_object
255 struct lu_object co_lu;
256 /** per-object-layer operations */
257 const struct cl_object_operations *co_ops;
258 /** offset of page slice in cl_page buffer */
263 * Description of the client object configuration. This is used for the
264 * creation of a new client object that is identified by a more state than
267 struct cl_object_conf {
269 struct lu_object_conf coc_lu;
272 * Object layout. This is consumed by lov.
274 struct lustre_md *coc_md;
276 * Description of particular stripe location in the
277 * cluster. This is consumed by osc.
279 struct lov_oinfo *coc_oinfo;
282 * VFS inode. This is consumed by vvp.
284 struct inode *coc_inode;
286 * Layout lock handle.
288 struct ldlm_lock *coc_lock;
290 * Operation to handle layout, OBJECT_CONF_XYZ.
296 /** configure layout, set up a new stripe, must be called while
297 * holding layout lock. */
299 /** invalidate the current stripe configuration due to losing
301 OBJECT_CONF_INVALIDATE = 1,
302 /** wait for old layout to go away so that new layout can be
308 * Operations implemented for each cl object layer.
310 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
312 struct cl_object_operations {
314 * Initialize page slice for this layer. Called top-to-bottom through
315 * every object layer when a new cl_page is instantiated. Layer
316 * keeping private per-page data, or requiring its own page operations
317 * vector should allocate these data here, and attach then to the page
318 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
321 * \retval NULL success.
323 * \retval ERR_PTR(errno) failure code.
325 * \retval valid-pointer pointer to already existing referenced page
326 * to be used instead of newly created.
328 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
329 struct cl_page *page, pgoff_t index);
331 * Initialize lock slice for this layer. Called top-to-bottom through
332 * every object layer when a new cl_lock is instantiated. Layer
333 * keeping private per-lock data, or requiring its own lock operations
334 * vector should allocate these data here, and attach then to the lock
335 * by calling cl_lock_slice_add(). Mandatory.
337 int (*coo_lock_init)(const struct lu_env *env,
338 struct cl_object *obj, struct cl_lock *lock,
339 const struct cl_io *io);
341 * Initialize io state for a given layer.
343 * called top-to-bottom once per io existence to initialize io
344 * state. If layer wants to keep some state for this type of io, it
345 * has to embed struct cl_io_slice in lu_env::le_ses, and register
346 * slice with cl_io_slice_add(). It is guaranteed that all threads
347 * participating in this io share the same session.
349 int (*coo_io_init)(const struct lu_env *env,
350 struct cl_object *obj, struct cl_io *io);
352 * Fill portion of \a attr that this layer controls. This method is
353 * called top-to-bottom through all object layers.
355 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
357 * \return 0: to continue
358 * \return +ve: to stop iterating through layers (but 0 is returned
359 * from enclosing cl_object_attr_get())
360 * \return -ve: to signal error
362 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
363 struct cl_attr *attr);
367 * \a valid is a bitmask composed from enum #cl_attr_valid, and
368 * indicating what attributes are to be set.
370 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
372 * \return the same convention as for
373 * cl_object_operations::coo_attr_get() is used.
375 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj,
376 const struct cl_attr *attr, unsigned valid);
378 * Update object configuration. Called top-to-bottom to modify object
381 * XXX error conditions and handling.
383 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
384 const struct cl_object_conf *conf);
386 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
387 * object. Layers are supposed to fill parts of \a lvb that will be
388 * shipped to the glimpse originator as a glimpse result.
390 * \see ccc_object_glimpse(), lovsub_object_glimpse(),
391 * \see osc_object_glimpse()
393 int (*coo_glimpse)(const struct lu_env *env,
394 const struct cl_object *obj, struct ost_lvb *lvb);
396 * Object prune method. Called when the layout is going to change on
397 * this object, therefore each layer has to clean up their cache,
398 * mainly pages and locks.
400 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
404 * Extended header for client object.
406 struct cl_object_header {
407 /** Standard lu_object_header. cl_object::co_lu::lo_header points
409 struct lu_object_header coh_lu;
411 * \todo XXX move locks below to the separate cache-lines, they are
412 * mostly useless otherwise.
415 /** Lock protecting lock list. */
416 spinlock_t coh_lock_guard;
418 /** List of cl_lock's granted for this object. */
419 cfs_list_t coh_locks;
422 * Parent object. It is assumed that an object has a well-defined
423 * parent, but not a well-defined child (there may be multiple
424 * sub-objects, for the same top-object). cl_object_header::coh_parent
425 * field allows certain code to be written generically, without
426 * limiting possible cl_object layouts unduly.
428 struct cl_object_header *coh_parent;
430 * Protects consistency between cl_attr of parent object and
431 * attributes of sub-objects, that the former is calculated ("merged")
434 * \todo XXX this can be read/write lock if needed.
436 spinlock_t coh_attr_guard;
438 * Size of cl_page + page slices
440 unsigned short coh_page_bufsize;
442 * Number of objects above this one: 0 for a top-object, 1 for its
445 unsigned char coh_nesting;
449 * Helper macro: iterate over all layers of the object \a obj, assigning every
450 * layer top-to-bottom to \a slice.
452 #define cl_object_for_each(slice, obj) \
453 cfs_list_for_each_entry((slice), \
454 &(obj)->co_lu.lo_header->loh_layers, \
457 * Helper macro: iterate over all layers of the object \a obj, assigning every
458 * layer bottom-to-top to \a slice.
460 #define cl_object_for_each_reverse(slice, obj) \
461 cfs_list_for_each_entry_reverse((slice), \
462 &(obj)->co_lu.lo_header->loh_layers, \
466 #define CL_PAGE_EOF ((pgoff_t)~0ull)
468 /** \addtogroup cl_page cl_page
472 * Layered client page.
474 * cl_page: represents a portion of a file, cached in the memory. All pages
475 * of the given file are of the same size, and are kept in the radix tree
476 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
477 * of the top-level file object are first class cl_objects, they have their
478 * own radix trees of pages and hence page is implemented as a sequence of
479 * struct cl_pages's, linked into double-linked list through
480 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
481 * corresponding radix tree at the corresponding logical offset.
483 * cl_page is associated with VM page of the hosting environment (struct
484 * page in Linux kernel, for example), struct page. It is assumed, that this
485 * association is implemented by one of cl_page layers (top layer in the
486 * current design) that
488 * - intercepts per-VM-page call-backs made by the environment (e.g.,
491 * - translates state (page flag bits) and locking between lustre and
494 * The association between cl_page and struct page is immutable and
495 * established when cl_page is created.
497 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
498 * this io an exclusive access to this page w.r.t. other io attempts and
499 * various events changing page state (such as transfer completion, or
500 * eviction of the page from the memory). Note, that in general cl_io
501 * cannot be identified with a particular thread, and page ownership is not
502 * exactly equal to the current thread holding a lock on the page. Layer
503 * implementing association between cl_page and struct page has to implement
504 * ownership on top of available synchronization mechanisms.
506 * While lustre client maintains the notion of an page ownership by io,
507 * hosting MM/VM usually has its own page concurrency control
508 * mechanisms. For example, in Linux, page access is synchronized by the
509 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
510 * takes care to acquire and release such locks as necessary around the
511 * calls to the file system methods (->readpage(), ->prepare_write(),
512 * ->commit_write(), etc.). This leads to the situation when there are two
513 * different ways to own a page in the client:
515 * - client code explicitly and voluntary owns the page (cl_page_own());
517 * - VM locks a page and then calls the client, that has "to assume"
518 * the ownership from the VM (cl_page_assume()).
520 * Dual methods to release ownership are cl_page_disown() and
521 * cl_page_unassume().
523 * cl_page is reference counted (cl_page::cp_ref). When reference counter
524 * drops to 0, the page is returned to the cache, unless it is in
525 * cl_page_state::CPS_FREEING state, in which case it is immediately
528 * The general logic guaranteeing the absence of "existential races" for
529 * pages is the following:
531 * - there are fixed known ways for a thread to obtain a new reference
534 * - by doing a lookup in the cl_object radix tree, protected by the
537 * - by starting from VM-locked struct page and following some
538 * hosting environment method (e.g., following ->private pointer in
539 * the case of Linux kernel), see cl_vmpage_page();
541 * - when the page enters cl_page_state::CPS_FREEING state, all these
542 * ways are severed with the proper synchronization
543 * (cl_page_delete());
545 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
548 * - no new references to the page in cl_page_state::CPS_FREEING state
549 * are allowed (checked in cl_page_get()).
551 * Together this guarantees that when last reference to a
552 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
553 * page, as neither references to it can be acquired at that point, nor
556 * cl_page is a state machine. States are enumerated in enum
557 * cl_page_state. Possible state transitions are enumerated in
558 * cl_page_state_set(). State transition process (i.e., actual changing of
559 * cl_page::cp_state field) is protected by the lock on the underlying VM
562 * Linux Kernel implementation.
564 * Binding between cl_page and struct page (which is a typedef for
565 * struct page) is implemented in the vvp layer. cl_page is attached to the
566 * ->private pointer of the struct page, together with the setting of
567 * PG_private bit in page->flags, and acquiring additional reference on the
568 * struct page (much like struct buffer_head, or any similar file system
569 * private data structures).
571 * PG_locked lock is used to implement both ownership and transfer
572 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
573 * states. No additional references are acquired for the duration of the
576 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
577 * write-out is "protected" by the special PG_writeback bit.
581 * States of cl_page. cl_page.c assumes particular order here.
583 * The page state machine is rather crude, as it doesn't recognize finer page
584 * states like "dirty" or "up to date". This is because such states are not
585 * always well defined for the whole stack (see, for example, the
586 * implementation of the read-ahead, that hides page up-to-dateness to track
587 * cache hits accurately). Such sub-states are maintained by the layers that
588 * are interested in them.
592 * Page is in the cache, un-owned. Page leaves cached state in the
595 * - [cl_page_state::CPS_OWNED] io comes across the page and
598 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
599 * req-formation engine decides that it wants to include this page
600 * into an cl_req being constructed, and yanks it from the cache;
602 * - [cl_page_state::CPS_FREEING] VM callback is executed to
603 * evict the page form the memory;
605 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
609 * Page is exclusively owned by some cl_io. Page may end up in this
610 * state as a result of
612 * - io creating new page and immediately owning it;
614 * - [cl_page_state::CPS_CACHED] io finding existing cached page
617 * - [cl_page_state::CPS_OWNED] io finding existing owned page
618 * and waiting for owner to release the page;
620 * Page leaves owned state in the following cases:
622 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
623 * the cache, doing nothing;
625 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
628 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
629 * transfer for this page;
631 * - [cl_page_state::CPS_FREEING] io decides to destroy this
632 * page (e.g., as part of truncate or extent lock cancellation).
634 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
638 * Page is being written out, as a part of a transfer. This state is
639 * entered when req-formation logic decided that it wants this page to
640 * be sent through the wire _now_. Specifically, it means that once
641 * this state is achieved, transfer completion handler (with either
642 * success or failure indication) is guaranteed to be executed against
643 * this page independently of any locks and any scheduling decisions
644 * made by the hosting environment (that effectively means that the
645 * page is never put into cl_page_state::CPS_PAGEOUT state "in
646 * advance". This property is mentioned, because it is important when
647 * reasoning about possible dead-locks in the system). The page can
648 * enter this state as a result of
650 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
651 * write-out of this page, or
653 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
654 * that it has enough dirty pages cached to issue a "good"
657 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
658 * is completed---it is moved into cl_page_state::CPS_CACHED state.
660 * Underlying VM page is locked for the duration of transfer.
662 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
666 * Page is being read in, as a part of a transfer. This is quite
667 * similar to the cl_page_state::CPS_PAGEOUT state, except that
668 * read-in is always "immediate"---there is no such thing a sudden
669 * construction of read cl_req from cached, presumably not up to date,
672 * Underlying VM page is locked for the duration of transfer.
674 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
678 * Page is being destroyed. This state is entered when client decides
679 * that page has to be deleted from its host object, as, e.g., a part
682 * Once this state is reached, there is no way to escape it.
684 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
691 /** Host page, the page is from the host inode which the cl_page
695 /** Transient page, the transient cl_page is used to bind a cl_page
696 * to vmpage which is not belonging to the same object of cl_page.
697 * it is used in DirectIO, lockless IO and liblustre. */
702 * Flags maintained for every cl_page.
706 * Set when pagein completes. Used for debugging (read completes at
707 * most once for a page).
709 CPF_READ_COMPLETED = 1 << 0
713 * Fields are protected by the lock on struct page, except for atomics and
716 * \invariant Data type invariants are in cl_page_invariant(). Basically:
717 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
718 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
719 * cl_page::cp_owner (when set).
722 /** Reference counter. */
724 /** An object this page is a part of. Immutable after creation. */
725 struct cl_object *cp_obj;
726 /** List of slices. Immutable after creation. */
727 cfs_list_t cp_layers;
728 struct page *cp_vmpage;
730 * Page state. This field is const to avoid accidental update, it is
731 * modified only internally within cl_page.c. Protected by a VM lock.
733 const enum cl_page_state cp_state;
734 /** Linkage of pages within group. Protected by cl_page::cp_mutex. */
736 /** Mutex serializing membership of a page in a batch. */
737 struct mutex cp_mutex;
738 /** Linkage of pages within cl_req. */
739 cfs_list_t cp_flight;
740 /** Transfer error. */
744 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
747 enum cl_page_type cp_type;
750 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
751 * by sub-io. Protected by a VM lock.
753 struct cl_io *cp_owner;
755 * Debug information, the task is owning the page.
757 struct task_struct *cp_task;
759 * Owning IO request in cl_page_state::CPS_PAGEOUT and
760 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
761 * the top-level pages. Protected by a VM lock.
763 struct cl_req *cp_req;
764 /** List of references to this page, for debugging. */
765 struct lu_ref cp_reference;
766 /** Link to an object, for debugging. */
767 struct lu_ref_link cp_obj_ref;
768 /** Link to a queue, for debugging. */
769 struct lu_ref_link cp_queue_ref;
770 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */
772 /** Assigned if doing a sync_io */
773 struct cl_sync_io *cp_sync_io;
777 * Per-layer part of cl_page.
779 * \see ccc_page, lov_page, osc_page
781 struct cl_page_slice {
782 struct cl_page *cpl_page;
785 * Object slice corresponding to this page slice. Immutable after
788 struct cl_object *cpl_obj;
789 const struct cl_page_operations *cpl_ops;
790 /** Linkage into cl_page::cp_layers. Immutable after creation. */
791 cfs_list_t cpl_linkage;
795 * Lock mode. For the client extent locks.
797 * \warning: cl_lock_mode_match() assumes particular ordering here.
802 * Mode of a lock that protects no data, and exists only as a
803 * placeholder. This is used for `glimpse' requests. A phantom lock
804 * might get promoted to real lock at some point.
813 * Requested transfer type.
823 * Per-layer page operations.
825 * Methods taking an \a io argument are for the activity happening in the
826 * context of given \a io. Page is assumed to be owned by that io, except for
827 * the obvious cases (like cl_page_operations::cpo_own()).
829 * \see vvp_page_ops, lov_page_ops, osc_page_ops
831 struct cl_page_operations {
833 * cl_page<->struct page methods. Only one layer in the stack has to
834 * implement these. Current code assumes that this functionality is
835 * provided by the topmost layer, see cl_page_disown0() as an example.
839 * Called when \a io acquires this page into the exclusive
840 * ownership. When this method returns, it is guaranteed that the is
841 * not owned by other io, and no transfer is going on against
845 * \see vvp_page_own(), lov_page_own()
847 int (*cpo_own)(const struct lu_env *env,
848 const struct cl_page_slice *slice,
849 struct cl_io *io, int nonblock);
850 /** Called when ownership it yielded. Optional.
852 * \see cl_page_disown()
853 * \see vvp_page_disown()
855 void (*cpo_disown)(const struct lu_env *env,
856 const struct cl_page_slice *slice, struct cl_io *io);
858 * Called for a page that is already "owned" by \a io from VM point of
861 * \see cl_page_assume()
862 * \see vvp_page_assume(), lov_page_assume()
864 void (*cpo_assume)(const struct lu_env *env,
865 const struct cl_page_slice *slice, struct cl_io *io);
866 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
867 * bottom-to-top when IO releases a page without actually unlocking
870 * \see cl_page_unassume()
871 * \see vvp_page_unassume()
873 void (*cpo_unassume)(const struct lu_env *env,
874 const struct cl_page_slice *slice,
877 * Announces whether the page contains valid data or not by \a uptodate.
879 * \see cl_page_export()
880 * \see vvp_page_export()
882 void (*cpo_export)(const struct lu_env *env,
883 const struct cl_page_slice *slice, int uptodate);
885 * Checks whether underlying VM page is locked (in the suitable
886 * sense). Used for assertions.
888 * \retval -EBUSY: page is protected by a lock of a given mode;
889 * \retval -ENODATA: page is not protected by a lock;
890 * \retval 0: this layer cannot decide. (Should never happen.)
892 int (*cpo_is_vmlocked)(const struct lu_env *env,
893 const struct cl_page_slice *slice);
899 * Called when page is truncated from the object. Optional.
901 * \see cl_page_discard()
902 * \see vvp_page_discard(), osc_page_discard()
904 void (*cpo_discard)(const struct lu_env *env,
905 const struct cl_page_slice *slice,
908 * Called when page is removed from the cache, and is about to being
909 * destroyed. Optional.
911 * \see cl_page_delete()
912 * \see vvp_page_delete(), osc_page_delete()
914 void (*cpo_delete)(const struct lu_env *env,
915 const struct cl_page_slice *slice);
916 /** Destructor. Frees resources and slice itself. */
917 void (*cpo_fini)(const struct lu_env *env,
918 struct cl_page_slice *slice);
921 * Checks whether the page is protected by a cl_lock. This is a
922 * per-layer method, because certain layers have ways to check for the
923 * lock much more efficiently than through the generic locks scan, or
924 * implement locking mechanisms separate from cl_lock, e.g.,
925 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
926 * being canceled, or scheduled for cancellation as soon as the last
927 * user goes away, too.
929 * \retval -EBUSY: page is protected by a lock of a given mode;
930 * \retval -ENODATA: page is not protected by a lock;
931 * \retval 0: this layer cannot decide.
933 * \see cl_page_is_under_lock()
935 int (*cpo_is_under_lock)(const struct lu_env *env,
936 const struct cl_page_slice *slice,
937 struct cl_io *io, pgoff_t *max);
940 * Optional debugging helper. Prints given page slice.
942 * \see cl_page_print()
944 int (*cpo_print)(const struct lu_env *env,
945 const struct cl_page_slice *slice,
946 void *cookie, lu_printer_t p);
950 * Transfer methods. See comment on cl_req for a description of
951 * transfer formation and life-cycle.
956 * Request type dependent vector of operations.
958 * Transfer operations depend on transfer mode (cl_req_type). To avoid
959 * passing transfer mode to each and every of these methods, and to
960 * avoid branching on request type inside of the methods, separate
961 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
962 * provided. That is, method invocation usually looks like
964 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
968 * Called when a page is submitted for a transfer as a part of
971 * \return 0 : page is eligible for submission;
972 * \return -EALREADY : skip this page;
973 * \return -ve : error.
975 * \see cl_page_prep()
977 int (*cpo_prep)(const struct lu_env *env,
978 const struct cl_page_slice *slice,
981 * Completion handler. This is guaranteed to be eventually
982 * fired after cl_page_operations::cpo_prep() or
983 * cl_page_operations::cpo_make_ready() call.
985 * This method can be called in a non-blocking context. It is
986 * guaranteed however, that the page involved and its object
987 * are pinned in memory (and, hence, calling cl_page_put() is
990 * \see cl_page_completion()
992 void (*cpo_completion)(const struct lu_env *env,
993 const struct cl_page_slice *slice,
996 * Called when cached page is about to be added to the
997 * cl_req as a part of req formation.
999 * \return 0 : proceed with this page;
1000 * \return -EAGAIN : skip this page;
1001 * \return -ve : error.
1003 * \see cl_page_make_ready()
1005 int (*cpo_make_ready)(const struct lu_env *env,
1006 const struct cl_page_slice *slice);
1009 * Tell transfer engine that only [to, from] part of a page should be
1012 * This is used for immediate transfers.
1014 * \todo XXX this is not very good interface. It would be much better
1015 * if all transfer parameters were supplied as arguments to
1016 * cl_io_operations::cio_submit() call, but it is not clear how to do
1017 * this for page queues.
1019 * \see cl_page_clip()
1021 void (*cpo_clip)(const struct lu_env *env,
1022 const struct cl_page_slice *slice,
1025 * \pre the page was queued for transferring.
1026 * \post page is removed from client's pending list, or -EBUSY
1027 * is returned if it has already been in transferring.
1029 * This is one of seldom page operation which is:
1030 * 0. called from top level;
1031 * 1. don't have vmpage locked;
1032 * 2. every layer should synchronize execution of its ->cpo_cancel()
1033 * with completion handlers. Osc uses client obd lock for this
1034 * purpose. Based on there is no vvp_page_cancel and
1035 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1037 * \see osc_page_cancel().
1039 int (*cpo_cancel)(const struct lu_env *env,
1040 const struct cl_page_slice *slice);
1042 * Write out a page by kernel. This is only called by ll_writepage
1045 * \see cl_page_flush()
1047 int (*cpo_flush)(const struct lu_env *env,
1048 const struct cl_page_slice *slice,
1054 * Helper macro, dumping detailed information about \a page into a log.
1056 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1058 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1059 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1060 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1061 CDEBUG(mask, format , ## __VA_ARGS__); \
1066 * Helper macro, dumping shorter information about \a page into a log.
1068 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1070 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1071 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1072 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1073 CDEBUG(mask, format , ## __VA_ARGS__); \
1077 static inline int __page_in_use(const struct cl_page *page, int refc)
1079 if (page->cp_type == CPT_CACHEABLE)
1081 LASSERT(atomic_read(&page->cp_ref) > 0);
1082 return (atomic_read(&page->cp_ref) > refc);
1084 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1085 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1087 static inline struct page *cl_page_vmpage(struct cl_page *page)
1089 LASSERT(page->cp_vmpage != NULL);
1090 return page->cp_vmpage;
1095 /** \addtogroup cl_lock cl_lock
1099 * Extent locking on the client.
1103 * The locking model of the new client code is built around
1107 * data-type representing an extent lock on a regular file. cl_lock is a
1108 * layered object (much like cl_object and cl_page), it consists of a header
1109 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1110 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1112 * All locks for a given object are linked into cl_object_header::coh_locks
1113 * list (protected by cl_object_header::coh_lock_guard spin-lock) through
1114 * cl_lock::cll_linkage. Currently this list is not sorted in any way. We can
1115 * sort it in starting lock offset, or use altogether different data structure
1118 * Typical cl_lock consists of the two layers:
1120 * - vvp_lock (vvp specific data), and
1121 * - lov_lock (lov specific data).
1123 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1124 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1126 * - lovsub_lock, and
1129 * Each sub-lock is associated with a cl_object (representing stripe
1130 * sub-object or the file to which top-level cl_lock is associated to), and is
1131 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1132 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1133 * is different from cl_page, that doesn't fan out (there is usually exactly
1134 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1135 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1139 * cl_lock is reference counted. When reference counter drops to 0, lock is
1140 * placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING
1141 * lock is destroyed when last reference is released. Referencing between
1142 * top-lock and its sub-locks is described in the lov documentation module.
1146 * Also, cl_lock is a state machine. This requires some clarification. One of
1147 * the goals of client IO re-write was to make IO path non-blocking, or at
1148 * least to make it easier to make it non-blocking in the future. Here
1149 * `non-blocking' means that when a system call (read, write, truncate)
1150 * reaches a situation where it has to wait for a communication with the
1151 * server, it should --instead of waiting-- remember its current state and
1152 * switch to some other work. E.g,. instead of waiting for a lock enqueue,
1153 * client should proceed doing IO on the next stripe, etc. Obviously this is
1154 * rather radical redesign, and it is not planned to be fully implemented at
1155 * this time, instead we are putting some infrastructure in place, that would
1156 * make it easier to do asynchronous non-blocking IO easier in the
1157 * future. Specifically, where old locking code goes to sleep (waiting for
1158 * enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When
1159 * enqueue reply comes, its completion handler signals that lock state-machine
1160 * is ready to transit to the next state. There is some generic code in
1161 * cl_lock.c that sleeps, waiting for these signals. As a result, for users of
1162 * this cl_lock.c code, it looks like locking is done in normal blocking
1163 * fashion, and it the same time it is possible to switch to the non-blocking
1164 * locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c
1167 * For a description of state machine states and transitions see enum
1170 * There are two ways to restrict a set of states which lock might move to:
1172 * - placing a "hold" on a lock guarantees that lock will not be moved
1173 * into cl_lock_state::CLS_FREEING state until hold is released. Hold
1174 * can be only acquired on a lock that is not in
1175 * cl_lock_state::CLS_FREEING. All holds on a lock are counted in
1176 * cl_lock::cll_holds. Hold protects lock from cancellation and
1177 * destruction. Requests to cancel and destroy a lock on hold will be
1178 * recorded, but only honored when last hold on a lock is released;
1180 * - placing a "user" on a lock guarantees that lock will not leave
1181 * cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING,
1182 * cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of
1183 * states, once it enters this set. That is, if a user is added onto a
1184 * lock in a state not from this set, it doesn't immediately enforce
1185 * lock to move to this set, but once lock enters this set it will
1186 * remain there until all users are removed. Lock users are counted in
1187 * cl_lock::cll_users.
1189 * User is used to assure that lock is not canceled or destroyed while
1190 * it is being enqueued, or actively used by some IO.
1192 * Currently, a user always comes with a hold (cl_lock_invariant()
1193 * checks that a number of holds is not less than a number of users).
1197 * This is how lock state-machine operates. struct cl_lock contains a mutex
1198 * cl_lock::cll_guard that protects struct fields.
1200 * - mutex is taken, and cl_lock::cll_state is examined.
1202 * - for every state there are possible target states where lock can move
1203 * into. They are tried in order. Attempts to move into next state are
1204 * done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try().
1206 * - if the transition can be performed immediately, state is changed,
1207 * and mutex is released.
1209 * - if the transition requires blocking, _try() function returns
1210 * cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to
1211 * sleep, waiting for possibility of lock state change. It is woken
1212 * up when some event occurs, that makes lock state change possible
1213 * (e.g., the reception of the reply from the server), and repeats
1216 * Top-lock and sub-lock has separate mutexes and the latter has to be taken
1217 * first to avoid dead-lock.
1219 * To see an example of interaction of all these issues, take a look at the
1220 * lov_cl.c:lov_lock_enqueue() function. It is called as a part of
1221 * cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by
1222 * advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note
1223 * also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It
1224 * also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be
1225 * done in parallel, rather than one after another (this is used for glimpse
1226 * locks, that cannot dead-lock).
1228 * INTERFACE AND USAGE
1230 * struct cl_lock_operations provide a number of call-backs that are invoked
1231 * when events of interest occurs. Layers can intercept and handle glimpse,
1232 * blocking, cancel ASTs and a reception of the reply from the server.
1234 * One important difference with the old client locking model is that new
1235 * client has a representation for the top-lock, whereas in the old code only
1236 * sub-locks existed as real data structures and file-level locks are
1237 * represented by "request sets" that are created and destroyed on each and
1238 * every lock creation.
1240 * Top-locks are cached, and can be found in the cache by the system calls. It
1241 * is possible that top-lock is in cache, but some of its sub-locks were
1242 * canceled and destroyed. In that case top-lock has to be enqueued again
1243 * before it can be used.
1245 * Overall process of the locking during IO operation is as following:
1247 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1248 * is called on each layer. Responsibility of this method is to add locks,
1249 * needed by a given layer into cl_io.ci_lockset.
1251 * - once locks for all layers were collected, they are sorted to avoid
1252 * dead-locks (cl_io_locks_sort()), and enqueued.
1254 * - when all locks are acquired, IO is performed;
1256 * - locks are released into cache.
1258 * Striping introduces major additional complexity into locking. The
1259 * fundamental problem is that it is generally unsafe to actively use (hold)
1260 * two locks on the different OST servers at the same time, as this introduces
1261 * inter-server dependency and can lead to cascading evictions.
1263 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1264 * that no multi-stripe locks are taken (note that this design abandons POSIX
1265 * read/write semantics). Such pieces ideally can be executed concurrently. At
1266 * the same time, certain types of IO cannot be sub-divived, without
1267 * sacrificing correctness. This includes:
1269 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1272 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1274 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1275 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1276 * has to be held together with the usual lock on [offset, offset + count].
1278 * As multi-stripe locks have to be allowed, it makes sense to cache them, so
1279 * that, for example, a sequence of O_APPEND writes can proceed quickly
1280 * without going down to the individual stripes to do lock matching. On the
1281 * other hand, multi-stripe locks shouldn't be used by normal read/write
1282 * calls. To achieve this, every layer can implement ->clo_fits_into() method,
1283 * that is called by lock matching code (cl_lock_lookup()), and that can be
1284 * used to selectively disable matching of certain locks for certain IOs. For
1285 * exmaple, lov layer implements lov_lock_fits_into() that allow multi-stripe
1286 * locks to be matched only for truncates and O_APPEND writes.
1288 * Interaction with DLM
1290 * In the expected setup, cl_lock is ultimately backed up by a collection of
1291 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1292 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1293 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1294 * description of interaction with DLM.
1300 struct cl_lock_descr {
1301 /** Object this lock is granted for. */
1302 struct cl_object *cld_obj;
1303 /** Index of the first page protected by this lock. */
1305 /** Index of the last page (inclusive) protected by this lock. */
1307 /** Group ID, for group lock */
1310 enum cl_lock_mode cld_mode;
1312 * flags to enqueue lock. A combination of bit-flags from
1313 * enum cl_enq_flags.
1315 __u32 cld_enq_flags;
1318 #define DDESCR "%s(%d):[%lu, %lu]"
1319 #define PDESCR(descr) \
1320 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1321 (descr)->cld_start, (descr)->cld_end
1323 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1326 * Lock state-machine states.
1331 * Possible state transitions:
1333 * +------------------>NEW
1335 * | | cl_enqueue_try()
1337 * | cl_unuse_try() V
1338 * | +--------------QUEUING (*)
1340 * | | | cl_enqueue_try()
1342 * | | cl_unuse_try() V
1343 * sub-lock | +-------------ENQUEUED (*)
1345 * | | | cl_wait_try()
1350 * | | HELD<---------+
1352 * | | | | cl_use_try()
1353 * | | cl_unuse_try() | |
1356 * | +------------>INTRANSIT (D) <--+
1358 * | cl_unuse_try() | | cached lock found
1359 * | | | cl_use_try()
1362 * +------------------CACHED---------+
1371 * In states marked with (*) transition to the same state (i.e., a loop
1372 * in the diagram) is possible.
1374 * (R) is the point where Receive call-back is invoked: it allows layers
1375 * to handle arrival of lock reply.
1377 * (C) is the point where Cancellation call-back is invoked.
1379 * (D) is the transit state which means the lock is changing.
1381 * Transition to FREEING state is possible from any other state in the
1382 * diagram in case of unrecoverable error.
1386 * These states are for individual cl_lock object. Top-lock and its sub-locks
1387 * can be in the different states. Another way to say this is that we have
1388 * nested state-machines.
1390 * Separate QUEUING and ENQUEUED states are needed to support non-blocking
1391 * operation for locks with multiple sub-locks. Imagine lock on a file F, that
1392 * intersects 3 stripes S0, S1, and S2. To enqueue F client has to send
1393 * enqueue to S0, wait for its completion, then send enqueue for S1, wait for
1394 * its completion and at last enqueue lock for S2, and wait for its
1395 * completion. In that case, top-lock is in QUEUING state while S0, S1 are
1396 * handled, and is in ENQUEUED state after enqueue to S2 has been sent (note
1397 * that in this case, sub-locks move from state to state, and top-lock remains
1398 * in the same state).
1400 enum cl_lock_state {
1402 * Lock that wasn't yet enqueued
1406 * Enqueue is in progress, blocking for some intermediate interaction
1407 * with the other side.
1411 * Lock is fully enqueued, waiting for server to reply when it is
1416 * Lock granted, actively used by some IO.
1420 * This state is used to mark the lock is being used, or unused.
1421 * We need this state because the lock may have several sublocks,
1422 * so it's impossible to have an atomic way to bring all sublocks
1423 * into CLS_HELD state at use case, or all sublocks to CLS_CACHED
1425 * If a thread is referring to a lock, and it sees the lock is in this
1426 * state, it must wait for the lock.
1427 * See state diagram for details.
1431 * Lock granted, not used.
1435 * Lock is being destroyed.
1441 enum cl_lock_flags {
1443 * lock has been cancelled. This flag is never cleared once set (by
1444 * cl_lock_cancel0()).
1446 CLF_CANCELLED = 1 << 0,
1447 /** cancellation is pending for this lock. */
1448 CLF_CANCELPEND = 1 << 1,
1449 /** destruction is pending for this lock. */
1450 CLF_DOOMED = 1 << 2,
1451 /** from enqueue RPC reply upcall. */
1452 CLF_FROM_UPCALL= 1 << 3,
1458 * Lock closure is a collection of locks (both top-locks and sub-locks) that
1459 * might be updated in a result of an operation on a certain lock (which lock
1460 * this is a closure of).
1462 * Closures are needed to guarantee dead-lock freedom in the presence of
1464 * - nested state-machines (top-lock state-machine composed of sub-lock
1465 * state-machines), and
1467 * - shared sub-locks.
1469 * Specifically, many operations, such as lock enqueue, wait, unlock,
1470 * etc. start from a top-lock, and then operate on a sub-locks of this
1471 * top-lock, holding a top-lock mutex. When sub-lock state changes as a result
1472 * of such operation, this change has to be propagated to all top-locks that
1473 * share this sub-lock. Obviously, no natural lock ordering (e.g.,
1474 * top-to-bottom or bottom-to-top) captures this scenario, so try-locking has
1475 * to be used. Lock closure systematizes this try-and-repeat logic.
1477 struct cl_lock_closure {
1479 * Lock that is mutexed when closure construction is started. When
1480 * closure in is `wait' mode (cl_lock_closure::clc_wait), mutex on
1481 * origin is released before waiting.
1483 struct cl_lock *clc_origin;
1485 * List of enclosed locks, so far. Locks are linked here through
1486 * cl_lock::cll_inclosure.
1488 cfs_list_t clc_list;
1490 * True iff closure is in a `wait' mode. This determines what
1491 * cl_lock_enclosure() does when a lock L to be added to the closure
1492 * is currently mutexed by some other thread.
1494 * If cl_lock_closure::clc_wait is not set, then closure construction
1495 * fails with CLO_REPEAT immediately.
1497 * In wait mode, cl_lock_enclosure() waits until next attempt to build
1498 * a closure might succeed. To this end it releases an origin mutex
1499 * (cl_lock_closure::clc_origin), that has to be the only lock mutex
1500 * owned by the current thread, and then waits on L mutex (by grabbing
1501 * it and immediately releasing), before returning CLO_REPEAT to the
1505 /** Number of locks in the closure. */
1510 * Layered client lock.
1513 /** Reference counter. */
1515 /** List of slices. Immutable after creation. */
1516 cfs_list_t cll_layers;
1518 * Linkage into cl_lock::cll_descr::cld_obj::coh_locks list. Protected
1519 * by cl_lock::cll_descr::cld_obj::coh_lock_guard.
1521 cfs_list_t cll_linkage;
1523 * Parameters of this lock. Protected by
1524 * cl_lock::cll_descr::cld_obj::coh_lock_guard nested within
1525 * cl_lock::cll_guard. Modified only on lock creation and in
1528 struct cl_lock_descr cll_descr;
1529 /** Protected by cl_lock::cll_guard. */
1530 enum cl_lock_state cll_state;
1531 /** signals state changes. */
1532 wait_queue_head_t cll_wq;
1534 * Recursive lock, most fields in cl_lock{} are protected by this.
1536 * Locking rules: this mutex is never held across network
1537 * communication, except when lock is being canceled.
1539 * Lock ordering: a mutex of a sub-lock is taken first, then a mutex
1540 * on a top-lock. Other direction is implemented through a
1541 * try-lock-repeat loop. Mutices of unrelated locks can be taken only
1544 * \see osc_lock_enqueue_wait(), lov_lock_cancel(), lov_sublock_wait().
1546 struct mutex cll_guard;
1547 struct task_struct *cll_guarder;
1551 * the owner for INTRANSIT state
1553 struct task_struct *cll_intransit_owner;
1556 * Number of holds on a lock. A hold prevents a lock from being
1557 * canceled and destroyed. Protected by cl_lock::cll_guard.
1559 * \see cl_lock_hold(), cl_lock_unhold(), cl_lock_release()
1563 * Number of lock users. Valid in cl_lock_state::CLS_HELD state
1564 * only. Lock user pins lock in CLS_HELD state. Protected by
1565 * cl_lock::cll_guard.
1567 * \see cl_wait(), cl_unuse().
1571 * Flag bit-mask. Values from enum cl_lock_flags. Updates are
1572 * protected by cl_lock::cll_guard.
1574 unsigned long cll_flags;
1576 * A linkage into a list of locks in a closure.
1578 * \see cl_lock_closure
1580 cfs_list_t cll_inclosure;
1582 * Confict lock at queuing time.
1584 struct cl_lock *cll_conflict;
1586 * A list of references to this lock, for debugging.
1588 struct lu_ref cll_reference;
1590 * A list of holds on this lock, for debugging.
1592 struct lu_ref cll_holders;
1594 * A reference for cl_lock::cll_descr::cld_obj. For debugging.
1596 struct lu_ref_link cll_obj_ref;
1597 #ifdef CONFIG_LOCKDEP
1598 /* "dep_map" name is assumed by lockdep.h macros. */
1599 struct lockdep_map dep_map;
1604 * Per-layer part of cl_lock
1606 * \see ccc_lock, lov_lock, lovsub_lock, osc_lock
1608 struct cl_lock_slice {
1609 struct cl_lock *cls_lock;
1610 /** Object slice corresponding to this lock slice. Immutable after
1612 struct cl_object *cls_obj;
1613 const struct cl_lock_operations *cls_ops;
1614 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1615 cfs_list_t cls_linkage;
1619 * Possible (non-error) return values of ->clo_{enqueue,wait,unlock}().
1621 * NOTE: lov_subresult() depends on ordering here.
1623 enum cl_lock_transition {
1624 /** operation cannot be completed immediately. Wait for state change. */
1626 /** operation had to release lock mutex, restart. */
1628 /** lower layer re-enqueued. */
1634 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1636 struct cl_lock_operations {
1638 * \name statemachine
1640 * State machine transitions. These 3 methods are called to transfer
1641 * lock from one state to another, as described in the commentary
1642 * above enum #cl_lock_state.
1644 * \retval 0 this layer has nothing more to do to before
1645 * transition to the target state happens;
1647 * \retval CLO_REPEAT method had to release and re-acquire cl_lock
1648 * mutex, repeat invocation of transition method
1649 * across all layers;
1651 * \retval CLO_WAIT this layer cannot move to the target state
1652 * immediately, as it has to wait for certain event
1653 * (e.g., the communication with the server). It
1654 * is guaranteed, that when the state transfer
1655 * becomes possible, cl_lock::cll_wq wait-queue
1656 * is signaled. Caller can wait for this event by
1657 * calling cl_lock_state_wait();
1659 * \retval -ve failure, abort state transition, move the lock
1660 * into cl_lock_state::CLS_FREEING state, and set
1661 * cl_lock::cll_error.
1663 * Once all layers voted to agree to transition (by returning 0), lock
1664 * is moved into corresponding target state. All state transition
1665 * methods are optional.
1669 * Attempts to enqueue the lock. Called top-to-bottom.
1671 * \see ccc_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1672 * \see osc_lock_enqueue()
1674 int (*clo_enqueue)(const struct lu_env *env,
1675 const struct cl_lock_slice *slice,
1676 struct cl_io *io, __u32 enqflags);
1678 * Attempts to wait for enqueue result. Called top-to-bottom.
1680 * \see ccc_lock_wait(), lov_lock_wait(), osc_lock_wait()
1682 int (*clo_wait)(const struct lu_env *env,
1683 const struct cl_lock_slice *slice);
1685 * Attempts to unlock the lock. Called bottom-to-top. In addition to
1686 * usual return values of lock state-machine methods, this can return
1687 * -ESTALE to indicate that lock cannot be returned to the cache, and
1688 * has to be re-initialized.
1689 * unuse is a one-shot operation, so it must NOT return CLO_WAIT.
1691 * \see ccc_lock_unuse(), lov_lock_unuse(), osc_lock_unuse()
1693 int (*clo_unuse)(const struct lu_env *env,
1694 const struct cl_lock_slice *slice);
1696 * Notifies layer that cached lock is started being used.
1698 * \pre lock->cll_state == CLS_CACHED
1700 * \see lov_lock_use(), osc_lock_use()
1702 int (*clo_use)(const struct lu_env *env,
1703 const struct cl_lock_slice *slice);
1704 /** @} statemachine */
1706 * A method invoked when lock state is changed (as a result of state
1707 * transition). This is used, for example, to track when the state of
1708 * a sub-lock changes, to propagate this change to the corresponding
1709 * top-lock. Optional
1711 * \see lovsub_lock_state()
1713 void (*clo_state)(const struct lu_env *env,
1714 const struct cl_lock_slice *slice,
1715 enum cl_lock_state st);
1717 * Returns true, iff given lock is suitable for the given io, idea
1718 * being, that there are certain "unsafe" locks, e.g., ones acquired
1719 * for O_APPEND writes, that we don't want to re-use for a normal
1720 * write, to avoid the danger of cascading evictions. Optional. Runs
1721 * under cl_object_header::coh_lock_guard.
1723 * XXX this should take more information about lock needed by
1724 * io. Probably lock description or something similar.
1726 * \see lov_fits_into()
1728 int (*clo_fits_into)(const struct lu_env *env,
1729 const struct cl_lock_slice *slice,
1730 const struct cl_lock_descr *need,
1731 const struct cl_io *io);
1734 * Asynchronous System Traps. All of then are optional, all are
1735 * executed bottom-to-top.
1740 * Cancellation callback. Cancel a lock voluntarily, or under
1741 * the request of server.
1743 void (*clo_cancel)(const struct lu_env *env,
1744 const struct cl_lock_slice *slice);
1746 * Lock weighting ast. Executed to estimate how precious this lock
1747 * is. The sum of results across all layers is used to determine
1748 * whether lock worth keeping in cache given present memory usage.
1750 * \see osc_lock_weigh(), vvp_lock_weigh(), lovsub_lock_weigh().
1752 unsigned long (*clo_weigh)(const struct lu_env *env,
1753 const struct cl_lock_slice *slice);
1757 * \see lovsub_lock_closure()
1759 int (*clo_closure)(const struct lu_env *env,
1760 const struct cl_lock_slice *slice,
1761 struct cl_lock_closure *closure);
1763 * Executed bottom-to-top when lock description changes (e.g., as a
1764 * result of server granting more generous lock than was requested).
1766 * \see lovsub_lock_modify()
1768 int (*clo_modify)(const struct lu_env *env,
1769 const struct cl_lock_slice *slice,
1770 const struct cl_lock_descr *updated);
1772 * Notifies layers (bottom-to-top) that lock is going to be
1773 * destroyed. Responsibility of layers is to prevent new references on
1774 * this lock from being acquired once this method returns.
1776 * This can be called multiple times due to the races.
1778 * \see cl_lock_delete()
1779 * \see osc_lock_delete(), lovsub_lock_delete()
1781 void (*clo_delete)(const struct lu_env *env,
1782 const struct cl_lock_slice *slice);
1784 * Destructor. Frees resources and the slice.
1786 * \see ccc_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1787 * \see osc_lock_fini()
1789 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1791 * Optional debugging helper. Prints given lock slice.
1793 int (*clo_print)(const struct lu_env *env,
1794 void *cookie, lu_printer_t p,
1795 const struct cl_lock_slice *slice);
1798 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1800 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1801 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1802 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1803 CDEBUG(mask, format , ## __VA_ARGS__); \
1807 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1811 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1817 /** \addtogroup cl_page_list cl_page_list
1818 * Page list used to perform collective operations on a group of pages.
1820 * Pages are added to the list one by one. cl_page_list acquires a reference
1821 * for every page in it. Page list is used to perform collective operations on
1824 * - submit pages for an immediate transfer,
1826 * - own pages on behalf of certain io (waiting for each page in turn),
1830 * When list is finalized, it releases references on all pages it still has.
1832 * \todo XXX concurrency control.
1836 struct cl_page_list {
1838 cfs_list_t pl_pages;
1839 struct task_struct *pl_owner;
1843 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1844 * contains an incoming page list and an outgoing page list.
1847 struct cl_page_list c2_qin;
1848 struct cl_page_list c2_qout;
1851 /** @} cl_page_list */
1853 /** \addtogroup cl_io cl_io
1858 * cl_io represents a high level I/O activity like
1859 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1862 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1863 * important distinction. We want to minimize number of calls to the allocator
1864 * in the fast path, e.g., in the case of read(2) when everything is cached:
1865 * client already owns the lock over region being read, and data are cached
1866 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1867 * per-layer io state is stored in the session, associated with the io, see
1868 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1869 * by using free-lists, see cl_env_get().
1871 * There is a small predefined number of possible io types, enumerated in enum
1874 * cl_io is a state machine, that can be advanced concurrently by the multiple
1875 * threads. It is up to these threads to control the concurrency and,
1876 * specifically, to detect when io is done, and its state can be safely
1879 * For read/write io overall execution plan is as following:
1881 * (0) initialize io state through all layers;
1883 * (1) loop: prepare chunk of work to do
1885 * (2) call all layers to collect locks they need to process current chunk
1887 * (3) sort all locks to avoid dead-locks, and acquire them
1889 * (4) process the chunk: call per-page methods
1890 * (cl_io_operations::cio_read_page() for read,
1891 * cl_io_operations::cio_prepare_write(),
1892 * cl_io_operations::cio_commit_write() for write)
1898 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1899 * address allocation efficiency issues mentioned above), and returns with the
1900 * special error condition from per-page method when current sub-io has to
1901 * block. This causes io loop to be repeated, and lov switches to the next
1902 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1907 /** read system call */
1909 /** write system call */
1911 /** truncate, utime system calls */
1914 * page fault handling
1918 * fsync system call handling
1919 * To write out a range of file
1923 * Miscellaneous io. This is used for occasional io activity that
1924 * doesn't fit into other types. Currently this is used for:
1926 * - cancellation of an extent lock. This io exists as a context
1927 * to write dirty pages from under the lock being canceled back
1930 * - VM induced page write-out. An io context for writing page out
1931 * for memory cleansing;
1933 * - glimpse. An io context to acquire glimpse lock.
1935 * - grouplock. An io context to acquire group lock.
1937 * CIT_MISC io is used simply as a context in which locks and pages
1938 * are manipulated. Such io has no internal "process", that is,
1939 * cl_io_loop() is never called for it.
1946 * States of cl_io state machine
1949 /** Not initialized. */
1953 /** IO iteration started. */
1957 /** Actual IO is in progress. */
1959 /** IO for the current iteration finished. */
1961 /** Locks released. */
1963 /** Iteration completed. */
1965 /** cl_io finalized. */
1970 * IO state private for a layer.
1972 * This is usually embedded into layer session data, rather than allocated
1975 * \see vvp_io, lov_io, osc_io, ccc_io
1977 struct cl_io_slice {
1978 struct cl_io *cis_io;
1979 /** corresponding object slice. Immutable after creation. */
1980 struct cl_object *cis_obj;
1981 /** io operations. Immutable after creation. */
1982 const struct cl_io_operations *cis_iop;
1984 * linkage into a list of all slices for a given cl_io, hanging off
1985 * cl_io::ci_layers. Immutable after creation.
1987 cfs_list_t cis_linkage;
1990 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1994 * Per-layer io operations.
1995 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1997 struct cl_io_operations {
1999 * Vector of io state transition methods for every io type.
2001 * \see cl_page_operations::io
2005 * Prepare io iteration at a given layer.
2007 * Called top-to-bottom at the beginning of each iteration of
2008 * "io loop" (if it makes sense for this type of io). Here
2009 * layer selects what work it will do during this iteration.
2011 * \see cl_io_operations::cio_iter_fini()
2013 int (*cio_iter_init) (const struct lu_env *env,
2014 const struct cl_io_slice *slice);
2016 * Finalize io iteration.
2018 * Called bottom-to-top at the end of each iteration of "io
2019 * loop". Here layers can decide whether IO has to be
2022 * \see cl_io_operations::cio_iter_init()
2024 void (*cio_iter_fini) (const struct lu_env *env,
2025 const struct cl_io_slice *slice);
2027 * Collect locks for the current iteration of io.
2029 * Called top-to-bottom to collect all locks necessary for
2030 * this iteration. This methods shouldn't actually enqueue
2031 * anything, instead it should post a lock through
2032 * cl_io_lock_add(). Once all locks are collected, they are
2033 * sorted and enqueued in the proper order.
2035 int (*cio_lock) (const struct lu_env *env,
2036 const struct cl_io_slice *slice);
2038 * Finalize unlocking.
2040 * Called bottom-to-top to finish layer specific unlocking
2041 * functionality, after generic code released all locks
2042 * acquired by cl_io_operations::cio_lock().
2044 void (*cio_unlock)(const struct lu_env *env,
2045 const struct cl_io_slice *slice);
2047 * Start io iteration.
2049 * Once all locks are acquired, called top-to-bottom to
2050 * commence actual IO. In the current implementation,
2051 * top-level vvp_io_{read,write}_start() does all the work
2052 * synchronously by calling generic_file_*(), so other layers
2053 * are called when everything is done.
2055 int (*cio_start)(const struct lu_env *env,
2056 const struct cl_io_slice *slice);
2058 * Called top-to-bottom at the end of io loop. Here layer
2059 * might wait for an unfinished asynchronous io.
2061 void (*cio_end) (const struct lu_env *env,
2062 const struct cl_io_slice *slice);
2064 * Called bottom-to-top to notify layers that read/write IO
2065 * iteration finished, with \a nob bytes transferred.
2067 void (*cio_advance)(const struct lu_env *env,
2068 const struct cl_io_slice *slice,
2071 * Called once per io, bottom-to-top to release io resources.
2073 void (*cio_fini) (const struct lu_env *env,
2074 const struct cl_io_slice *slice);
2078 * Submit pages from \a queue->c2_qin for IO, and move
2079 * successfully submitted pages into \a queue->c2_qout. Return
2080 * non-zero if failed to submit even the single page. If
2081 * submission failed after some pages were moved into \a
2082 * queue->c2_qout, completion callback with non-zero ioret is
2085 int (*cio_submit)(const struct lu_env *env,
2086 const struct cl_io_slice *slice,
2087 enum cl_req_type crt,
2088 struct cl_2queue *queue);
2090 * Queue async page for write.
2091 * The difference between cio_submit and cio_queue is that
2092 * cio_submit is for urgent request.
2094 int (*cio_commit_async)(const struct lu_env *env,
2095 const struct cl_io_slice *slice,
2096 struct cl_page_list *queue, int from, int to,
2099 * Read missing page.
2101 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
2102 * method, when it hits not-up-to-date page in the range. Optional.
2104 * \pre io->ci_type == CIT_READ
2106 int (*cio_read_page)(const struct lu_env *env,
2107 const struct cl_io_slice *slice,
2108 const struct cl_page_slice *page);
2110 * Optional debugging helper. Print given io slice.
2112 int (*cio_print)(const struct lu_env *env, void *cookie,
2113 lu_printer_t p, const struct cl_io_slice *slice);
2117 * Flags to lock enqueue procedure.
2122 * instruct server to not block, if conflicting lock is found. Instead
2123 * -EWOULDBLOCK is returned immediately.
2125 CEF_NONBLOCK = 0x00000001,
2127 * take lock asynchronously (out of order), as it cannot
2128 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
2130 CEF_ASYNC = 0x00000002,
2132 * tell the server to instruct (though a flag in the blocking ast) an
2133 * owner of the conflicting lock, that it can drop dirty pages
2134 * protected by this lock, without sending them to the server.
2136 CEF_DISCARD_DATA = 0x00000004,
2138 * tell the sub layers that it must be a `real' lock. This is used for
2139 * mmapped-buffer locks and glimpse locks that must be never converted
2140 * into lockless mode.
2142 * \see vvp_mmap_locks(), cl_glimpse_lock().
2144 CEF_MUST = 0x00000008,
2146 * tell the sub layers that never request a `real' lock. This flag is
2147 * not used currently.
2149 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
2150 * conversion policy: ci_lockreq describes generic information of lock
2151 * requirement for this IO, especially for locks which belong to the
2152 * object doing IO; however, lock itself may have precise requirements
2153 * that are described by the enqueue flags.
2155 CEF_NEVER = 0x00000010,
2157 * for async glimpse lock.
2159 CEF_AGL = 0x00000020,
2161 * mask of enq_flags.
2163 CEF_MASK = 0x0000003f,
2167 * Link between lock and io. Intermediate structure is needed, because the
2168 * same lock can be part of multiple io's simultaneously.
2170 struct cl_io_lock_link {
2171 /** linkage into one of cl_lockset lists. */
2172 cfs_list_t cill_linkage;
2173 struct cl_lock_descr cill_descr;
2174 struct cl_lock *cill_lock;
2175 /** optional destructor */
2176 void (*cill_fini)(const struct lu_env *env,
2177 struct cl_io_lock_link *link);
2181 * Lock-set represents a collection of locks, that io needs at a
2182 * time. Generally speaking, client tries to avoid holding multiple locks when
2185 * - holding extent locks over multiple ost's introduces the danger of
2186 * "cascading timeouts";
2188 * - holding multiple locks over the same ost is still dead-lock prone,
2189 * see comment in osc_lock_enqueue(),
2191 * but there are certain situations where this is unavoidable:
2193 * - O_APPEND writes have to take [0, EOF] lock for correctness;
2195 * - truncate has to take [new-size, EOF] lock for correctness;
2197 * - SNS has to take locks across full stripe for correctness;
2199 * - in the case when user level buffer, supplied to {read,write}(file0),
2200 * is a part of a memory mapped lustre file, client has to take a dlm
2201 * locks on file0, and all files that back up the buffer (or a part of
2202 * the buffer, that is being processed in the current chunk, in any
2203 * case, there are situations where at least 2 locks are necessary).
2205 * In such cases we at least try to take locks in the same consistent
2206 * order. To this end, all locks are first collected, then sorted, and then
2210 /** locks to be acquired. */
2211 cfs_list_t cls_todo;
2212 /** locks currently being processed. */
2213 cfs_list_t cls_curr;
2214 /** locks acquired. */
2215 cfs_list_t cls_done;
2219 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
2220 * but 'req' is always to be thought as 'request' :-)
2222 enum cl_io_lock_dmd {
2223 /** Always lock data (e.g., O_APPEND). */
2225 /** Layers are free to decide between local and global locking. */
2227 /** Never lock: there is no cache (e.g., liblustre). */
2231 enum cl_fsync_mode {
2232 /** start writeback, do not wait for them to finish */
2234 /** start writeback and wait for them to finish */
2236 /** discard all of dirty pages in a specific file range */
2237 CL_FSYNC_DISCARD = 2,
2238 /** start writeback and make sure they have reached storage before
2239 * return. OST_SYNC RPC must be issued and finished */
2243 struct cl_io_rw_common {
2253 * cl_io is shared by all threads participating in this IO (in current
2254 * implementation only one thread advances IO, but parallel IO design and
2255 * concurrent copy_*_user() require multiple threads acting on the same IO. It
2256 * is up to these threads to serialize their activities, including updates to
2257 * mutable cl_io fields.
2260 /** type of this IO. Immutable after creation. */
2261 enum cl_io_type ci_type;
2262 /** current state of cl_io state machine. */
2263 enum cl_io_state ci_state;
2264 /** main object this io is against. Immutable after creation. */
2265 struct cl_object *ci_obj;
2267 * Upper layer io, of which this io is a part of. Immutable after
2270 struct cl_io *ci_parent;
2271 /** List of slices. Immutable after creation. */
2272 cfs_list_t ci_layers;
2273 /** list of locks (to be) acquired by this io. */
2274 struct cl_lockset ci_lockset;
2275 /** lock requirements, this is just a help info for sublayers. */
2276 enum cl_io_lock_dmd ci_lockreq;
2279 struct cl_io_rw_common rd;
2282 struct cl_io_rw_common wr;
2286 struct cl_io_rw_common ci_rw;
2287 struct cl_setattr_io {
2288 struct ost_lvb sa_attr;
2289 unsigned int sa_valid;
2290 struct obd_capa *sa_capa;
2292 struct cl_fault_io {
2293 /** page index within file. */
2295 /** bytes valid byte on a faulted page. */
2297 /** writable page? for nopage() only */
2299 /** page of an executable? */
2301 /** page_mkwrite() */
2303 /** resulting page */
2304 struct cl_page *ft_page;
2306 struct cl_fsync_io {
2309 struct obd_capa *fi_capa;
2310 /** file system level fid */
2311 struct lu_fid *fi_fid;
2312 enum cl_fsync_mode fi_mode;
2313 /* how many pages were written/discarded */
2314 unsigned int fi_nr_written;
2317 struct cl_2queue ci_queue;
2320 unsigned int ci_continue:1,
2322 * This io has held grouplock, to inform sublayers that
2323 * don't do lockless i/o.
2327 * The whole IO need to be restarted because layout has been changed
2331 * to not refresh layout - the IO issuer knows that the layout won't
2332 * change(page operations, layout change causes all page to be
2333 * discarded), or it doesn't matter if it changes(sync).
2337 * Check if layout changed after the IO finishes. Mainly for HSM
2338 * requirement. If IO occurs to openning files, it doesn't need to
2339 * verify layout because HSM won't release openning files.
2340 * Right now, only two opertaions need to verify layout: glimpse
2345 * file is released, restore has to to be triggered by vvp layer
2347 ci_restore_needed:1,
2353 * Number of pages owned by this IO. For invariant checking.
2355 unsigned ci_owned_nr;
2360 /** \addtogroup cl_req cl_req
2365 * There are two possible modes of transfer initiation on the client:
2367 * - immediate transfer: this is started when a high level io wants a page
2368 * or a collection of pages to be transferred right away. Examples:
2369 * read-ahead, synchronous read in the case of non-page aligned write,
2370 * page write-out as a part of extent lock cancellation, page write-out
2371 * as a part of memory cleansing. Immediate transfer can be both
2372 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
2374 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
2375 * when io wants to transfer a page to the server some time later, when
2376 * it can be done efficiently. Example: pages dirtied by the write(2)
2379 * In any case, transfer takes place in the form of a cl_req, which is a
2380 * representation for a network RPC.
2382 * Pages queued for an opportunistic transfer are cached until it is decided
2383 * that efficient RPC can be composed of them. This decision is made by "a
2384 * req-formation engine", currently implemented as a part of osc
2385 * layer. Req-formation depends on many factors: the size of the resulting
2386 * RPC, whether or not multi-object RPCs are supported by the server,
2387 * max-rpc-in-flight limitations, size of the dirty cache, etc.
2389 * For the immediate transfer io submits a cl_page_list, that req-formation
2390 * engine slices into cl_req's, possibly adding cached pages to some of
2391 * the resulting req's.
2393 * Whenever a page from cl_page_list is added to a newly constructed req, its
2394 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
2395 * page state is atomically changed from cl_page_state::CPS_OWNED to
2396 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
2397 * is zeroed, and cl_page::cp_req is set to the
2398 * req. cl_page_operations::cpo_prep() method at the particular layer might
2399 * return -EALREADY to indicate that it does not need to submit this page
2400 * at all. This is possible, for example, if page, submitted for read,
2401 * became up-to-date in the meantime; and for write, the page don't have
2402 * dirty bit marked. \see cl_io_submit_rw()
2404 * Whenever a cached page is added to a newly constructed req, its
2405 * cl_page_operations::cpo_make_ready() layer methods are called. At that
2406 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
2407 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
2408 * req. cl_page_operations::cpo_make_ready() method at the particular layer
2409 * might return -EAGAIN to indicate that this page is not eligible for the
2410 * transfer right now.
2414 * Plan is to divide transfers into "priority bands" (indicated when
2415 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
2416 * and allow glueing of cached pages to immediate transfers only within single
2417 * band. This would make high priority transfers (like lock cancellation or
2418 * memory pressure induced write-out) really high priority.
2423 * Per-transfer attributes.
2425 struct cl_req_attr {
2426 /** Generic attributes for the server consumption. */
2427 struct obdo *cra_oa;
2429 struct obd_capa *cra_capa;
2431 char cra_jobid[JOBSTATS_JOBID_SIZE];
2435 * Transfer request operations definable at every layer.
2437 * Concurrency: transfer formation engine synchronizes calls to all transfer
2440 struct cl_req_operations {
2442 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
2443 * complete (all pages are added).
2445 * \see osc_req_prep()
2447 int (*cro_prep)(const struct lu_env *env,
2448 const struct cl_req_slice *slice);
2450 * Called top-to-bottom to fill in \a oa fields. This is called twice
2451 * with different flags, see bug 10150 and osc_build_req().
2453 * \param obj an object from cl_req which attributes are to be set in
2456 * \param oa struct obdo where attributes are placed
2458 * \param flags \a oa fields to be filled.
2460 void (*cro_attr_set)(const struct lu_env *env,
2461 const struct cl_req_slice *slice,
2462 const struct cl_object *obj,
2463 struct cl_req_attr *attr, obd_valid flags);
2465 * Called top-to-bottom from cl_req_completion() to notify layers that
2466 * transfer completed. Has to free all state allocated by
2467 * cl_device_operations::cdo_req_init().
2469 void (*cro_completion)(const struct lu_env *env,
2470 const struct cl_req_slice *slice, int ioret);
2474 * A per-object state that (potentially multi-object) transfer request keeps.
2477 /** object itself */
2478 struct cl_object *ro_obj;
2479 /** reference to cl_req_obj::ro_obj. For debugging. */
2480 struct lu_ref_link ro_obj_ref;
2481 /* something else? Number of pages for a given object? */
2487 * Transfer requests are not reference counted, because IO sub-system owns
2488 * them exclusively and knows when to free them.
2492 * cl_req is created by cl_req_alloc() that calls
2493 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2494 * state in every layer.
2496 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2497 * contains pages for.
2499 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2500 * called top-to-bottom. At that point layers can modify req, let it pass, or
2501 * deny it completely. This is to support things like SNS that have transfer
2502 * ordering requirements invisible to the individual req-formation engine.
2504 * On transfer completion (or transfer timeout, or failure to initiate the
2505 * transfer of an allocated req), cl_req_operations::cro_completion() method
2506 * is called, after execution of cl_page_operations::cpo_completion() of all
2510 enum cl_req_type crq_type;
2511 /** A list of pages being transfered */
2512 cfs_list_t crq_pages;
2513 /** Number of pages in cl_req::crq_pages */
2514 unsigned crq_nrpages;
2515 /** An array of objects which pages are in ->crq_pages */
2516 struct cl_req_obj *crq_o;
2517 /** Number of elements in cl_req::crq_objs[] */
2518 unsigned crq_nrobjs;
2519 cfs_list_t crq_layers;
2523 * Per-layer state for request.
2525 struct cl_req_slice {
2526 struct cl_req *crs_req;
2527 struct cl_device *crs_dev;
2528 cfs_list_t crs_linkage;
2529 const struct cl_req_operations *crs_ops;
2534 enum cache_stats_item {
2535 /** how many cache lookups were performed */
2537 /** how many times cache lookup resulted in a hit */
2539 /** how many entities are in the cache right now */
2541 /** how many entities in the cache are actively used (and cannot be
2542 * evicted) right now */
2544 /** how many entities were created at all */
2549 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2552 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2554 struct cache_stats {
2555 const char *cs_name;
2556 atomic_t cs_stats[CS_NR];
2559 /** These are not exported so far */
2560 void cache_stats_init (struct cache_stats *cs, const char *name);
2563 * Client-side site. This represents particular client stack. "Global"
2564 * variables should (directly or indirectly) be added here to allow multiple
2565 * clients to co-exist in the single address space.
2568 struct lu_site cs_lu;
2570 * Statistical counters. Atomics do not scale, something better like
2571 * per-cpu counters is needed.
2573 * These are exported as /proc/fs/lustre/llite/.../site
2575 * When interpreting keep in mind that both sub-locks (and sub-pages)
2576 * and top-locks (and top-pages) are accounted here.
2578 struct cache_stats cs_pages;
2579 struct cache_stats cs_locks;
2580 atomic_t cs_pages_state[CPS_NR];
2581 atomic_t cs_locks_state[CLS_NR];
2584 int cl_site_init(struct cl_site *s, struct cl_device *top);
2585 void cl_site_fini(struct cl_site *s);
2586 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2589 * Output client site statistical counters into a buffer. Suitable for
2590 * ll_rd_*()-style functions.
2592 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2597 * Type conversion and accessory functions.
2601 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2603 return container_of(site, struct cl_site, cs_lu);
2606 static inline int lu_device_is_cl(const struct lu_device *d)
2608 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2611 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2613 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2614 return container_of0(d, struct cl_device, cd_lu_dev);
2617 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2619 return &d->cd_lu_dev;
2622 static inline struct cl_object *lu2cl(const struct lu_object *o)
2624 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2625 return container_of0(o, struct cl_object, co_lu);
2628 static inline const struct cl_object_conf *
2629 lu2cl_conf(const struct lu_object_conf *conf)
2631 return container_of0(conf, struct cl_object_conf, coc_lu);
2634 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2636 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2639 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2641 return container_of0(h, struct cl_object_header, coh_lu);
2644 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2646 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2650 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2652 return luh2coh(obj->co_lu.lo_header);
2655 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2657 return lu_device_init(&d->cd_lu_dev, t);
2660 static inline void cl_device_fini(struct cl_device *d)
2662 lu_device_fini(&d->cd_lu_dev);
2665 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2666 struct cl_object *obj, pgoff_t index,
2667 const struct cl_page_operations *ops);
2668 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2669 struct cl_object *obj,
2670 const struct cl_lock_operations *ops);
2671 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2672 struct cl_object *obj, const struct cl_io_operations *ops);
2673 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2674 struct cl_device *dev,
2675 const struct cl_req_operations *ops);
2678 /** \defgroup cl_object cl_object
2680 struct cl_object *cl_object_top (struct cl_object *o);
2681 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2682 const struct lu_fid *fid,
2683 const struct cl_object_conf *c);
2685 int cl_object_header_init(struct cl_object_header *h);
2686 void cl_object_header_fini(struct cl_object_header *h);
2687 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2688 void cl_object_get (struct cl_object *o);
2689 void cl_object_attr_lock (struct cl_object *o);
2690 void cl_object_attr_unlock(struct cl_object *o);
2691 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2692 struct cl_attr *attr);
2693 int cl_object_attr_set (const struct lu_env *env, struct cl_object *obj,
2694 const struct cl_attr *attr, unsigned valid);
2695 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2696 struct ost_lvb *lvb);
2697 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2698 const struct cl_object_conf *conf);
2699 void cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2700 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2701 int cl_object_has_locks (struct cl_object *obj);
2704 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2706 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2708 return cl_object_header(o0) == cl_object_header(o1);
2711 static inline void cl_object_page_init(struct cl_object *clob, int size)
2713 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2714 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2717 static inline void *cl_object_page_slice(struct cl_object *clob,
2718 struct cl_page *page)
2720 return (void *)((char *)page + clob->co_slice_off);
2724 * Return refcount of cl_object.
2726 static inline int cl_object_refc(struct cl_object *clob)
2728 struct lu_object_header *header = clob->co_lu.lo_header;
2729 return atomic_read(&header->loh_ref);
2734 /** \defgroup cl_page cl_page
2742 /* callback of cl_page_gang_lookup() */
2744 struct cl_page *cl_page_find (const struct lu_env *env,
2745 struct cl_object *obj,
2746 pgoff_t idx, struct page *vmpage,
2747 enum cl_page_type type);
2748 struct cl_page *cl_page_alloc (const struct lu_env *env,
2749 struct cl_object *o, pgoff_t ind,
2750 struct page *vmpage,
2751 enum cl_page_type type);
2752 void cl_page_get (struct cl_page *page);
2753 void cl_page_put (const struct lu_env *env,
2754 struct cl_page *page);
2755 void cl_page_print (const struct lu_env *env, void *cookie,
2756 lu_printer_t printer,
2757 const struct cl_page *pg);
2758 void cl_page_header_print(const struct lu_env *env, void *cookie,
2759 lu_printer_t printer,
2760 const struct cl_page *pg);
2761 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2762 struct cl_page *cl_page_top (struct cl_page *page);
2764 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2765 const struct lu_device_type *dtype);
2770 * Functions dealing with the ownership of page by io.
2774 int cl_page_own (const struct lu_env *env,
2775 struct cl_io *io, struct cl_page *page);
2776 int cl_page_own_try (const struct lu_env *env,
2777 struct cl_io *io, struct cl_page *page);
2778 void cl_page_assume (const struct lu_env *env,
2779 struct cl_io *io, struct cl_page *page);
2780 void cl_page_unassume (const struct lu_env *env,
2781 struct cl_io *io, struct cl_page *pg);
2782 void cl_page_disown (const struct lu_env *env,
2783 struct cl_io *io, struct cl_page *page);
2784 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2791 * Functions dealing with the preparation of a page for a transfer, and
2792 * tracking transfer state.
2795 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2796 struct cl_page *pg, enum cl_req_type crt);
2797 void cl_page_completion (const struct lu_env *env,
2798 struct cl_page *pg, enum cl_req_type crt, int ioret);
2799 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2800 enum cl_req_type crt);
2801 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2802 struct cl_page *pg, enum cl_req_type crt);
2803 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2805 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2806 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2807 struct cl_page *pg);
2813 * \name helper routines
2814 * Functions to discard, delete and export a cl_page.
2817 void cl_page_discard (const struct lu_env *env, struct cl_io *io,
2818 struct cl_page *pg);
2819 void cl_page_delete (const struct lu_env *env, struct cl_page *pg);
2820 int cl_page_is_vmlocked (const struct lu_env *env,
2821 const struct cl_page *pg);
2822 void cl_page_export (const struct lu_env *env,
2823 struct cl_page *pg, int uptodate);
2824 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2825 struct cl_page *page, pgoff_t *max_index);
2826 loff_t cl_offset (const struct cl_object *obj, pgoff_t idx);
2827 pgoff_t cl_index (const struct cl_object *obj, loff_t offset);
2828 int cl_page_size (const struct cl_object *obj);
2829 int cl_pages_prune (const struct lu_env *env, struct cl_object *obj);
2831 void cl_lock_print (const struct lu_env *env, void *cookie,
2832 lu_printer_t printer, const struct cl_lock *lock);
2833 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2834 lu_printer_t printer,
2835 const struct cl_lock_descr *descr);
2840 /** \defgroup cl_lock cl_lock
2843 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2844 const struct cl_lock_descr *need,
2845 const char *scope, const void *source);
2846 struct cl_lock *cl_lock_peek(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_request(const struct lu_env *env, struct cl_io *io,
2850 const struct cl_lock_descr *need,
2851 const char *scope, const void *source);
2852 struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env,
2853 struct cl_object *obj, pgoff_t index,
2854 struct cl_lock *except, int pending,
2856 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2857 const struct lu_device_type *dtype);
2859 void cl_lock_get (struct cl_lock *lock);
2860 void cl_lock_get_trust (struct cl_lock *lock);
2861 void cl_lock_put (const struct lu_env *env, struct cl_lock *lock);
2862 void cl_lock_hold_add (const struct lu_env *env, struct cl_lock *lock,
2863 const char *scope, const void *source);
2864 void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
2865 const char *scope, const void *source);
2866 void cl_lock_unhold (const struct lu_env *env, struct cl_lock *lock,
2867 const char *scope, const void *source);
2868 void cl_lock_release (const struct lu_env *env, struct cl_lock *lock,
2869 const char *scope, const void *source);
2870 void cl_lock_user_add (const struct lu_env *env, struct cl_lock *lock);
2871 void cl_lock_user_del (const struct lu_env *env, struct cl_lock *lock);
2873 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
2874 struct cl_lock *lock);
2875 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
2876 enum cl_lock_state state);
2877 int cl_lock_is_intransit(struct cl_lock *lock);
2879 int cl_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock,
2882 /** \name statemachine statemachine
2883 * Interface to lock state machine consists of 3 parts:
2885 * - "try" functions that attempt to effect a state transition. If state
2886 * transition is not possible right now (e.g., if it has to wait for some
2887 * asynchronous event to occur), these functions return
2888 * cl_lock_transition::CLO_WAIT.
2890 * - "non-try" functions that implement synchronous blocking interface on
2891 * top of non-blocking "try" functions. These functions repeatedly call
2892 * corresponding "try" versions, and if state transition is not possible
2893 * immediately, wait for lock state change.
2895 * - methods from cl_lock_operations, called by "try" functions. Lock can
2896 * be advanced to the target state only when all layers voted that they
2897 * are ready for this transition. "Try" functions call methods under lock
2898 * mutex. If a layer had to release a mutex, it re-acquires it and returns
2899 * cl_lock_transition::CLO_REPEAT, causing "try" function to call all
2902 * TRY NON-TRY METHOD FINAL STATE
2904 * cl_enqueue_try() cl_enqueue() cl_lock_operations::clo_enqueue() CLS_ENQUEUED
2906 * cl_wait_try() cl_wait() cl_lock_operations::clo_wait() CLS_HELD
2908 * cl_unuse_try() cl_unuse() cl_lock_operations::clo_unuse() CLS_CACHED
2910 * cl_use_try() NONE cl_lock_operations::clo_use() CLS_HELD
2914 int cl_enqueue (const struct lu_env *env, struct cl_lock *lock,
2915 struct cl_io *io, __u32 flags);
2916 int cl_wait (const struct lu_env *env, struct cl_lock *lock);
2917 void cl_unuse (const struct lu_env *env, struct cl_lock *lock);
2918 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
2919 struct cl_io *io, __u32 flags);
2920 int cl_unuse_try (const struct lu_env *env, struct cl_lock *lock);
2921 int cl_wait_try (const struct lu_env *env, struct cl_lock *lock);
2922 int cl_use_try (const struct lu_env *env, struct cl_lock *lock, int atomic);
2924 /** @} statemachine */
2926 void cl_lock_signal (const struct lu_env *env, struct cl_lock *lock);
2927 int cl_lock_state_wait (const struct lu_env *env, struct cl_lock *lock);
2928 void cl_lock_state_set (const struct lu_env *env, struct cl_lock *lock,
2929 enum cl_lock_state state);
2930 int cl_queue_match (const cfs_list_t *queue,
2931 const struct cl_lock_descr *need);
2933 void cl_lock_mutex_get (const struct lu_env *env, struct cl_lock *lock);
2934 int cl_lock_mutex_try (const struct lu_env *env, struct cl_lock *lock);
2935 void cl_lock_mutex_put (const struct lu_env *env, struct cl_lock *lock);
2936 int cl_lock_is_mutexed (struct cl_lock *lock);
2937 int cl_lock_nr_mutexed (const struct lu_env *env);
2938 int cl_lock_discard_pages(const struct lu_env *env, struct cl_lock *lock);
2939 int cl_lock_ext_match (const struct cl_lock_descr *has,
2940 const struct cl_lock_descr *need);
2941 int cl_lock_descr_match(const struct cl_lock_descr *has,
2942 const struct cl_lock_descr *need);
2943 int cl_lock_mode_match (enum cl_lock_mode has, enum cl_lock_mode need);
2944 int cl_lock_modify (const struct lu_env *env, struct cl_lock *lock,
2945 const struct cl_lock_descr *desc);
2947 void cl_lock_closure_init (const struct lu_env *env,
2948 struct cl_lock_closure *closure,
2949 struct cl_lock *origin, int wait);
2950 void cl_lock_closure_fini (struct cl_lock_closure *closure);
2951 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
2952 struct cl_lock_closure *closure);
2953 void cl_lock_disclosure (const struct lu_env *env,
2954 struct cl_lock_closure *closure);
2955 int cl_lock_enclosure (const struct lu_env *env, struct cl_lock *lock,
2956 struct cl_lock_closure *closure);
2958 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2959 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock);
2960 void cl_lock_error (const struct lu_env *env, struct cl_lock *lock, int error);
2961 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int wait);
2963 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock);
2967 /** \defgroup cl_io cl_io
2970 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2971 enum cl_io_type iot, struct cl_object *obj);
2972 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2973 enum cl_io_type iot, struct cl_object *obj);
2974 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2975 enum cl_io_type iot, loff_t pos, size_t count);
2976 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2978 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2979 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2980 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2981 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2982 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2983 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2984 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2985 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2986 struct cl_io_lock_link *link);
2987 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2988 struct cl_lock_descr *descr);
2989 int cl_io_read_page (const struct lu_env *env, struct cl_io *io,
2990 struct cl_page *page);
2991 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2992 enum cl_req_type iot, struct cl_2queue *queue);
2993 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2994 enum cl_req_type iot, struct cl_2queue *queue,
2996 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2997 struct cl_page_list *queue, int from, int to,
2999 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
3001 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
3002 struct cl_page_list *queue);
3003 int cl_io_is_going (const struct lu_env *env);
3006 * True, iff \a io is an O_APPEND write(2).
3008 static inline int cl_io_is_append(const struct cl_io *io)
3010 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
3013 static inline int cl_io_is_sync_write(const struct cl_io *io)
3015 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
3018 static inline int cl_io_is_mkwrite(const struct cl_io *io)
3020 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
3024 * True, iff \a io is a truncate(2).
3026 static inline int cl_io_is_trunc(const struct cl_io *io)
3028 return io->ci_type == CIT_SETATTR &&
3029 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
3032 struct cl_io *cl_io_top(struct cl_io *io);
3034 void cl_io_print(const struct lu_env *env, void *cookie,
3035 lu_printer_t printer, const struct cl_io *io);
3037 #define CL_IO_SLICE_CLEAN(foo_io, base) \
3039 typeof(foo_io) __foo_io = (foo_io); \
3041 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
3042 memset(&__foo_io->base + 1, 0, \
3043 (sizeof *__foo_io) - sizeof __foo_io->base); \
3048 /** \defgroup cl_page_list cl_page_list
3052 * Last page in the page list.
3054 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
3056 LASSERT(plist->pl_nr > 0);
3057 return cfs_list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
3060 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
3062 LASSERT(plist->pl_nr > 0);
3063 return cfs_list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
3067 * Iterate over pages in a page list.
3069 #define cl_page_list_for_each(page, list) \
3070 cfs_list_for_each_entry((page), &(list)->pl_pages, cp_batch)
3073 * Iterate over pages in a page list, taking possible removals into account.
3075 #define cl_page_list_for_each_safe(page, temp, list) \
3076 cfs_list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
3078 void cl_page_list_init (struct cl_page_list *plist);
3079 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
3080 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
3081 struct cl_page *page);
3082 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
3083 struct cl_page *page);
3084 void cl_page_list_splice (struct cl_page_list *list,
3085 struct cl_page_list *head);
3086 void cl_page_list_del (const struct lu_env *env,
3087 struct cl_page_list *plist, struct cl_page *page);
3088 void cl_page_list_disown (const struct lu_env *env,
3089 struct cl_io *io, struct cl_page_list *plist);
3090 int cl_page_list_own (const struct lu_env *env,
3091 struct cl_io *io, struct cl_page_list *plist);
3092 void cl_page_list_assume (const struct lu_env *env,
3093 struct cl_io *io, struct cl_page_list *plist);
3094 void cl_page_list_discard(const struct lu_env *env,
3095 struct cl_io *io, struct cl_page_list *plist);
3096 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
3098 void cl_2queue_init (struct cl_2queue *queue);
3099 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
3100 void cl_2queue_disown (const struct lu_env *env,
3101 struct cl_io *io, struct cl_2queue *queue);
3102 void cl_2queue_assume (const struct lu_env *env,
3103 struct cl_io *io, struct cl_2queue *queue);
3104 void cl_2queue_discard (const struct lu_env *env,
3105 struct cl_io *io, struct cl_2queue *queue);
3106 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
3107 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
3109 /** @} cl_page_list */
3111 /** \defgroup cl_req cl_req
3113 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
3114 enum cl_req_type crt, int nr_objects);
3116 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
3117 struct cl_page *page);
3118 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
3119 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
3120 void cl_req_attr_set (const struct lu_env *env, struct cl_req *req,
3121 struct cl_req_attr *attr, obd_valid flags);
3122 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
3124 /** \defgroup cl_sync_io cl_sync_io
3128 * Anchor for synchronous transfer. This is allocated on a stack by thread
3129 * doing synchronous transfer, and a pointer to this structure is set up in
3130 * every page submitted for transfer. Transfer completion routine updates
3131 * anchor and wakes up waiting thread when transfer is complete.
3134 /** number of pages yet to be transferred. */
3135 atomic_t csi_sync_nr;
3138 /** barrier of destroy this structure */
3139 atomic_t csi_barrier;
3140 /** completion to be signaled when transfer is complete. */
3141 wait_queue_head_t csi_waitq;
3144 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages);
3145 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
3146 struct cl_page_list *queue, struct cl_sync_io *anchor,
3148 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret);
3150 /** @} cl_sync_io */
3154 /** \defgroup cl_env cl_env
3156 * lu_env handling for a client.
3158 * lu_env is an environment within which lustre code executes. Its major part
3159 * is lu_context---a fast memory allocation mechanism that is used to conserve
3160 * precious kernel stack space. Originally lu_env was designed for a server,
3163 * - there is a (mostly) fixed number of threads, and
3165 * - call chains have no non-lustre portions inserted between lustre code.
3167 * On a client both these assumtpion fails, because every user thread can
3168 * potentially execute lustre code as part of a system call, and lustre calls
3169 * into VFS or MM that call back into lustre.
3171 * To deal with that, cl_env wrapper functions implement the following
3174 * - allocation and destruction of environment is amortized by caching no
3175 * longer used environments instead of destroying them;
3177 * - there is a notion of "current" environment, attached to the kernel
3178 * data structure representing current thread Top-level lustre code
3179 * allocates an environment and makes it current, then calls into
3180 * non-lustre code, that in turn calls lustre back. Low-level lustre
3181 * code thus called can fetch environment created by the top-level code
3182 * and reuse it, avoiding additional environment allocation.
3183 * Right now, three interfaces can attach the cl_env to running thread:
3186 * - cl_env_reexit(cl_env_reenter had to be called priorly)
3188 * \see lu_env, lu_context, lu_context_key
3191 struct cl_env_nest {
3196 struct lu_env *cl_env_peek (int *refcheck);
3197 struct lu_env *cl_env_get (int *refcheck);
3198 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
3199 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
3200 void cl_env_put (struct lu_env *env, int *refcheck);
3201 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
3202 void *cl_env_reenter (void);
3203 void cl_env_reexit (void *cookie);
3204 void cl_env_implant (struct lu_env *env, int *refcheck);
3205 void cl_env_unplant (struct lu_env *env, int *refcheck);
3206 unsigned cl_env_cache_purge(unsigned nr);
3207 struct lu_env *cl_env_percpu_get (void);
3208 void cl_env_percpu_put (struct lu_env *env);
3215 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
3216 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
3218 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
3219 struct lu_device_type *ldt,
3220 struct lu_device *next);
3223 int cl_global_init(void);
3224 void cl_global_fini(void);
3226 #endif /* _LINUX_CL_OBJECT_H */