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>
103 # include <linux/mutex.h>
104 # include <linux/radix-tree.h>
110 struct cl_device_operations;
113 struct cl_object_page_operations;
114 struct cl_object_lock_operations;
117 struct cl_page_slice;
119 struct cl_lock_slice;
121 struct cl_lock_operations;
122 struct cl_page_operations;
131 * Operations for each data device in the client stack.
133 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
135 struct cl_device_operations {
137 * Initialize cl_req. This method is called top-to-bottom on all
138 * devices in the stack to get them a chance to allocate layer-private
139 * data, and to attach them to the cl_req by calling
140 * cl_req_slice_add().
142 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
143 * \see ccc_req_init()
145 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
150 * Device in the client stack.
152 * \see ccc_device, lov_device, lovsub_device, osc_device
156 struct lu_device cd_lu_dev;
157 /** Per-layer operation vector. */
158 const struct cl_device_operations *cd_ops;
161 /** \addtogroup cl_object cl_object
164 * "Data attributes" of cl_object. Data attributes can be updated
165 * independently for a sub-object, and top-object's attributes are calculated
166 * from sub-objects' ones.
169 /** Object size, in bytes */
172 * Known minimal size, in bytes.
174 * This is only valid when at least one DLM lock is held.
177 /** Modification time. Measured in seconds since epoch. */
179 /** Access time. Measured in seconds since epoch. */
181 /** Change time. Measured in seconds since epoch. */
184 * Blocks allocated to this cl_object on the server file system.
186 * \todo XXX An interface for block size is needed.
190 * User identifier for quota purposes.
194 * Group identifier for quota purposes.
200 * Fields in cl_attr that are being set.
214 * Sub-class of lu_object with methods common for objects on the client
217 * cl_object: represents a regular file system object, both a file and a
218 * stripe. cl_object is based on lu_object: it is identified by a fid,
219 * layered, cached, hashed, and lrued. Important distinction with the server
220 * side, where md_object and dt_object are used, is that cl_object "fans out"
221 * at the lov/sns level: depending on the file layout, single file is
222 * represented as a set of "sub-objects" (stripes). At the implementation
223 * level, struct lov_object contains an array of cl_objects. Each sub-object
224 * is a full-fledged cl_object, having its fid, living in the lru and hash
227 * This leads to the next important difference with the server side: on the
228 * client, it's quite usual to have objects with the different sequence of
229 * layers. For example, typical top-object is composed of the following
235 * whereas its sub-objects are composed of
240 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
241 * track of the object-subobject relationship.
243 * Sub-objects are not cached independently: when top-object is about to
244 * be discarded from the memory, all its sub-objects are torn-down and
247 * \see ccc_object, lov_object, lovsub_object, osc_object
251 struct lu_object co_lu;
252 /** per-object-layer operations */
253 const struct cl_object_operations *co_ops;
254 /** offset of page slice in cl_page buffer */
259 * Description of the client object configuration. This is used for the
260 * creation of a new client object that is identified by a more state than
263 struct cl_object_conf {
265 struct lu_object_conf coc_lu;
268 * Object layout. This is consumed by lov.
270 struct lustre_md *coc_md;
272 * Description of particular stripe location in the
273 * cluster. This is consumed by osc.
275 struct lov_oinfo *coc_oinfo;
278 * VFS inode. This is consumed by vvp.
280 struct inode *coc_inode;
282 * Layout lock handle.
284 struct ldlm_lock *coc_lock;
286 * Operation to handle layout, OBJECT_CONF_XYZ.
292 /** configure layout, set up a new stripe, must be called while
293 * holding layout lock. */
295 /** invalidate the current stripe configuration due to losing
297 OBJECT_CONF_INVALIDATE = 1,
298 /** wait for old layout to go away so that new layout can be
304 * Operations implemented for each cl object layer.
306 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
308 struct cl_object_operations {
310 * Initialize page slice for this layer. Called top-to-bottom through
311 * every object layer when a new cl_page is instantiated. Layer
312 * keeping private per-page data, or requiring its own page operations
313 * vector should allocate these data here, and attach then to the page
314 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
317 * \retval NULL success.
319 * \retval ERR_PTR(errno) failure code.
321 * \retval valid-pointer pointer to already existing referenced page
322 * to be used instead of newly created.
324 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
325 struct cl_page *page, struct page *vmpage);
327 * Initialize lock slice for this layer. Called top-to-bottom through
328 * every object layer when a new cl_lock is instantiated. Layer
329 * keeping private per-lock data, or requiring its own lock operations
330 * vector should allocate these data here, and attach then to the lock
331 * by calling cl_lock_slice_add(). Mandatory.
333 int (*coo_lock_init)(const struct lu_env *env,
334 struct cl_object *obj, struct cl_lock *lock,
335 const struct cl_io *io);
337 * Initialize io state for a given layer.
339 * called top-to-bottom once per io existence to initialize io
340 * state. If layer wants to keep some state for this type of io, it
341 * has to embed struct cl_io_slice in lu_env::le_ses, and register
342 * slice with cl_io_slice_add(). It is guaranteed that all threads
343 * participating in this io share the same session.
345 int (*coo_io_init)(const struct lu_env *env,
346 struct cl_object *obj, struct cl_io *io);
348 * Fill portion of \a attr that this layer controls. This method is
349 * called top-to-bottom through all object layers.
351 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
353 * \return 0: to continue
354 * \return +ve: to stop iterating through layers (but 0 is returned
355 * from enclosing cl_object_attr_get())
356 * \return -ve: to signal error
358 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
359 struct cl_attr *attr);
363 * \a valid is a bitmask composed from enum #cl_attr_valid, and
364 * indicating what attributes are to be set.
366 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
368 * \return the same convention as for
369 * cl_object_operations::coo_attr_get() is used.
371 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj,
372 const struct cl_attr *attr, unsigned valid);
374 * Update object configuration. Called top-to-bottom to modify object
377 * XXX error conditions and handling.
379 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
380 const struct cl_object_conf *conf);
382 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
383 * object. Layers are supposed to fill parts of \a lvb that will be
384 * shipped to the glimpse originator as a glimpse result.
386 * \see ccc_object_glimpse(), lovsub_object_glimpse(),
387 * \see osc_object_glimpse()
389 int (*coo_glimpse)(const struct lu_env *env,
390 const struct cl_object *obj, struct ost_lvb *lvb);
394 * Extended header for client object.
396 struct cl_object_header {
397 /** Standard lu_object_header. cl_object::co_lu::lo_header points
399 struct lu_object_header coh_lu;
401 * \todo XXX move locks below to the separate cache-lines, they are
402 * mostly useless otherwise.
405 /** Lock protecting page tree. */
406 spinlock_t coh_page_guard;
407 /** Lock protecting lock list. */
408 spinlock_t coh_lock_guard;
410 /** Radix tree of cl_page's, cached for this object. */
411 struct radix_tree_root coh_tree;
412 /** # of pages in radix tree. */
413 unsigned long coh_pages;
414 /** List of cl_lock's granted for this object. */
415 cfs_list_t coh_locks;
418 * Parent object. It is assumed that an object has a well-defined
419 * parent, but not a well-defined child (there may be multiple
420 * sub-objects, for the same top-object). cl_object_header::coh_parent
421 * field allows certain code to be written generically, without
422 * limiting possible cl_object layouts unduly.
424 struct cl_object_header *coh_parent;
426 * Protects consistency between cl_attr of parent object and
427 * attributes of sub-objects, that the former is calculated ("merged")
430 * \todo XXX this can be read/write lock if needed.
432 spinlock_t coh_attr_guard;
434 * Size of cl_page + page slices
436 unsigned short coh_page_bufsize;
438 * Number of objects above this one: 0 for a top-object, 1 for its
441 unsigned char coh_nesting;
445 * Helper macro: iterate over all layers of the object \a obj, assigning every
446 * layer top-to-bottom to \a slice.
448 #define cl_object_for_each(slice, obj) \
449 cfs_list_for_each_entry((slice), \
450 &(obj)->co_lu.lo_header->loh_layers, \
453 * Helper macro: iterate over all layers of the object \a obj, assigning every
454 * layer bottom-to-top to \a slice.
456 #define cl_object_for_each_reverse(slice, obj) \
457 cfs_list_for_each_entry_reverse((slice), \
458 &(obj)->co_lu.lo_header->loh_layers, \
463 #define pgoff_t unsigned long
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 /** Logical page index within the object. Immutable after creation. */
728 /** List of slices. Immutable after creation. */
729 cfs_list_t cp_layers;
730 /** Parent page, NULL for top-level page. Immutable after creation. */
731 struct cl_page *cp_parent;
732 /** Lower-layer page. NULL for bottommost page. Immutable after
734 struct cl_page *cp_child;
736 * Page state. This field is const to avoid accidental update, it is
737 * modified only internally within cl_page.c. Protected by a VM lock.
739 const enum cl_page_state cp_state;
740 /** Linkage of pages within group. Protected by cl_page::cp_mutex. */
742 /** Mutex serializing membership of a page in a batch. */
743 struct mutex cp_mutex;
744 /** Linkage of pages within cl_req. */
745 cfs_list_t cp_flight;
746 /** Transfer error. */
750 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
753 enum cl_page_type cp_type;
756 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
757 * by sub-io. Protected by a VM lock.
759 struct cl_io *cp_owner;
761 * Debug information, the task is owning the page.
763 struct task_struct *cp_task;
765 * Owning IO request in cl_page_state::CPS_PAGEOUT and
766 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
767 * the top-level pages. Protected by a VM lock.
769 struct cl_req *cp_req;
770 /** List of references to this page, for debugging. */
771 struct lu_ref cp_reference;
772 /** Link to an object, for debugging. */
773 struct lu_ref_link cp_obj_ref;
774 /** Link to a queue, for debugging. */
775 struct lu_ref_link cp_queue_ref;
776 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */
778 /** Assigned if doing a sync_io */
779 struct cl_sync_io *cp_sync_io;
783 * Per-layer part of cl_page.
785 * \see ccc_page, lov_page, osc_page
787 struct cl_page_slice {
788 struct cl_page *cpl_page;
790 * Object slice corresponding to this page slice. Immutable after
793 struct cl_object *cpl_obj;
794 const struct cl_page_operations *cpl_ops;
795 /** Linkage into cl_page::cp_layers. Immutable after creation. */
796 cfs_list_t cpl_linkage;
800 * Lock mode. For the client extent locks.
802 * \warning: cl_lock_mode_match() assumes particular ordering here.
807 * Mode of a lock that protects no data, and exists only as a
808 * placeholder. This is used for `glimpse' requests. A phantom lock
809 * might get promoted to real lock at some point.
818 * Requested transfer type.
828 * Per-layer page operations.
830 * Methods taking an \a io argument are for the activity happening in the
831 * context of given \a io. Page is assumed to be owned by that io, except for
832 * the obvious cases (like cl_page_operations::cpo_own()).
834 * \see vvp_page_ops, lov_page_ops, osc_page_ops
836 struct cl_page_operations {
838 * cl_page<->struct page methods. Only one layer in the stack has to
839 * implement these. Current code assumes that this functionality is
840 * provided by the topmost layer, see cl_page_disown0() as an example.
844 * \return the underlying VM page. Optional.
846 struct page *(*cpo_vmpage)(const struct lu_env *env,
847 const struct cl_page_slice *slice);
849 * Called when \a io acquires this page into the exclusive
850 * ownership. When this method returns, it is guaranteed that the is
851 * not owned by other io, and no transfer is going on against
855 * \see vvp_page_own(), lov_page_own()
857 int (*cpo_own)(const struct lu_env *env,
858 const struct cl_page_slice *slice,
859 struct cl_io *io, int nonblock);
860 /** Called when ownership it yielded. Optional.
862 * \see cl_page_disown()
863 * \see vvp_page_disown()
865 void (*cpo_disown)(const struct lu_env *env,
866 const struct cl_page_slice *slice, struct cl_io *io);
868 * Called for a page that is already "owned" by \a io from VM point of
871 * \see cl_page_assume()
872 * \see vvp_page_assume(), lov_page_assume()
874 void (*cpo_assume)(const struct lu_env *env,
875 const struct cl_page_slice *slice, struct cl_io *io);
876 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
877 * bottom-to-top when IO releases a page without actually unlocking
880 * \see cl_page_unassume()
881 * \see vvp_page_unassume()
883 void (*cpo_unassume)(const struct lu_env *env,
884 const struct cl_page_slice *slice,
887 * Announces whether the page contains valid data or not by \a uptodate.
889 * \see cl_page_export()
890 * \see vvp_page_export()
892 void (*cpo_export)(const struct lu_env *env,
893 const struct cl_page_slice *slice, int uptodate);
895 * Unmaps page from the user space (if it is mapped).
897 * \see cl_page_unmap()
898 * \see vvp_page_unmap()
900 int (*cpo_unmap)(const struct lu_env *env,
901 const struct cl_page_slice *slice, struct cl_io *io);
903 * Checks whether underlying VM page is locked (in the suitable
904 * sense). Used for assertions.
906 * \retval -EBUSY: page is protected by a lock of a given mode;
907 * \retval -ENODATA: page is not protected by a lock;
908 * \retval 0: this layer cannot decide. (Should never happen.)
910 int (*cpo_is_vmlocked)(const struct lu_env *env,
911 const struct cl_page_slice *slice);
917 * Called when page is truncated from the object. Optional.
919 * \see cl_page_discard()
920 * \see vvp_page_discard(), osc_page_discard()
922 void (*cpo_discard)(const struct lu_env *env,
923 const struct cl_page_slice *slice,
926 * Called when page is removed from the cache, and is about to being
927 * destroyed. Optional.
929 * \see cl_page_delete()
930 * \see vvp_page_delete(), osc_page_delete()
932 void (*cpo_delete)(const struct lu_env *env,
933 const struct cl_page_slice *slice);
934 /** Destructor. Frees resources and slice itself. */
935 void (*cpo_fini)(const struct lu_env *env,
936 struct cl_page_slice *slice);
939 * Checks whether the page is protected by a cl_lock. This is a
940 * per-layer method, because certain layers have ways to check for the
941 * lock much more efficiently than through the generic locks scan, or
942 * implement locking mechanisms separate from cl_lock, e.g.,
943 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
944 * being canceled, or scheduled for cancellation as soon as the last
945 * user goes away, too.
947 * \retval -EBUSY: page is protected by a lock of a given mode;
948 * \retval -ENODATA: page is not protected by a lock;
949 * \retval 0: this layer cannot decide.
951 * \see cl_page_is_under_lock()
953 int (*cpo_is_under_lock)(const struct lu_env *env,
954 const struct cl_page_slice *slice,
958 * Optional debugging helper. Prints given page slice.
960 * \see cl_page_print()
962 int (*cpo_print)(const struct lu_env *env,
963 const struct cl_page_slice *slice,
964 void *cookie, lu_printer_t p);
968 * Transfer methods. See comment on cl_req for a description of
969 * transfer formation and life-cycle.
974 * Request type dependent vector of operations.
976 * Transfer operations depend on transfer mode (cl_req_type). To avoid
977 * passing transfer mode to each and every of these methods, and to
978 * avoid branching on request type inside of the methods, separate
979 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
980 * provided. That is, method invocation usually looks like
982 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
986 * Called when a page is submitted for a transfer as a part of
989 * \return 0 : page is eligible for submission;
990 * \return -EALREADY : skip this page;
991 * \return -ve : error.
993 * \see cl_page_prep()
995 int (*cpo_prep)(const struct lu_env *env,
996 const struct cl_page_slice *slice,
999 * Completion handler. This is guaranteed to be eventually
1000 * fired after cl_page_operations::cpo_prep() or
1001 * cl_page_operations::cpo_make_ready() call.
1003 * This method can be called in a non-blocking context. It is
1004 * guaranteed however, that the page involved and its object
1005 * are pinned in memory (and, hence, calling cl_page_put() is
1008 * \see cl_page_completion()
1010 void (*cpo_completion)(const struct lu_env *env,
1011 const struct cl_page_slice *slice,
1014 * Called when cached page is about to be added to the
1015 * cl_req as a part of req formation.
1017 * \return 0 : proceed with this page;
1018 * \return -EAGAIN : skip this page;
1019 * \return -ve : error.
1021 * \see cl_page_make_ready()
1023 int (*cpo_make_ready)(const struct lu_env *env,
1024 const struct cl_page_slice *slice);
1026 * Announce that this page is to be written out
1027 * opportunistically, that is, page is dirty, it is not
1028 * necessary to start write-out transfer right now, but
1029 * eventually page has to be written out.
1031 * Main caller of this is the write path (see
1032 * vvp_io_commit_write()), using this method to build a
1033 * "transfer cache" from which large transfers are then
1034 * constructed by the req-formation engine.
1036 * \todo XXX it would make sense to add page-age tracking
1037 * semantics here, and to oblige the req-formation engine to
1038 * send the page out not later than it is too old.
1040 * \see cl_page_cache_add()
1042 int (*cpo_cache_add)(const struct lu_env *env,
1043 const struct cl_page_slice *slice,
1047 * Tell transfer engine that only [to, from] part of a page should be
1050 * This is used for immediate transfers.
1052 * \todo XXX this is not very good interface. It would be much better
1053 * if all transfer parameters were supplied as arguments to
1054 * cl_io_operations::cio_submit() call, but it is not clear how to do
1055 * this for page queues.
1057 * \see cl_page_clip()
1059 void (*cpo_clip)(const struct lu_env *env,
1060 const struct cl_page_slice *slice,
1063 * \pre the page was queued for transferring.
1064 * \post page is removed from client's pending list, or -EBUSY
1065 * is returned if it has already been in transferring.
1067 * This is one of seldom page operation which is:
1068 * 0. called from top level;
1069 * 1. don't have vmpage locked;
1070 * 2. every layer should synchronize execution of its ->cpo_cancel()
1071 * with completion handlers. Osc uses client obd lock for this
1072 * purpose. Based on there is no vvp_page_cancel and
1073 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1075 * \see osc_page_cancel().
1077 int (*cpo_cancel)(const struct lu_env *env,
1078 const struct cl_page_slice *slice);
1080 * Write out a page by kernel. This is only called by ll_writepage
1083 * \see cl_page_flush()
1085 int (*cpo_flush)(const struct lu_env *env,
1086 const struct cl_page_slice *slice,
1092 * Helper macro, dumping detailed information about \a page into a log.
1094 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1096 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1098 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1099 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1100 CDEBUG(mask, format , ## __VA_ARGS__); \
1105 * Helper macro, dumping shorter information about \a page into a log.
1107 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1109 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1111 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1112 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1113 CDEBUG(mask, format , ## __VA_ARGS__); \
1117 static inline int __page_in_use(const struct cl_page *page, int refc)
1119 if (page->cp_type == CPT_CACHEABLE)
1121 LASSERT(cfs_atomic_read(&page->cp_ref) > 0);
1122 return (cfs_atomic_read(&page->cp_ref) > refc);
1124 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1125 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1129 /** \addtogroup cl_lock cl_lock
1133 * Extent locking on the client.
1137 * The locking model of the new client code is built around
1141 * data-type representing an extent lock on a regular file. cl_lock is a
1142 * layered object (much like cl_object and cl_page), it consists of a header
1143 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1144 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1146 * All locks for a given object are linked into cl_object_header::coh_locks
1147 * list (protected by cl_object_header::coh_lock_guard spin-lock) through
1148 * cl_lock::cll_linkage. Currently this list is not sorted in any way. We can
1149 * sort it in starting lock offset, or use altogether different data structure
1152 * Typical cl_lock consists of the two layers:
1154 * - vvp_lock (vvp specific data), and
1155 * - lov_lock (lov specific data).
1157 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1158 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1160 * - lovsub_lock, and
1163 * Each sub-lock is associated with a cl_object (representing stripe
1164 * sub-object or the file to which top-level cl_lock is associated to), and is
1165 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1166 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1167 * is different from cl_page, that doesn't fan out (there is usually exactly
1168 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1169 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1173 * cl_lock is reference counted. When reference counter drops to 0, lock is
1174 * placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING
1175 * lock is destroyed when last reference is released. Referencing between
1176 * top-lock and its sub-locks is described in the lov documentation module.
1180 * Also, cl_lock is a state machine. This requires some clarification. One of
1181 * the goals of client IO re-write was to make IO path non-blocking, or at
1182 * least to make it easier to make it non-blocking in the future. Here
1183 * `non-blocking' means that when a system call (read, write, truncate)
1184 * reaches a situation where it has to wait for a communication with the
1185 * server, it should --instead of waiting-- remember its current state and
1186 * switch to some other work. E.g,. instead of waiting for a lock enqueue,
1187 * client should proceed doing IO on the next stripe, etc. Obviously this is
1188 * rather radical redesign, and it is not planned to be fully implemented at
1189 * this time, instead we are putting some infrastructure in place, that would
1190 * make it easier to do asynchronous non-blocking IO easier in the
1191 * future. Specifically, where old locking code goes to sleep (waiting for
1192 * enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When
1193 * enqueue reply comes, its completion handler signals that lock state-machine
1194 * is ready to transit to the next state. There is some generic code in
1195 * cl_lock.c that sleeps, waiting for these signals. As a result, for users of
1196 * this cl_lock.c code, it looks like locking is done in normal blocking
1197 * fashion, and it the same time it is possible to switch to the non-blocking
1198 * locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c
1201 * For a description of state machine states and transitions see enum
1204 * There are two ways to restrict a set of states which lock might move to:
1206 * - placing a "hold" on a lock guarantees that lock will not be moved
1207 * into cl_lock_state::CLS_FREEING state until hold is released. Hold
1208 * can be only acquired on a lock that is not in
1209 * cl_lock_state::CLS_FREEING. All holds on a lock are counted in
1210 * cl_lock::cll_holds. Hold protects lock from cancellation and
1211 * destruction. Requests to cancel and destroy a lock on hold will be
1212 * recorded, but only honored when last hold on a lock is released;
1214 * - placing a "user" on a lock guarantees that lock will not leave
1215 * cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING,
1216 * cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of
1217 * states, once it enters this set. That is, if a user is added onto a
1218 * lock in a state not from this set, it doesn't immediately enforce
1219 * lock to move to this set, but once lock enters this set it will
1220 * remain there until all users are removed. Lock users are counted in
1221 * cl_lock::cll_users.
1223 * User is used to assure that lock is not canceled or destroyed while
1224 * it is being enqueued, or actively used by some IO.
1226 * Currently, a user always comes with a hold (cl_lock_invariant()
1227 * checks that a number of holds is not less than a number of users).
1231 * This is how lock state-machine operates. struct cl_lock contains a mutex
1232 * cl_lock::cll_guard that protects struct fields.
1234 * - mutex is taken, and cl_lock::cll_state is examined.
1236 * - for every state there are possible target states where lock can move
1237 * into. They are tried in order. Attempts to move into next state are
1238 * done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try().
1240 * - if the transition can be performed immediately, state is changed,
1241 * and mutex is released.
1243 * - if the transition requires blocking, _try() function returns
1244 * cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to
1245 * sleep, waiting for possibility of lock state change. It is woken
1246 * up when some event occurs, that makes lock state change possible
1247 * (e.g., the reception of the reply from the server), and repeats
1250 * Top-lock and sub-lock has separate mutexes and the latter has to be taken
1251 * first to avoid dead-lock.
1253 * To see an example of interaction of all these issues, take a look at the
1254 * lov_cl.c:lov_lock_enqueue() function. It is called as a part of
1255 * cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by
1256 * advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note
1257 * also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It
1258 * also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be
1259 * done in parallel, rather than one after another (this is used for glimpse
1260 * locks, that cannot dead-lock).
1262 * INTERFACE AND USAGE
1264 * struct cl_lock_operations provide a number of call-backs that are invoked
1265 * when events of interest occurs. Layers can intercept and handle glimpse,
1266 * blocking, cancel ASTs and a reception of the reply from the server.
1268 * One important difference with the old client locking model is that new
1269 * client has a representation for the top-lock, whereas in the old code only
1270 * sub-locks existed as real data structures and file-level locks are
1271 * represented by "request sets" that are created and destroyed on each and
1272 * every lock creation.
1274 * Top-locks are cached, and can be found in the cache by the system calls. It
1275 * is possible that top-lock is in cache, but some of its sub-locks were
1276 * canceled and destroyed. In that case top-lock has to be enqueued again
1277 * before it can be used.
1279 * Overall process of the locking during IO operation is as following:
1281 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1282 * is called on each layer. Responsibility of this method is to add locks,
1283 * needed by a given layer into cl_io.ci_lockset.
1285 * - once locks for all layers were collected, they are sorted to avoid
1286 * dead-locks (cl_io_locks_sort()), and enqueued.
1288 * - when all locks are acquired, IO is performed;
1290 * - locks are released into cache.
1292 * Striping introduces major additional complexity into locking. The
1293 * fundamental problem is that it is generally unsafe to actively use (hold)
1294 * two locks on the different OST servers at the same time, as this introduces
1295 * inter-server dependency and can lead to cascading evictions.
1297 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1298 * that no multi-stripe locks are taken (note that this design abandons POSIX
1299 * read/write semantics). Such pieces ideally can be executed concurrently. At
1300 * the same time, certain types of IO cannot be sub-divived, without
1301 * sacrificing correctness. This includes:
1303 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1306 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1308 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1309 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1310 * has to be held together with the usual lock on [offset, offset + count].
1312 * As multi-stripe locks have to be allowed, it makes sense to cache them, so
1313 * that, for example, a sequence of O_APPEND writes can proceed quickly
1314 * without going down to the individual stripes to do lock matching. On the
1315 * other hand, multi-stripe locks shouldn't be used by normal read/write
1316 * calls. To achieve this, every layer can implement ->clo_fits_into() method,
1317 * that is called by lock matching code (cl_lock_lookup()), and that can be
1318 * used to selectively disable matching of certain locks for certain IOs. For
1319 * exmaple, lov layer implements lov_lock_fits_into() that allow multi-stripe
1320 * locks to be matched only for truncates and O_APPEND writes.
1322 * Interaction with DLM
1324 * In the expected setup, cl_lock is ultimately backed up by a collection of
1325 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1326 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1327 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1328 * description of interaction with DLM.
1334 struct cl_lock_descr {
1335 /** Object this lock is granted for. */
1336 struct cl_object *cld_obj;
1337 /** Index of the first page protected by this lock. */
1339 /** Index of the last page (inclusive) protected by this lock. */
1341 /** Group ID, for group lock */
1344 enum cl_lock_mode cld_mode;
1346 * flags to enqueue lock. A combination of bit-flags from
1347 * enum cl_enq_flags.
1349 __u32 cld_enq_flags;
1352 #define DDESCR "%s(%d):[%lu, %lu]"
1353 #define PDESCR(descr) \
1354 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1355 (descr)->cld_start, (descr)->cld_end
1357 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1360 * Lock state-machine states.
1365 * Possible state transitions:
1367 * +------------------>NEW
1369 * | | cl_enqueue_try()
1371 * | cl_unuse_try() V
1372 * | +--------------QUEUING (*)
1374 * | | | cl_enqueue_try()
1376 * | | cl_unuse_try() V
1377 * sub-lock | +-------------ENQUEUED (*)
1379 * | | | cl_wait_try()
1384 * | | HELD<---------+
1386 * | | | | cl_use_try()
1387 * | | cl_unuse_try() | |
1390 * | +------------>INTRANSIT (D) <--+
1392 * | cl_unuse_try() | | cached lock found
1393 * | | | cl_use_try()
1396 * +------------------CACHED---------+
1405 * In states marked with (*) transition to the same state (i.e., a loop
1406 * in the diagram) is possible.
1408 * (R) is the point where Receive call-back is invoked: it allows layers
1409 * to handle arrival of lock reply.
1411 * (C) is the point where Cancellation call-back is invoked.
1413 * (D) is the transit state which means the lock is changing.
1415 * Transition to FREEING state is possible from any other state in the
1416 * diagram in case of unrecoverable error.
1420 * These states are for individual cl_lock object. Top-lock and its sub-locks
1421 * can be in the different states. Another way to say this is that we have
1422 * nested state-machines.
1424 * Separate QUEUING and ENQUEUED states are needed to support non-blocking
1425 * operation for locks with multiple sub-locks. Imagine lock on a file F, that
1426 * intersects 3 stripes S0, S1, and S2. To enqueue F client has to send
1427 * enqueue to S0, wait for its completion, then send enqueue for S1, wait for
1428 * its completion and at last enqueue lock for S2, and wait for its
1429 * completion. In that case, top-lock is in QUEUING state while S0, S1 are
1430 * handled, and is in ENQUEUED state after enqueue to S2 has been sent (note
1431 * that in this case, sub-locks move from state to state, and top-lock remains
1432 * in the same state).
1434 enum cl_lock_state {
1436 * Lock that wasn't yet enqueued
1440 * Enqueue is in progress, blocking for some intermediate interaction
1441 * with the other side.
1445 * Lock is fully enqueued, waiting for server to reply when it is
1450 * Lock granted, actively used by some IO.
1454 * This state is used to mark the lock is being used, or unused.
1455 * We need this state because the lock may have several sublocks,
1456 * so it's impossible to have an atomic way to bring all sublocks
1457 * into CLS_HELD state at use case, or all sublocks to CLS_CACHED
1459 * If a thread is referring to a lock, and it sees the lock is in this
1460 * state, it must wait for the lock.
1461 * See state diagram for details.
1465 * Lock granted, not used.
1469 * Lock is being destroyed.
1475 enum cl_lock_flags {
1477 * lock has been cancelled. This flag is never cleared once set (by
1478 * cl_lock_cancel0()).
1480 CLF_CANCELLED = 1 << 0,
1481 /** cancellation is pending for this lock. */
1482 CLF_CANCELPEND = 1 << 1,
1483 /** destruction is pending for this lock. */
1484 CLF_DOOMED = 1 << 2,
1485 /** from enqueue RPC reply upcall. */
1486 CLF_FROM_UPCALL= 1 << 3,
1492 * Lock closure is a collection of locks (both top-locks and sub-locks) that
1493 * might be updated in a result of an operation on a certain lock (which lock
1494 * this is a closure of).
1496 * Closures are needed to guarantee dead-lock freedom in the presence of
1498 * - nested state-machines (top-lock state-machine composed of sub-lock
1499 * state-machines), and
1501 * - shared sub-locks.
1503 * Specifically, many operations, such as lock enqueue, wait, unlock,
1504 * etc. start from a top-lock, and then operate on a sub-locks of this
1505 * top-lock, holding a top-lock mutex. When sub-lock state changes as a result
1506 * of such operation, this change has to be propagated to all top-locks that
1507 * share this sub-lock. Obviously, no natural lock ordering (e.g.,
1508 * top-to-bottom or bottom-to-top) captures this scenario, so try-locking has
1509 * to be used. Lock closure systematizes this try-and-repeat logic.
1511 struct cl_lock_closure {
1513 * Lock that is mutexed when closure construction is started. When
1514 * closure in is `wait' mode (cl_lock_closure::clc_wait), mutex on
1515 * origin is released before waiting.
1517 struct cl_lock *clc_origin;
1519 * List of enclosed locks, so far. Locks are linked here through
1520 * cl_lock::cll_inclosure.
1522 cfs_list_t clc_list;
1524 * True iff closure is in a `wait' mode. This determines what
1525 * cl_lock_enclosure() does when a lock L to be added to the closure
1526 * is currently mutexed by some other thread.
1528 * If cl_lock_closure::clc_wait is not set, then closure construction
1529 * fails with CLO_REPEAT immediately.
1531 * In wait mode, cl_lock_enclosure() waits until next attempt to build
1532 * a closure might succeed. To this end it releases an origin mutex
1533 * (cl_lock_closure::clc_origin), that has to be the only lock mutex
1534 * owned by the current thread, and then waits on L mutex (by grabbing
1535 * it and immediately releasing), before returning CLO_REPEAT to the
1539 /** Number of locks in the closure. */
1544 * Layered client lock.
1547 /** Reference counter. */
1548 cfs_atomic_t cll_ref;
1549 /** List of slices. Immutable after creation. */
1550 cfs_list_t cll_layers;
1552 * Linkage into cl_lock::cll_descr::cld_obj::coh_locks list. Protected
1553 * by cl_lock::cll_descr::cld_obj::coh_lock_guard.
1555 cfs_list_t cll_linkage;
1557 * Parameters of this lock. Protected by
1558 * cl_lock::cll_descr::cld_obj::coh_lock_guard nested within
1559 * cl_lock::cll_guard. Modified only on lock creation and in
1562 struct cl_lock_descr cll_descr;
1563 /** Protected by cl_lock::cll_guard. */
1564 enum cl_lock_state cll_state;
1565 /** signals state changes. */
1566 wait_queue_head_t cll_wq;
1568 * Recursive lock, most fields in cl_lock{} are protected by this.
1570 * Locking rules: this mutex is never held across network
1571 * communication, except when lock is being canceled.
1573 * Lock ordering: a mutex of a sub-lock is taken first, then a mutex
1574 * on a top-lock. Other direction is implemented through a
1575 * try-lock-repeat loop. Mutices of unrelated locks can be taken only
1578 * \see osc_lock_enqueue_wait(), lov_lock_cancel(), lov_sublock_wait().
1580 struct mutex cll_guard;
1581 struct task_struct *cll_guarder;
1585 * the owner for INTRANSIT state
1587 struct task_struct *cll_intransit_owner;
1590 * Number of holds on a lock. A hold prevents a lock from being
1591 * canceled and destroyed. Protected by cl_lock::cll_guard.
1593 * \see cl_lock_hold(), cl_lock_unhold(), cl_lock_release()
1597 * Number of lock users. Valid in cl_lock_state::CLS_HELD state
1598 * only. Lock user pins lock in CLS_HELD state. Protected by
1599 * cl_lock::cll_guard.
1601 * \see cl_wait(), cl_unuse().
1605 * Flag bit-mask. Values from enum cl_lock_flags. Updates are
1606 * protected by cl_lock::cll_guard.
1608 unsigned long cll_flags;
1610 * A linkage into a list of locks in a closure.
1612 * \see cl_lock_closure
1614 cfs_list_t cll_inclosure;
1616 * Confict lock at queuing time.
1618 struct cl_lock *cll_conflict;
1620 * A list of references to this lock, for debugging.
1622 struct lu_ref cll_reference;
1624 * A list of holds on this lock, for debugging.
1626 struct lu_ref cll_holders;
1628 * A reference for cl_lock::cll_descr::cld_obj. For debugging.
1630 struct lu_ref_link cll_obj_ref;
1631 #ifdef CONFIG_LOCKDEP
1632 /* "dep_map" name is assumed by lockdep.h macros. */
1633 struct lockdep_map dep_map;
1638 * Per-layer part of cl_lock
1640 * \see ccc_lock, lov_lock, lovsub_lock, osc_lock
1642 struct cl_lock_slice {
1643 struct cl_lock *cls_lock;
1644 /** Object slice corresponding to this lock slice. Immutable after
1646 struct cl_object *cls_obj;
1647 const struct cl_lock_operations *cls_ops;
1648 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1649 cfs_list_t cls_linkage;
1653 * Possible (non-error) return values of ->clo_{enqueue,wait,unlock}().
1655 * NOTE: lov_subresult() depends on ordering here.
1657 enum cl_lock_transition {
1658 /** operation cannot be completed immediately. Wait for state change. */
1660 /** operation had to release lock mutex, restart. */
1662 /** lower layer re-enqueued. */
1668 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1670 struct cl_lock_operations {
1672 * \name statemachine
1674 * State machine transitions. These 3 methods are called to transfer
1675 * lock from one state to another, as described in the commentary
1676 * above enum #cl_lock_state.
1678 * \retval 0 this layer has nothing more to do to before
1679 * transition to the target state happens;
1681 * \retval CLO_REPEAT method had to release and re-acquire cl_lock
1682 * mutex, repeat invocation of transition method
1683 * across all layers;
1685 * \retval CLO_WAIT this layer cannot move to the target state
1686 * immediately, as it has to wait for certain event
1687 * (e.g., the communication with the server). It
1688 * is guaranteed, that when the state transfer
1689 * becomes possible, cl_lock::cll_wq wait-queue
1690 * is signaled. Caller can wait for this event by
1691 * calling cl_lock_state_wait();
1693 * \retval -ve failure, abort state transition, move the lock
1694 * into cl_lock_state::CLS_FREEING state, and set
1695 * cl_lock::cll_error.
1697 * Once all layers voted to agree to transition (by returning 0), lock
1698 * is moved into corresponding target state. All state transition
1699 * methods are optional.
1703 * Attempts to enqueue the lock. Called top-to-bottom.
1705 * \see ccc_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1706 * \see osc_lock_enqueue()
1708 int (*clo_enqueue)(const struct lu_env *env,
1709 const struct cl_lock_slice *slice,
1710 struct cl_io *io, __u32 enqflags);
1712 * Attempts to wait for enqueue result. Called top-to-bottom.
1714 * \see ccc_lock_wait(), lov_lock_wait(), osc_lock_wait()
1716 int (*clo_wait)(const struct lu_env *env,
1717 const struct cl_lock_slice *slice);
1719 * Attempts to unlock the lock. Called bottom-to-top. In addition to
1720 * usual return values of lock state-machine methods, this can return
1721 * -ESTALE to indicate that lock cannot be returned to the cache, and
1722 * has to be re-initialized.
1723 * unuse is a one-shot operation, so it must NOT return CLO_WAIT.
1725 * \see ccc_lock_unuse(), lov_lock_unuse(), osc_lock_unuse()
1727 int (*clo_unuse)(const struct lu_env *env,
1728 const struct cl_lock_slice *slice);
1730 * Notifies layer that cached lock is started being used.
1732 * \pre lock->cll_state == CLS_CACHED
1734 * \see lov_lock_use(), osc_lock_use()
1736 int (*clo_use)(const struct lu_env *env,
1737 const struct cl_lock_slice *slice);
1738 /** @} statemachine */
1740 * A method invoked when lock state is changed (as a result of state
1741 * transition). This is used, for example, to track when the state of
1742 * a sub-lock changes, to propagate this change to the corresponding
1743 * top-lock. Optional
1745 * \see lovsub_lock_state()
1747 void (*clo_state)(const struct lu_env *env,
1748 const struct cl_lock_slice *slice,
1749 enum cl_lock_state st);
1751 * Returns true, iff given lock is suitable for the given io, idea
1752 * being, that there are certain "unsafe" locks, e.g., ones acquired
1753 * for O_APPEND writes, that we don't want to re-use for a normal
1754 * write, to avoid the danger of cascading evictions. Optional. Runs
1755 * under cl_object_header::coh_lock_guard.
1757 * XXX this should take more information about lock needed by
1758 * io. Probably lock description or something similar.
1760 * \see lov_fits_into()
1762 int (*clo_fits_into)(const struct lu_env *env,
1763 const struct cl_lock_slice *slice,
1764 const struct cl_lock_descr *need,
1765 const struct cl_io *io);
1768 * Asynchronous System Traps. All of then are optional, all are
1769 * executed bottom-to-top.
1774 * Cancellation callback. Cancel a lock voluntarily, or under
1775 * the request of server.
1777 void (*clo_cancel)(const struct lu_env *env,
1778 const struct cl_lock_slice *slice);
1780 * Lock weighting ast. Executed to estimate how precious this lock
1781 * is. The sum of results across all layers is used to determine
1782 * whether lock worth keeping in cache given present memory usage.
1784 * \see osc_lock_weigh(), vvp_lock_weigh(), lovsub_lock_weigh().
1786 unsigned long (*clo_weigh)(const struct lu_env *env,
1787 const struct cl_lock_slice *slice);
1791 * \see lovsub_lock_closure()
1793 int (*clo_closure)(const struct lu_env *env,
1794 const struct cl_lock_slice *slice,
1795 struct cl_lock_closure *closure);
1797 * Executed bottom-to-top when lock description changes (e.g., as a
1798 * result of server granting more generous lock than was requested).
1800 * \see lovsub_lock_modify()
1802 int (*clo_modify)(const struct lu_env *env,
1803 const struct cl_lock_slice *slice,
1804 const struct cl_lock_descr *updated);
1806 * Notifies layers (bottom-to-top) that lock is going to be
1807 * destroyed. Responsibility of layers is to prevent new references on
1808 * this lock from being acquired once this method returns.
1810 * This can be called multiple times due to the races.
1812 * \see cl_lock_delete()
1813 * \see osc_lock_delete(), lovsub_lock_delete()
1815 void (*clo_delete)(const struct lu_env *env,
1816 const struct cl_lock_slice *slice);
1818 * Destructor. Frees resources and the slice.
1820 * \see ccc_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1821 * \see osc_lock_fini()
1823 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1825 * Optional debugging helper. Prints given lock slice.
1827 int (*clo_print)(const struct lu_env *env,
1828 void *cookie, lu_printer_t p,
1829 const struct cl_lock_slice *slice);
1832 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1834 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1836 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1837 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1838 CDEBUG(mask, format , ## __VA_ARGS__); \
1842 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1846 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1852 /** \addtogroup cl_page_list cl_page_list
1853 * Page list used to perform collective operations on a group of pages.
1855 * Pages are added to the list one by one. cl_page_list acquires a reference
1856 * for every page in it. Page list is used to perform collective operations on
1859 * - submit pages for an immediate transfer,
1861 * - own pages on behalf of certain io (waiting for each page in turn),
1865 * When list is finalized, it releases references on all pages it still has.
1867 * \todo XXX concurrency control.
1871 struct cl_page_list {
1873 cfs_list_t pl_pages;
1874 struct task_struct *pl_owner;
1878 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1879 * contains an incoming page list and an outgoing page list.
1882 struct cl_page_list c2_qin;
1883 struct cl_page_list c2_qout;
1886 /** @} cl_page_list */
1888 /** \addtogroup cl_io cl_io
1893 * cl_io represents a high level I/O activity like
1894 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1897 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1898 * important distinction. We want to minimize number of calls to the allocator
1899 * in the fast path, e.g., in the case of read(2) when everything is cached:
1900 * client already owns the lock over region being read, and data are cached
1901 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1902 * per-layer io state is stored in the session, associated with the io, see
1903 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1904 * by using free-lists, see cl_env_get().
1906 * There is a small predefined number of possible io types, enumerated in enum
1909 * cl_io is a state machine, that can be advanced concurrently by the multiple
1910 * threads. It is up to these threads to control the concurrency and,
1911 * specifically, to detect when io is done, and its state can be safely
1914 * For read/write io overall execution plan is as following:
1916 * (0) initialize io state through all layers;
1918 * (1) loop: prepare chunk of work to do
1920 * (2) call all layers to collect locks they need to process current chunk
1922 * (3) sort all locks to avoid dead-locks, and acquire them
1924 * (4) process the chunk: call per-page methods
1925 * (cl_io_operations::cio_read_page() for read,
1926 * cl_io_operations::cio_prepare_write(),
1927 * cl_io_operations::cio_commit_write() for write)
1933 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1934 * address allocation efficiency issues mentioned above), and returns with the
1935 * special error condition from per-page method when current sub-io has to
1936 * block. This causes io loop to be repeated, and lov switches to the next
1937 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1942 /** read system call */
1944 /** write system call */
1946 /** truncate, utime system calls */
1949 * page fault handling
1953 * fsync system call handling
1954 * To write out a range of file
1958 * Miscellaneous io. This is used for occasional io activity that
1959 * doesn't fit into other types. Currently this is used for:
1961 * - cancellation of an extent lock. This io exists as a context
1962 * to write dirty pages from under the lock being canceled back
1965 * - VM induced page write-out. An io context for writing page out
1966 * for memory cleansing;
1968 * - glimpse. An io context to acquire glimpse lock.
1970 * - grouplock. An io context to acquire group lock.
1972 * CIT_MISC io is used simply as a context in which locks and pages
1973 * are manipulated. Such io has no internal "process", that is,
1974 * cl_io_loop() is never called for it.
1981 * States of cl_io state machine
1984 /** Not initialized. */
1988 /** IO iteration started. */
1992 /** Actual IO is in progress. */
1994 /** IO for the current iteration finished. */
1996 /** Locks released. */
1998 /** Iteration completed. */
2000 /** cl_io finalized. */
2005 * IO state private for a layer.
2007 * This is usually embedded into layer session data, rather than allocated
2010 * \see vvp_io, lov_io, osc_io, ccc_io
2012 struct cl_io_slice {
2013 struct cl_io *cis_io;
2014 /** corresponding object slice. Immutable after creation. */
2015 struct cl_object *cis_obj;
2016 /** io operations. Immutable after creation. */
2017 const struct cl_io_operations *cis_iop;
2019 * linkage into a list of all slices for a given cl_io, hanging off
2020 * cl_io::ci_layers. Immutable after creation.
2022 cfs_list_t cis_linkage;
2027 * Per-layer io operations.
2028 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
2030 struct cl_io_operations {
2032 * Vector of io state transition methods for every io type.
2034 * \see cl_page_operations::io
2038 * Prepare io iteration at a given layer.
2040 * Called top-to-bottom at the beginning of each iteration of
2041 * "io loop" (if it makes sense for this type of io). Here
2042 * layer selects what work it will do during this iteration.
2044 * \see cl_io_operations::cio_iter_fini()
2046 int (*cio_iter_init) (const struct lu_env *env,
2047 const struct cl_io_slice *slice);
2049 * Finalize io iteration.
2051 * Called bottom-to-top at the end of each iteration of "io
2052 * loop". Here layers can decide whether IO has to be
2055 * \see cl_io_operations::cio_iter_init()
2057 void (*cio_iter_fini) (const struct lu_env *env,
2058 const struct cl_io_slice *slice);
2060 * Collect locks for the current iteration of io.
2062 * Called top-to-bottom to collect all locks necessary for
2063 * this iteration. This methods shouldn't actually enqueue
2064 * anything, instead it should post a lock through
2065 * cl_io_lock_add(). Once all locks are collected, they are
2066 * sorted and enqueued in the proper order.
2068 int (*cio_lock) (const struct lu_env *env,
2069 const struct cl_io_slice *slice);
2071 * Finalize unlocking.
2073 * Called bottom-to-top to finish layer specific unlocking
2074 * functionality, after generic code released all locks
2075 * acquired by cl_io_operations::cio_lock().
2077 void (*cio_unlock)(const struct lu_env *env,
2078 const struct cl_io_slice *slice);
2080 * Start io iteration.
2082 * Once all locks are acquired, called top-to-bottom to
2083 * commence actual IO. In the current implementation,
2084 * top-level vvp_io_{read,write}_start() does all the work
2085 * synchronously by calling generic_file_*(), so other layers
2086 * are called when everything is done.
2088 int (*cio_start)(const struct lu_env *env,
2089 const struct cl_io_slice *slice);
2091 * Called top-to-bottom at the end of io loop. Here layer
2092 * might wait for an unfinished asynchronous io.
2094 void (*cio_end) (const struct lu_env *env,
2095 const struct cl_io_slice *slice);
2097 * Called bottom-to-top to notify layers that read/write IO
2098 * iteration finished, with \a nob bytes transferred.
2100 void (*cio_advance)(const struct lu_env *env,
2101 const struct cl_io_slice *slice,
2104 * Called once per io, bottom-to-top to release io resources.
2106 void (*cio_fini) (const struct lu_env *env,
2107 const struct cl_io_slice *slice);
2111 * Submit pages from \a queue->c2_qin for IO, and move
2112 * successfully submitted pages into \a queue->c2_qout. Return
2113 * non-zero if failed to submit even the single page. If
2114 * submission failed after some pages were moved into \a
2115 * queue->c2_qout, completion callback with non-zero ioret is
2118 int (*cio_submit)(const struct lu_env *env,
2119 const struct cl_io_slice *slice,
2120 enum cl_req_type crt,
2121 struct cl_2queue *queue);
2124 * Read missing page.
2126 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
2127 * method, when it hits not-up-to-date page in the range. Optional.
2129 * \pre io->ci_type == CIT_READ
2131 int (*cio_read_page)(const struct lu_env *env,
2132 const struct cl_io_slice *slice,
2133 const struct cl_page_slice *page);
2135 * Prepare write of a \a page. Called bottom-to-top by a top-level
2136 * cl_io_operations::op[CIT_WRITE]::cio_start() to prepare page for
2137 * get data from user-level buffer.
2139 * \pre io->ci_type == CIT_WRITE
2141 * \see vvp_io_prepare_write(), lov_io_prepare_write(),
2142 * osc_io_prepare_write().
2144 int (*cio_prepare_write)(const struct lu_env *env,
2145 const struct cl_io_slice *slice,
2146 const struct cl_page_slice *page,
2147 unsigned from, unsigned to);
2150 * \pre io->ci_type == CIT_WRITE
2152 * \see vvp_io_commit_write(), lov_io_commit_write(),
2153 * osc_io_commit_write().
2155 int (*cio_commit_write)(const struct lu_env *env,
2156 const struct cl_io_slice *slice,
2157 const struct cl_page_slice *page,
2158 unsigned from, unsigned to);
2160 * Optional debugging helper. Print given io slice.
2162 int (*cio_print)(const struct lu_env *env, void *cookie,
2163 lu_printer_t p, const struct cl_io_slice *slice);
2167 * Flags to lock enqueue procedure.
2172 * instruct server to not block, if conflicting lock is found. Instead
2173 * -EWOULDBLOCK is returned immediately.
2175 CEF_NONBLOCK = 0x00000001,
2177 * take lock asynchronously (out of order), as it cannot
2178 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
2180 CEF_ASYNC = 0x00000002,
2182 * tell the server to instruct (though a flag in the blocking ast) an
2183 * owner of the conflicting lock, that it can drop dirty pages
2184 * protected by this lock, without sending them to the server.
2186 CEF_DISCARD_DATA = 0x00000004,
2188 * tell the sub layers that it must be a `real' lock. This is used for
2189 * mmapped-buffer locks and glimpse locks that must be never converted
2190 * into lockless mode.
2192 * \see vvp_mmap_locks(), cl_glimpse_lock().
2194 CEF_MUST = 0x00000008,
2196 * tell the sub layers that never request a `real' lock. This flag is
2197 * not used currently.
2199 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
2200 * conversion policy: ci_lockreq describes generic information of lock
2201 * requirement for this IO, especially for locks which belong to the
2202 * object doing IO; however, lock itself may have precise requirements
2203 * that are described by the enqueue flags.
2205 CEF_NEVER = 0x00000010,
2207 * for async glimpse lock.
2209 CEF_AGL = 0x00000020,
2211 * mask of enq_flags.
2213 CEF_MASK = 0x0000003f,
2217 * Link between lock and io. Intermediate structure is needed, because the
2218 * same lock can be part of multiple io's simultaneously.
2220 struct cl_io_lock_link {
2221 /** linkage into one of cl_lockset lists. */
2222 cfs_list_t cill_linkage;
2223 struct cl_lock_descr cill_descr;
2224 struct cl_lock *cill_lock;
2225 /** optional destructor */
2226 void (*cill_fini)(const struct lu_env *env,
2227 struct cl_io_lock_link *link);
2231 * Lock-set represents a collection of locks, that io needs at a
2232 * time. Generally speaking, client tries to avoid holding multiple locks when
2235 * - holding extent locks over multiple ost's introduces the danger of
2236 * "cascading timeouts";
2238 * - holding multiple locks over the same ost is still dead-lock prone,
2239 * see comment in osc_lock_enqueue(),
2241 * but there are certain situations where this is unavoidable:
2243 * - O_APPEND writes have to take [0, EOF] lock for correctness;
2245 * - truncate has to take [new-size, EOF] lock for correctness;
2247 * - SNS has to take locks across full stripe for correctness;
2249 * - in the case when user level buffer, supplied to {read,write}(file0),
2250 * is a part of a memory mapped lustre file, client has to take a dlm
2251 * locks on file0, and all files that back up the buffer (or a part of
2252 * the buffer, that is being processed in the current chunk, in any
2253 * case, there are situations where at least 2 locks are necessary).
2255 * In such cases we at least try to take locks in the same consistent
2256 * order. To this end, all locks are first collected, then sorted, and then
2260 /** locks to be acquired. */
2261 cfs_list_t cls_todo;
2262 /** locks currently being processed. */
2263 cfs_list_t cls_curr;
2264 /** locks acquired. */
2265 cfs_list_t cls_done;
2269 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
2270 * but 'req' is always to be thought as 'request' :-)
2272 enum cl_io_lock_dmd {
2273 /** Always lock data (e.g., O_APPEND). */
2275 /** Layers are free to decide between local and global locking. */
2277 /** Never lock: there is no cache (e.g., liblustre). */
2281 enum cl_fsync_mode {
2282 /** start writeback, do not wait for them to finish */
2284 /** start writeback and wait for them to finish */
2286 /** discard all of dirty pages in a specific file range */
2287 CL_FSYNC_DISCARD = 2,
2288 /** start writeback and make sure they have reached storage before
2289 * return. OST_SYNC RPC must be issued and finished */
2293 struct cl_io_rw_common {
2303 * cl_io is shared by all threads participating in this IO (in current
2304 * implementation only one thread advances IO, but parallel IO design and
2305 * concurrent copy_*_user() require multiple threads acting on the same IO. It
2306 * is up to these threads to serialize their activities, including updates to
2307 * mutable cl_io fields.
2310 /** type of this IO. Immutable after creation. */
2311 enum cl_io_type ci_type;
2312 /** current state of cl_io state machine. */
2313 enum cl_io_state ci_state;
2314 /** main object this io is against. Immutable after creation. */
2315 struct cl_object *ci_obj;
2317 * Upper layer io, of which this io is a part of. Immutable after
2320 struct cl_io *ci_parent;
2321 /** List of slices. Immutable after creation. */
2322 cfs_list_t ci_layers;
2323 /** list of locks (to be) acquired by this io. */
2324 struct cl_lockset ci_lockset;
2325 /** lock requirements, this is just a help info for sublayers. */
2326 enum cl_io_lock_dmd ci_lockreq;
2329 struct cl_io_rw_common rd;
2332 struct cl_io_rw_common wr;
2336 struct cl_io_rw_common ci_rw;
2337 struct cl_setattr_io {
2338 struct ost_lvb sa_attr;
2339 unsigned int sa_valid;
2340 struct obd_capa *sa_capa;
2342 struct cl_fault_io {
2343 /** page index within file. */
2345 /** bytes valid byte on a faulted page. */
2347 /** writable page? for nopage() only */
2349 /** page of an executable? */
2351 /** page_mkwrite() */
2353 /** resulting page */
2354 struct cl_page *ft_page;
2356 struct cl_fsync_io {
2359 struct obd_capa *fi_capa;
2360 /** file system level fid */
2361 struct lu_fid *fi_fid;
2362 enum cl_fsync_mode fi_mode;
2363 /* how many pages were written/discarded */
2364 unsigned int fi_nr_written;
2367 struct cl_2queue ci_queue;
2370 unsigned int ci_continue:1,
2372 * This io has held grouplock, to inform sublayers that
2373 * don't do lockless i/o.
2377 * The whole IO need to be restarted because layout has been changed
2381 * to not refresh layout - the IO issuer knows that the layout won't
2382 * change(page operations, layout change causes all page to be
2383 * discarded), or it doesn't matter if it changes(sync).
2387 * Check if layout changed after the IO finishes. Mainly for HSM
2388 * requirement. If IO occurs to openning files, it doesn't need to
2389 * verify layout because HSM won't release openning files.
2390 * Right now, only two opertaions need to verify layout: glimpse
2395 * file is released, restore has to to be triggered by vvp layer
2397 ci_restore_needed:1,
2403 * Number of pages owned by this IO. For invariant checking.
2405 unsigned ci_owned_nr;
2410 /** \addtogroup cl_req cl_req
2415 * There are two possible modes of transfer initiation on the client:
2417 * - immediate transfer: this is started when a high level io wants a page
2418 * or a collection of pages to be transferred right away. Examples:
2419 * read-ahead, synchronous read in the case of non-page aligned write,
2420 * page write-out as a part of extent lock cancellation, page write-out
2421 * as a part of memory cleansing. Immediate transfer can be both
2422 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
2424 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
2425 * when io wants to transfer a page to the server some time later, when
2426 * it can be done efficiently. Example: pages dirtied by the write(2)
2429 * In any case, transfer takes place in the form of a cl_req, which is a
2430 * representation for a network RPC.
2432 * Pages queued for an opportunistic transfer are cached until it is decided
2433 * that efficient RPC can be composed of them. This decision is made by "a
2434 * req-formation engine", currently implemented as a part of osc
2435 * layer. Req-formation depends on many factors: the size of the resulting
2436 * RPC, whether or not multi-object RPCs are supported by the server,
2437 * max-rpc-in-flight limitations, size of the dirty cache, etc.
2439 * For the immediate transfer io submits a cl_page_list, that req-formation
2440 * engine slices into cl_req's, possibly adding cached pages to some of
2441 * the resulting req's.
2443 * Whenever a page from cl_page_list is added to a newly constructed req, its
2444 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
2445 * page state is atomically changed from cl_page_state::CPS_OWNED to
2446 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
2447 * is zeroed, and cl_page::cp_req is set to the
2448 * req. cl_page_operations::cpo_prep() method at the particular layer might
2449 * return -EALREADY to indicate that it does not need to submit this page
2450 * at all. This is possible, for example, if page, submitted for read,
2451 * became up-to-date in the meantime; and for write, the page don't have
2452 * dirty bit marked. \see cl_io_submit_rw()
2454 * Whenever a cached page is added to a newly constructed req, its
2455 * cl_page_operations::cpo_make_ready() layer methods are called. At that
2456 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
2457 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
2458 * req. cl_page_operations::cpo_make_ready() method at the particular layer
2459 * might return -EAGAIN to indicate that this page is not eligible for the
2460 * transfer right now.
2464 * Plan is to divide transfers into "priority bands" (indicated when
2465 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
2466 * and allow glueing of cached pages to immediate transfers only within single
2467 * band. This would make high priority transfers (like lock cancellation or
2468 * memory pressure induced write-out) really high priority.
2473 * Per-transfer attributes.
2475 struct cl_req_attr {
2476 /** Generic attributes for the server consumption. */
2477 struct obdo *cra_oa;
2479 struct obd_capa *cra_capa;
2481 char cra_jobid[JOBSTATS_JOBID_SIZE];
2485 * Transfer request operations definable at every layer.
2487 * Concurrency: transfer formation engine synchronizes calls to all transfer
2490 struct cl_req_operations {
2492 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
2493 * complete (all pages are added).
2495 * \see osc_req_prep()
2497 int (*cro_prep)(const struct lu_env *env,
2498 const struct cl_req_slice *slice);
2500 * Called top-to-bottom to fill in \a oa fields. This is called twice
2501 * with different flags, see bug 10150 and osc_build_req().
2503 * \param obj an object from cl_req which attributes are to be set in
2506 * \param oa struct obdo where attributes are placed
2508 * \param flags \a oa fields to be filled.
2510 void (*cro_attr_set)(const struct lu_env *env,
2511 const struct cl_req_slice *slice,
2512 const struct cl_object *obj,
2513 struct cl_req_attr *attr, obd_valid flags);
2515 * Called top-to-bottom from cl_req_completion() to notify layers that
2516 * transfer completed. Has to free all state allocated by
2517 * cl_device_operations::cdo_req_init().
2519 void (*cro_completion)(const struct lu_env *env,
2520 const struct cl_req_slice *slice, int ioret);
2524 * A per-object state that (potentially multi-object) transfer request keeps.
2527 /** object itself */
2528 struct cl_object *ro_obj;
2529 /** reference to cl_req_obj::ro_obj. For debugging. */
2530 struct lu_ref_link ro_obj_ref;
2531 /* something else? Number of pages for a given object? */
2537 * Transfer requests are not reference counted, because IO sub-system owns
2538 * them exclusively and knows when to free them.
2542 * cl_req is created by cl_req_alloc() that calls
2543 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2544 * state in every layer.
2546 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2547 * contains pages for.
2549 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2550 * called top-to-bottom. At that point layers can modify req, let it pass, or
2551 * deny it completely. This is to support things like SNS that have transfer
2552 * ordering requirements invisible to the individual req-formation engine.
2554 * On transfer completion (or transfer timeout, or failure to initiate the
2555 * transfer of an allocated req), cl_req_operations::cro_completion() method
2556 * is called, after execution of cl_page_operations::cpo_completion() of all
2560 enum cl_req_type crq_type;
2561 /** A list of pages being transfered */
2562 cfs_list_t crq_pages;
2563 /** Number of pages in cl_req::crq_pages */
2564 unsigned crq_nrpages;
2565 /** An array of objects which pages are in ->crq_pages */
2566 struct cl_req_obj *crq_o;
2567 /** Number of elements in cl_req::crq_objs[] */
2568 unsigned crq_nrobjs;
2569 cfs_list_t crq_layers;
2573 * Per-layer state for request.
2575 struct cl_req_slice {
2576 struct cl_req *crs_req;
2577 struct cl_device *crs_dev;
2578 cfs_list_t crs_linkage;
2579 const struct cl_req_operations *crs_ops;
2584 enum cache_stats_item {
2585 /** how many cache lookups were performed */
2587 /** how many times cache lookup resulted in a hit */
2589 /** how many entities are in the cache right now */
2591 /** how many entities in the cache are actively used (and cannot be
2592 * evicted) right now */
2594 /** how many entities were created at all */
2599 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2602 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2604 struct cache_stats {
2605 const char *cs_name;
2606 cfs_atomic_t cs_stats[CS_NR];
2609 /** These are not exported so far */
2610 void cache_stats_init (struct cache_stats *cs, const char *name);
2611 int cache_stats_print(const struct cache_stats *cs,
2612 char *page, int count, int header);
2615 * Client-side site. This represents particular client stack. "Global"
2616 * variables should (directly or indirectly) be added here to allow multiple
2617 * clients to co-exist in the single address space.
2620 struct lu_site cs_lu;
2622 * Statistical counters. Atomics do not scale, something better like
2623 * per-cpu counters is needed.
2625 * These are exported as /proc/fs/lustre/llite/.../site
2627 * When interpreting keep in mind that both sub-locks (and sub-pages)
2628 * and top-locks (and top-pages) are accounted here.
2630 struct cache_stats cs_pages;
2631 struct cache_stats cs_locks;
2632 cfs_atomic_t cs_pages_state[CPS_NR];
2633 cfs_atomic_t cs_locks_state[CLS_NR];
2636 int cl_site_init (struct cl_site *s, struct cl_device *top);
2637 void cl_site_fini (struct cl_site *s);
2638 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2641 * Output client site statistical counters into a buffer. Suitable for
2642 * ll_rd_*()-style functions.
2644 int cl_site_stats_print(const struct cl_site *s, char *page, int count);
2649 * Type conversion and accessory functions.
2653 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2655 return container_of(site, struct cl_site, cs_lu);
2658 static inline int lu_device_is_cl(const struct lu_device *d)
2660 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2663 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2665 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2666 return container_of0(d, struct cl_device, cd_lu_dev);
2669 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2671 return &d->cd_lu_dev;
2674 static inline struct cl_object *lu2cl(const struct lu_object *o)
2676 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2677 return container_of0(o, struct cl_object, co_lu);
2680 static inline const struct cl_object_conf *
2681 lu2cl_conf(const struct lu_object_conf *conf)
2683 return container_of0(conf, struct cl_object_conf, coc_lu);
2686 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2688 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2691 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2693 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2694 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2697 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2699 return container_of0(h, struct cl_object_header, coh_lu);
2702 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2704 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2708 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2710 return luh2coh(obj->co_lu.lo_header);
2713 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2715 return lu_device_init(&d->cd_lu_dev, t);
2718 static inline void cl_device_fini(struct cl_device *d)
2720 lu_device_fini(&d->cd_lu_dev);
2723 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2724 struct cl_object *obj,
2725 const struct cl_page_operations *ops);
2726 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2727 struct cl_object *obj,
2728 const struct cl_lock_operations *ops);
2729 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2730 struct cl_object *obj, const struct cl_io_operations *ops);
2731 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2732 struct cl_device *dev,
2733 const struct cl_req_operations *ops);
2736 /** \defgroup cl_object cl_object
2738 struct cl_object *cl_object_top (struct cl_object *o);
2739 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2740 const struct lu_fid *fid,
2741 const struct cl_object_conf *c);
2743 int cl_object_header_init(struct cl_object_header *h);
2744 void cl_object_header_fini(struct cl_object_header *h);
2745 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2746 void cl_object_get (struct cl_object *o);
2747 void cl_object_attr_lock (struct cl_object *o);
2748 void cl_object_attr_unlock(struct cl_object *o);
2749 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2750 struct cl_attr *attr);
2751 int cl_object_attr_set (const struct lu_env *env, struct cl_object *obj,
2752 const struct cl_attr *attr, unsigned valid);
2753 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2754 struct ost_lvb *lvb);
2755 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2756 const struct cl_object_conf *conf);
2757 void cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2758 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2759 int cl_object_has_locks (struct cl_object *obj);
2762 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2764 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2766 return cl_object_header(o0) == cl_object_header(o1);
2769 static inline void cl_object_page_init(struct cl_object *clob, int size)
2771 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2772 cl_object_header(clob)->coh_page_bufsize += ALIGN(size, 8);
2775 static inline void *cl_object_page_slice(struct cl_object *clob,
2776 struct cl_page *page)
2778 return (void *)((char *)page + clob->co_slice_off);
2783 /** \defgroup cl_page cl_page
2792 /* callback of cl_page_gang_lookup() */
2793 typedef int (*cl_page_gang_cb_t) (const struct lu_env *, struct cl_io *,
2794 struct cl_page *, void *);
2795 int cl_page_gang_lookup (const struct lu_env *env,
2796 struct cl_object *obj,
2798 pgoff_t start, pgoff_t end,
2799 cl_page_gang_cb_t cb, void *cbdata);
2800 struct cl_page *cl_page_lookup (struct cl_object_header *hdr,
2802 struct cl_page *cl_page_find (const struct lu_env *env,
2803 struct cl_object *obj,
2804 pgoff_t idx, struct page *vmpage,
2805 enum cl_page_type type);
2806 struct cl_page *cl_page_find_sub (const struct lu_env *env,
2807 struct cl_object *obj,
2808 pgoff_t idx, struct page *vmpage,
2809 struct cl_page *parent);
2810 void cl_page_get (struct cl_page *page);
2811 void cl_page_put (const struct lu_env *env,
2812 struct cl_page *page);
2813 void cl_page_print (const struct lu_env *env, void *cookie,
2814 lu_printer_t printer,
2815 const struct cl_page *pg);
2816 void cl_page_header_print(const struct lu_env *env, void *cookie,
2817 lu_printer_t printer,
2818 const struct cl_page *pg);
2819 struct page *cl_page_vmpage (const struct lu_env *env,
2820 struct cl_page *page);
2821 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2822 struct cl_page *cl_page_top (struct cl_page *page);
2824 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2825 const struct lu_device_type *dtype);
2830 * Functions dealing with the ownership of page by io.
2834 int cl_page_own (const struct lu_env *env,
2835 struct cl_io *io, struct cl_page *page);
2836 int cl_page_own_try (const struct lu_env *env,
2837 struct cl_io *io, struct cl_page *page);
2838 void cl_page_assume (const struct lu_env *env,
2839 struct cl_io *io, struct cl_page *page);
2840 void cl_page_unassume (const struct lu_env *env,
2841 struct cl_io *io, struct cl_page *pg);
2842 void cl_page_disown (const struct lu_env *env,
2843 struct cl_io *io, struct cl_page *page);
2844 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2851 * Functions dealing with the preparation of a page for a transfer, and
2852 * tracking transfer state.
2855 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2856 struct cl_page *pg, enum cl_req_type crt);
2857 void cl_page_completion (const struct lu_env *env,
2858 struct cl_page *pg, enum cl_req_type crt, int ioret);
2859 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2860 enum cl_req_type crt);
2861 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2862 struct cl_page *pg, enum cl_req_type crt);
2863 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2865 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2866 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2867 struct cl_page *pg);
2873 * \name helper routines
2874 * Functions to discard, delete and export a cl_page.
2877 void cl_page_discard (const struct lu_env *env, struct cl_io *io,
2878 struct cl_page *pg);
2879 void cl_page_delete (const struct lu_env *env, struct cl_page *pg);
2880 int cl_page_unmap (const struct lu_env *env, struct cl_io *io,
2881 struct cl_page *pg);
2882 int cl_page_is_vmlocked (const struct lu_env *env,
2883 const struct cl_page *pg);
2884 void cl_page_export (const struct lu_env *env,
2885 struct cl_page *pg, int uptodate);
2886 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2887 struct cl_page *page);
2888 loff_t cl_offset (const struct cl_object *obj, pgoff_t idx);
2889 pgoff_t cl_index (const struct cl_object *obj, loff_t offset);
2890 int cl_page_size (const struct cl_object *obj);
2891 int cl_pages_prune (const struct lu_env *env, struct cl_object *obj);
2893 void cl_lock_print (const struct lu_env *env, void *cookie,
2894 lu_printer_t printer, const struct cl_lock *lock);
2895 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2896 lu_printer_t printer,
2897 const struct cl_lock_descr *descr);
2902 /** \defgroup cl_lock cl_lock
2905 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2906 const struct cl_lock_descr *need,
2907 const char *scope, const void *source);
2908 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
2909 const struct cl_lock_descr *need,
2910 const char *scope, const void *source);
2911 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2912 const struct cl_lock_descr *need,
2913 const char *scope, const void *source);
2914 struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env,
2915 struct cl_object *obj, pgoff_t index,
2916 struct cl_lock *except, int pending,
2918 static inline struct cl_lock *cl_lock_at_page(const struct lu_env *env,
2919 struct cl_object *obj,
2920 struct cl_page *page,
2921 struct cl_lock *except,
2922 int pending, int canceld)
2924 LASSERT(cl_object_header(obj) == cl_object_header(page->cp_obj));
2925 return cl_lock_at_pgoff(env, obj, page->cp_index, except,
2929 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2930 const struct lu_device_type *dtype);
2932 void cl_lock_get (struct cl_lock *lock);
2933 void cl_lock_get_trust (struct cl_lock *lock);
2934 void cl_lock_put (const struct lu_env *env, struct cl_lock *lock);
2935 void cl_lock_hold_add (const struct lu_env *env, struct cl_lock *lock,
2936 const char *scope, const void *source);
2937 void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
2938 const char *scope, const void *source);
2939 void cl_lock_unhold (const struct lu_env *env, struct cl_lock *lock,
2940 const char *scope, const void *source);
2941 void cl_lock_release (const struct lu_env *env, struct cl_lock *lock,
2942 const char *scope, const void *source);
2943 void cl_lock_user_add (const struct lu_env *env, struct cl_lock *lock);
2944 void cl_lock_user_del (const struct lu_env *env, struct cl_lock *lock);
2946 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
2947 struct cl_lock *lock);
2948 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
2949 enum cl_lock_state state);
2950 int cl_lock_is_intransit(struct cl_lock *lock);
2952 int cl_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock,
2955 /** \name statemachine statemachine
2956 * Interface to lock state machine consists of 3 parts:
2958 * - "try" functions that attempt to effect a state transition. If state
2959 * transition is not possible right now (e.g., if it has to wait for some
2960 * asynchronous event to occur), these functions return
2961 * cl_lock_transition::CLO_WAIT.
2963 * - "non-try" functions that implement synchronous blocking interface on
2964 * top of non-blocking "try" functions. These functions repeatedly call
2965 * corresponding "try" versions, and if state transition is not possible
2966 * immediately, wait for lock state change.
2968 * - methods from cl_lock_operations, called by "try" functions. Lock can
2969 * be advanced to the target state only when all layers voted that they
2970 * are ready for this transition. "Try" functions call methods under lock
2971 * mutex. If a layer had to release a mutex, it re-acquires it and returns
2972 * cl_lock_transition::CLO_REPEAT, causing "try" function to call all
2975 * TRY NON-TRY METHOD FINAL STATE
2977 * cl_enqueue_try() cl_enqueue() cl_lock_operations::clo_enqueue() CLS_ENQUEUED
2979 * cl_wait_try() cl_wait() cl_lock_operations::clo_wait() CLS_HELD
2981 * cl_unuse_try() cl_unuse() cl_lock_operations::clo_unuse() CLS_CACHED
2983 * cl_use_try() NONE cl_lock_operations::clo_use() CLS_HELD
2987 int cl_enqueue (const struct lu_env *env, struct cl_lock *lock,
2988 struct cl_io *io, __u32 flags);
2989 int cl_wait (const struct lu_env *env, struct cl_lock *lock);
2990 void cl_unuse (const struct lu_env *env, struct cl_lock *lock);
2991 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
2992 struct cl_io *io, __u32 flags);
2993 int cl_unuse_try (const struct lu_env *env, struct cl_lock *lock);
2994 int cl_wait_try (const struct lu_env *env, struct cl_lock *lock);
2995 int cl_use_try (const struct lu_env *env, struct cl_lock *lock, int atomic);
2997 /** @} statemachine */
2999 void cl_lock_signal (const struct lu_env *env, struct cl_lock *lock);
3000 int cl_lock_state_wait (const struct lu_env *env, struct cl_lock *lock);
3001 void cl_lock_state_set (const struct lu_env *env, struct cl_lock *lock,
3002 enum cl_lock_state state);
3003 int cl_queue_match (const cfs_list_t *queue,
3004 const struct cl_lock_descr *need);
3006 void cl_lock_mutex_get (const struct lu_env *env, struct cl_lock *lock);
3007 int cl_lock_mutex_try (const struct lu_env *env, struct cl_lock *lock);
3008 void cl_lock_mutex_put (const struct lu_env *env, struct cl_lock *lock);
3009 int cl_lock_is_mutexed (struct cl_lock *lock);
3010 int cl_lock_nr_mutexed (const struct lu_env *env);
3011 int cl_lock_discard_pages(const struct lu_env *env, struct cl_lock *lock);
3012 int cl_lock_ext_match (const struct cl_lock_descr *has,
3013 const struct cl_lock_descr *need);
3014 int cl_lock_descr_match(const struct cl_lock_descr *has,
3015 const struct cl_lock_descr *need);
3016 int cl_lock_mode_match (enum cl_lock_mode has, enum cl_lock_mode need);
3017 int cl_lock_modify (const struct lu_env *env, struct cl_lock *lock,
3018 const struct cl_lock_descr *desc);
3020 void cl_lock_closure_init (const struct lu_env *env,
3021 struct cl_lock_closure *closure,
3022 struct cl_lock *origin, int wait);
3023 void cl_lock_closure_fini (struct cl_lock_closure *closure);
3024 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
3025 struct cl_lock_closure *closure);
3026 void cl_lock_disclosure (const struct lu_env *env,
3027 struct cl_lock_closure *closure);
3028 int cl_lock_enclosure (const struct lu_env *env, struct cl_lock *lock,
3029 struct cl_lock_closure *closure);
3031 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
3032 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock);
3033 void cl_lock_error (const struct lu_env *env, struct cl_lock *lock, int error);
3034 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int wait);
3036 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock);
3040 /** \defgroup cl_io cl_io
3043 int cl_io_init (const struct lu_env *env, struct cl_io *io,
3044 enum cl_io_type iot, struct cl_object *obj);
3045 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
3046 enum cl_io_type iot, struct cl_object *obj);
3047 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
3048 enum cl_io_type iot, loff_t pos, size_t count);
3049 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
3051 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
3052 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
3053 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
3054 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
3055 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
3056 int cl_io_start (const struct lu_env *env, struct cl_io *io);
3057 void cl_io_end (const struct lu_env *env, struct cl_io *io);
3058 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
3059 struct cl_io_lock_link *link);
3060 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
3061 struct cl_lock_descr *descr);
3062 int cl_io_read_page (const struct lu_env *env, struct cl_io *io,
3063 struct cl_page *page);
3064 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
3065 struct cl_page *page, unsigned from, unsigned to);
3066 int cl_io_commit_write (const struct lu_env *env, struct cl_io *io,
3067 struct cl_page *page, unsigned from, unsigned to);
3068 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
3069 enum cl_req_type iot, struct cl_2queue *queue);
3070 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
3071 enum cl_req_type iot, struct cl_2queue *queue,
3073 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
3075 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
3076 struct cl_page_list *queue);
3077 int cl_io_is_going (const struct lu_env *env);
3080 * True, iff \a io is an O_APPEND write(2).
3082 static inline int cl_io_is_append(const struct cl_io *io)
3084 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
3087 static inline int cl_io_is_sync_write(const struct cl_io *io)
3089 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
3092 static inline int cl_io_is_mkwrite(const struct cl_io *io)
3094 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
3098 * True, iff \a io is a truncate(2).
3100 static inline int cl_io_is_trunc(const struct cl_io *io)
3102 return io->ci_type == CIT_SETATTR &&
3103 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
3106 struct cl_io *cl_io_top(struct cl_io *io);
3108 void cl_io_print(const struct lu_env *env, void *cookie,
3109 lu_printer_t printer, const struct cl_io *io);
3111 #define CL_IO_SLICE_CLEAN(foo_io, base) \
3113 typeof(foo_io) __foo_io = (foo_io); \
3115 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
3116 memset(&__foo_io->base + 1, 0, \
3117 (sizeof *__foo_io) - sizeof __foo_io->base); \
3122 /** \defgroup cl_page_list cl_page_list
3126 * Last page in the page list.
3128 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
3130 LASSERT(plist->pl_nr > 0);
3131 return cfs_list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
3135 * Iterate over pages in a page list.
3137 #define cl_page_list_for_each(page, list) \
3138 cfs_list_for_each_entry((page), &(list)->pl_pages, cp_batch)
3141 * Iterate over pages in a page list, taking possible removals into account.
3143 #define cl_page_list_for_each_safe(page, temp, list) \
3144 cfs_list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
3146 void cl_page_list_init (struct cl_page_list *plist);
3147 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
3148 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
3149 struct cl_page *page);
3150 void cl_page_list_splice (struct cl_page_list *list,
3151 struct cl_page_list *head);
3152 void cl_page_list_del (const struct lu_env *env,
3153 struct cl_page_list *plist, struct cl_page *page);
3154 void cl_page_list_disown (const struct lu_env *env,
3155 struct cl_io *io, struct cl_page_list *plist);
3156 int cl_page_list_own (const struct lu_env *env,
3157 struct cl_io *io, struct cl_page_list *plist);
3158 void cl_page_list_assume (const struct lu_env *env,
3159 struct cl_io *io, struct cl_page_list *plist);
3160 void cl_page_list_discard(const struct lu_env *env,
3161 struct cl_io *io, struct cl_page_list *plist);
3162 int cl_page_list_unmap (const struct lu_env *env,
3163 struct cl_io *io, struct cl_page_list *plist);
3164 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
3166 void cl_2queue_init (struct cl_2queue *queue);
3167 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
3168 void cl_2queue_disown (const struct lu_env *env,
3169 struct cl_io *io, struct cl_2queue *queue);
3170 void cl_2queue_assume (const struct lu_env *env,
3171 struct cl_io *io, struct cl_2queue *queue);
3172 void cl_2queue_discard (const struct lu_env *env,
3173 struct cl_io *io, struct cl_2queue *queue);
3174 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
3175 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
3177 /** @} cl_page_list */
3179 /** \defgroup cl_req cl_req
3181 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
3182 enum cl_req_type crt, int nr_objects);
3184 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
3185 struct cl_page *page);
3186 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
3187 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
3188 void cl_req_attr_set (const struct lu_env *env, struct cl_req *req,
3189 struct cl_req_attr *attr, obd_valid flags);
3190 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
3192 /** \defgroup cl_sync_io cl_sync_io
3196 * Anchor for synchronous transfer. This is allocated on a stack by thread
3197 * doing synchronous transfer, and a pointer to this structure is set up in
3198 * every page submitted for transfer. Transfer completion routine updates
3199 * anchor and wakes up waiting thread when transfer is complete.
3202 /** number of pages yet to be transferred. */
3203 cfs_atomic_t csi_sync_nr;
3206 /** barrier of destroy this structure */
3207 cfs_atomic_t csi_barrier;
3208 /** completion to be signaled when transfer is complete. */
3209 wait_queue_head_t csi_waitq;
3212 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages);
3213 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
3214 struct cl_page_list *queue, struct cl_sync_io *anchor,
3216 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret);
3218 /** @} cl_sync_io */
3222 /** \defgroup cl_env cl_env
3224 * lu_env handling for a client.
3226 * lu_env is an environment within which lustre code executes. Its major part
3227 * is lu_context---a fast memory allocation mechanism that is used to conserve
3228 * precious kernel stack space. Originally lu_env was designed for a server,
3231 * - there is a (mostly) fixed number of threads, and
3233 * - call chains have no non-lustre portions inserted between lustre code.
3235 * On a client both these assumtpion fails, because every user thread can
3236 * potentially execute lustre code as part of a system call, and lustre calls
3237 * into VFS or MM that call back into lustre.
3239 * To deal with that, cl_env wrapper functions implement the following
3242 * - allocation and destruction of environment is amortized by caching no
3243 * longer used environments instead of destroying them;
3245 * - there is a notion of "current" environment, attached to the kernel
3246 * data structure representing current thread Top-level lustre code
3247 * allocates an environment and makes it current, then calls into
3248 * non-lustre code, that in turn calls lustre back. Low-level lustre
3249 * code thus called can fetch environment created by the top-level code
3250 * and reuse it, avoiding additional environment allocation.
3251 * Right now, three interfaces can attach the cl_env to running thread:
3254 * - cl_env_reexit(cl_env_reenter had to be called priorly)
3256 * \see lu_env, lu_context, lu_context_key
3259 struct cl_env_nest {
3264 struct lu_env *cl_env_peek (int *refcheck);
3265 struct lu_env *cl_env_get (int *refcheck);
3266 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
3267 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
3268 void cl_env_put (struct lu_env *env, int *refcheck);
3269 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
3270 void *cl_env_reenter (void);
3271 void cl_env_reexit (void *cookie);
3272 void cl_env_implant (struct lu_env *env, int *refcheck);
3273 void cl_env_unplant (struct lu_env *env, int *refcheck);
3280 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
3281 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
3283 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
3284 struct lu_device_type *ldt,
3285 struct lu_device *next);
3288 int cl_global_init(void);
3289 void cl_global_fini(void);
3291 #endif /* _LINUX_CL_OBJECT_H */