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, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 #ifndef _LUSTRE_CL_OBJECT_H
37 #define _LUSTRE_CL_OBJECT_H
39 /** \defgroup clio clio
41 * Client objects implement io operations and cache pages.
43 * Examples: lov and osc are implementations of cl interface.
45 * Big Theory Statement.
49 * Client implementation is based on the following data-types:
55 * - cl_lock represents an extent lock on an object.
57 * - cl_io represents high-level i/o activity such as whole read/write
58 * system call, or write-out of pages from under the lock being
59 * canceled. cl_io has sub-ios that can be stopped and resumed
60 * independently, thus achieving high degree of transfer
61 * parallelism. Single cl_io can be advanced forward by
62 * the multiple threads (although in the most usual case of
63 * read/write system call it is associated with the single user
64 * thread, that issued the system call).
66 * - cl_req represents a collection of pages for a transfer. cl_req is
67 * constructed by req-forming engine that tries to saturate
68 * transport with large and continuous transfers.
72 * - to avoid confusion high-level I/O operation like read or write system
73 * call is referred to as "an io", whereas low-level I/O operation, like
74 * RPC, is referred to as "a transfer"
76 * - "generic code" means generic (not file system specific) code in the
77 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
78 * is not layer specific.
84 * - cl_object_header::coh_page_guard
85 * - cl_object_header::coh_lock_guard
88 * See the top comment in cl_object.c for the description of overall locking and
89 * reference-counting design.
91 * See comments below for the description of i/o, page, and dlm-locking
98 * super-class definitions.
100 #include <lu_object.h>
102 # include <linux/mutex.h>
103 # include <linux/radix-tree.h>
108 struct cl_device_operations;
111 struct cl_object_page_operations;
112 struct cl_object_lock_operations;
115 struct cl_page_slice;
117 struct cl_lock_slice;
119 struct cl_lock_operations;
120 struct cl_page_operations;
129 * Operations for each data device in the client stack.
131 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
133 struct cl_device_operations {
135 * Initialize cl_req. This method is called top-to-bottom on all
136 * devices in the stack to get them a chance to allocate layer-private
137 * data, and to attach them to the cl_req by calling
138 * cl_req_slice_add().
140 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
141 * \see ccc_req_init()
143 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
148 * Device in the client stack.
150 * \see ccc_device, lov_device, lovsub_device, osc_device
154 struct lu_device cd_lu_dev;
155 /** Per-layer operation vector. */
156 const struct cl_device_operations *cd_ops;
159 /** \addtogroup cl_object cl_object
162 * "Data attributes" of cl_object. Data attributes can be updated
163 * independently for a sub-object, and top-object's attributes are calculated
164 * from sub-objects' ones.
167 /** Object size, in bytes */
170 * Known minimal size, in bytes.
172 * This is only valid when at least one DLM lock is held.
175 /** Modification time. Measured in seconds since epoch. */
177 /** Access time. Measured in seconds since epoch. */
179 /** Change time. Measured in seconds since epoch. */
182 * Blocks allocated to this cl_object on the server file system.
184 * \todo XXX An interface for block size is needed.
188 * User identifier for quota purposes.
192 * Group identifier for quota purposes.
198 * Fields in cl_attr that are being set.
212 * Sub-class of lu_object with methods common for objects on the client
215 * cl_object: represents a regular file system object, both a file and a
216 * stripe. cl_object is based on lu_object: it is identified by a fid,
217 * layered, cached, hashed, and lrued. Important distinction with the server
218 * side, where md_object and dt_object are used, is that cl_object "fans out"
219 * at the lov/sns level: depending on the file layout, single file is
220 * represented as a set of "sub-objects" (stripes). At the implementation
221 * level, struct lov_object contains an array of cl_objects. Each sub-object
222 * is a full-fledged cl_object, having its fid, living in the lru and hash
225 * This leads to the next important difference with the server side: on the
226 * client, it's quite usual to have objects with the different sequence of
227 * layers. For example, typical top-object is composed of the following
233 * whereas its sub-objects are composed of
238 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
239 * track of the object-subobject relationship.
241 * Sub-objects are not cached independently: when top-object is about to
242 * be discarded from the memory, all its sub-objects are torn-down and
245 * \see ccc_object, lov_object, lovsub_object, osc_object
249 struct lu_object co_lu;
250 /** per-object-layer operations */
251 const struct cl_object_operations *co_ops;
252 /** offset of page slice in cl_page buffer */
257 * Description of the client object configuration. This is used for the
258 * creation of a new client object that is identified by a more state than
261 struct cl_object_conf {
263 struct lu_object_conf coc_lu;
266 * Object layout. This is consumed by lov.
268 struct lustre_md *coc_md;
270 * Description of particular stripe location in the
271 * cluster. This is consumed by osc.
273 struct lov_oinfo *coc_oinfo;
276 * VFS inode. This is consumed by vvp.
278 struct inode *coc_inode;
280 * Layout lock handle.
282 struct ldlm_lock *coc_lock;
284 * Operation to handle layout, OBJECT_CONF_XYZ.
290 /** configure layout, set up a new stripe, must be called while
291 * holding layout lock. */
293 /** invalidate the current stripe configuration due to losing
295 OBJECT_CONF_INVALIDATE = 1,
296 /** wait for old layout to go away so that new layout can be
302 * Operations implemented for each cl object layer.
304 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
306 struct cl_object_operations {
308 * Initialize page slice for this layer. Called top-to-bottom through
309 * every object layer when a new cl_page is instantiated. Layer
310 * keeping private per-page data, or requiring its own page operations
311 * vector should allocate these data here, and attach then to the page
312 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
315 * \retval NULL success.
317 * \retval ERR_PTR(errno) failure code.
319 * \retval valid-pointer pointer to already existing referenced page
320 * to be used instead of newly created.
322 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
323 struct cl_page *page, struct page *vmpage);
325 * Initialize lock slice for this layer. Called top-to-bottom through
326 * every object layer when a new cl_lock is instantiated. Layer
327 * keeping private per-lock data, or requiring its own lock operations
328 * vector should allocate these data here, and attach then to the lock
329 * by calling cl_lock_slice_add(). Mandatory.
331 int (*coo_lock_init)(const struct lu_env *env,
332 struct cl_object *obj, struct cl_lock *lock,
333 const struct cl_io *io);
335 * Initialize io state for a given layer.
337 * called top-to-bottom once per io existence to initialize io
338 * state. If layer wants to keep some state for this type of io, it
339 * has to embed struct cl_io_slice in lu_env::le_ses, and register
340 * slice with cl_io_slice_add(). It is guaranteed that all threads
341 * participating in this io share the same session.
343 int (*coo_io_init)(const struct lu_env *env,
344 struct cl_object *obj, struct cl_io *io);
346 * Fill portion of \a attr that this layer controls. This method is
347 * called top-to-bottom through all object layers.
349 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
351 * \return 0: to continue
352 * \return +ve: to stop iterating through layers (but 0 is returned
353 * from enclosing cl_object_attr_get())
354 * \return -ve: to signal error
356 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
357 struct cl_attr *attr);
361 * \a valid is a bitmask composed from enum #cl_attr_valid, and
362 * indicating what attributes are to be set.
364 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
366 * \return the same convention as for
367 * cl_object_operations::coo_attr_get() is used.
369 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj,
370 const struct cl_attr *attr, unsigned valid);
372 * Update object configuration. Called top-to-bottom to modify object
375 * XXX error conditions and handling.
377 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
378 const struct cl_object_conf *conf);
380 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
381 * object. Layers are supposed to fill parts of \a lvb that will be
382 * shipped to the glimpse originator as a glimpse result.
384 * \see ccc_object_glimpse(), lovsub_object_glimpse(),
385 * \see osc_object_glimpse()
387 int (*coo_glimpse)(const struct lu_env *env,
388 const struct cl_object *obj, struct ost_lvb *lvb);
392 * Extended header for client object.
394 struct cl_object_header {
395 /** Standard lu_object_header. cl_object::co_lu::lo_header points
397 struct lu_object_header coh_lu;
399 * \todo XXX move locks below to the separate cache-lines, they are
400 * mostly useless otherwise.
403 /** Lock protecting page tree. */
404 spinlock_t coh_page_guard;
405 /** Lock protecting lock list. */
406 spinlock_t coh_lock_guard;
408 /** Radix tree of cl_page's, cached for this object. */
409 struct radix_tree_root coh_tree;
410 /** # of pages in radix tree. */
411 unsigned long coh_pages;
412 /** List of cl_lock's granted for this object. */
413 struct list_head coh_locks;
416 * Parent object. It is assumed that an object has a well-defined
417 * parent, but not a well-defined child (there may be multiple
418 * sub-objects, for the same top-object). cl_object_header::coh_parent
419 * field allows certain code to be written generically, without
420 * limiting possible cl_object layouts unduly.
422 struct cl_object_header *coh_parent;
424 * Protects consistency between cl_attr of parent object and
425 * attributes of sub-objects, that the former is calculated ("merged")
428 * \todo XXX this can be read/write lock if needed.
430 spinlock_t coh_attr_guard;
432 * Size of cl_page + page slices
434 unsigned short coh_page_bufsize;
436 * Number of objects above this one: 0 for a top-object, 1 for its
439 unsigned char coh_nesting;
443 * Helper macro: iterate over all layers of the object \a obj, assigning every
444 * layer top-to-bottom to \a slice.
446 #define cl_object_for_each(slice, obj) \
447 list_for_each_entry((slice), \
448 &(obj)->co_lu.lo_header->loh_layers, \
451 * Helper macro: iterate over all layers of the object \a obj, assigning every
452 * layer bottom-to-top to \a slice.
454 #define cl_object_for_each_reverse(slice, obj) \
455 list_for_each_entry_reverse((slice), \
456 &(obj)->co_lu.lo_header->loh_layers, \
461 #define pgoff_t unsigned long
464 #define CL_PAGE_EOF ((pgoff_t)~0ull)
466 /** \addtogroup cl_page cl_page
470 * Layered client page.
472 * cl_page: represents a portion of a file, cached in the memory. All pages
473 * of the given file are of the same size, and are kept in the radix tree
474 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
475 * of the top-level file object are first class cl_objects, they have their
476 * own radix trees of pages and hence page is implemented as a sequence of
477 * struct cl_pages's, linked into double-linked list through
478 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
479 * corresponding radix tree at the corresponding logical offset.
481 * cl_page is associated with VM page of the hosting environment (struct
482 * page in Linux kernel, for example), struct page. It is assumed, that this
483 * association is implemented by one of cl_page layers (top layer in the
484 * current design) that
486 * - intercepts per-VM-page call-backs made by the environment (e.g.,
489 * - translates state (page flag bits) and locking between lustre and
492 * The association between cl_page and struct page is immutable and
493 * established when cl_page is created.
495 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
496 * this io an exclusive access to this page w.r.t. other io attempts and
497 * various events changing page state (such as transfer completion, or
498 * eviction of the page from the memory). Note, that in general cl_io
499 * cannot be identified with a particular thread, and page ownership is not
500 * exactly equal to the current thread holding a lock on the page. Layer
501 * implementing association between cl_page and struct page has to implement
502 * ownership on top of available synchronization mechanisms.
504 * While lustre client maintains the notion of an page ownership by io,
505 * hosting MM/VM usually has its own page concurrency control
506 * mechanisms. For example, in Linux, page access is synchronized by the
507 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
508 * takes care to acquire and release such locks as necessary around the
509 * calls to the file system methods (->readpage(), ->prepare_write(),
510 * ->commit_write(), etc.). This leads to the situation when there are two
511 * different ways to own a page in the client:
513 * - client code explicitly and voluntary owns the page (cl_page_own());
515 * - VM locks a page and then calls the client, that has "to assume"
516 * the ownership from the VM (cl_page_assume()).
518 * Dual methods to release ownership are cl_page_disown() and
519 * cl_page_unassume().
521 * cl_page is reference counted (cl_page::cp_ref). When reference counter
522 * drops to 0, the page is returned to the cache, unless it is in
523 * cl_page_state::CPS_FREEING state, in which case it is immediately
526 * The general logic guaranteeing the absence of "existential races" for
527 * pages is the following:
529 * - there are fixed known ways for a thread to obtain a new reference
532 * - by doing a lookup in the cl_object radix tree, protected by the
535 * - by starting from VM-locked struct page and following some
536 * hosting environment method (e.g., following ->private pointer in
537 * the case of Linux kernel), see cl_vmpage_page();
539 * - when the page enters cl_page_state::CPS_FREEING state, all these
540 * ways are severed with the proper synchronization
541 * (cl_page_delete());
543 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
546 * - no new references to the page in cl_page_state::CPS_FREEING state
547 * are allowed (checked in cl_page_get()).
549 * Together this guarantees that when last reference to a
550 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
551 * page, as neither references to it can be acquired at that point, nor
554 * cl_page is a state machine. States are enumerated in enum
555 * cl_page_state. Possible state transitions are enumerated in
556 * cl_page_state_set(). State transition process (i.e., actual changing of
557 * cl_page::cp_state field) is protected by the lock on the underlying VM
560 * Linux Kernel implementation.
562 * Binding between cl_page and struct page (which is a typedef for
563 * struct page) is implemented in the vvp layer. cl_page is attached to the
564 * ->private pointer of the struct page, together with the setting of
565 * PG_private bit in page->flags, and acquiring additional reference on the
566 * struct page (much like struct buffer_head, or any similar file system
567 * private data structures).
569 * PG_locked lock is used to implement both ownership and transfer
570 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
571 * states. No additional references are acquired for the duration of the
574 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
575 * write-out is "protected" by the special PG_writeback bit.
579 * States of cl_page. cl_page.c assumes particular order here.
581 * The page state machine is rather crude, as it doesn't recognize finer page
582 * states like "dirty" or "up to date". This is because such states are not
583 * always well defined for the whole stack (see, for example, the
584 * implementation of the read-ahead, that hides page up-to-dateness to track
585 * cache hits accurately). Such sub-states are maintained by the layers that
586 * are interested in them.
590 * Page is in the cache, un-owned. Page leaves cached state in the
593 * - [cl_page_state::CPS_OWNED] io comes across the page and
596 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
597 * req-formation engine decides that it wants to include this page
598 * into an cl_req being constructed, and yanks it from the cache;
600 * - [cl_page_state::CPS_FREEING] VM callback is executed to
601 * evict the page form the memory;
603 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
607 * Page is exclusively owned by some cl_io. Page may end up in this
608 * state as a result of
610 * - io creating new page and immediately owning it;
612 * - [cl_page_state::CPS_CACHED] io finding existing cached page
615 * - [cl_page_state::CPS_OWNED] io finding existing owned page
616 * and waiting for owner to release the page;
618 * Page leaves owned state in the following cases:
620 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
621 * the cache, doing nothing;
623 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
626 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
627 * transfer for this page;
629 * - [cl_page_state::CPS_FREEING] io decides to destroy this
630 * page (e.g., as part of truncate or extent lock cancellation).
632 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
636 * Page is being written out, as a part of a transfer. This state is
637 * entered when req-formation logic decided that it wants this page to
638 * be sent through the wire _now_. Specifically, it means that once
639 * this state is achieved, transfer completion handler (with either
640 * success or failure indication) is guaranteed to be executed against
641 * this page independently of any locks and any scheduling decisions
642 * made by the hosting environment (that effectively means that the
643 * page is never put into cl_page_state::CPS_PAGEOUT state "in
644 * advance". This property is mentioned, because it is important when
645 * reasoning about possible dead-locks in the system). The page can
646 * enter this state as a result of
648 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
649 * write-out of this page, or
651 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
652 * that it has enough dirty pages cached to issue a "good"
655 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
656 * is completed---it is moved into cl_page_state::CPS_CACHED state.
658 * Underlying VM page is locked for the duration of transfer.
660 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
664 * Page is being read in, as a part of a transfer. This is quite
665 * similar to the cl_page_state::CPS_PAGEOUT state, except that
666 * read-in is always "immediate"---there is no such thing a sudden
667 * construction of read cl_req from cached, presumably not up to date,
670 * Underlying VM page is locked for the duration of transfer.
672 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
676 * Page is being destroyed. This state is entered when client decides
677 * that page has to be deleted from its host object, as, e.g., a part
680 * Once this state is reached, there is no way to escape it.
682 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
689 /** Host page, the page is from the host inode which the cl_page
693 /** Transient page, the transient cl_page is used to bind a cl_page
694 * to vmpage which is not belonging to the same object of cl_page.
695 * it is used in DirectIO, lockless IO and liblustre. */
700 * Flags maintained for every cl_page.
704 * Set when pagein completes. Used for debugging (read completes at
705 * most once for a page).
707 CPF_READ_COMPLETED = 1 << 0
711 * Fields are protected by the lock on struct page, except for atomics and
714 * \invariant Data type invariants are in cl_page_invariant(). Basically:
715 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
716 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
717 * cl_page::cp_owner (when set).
720 /** Reference counter. */
722 /** An object this page is a part of. Immutable after creation. */
723 struct cl_object *cp_obj;
724 /** Logical page index within the object. Immutable after creation. */
726 /** List of slices. Immutable after creation. */
727 struct list_head cp_layers;
728 /** Parent page, NULL for top-level page. Immutable after creation. */
729 struct cl_page *cp_parent;
730 /** Lower-layer page. NULL for bottommost page. Immutable after
732 struct cl_page *cp_child;
734 * Page state. This field is const to avoid accidental update, it is
735 * modified only internally within cl_page.c. Protected by a VM lock.
737 const enum cl_page_state cp_state;
738 /** Linkage of pages within group. Protected by cl_page::cp_mutex. */
739 struct list_head cp_batch;
740 /** Mutex serializing membership of a page in a batch. */
741 struct mutex cp_mutex;
742 /** Linkage of pages within cl_req. */
743 struct list_head cp_flight;
744 /** Transfer error. */
748 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
751 enum cl_page_type cp_type;
754 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
755 * by sub-io. Protected by a VM lock.
757 struct cl_io *cp_owner;
759 * Debug information, the task is owning the page.
761 struct task_struct *cp_task;
763 * Owning IO request in cl_page_state::CPS_PAGEOUT and
764 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
765 * the top-level pages. Protected by a VM lock.
767 struct cl_req *cp_req;
768 /** List of references to this page, for debugging. */
769 struct lu_ref cp_reference;
770 /** Link to an object, for debugging. */
771 struct lu_ref_link cp_obj_ref;
772 /** Link to a queue, for debugging. */
773 struct lu_ref_link cp_queue_ref;
774 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */
776 /** Assigned if doing a sync_io */
777 struct cl_sync_io *cp_sync_io;
781 * Per-layer part of cl_page.
783 * \see ccc_page, lov_page, osc_page
785 struct cl_page_slice {
786 struct cl_page *cpl_page;
788 * Object slice corresponding to this page slice. Immutable after
791 struct cl_object *cpl_obj;
792 const struct cl_page_operations *cpl_ops;
793 /** Linkage into cl_page::cp_layers. Immutable after creation. */
794 struct list_head cpl_linkage;
798 * Lock mode. For the client extent locks.
800 * \warning: cl_lock_mode_match() assumes particular ordering here.
805 * Mode of a lock that protects no data, and exists only as a
806 * placeholder. This is used for `glimpse' requests. A phantom lock
807 * might get promoted to real lock at some point.
816 * Requested transfer type.
826 * Per-layer page operations.
828 * Methods taking an \a io argument are for the activity happening in the
829 * context of given \a io. Page is assumed to be owned by that io, except for
830 * the obvious cases (like cl_page_operations::cpo_own()).
832 * \see vvp_page_ops, lov_page_ops, osc_page_ops
834 struct cl_page_operations {
836 * cl_page<->struct page methods. Only one layer in the stack has to
837 * implement these. Current code assumes that this functionality is
838 * provided by the topmost layer, see cl_page_disown0() as an example.
842 * \return the underlying VM page. Optional.
844 struct page *(*cpo_vmpage)(const struct lu_env *env,
845 const struct cl_page_slice *slice);
847 * Called when \a io acquires this page into the exclusive
848 * ownership. When this method returns, it is guaranteed that the is
849 * not owned by other io, and no transfer is going on against
853 * \see vvp_page_own(), lov_page_own()
855 int (*cpo_own)(const struct lu_env *env,
856 const struct cl_page_slice *slice,
857 struct cl_io *io, int nonblock);
858 /** Called when ownership it yielded. Optional.
860 * \see cl_page_disown()
861 * \see vvp_page_disown()
863 void (*cpo_disown)(const struct lu_env *env,
864 const struct cl_page_slice *slice, struct cl_io *io);
866 * Called for a page that is already "owned" by \a io from VM point of
869 * \see cl_page_assume()
870 * \see vvp_page_assume(), lov_page_assume()
872 void (*cpo_assume)(const struct lu_env *env,
873 const struct cl_page_slice *slice, struct cl_io *io);
874 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
875 * bottom-to-top when IO releases a page without actually unlocking
878 * \see cl_page_unassume()
879 * \see vvp_page_unassume()
881 void (*cpo_unassume)(const struct lu_env *env,
882 const struct cl_page_slice *slice,
885 * Announces whether the page contains valid data or not by \a uptodate.
887 * \see cl_page_export()
888 * \see vvp_page_export()
890 void (*cpo_export)(const struct lu_env *env,
891 const struct cl_page_slice *slice, int uptodate);
893 * Unmaps page from the user space (if it is mapped).
895 * \see cl_page_unmap()
896 * \see vvp_page_unmap()
898 int (*cpo_unmap)(const struct lu_env *env,
899 const struct cl_page_slice *slice, struct cl_io *io);
901 * Checks whether underlying VM page is locked (in the suitable
902 * sense). Used for assertions.
904 * \retval -EBUSY: page is protected by a lock of a given mode;
905 * \retval -ENODATA: page is not protected by a lock;
906 * \retval 0: this layer cannot decide. (Should never happen.)
908 int (*cpo_is_vmlocked)(const struct lu_env *env,
909 const struct cl_page_slice *slice);
915 * Called when page is truncated from the object. Optional.
917 * \see cl_page_discard()
918 * \see vvp_page_discard(), osc_page_discard()
920 void (*cpo_discard)(const struct lu_env *env,
921 const struct cl_page_slice *slice,
924 * Called when page is removed from the cache, and is about to being
925 * destroyed. Optional.
927 * \see cl_page_delete()
928 * \see vvp_page_delete(), osc_page_delete()
930 void (*cpo_delete)(const struct lu_env *env,
931 const struct cl_page_slice *slice);
932 /** Destructor. Frees resources and slice itself. */
933 void (*cpo_fini)(const struct lu_env *env,
934 struct cl_page_slice *slice);
937 * Checks whether the page is protected by a cl_lock. This is a
938 * per-layer method, because certain layers have ways to check for the
939 * lock much more efficiently than through the generic locks scan, or
940 * implement locking mechanisms separate from cl_lock, e.g.,
941 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
942 * being canceled, or scheduled for cancellation as soon as the last
943 * user goes away, too.
945 * \retval -EBUSY: page is protected by a lock of a given mode;
946 * \retval -ENODATA: page is not protected by a lock;
947 * \retval 0: this layer cannot decide.
949 * \see cl_page_is_under_lock()
951 int (*cpo_is_under_lock)(const struct lu_env *env,
952 const struct cl_page_slice *slice,
956 * Optional debugging helper. Prints given page slice.
958 * \see cl_page_print()
960 int (*cpo_print)(const struct lu_env *env,
961 const struct cl_page_slice *slice,
962 void *cookie, lu_printer_t p);
966 * Transfer methods. See comment on cl_req for a description of
967 * transfer formation and life-cycle.
972 * Request type dependent vector of operations.
974 * Transfer operations depend on transfer mode (cl_req_type). To avoid
975 * passing transfer mode to each and every of these methods, and to
976 * avoid branching on request type inside of the methods, separate
977 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
978 * provided. That is, method invocation usually looks like
980 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
984 * Called when a page is submitted for a transfer as a part of
987 * \return 0 : page is eligible for submission;
988 * \return -EALREADY : skip this page;
989 * \return -ve : error.
991 * \see cl_page_prep()
993 int (*cpo_prep)(const struct lu_env *env,
994 const struct cl_page_slice *slice,
997 * Completion handler. This is guaranteed to be eventually
998 * fired after cl_page_operations::cpo_prep() or
999 * cl_page_operations::cpo_make_ready() call.
1001 * This method can be called in a non-blocking context. It is
1002 * guaranteed however, that the page involved and its object
1003 * are pinned in memory (and, hence, calling cl_page_put() is
1006 * \see cl_page_completion()
1008 void (*cpo_completion)(const struct lu_env *env,
1009 const struct cl_page_slice *slice,
1012 * Called when cached page is about to be added to the
1013 * cl_req as a part of req formation.
1015 * \return 0 : proceed with this page;
1016 * \return -EAGAIN : skip this page;
1017 * \return -ve : error.
1019 * \see cl_page_make_ready()
1021 int (*cpo_make_ready)(const struct lu_env *env,
1022 const struct cl_page_slice *slice);
1024 * Announce that this page is to be written out
1025 * opportunistically, that is, page is dirty, it is not
1026 * necessary to start write-out transfer right now, but
1027 * eventually page has to be written out.
1029 * Main caller of this is the write path (see
1030 * vvp_io_commit_write()), using this method to build a
1031 * "transfer cache" from which large transfers are then
1032 * constructed by the req-formation engine.
1034 * \todo XXX it would make sense to add page-age tracking
1035 * semantics here, and to oblige the req-formation engine to
1036 * send the page out not later than it is too old.
1038 * \see cl_page_cache_add()
1040 int (*cpo_cache_add)(const struct lu_env *env,
1041 const struct cl_page_slice *slice,
1045 * Tell transfer engine that only [to, from] part of a page should be
1048 * This is used for immediate transfers.
1050 * \todo XXX this is not very good interface. It would be much better
1051 * if all transfer parameters were supplied as arguments to
1052 * cl_io_operations::cio_submit() call, but it is not clear how to do
1053 * this for page queues.
1055 * \see cl_page_clip()
1057 void (*cpo_clip)(const struct lu_env *env,
1058 const struct cl_page_slice *slice,
1061 * \pre the page was queued for transferring.
1062 * \post page is removed from client's pending list, or -EBUSY
1063 * is returned if it has already been in transferring.
1065 * This is one of seldom page operation which is:
1066 * 0. called from top level;
1067 * 1. don't have vmpage locked;
1068 * 2. every layer should synchronize execution of its ->cpo_cancel()
1069 * with completion handlers. Osc uses client obd lock for this
1070 * purpose. Based on there is no vvp_page_cancel and
1071 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1073 * \see osc_page_cancel().
1075 int (*cpo_cancel)(const struct lu_env *env,
1076 const struct cl_page_slice *slice);
1078 * Write out a page by kernel. This is only called by ll_writepage
1081 * \see cl_page_flush()
1083 int (*cpo_flush)(const struct lu_env *env,
1084 const struct cl_page_slice *slice,
1090 * Helper macro, dumping detailed information about \a page into a log.
1092 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1094 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1096 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1097 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1098 CDEBUG(mask, format , ## __VA_ARGS__); \
1103 * Helper macro, dumping shorter information about \a page into a log.
1105 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1107 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1109 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1110 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1111 CDEBUG(mask, format , ## __VA_ARGS__); \
1115 static inline int __page_in_use(const struct cl_page *page, int refc)
1117 if (page->cp_type == CPT_CACHEABLE)
1119 LASSERT(atomic_read(&page->cp_ref) > 0);
1120 return (atomic_read(&page->cp_ref) > refc);
1122 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1123 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1127 /** \addtogroup cl_lock cl_lock
1131 * Extent locking on the client.
1135 * The locking model of the new client code is built around
1139 * data-type representing an extent lock on a regular file. cl_lock is a
1140 * layered object (much like cl_object and cl_page), it consists of a header
1141 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1142 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1144 * All locks for a given object are linked into cl_object_header::coh_locks
1145 * list (protected by cl_object_header::coh_lock_guard spin-lock) through
1146 * cl_lock::cll_linkage. Currently this list is not sorted in any way. We can
1147 * sort it in starting lock offset, or use altogether different data structure
1150 * Typical cl_lock consists of the two layers:
1152 * - vvp_lock (vvp specific data), and
1153 * - lov_lock (lov specific data).
1155 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1156 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1158 * - lovsub_lock, and
1161 * Each sub-lock is associated with a cl_object (representing stripe
1162 * sub-object or the file to which top-level cl_lock is associated to), and is
1163 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1164 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1165 * is different from cl_page, that doesn't fan out (there is usually exactly
1166 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1167 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1171 * cl_lock is reference counted. When reference counter drops to 0, lock is
1172 * placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING
1173 * lock is destroyed when last reference is released. Referencing between
1174 * top-lock and its sub-locks is described in the lov documentation module.
1178 * Also, cl_lock is a state machine. This requires some clarification. One of
1179 * the goals of client IO re-write was to make IO path non-blocking, or at
1180 * least to make it easier to make it non-blocking in the future. Here
1181 * `non-blocking' means that when a system call (read, write, truncate)
1182 * reaches a situation where it has to wait for a communication with the
1183 * server, it should --instead of waiting-- remember its current state and
1184 * switch to some other work. E.g,. instead of waiting for a lock enqueue,
1185 * client should proceed doing IO on the next stripe, etc. Obviously this is
1186 * rather radical redesign, and it is not planned to be fully implemented at
1187 * this time, instead we are putting some infrastructure in place, that would
1188 * make it easier to do asynchronous non-blocking IO easier in the
1189 * future. Specifically, where old locking code goes to sleep (waiting for
1190 * enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When
1191 * enqueue reply comes, its completion handler signals that lock state-machine
1192 * is ready to transit to the next state. There is some generic code in
1193 * cl_lock.c that sleeps, waiting for these signals. As a result, for users of
1194 * this cl_lock.c code, it looks like locking is done in normal blocking
1195 * fashion, and it the same time it is possible to switch to the non-blocking
1196 * locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c
1199 * For a description of state machine states and transitions see enum
1202 * There are two ways to restrict a set of states which lock might move to:
1204 * - placing a "hold" on a lock guarantees that lock will not be moved
1205 * into cl_lock_state::CLS_FREEING state until hold is released. Hold
1206 * can be only acquired on a lock that is not in
1207 * cl_lock_state::CLS_FREEING. All holds on a lock are counted in
1208 * cl_lock::cll_holds. Hold protects lock from cancellation and
1209 * destruction. Requests to cancel and destroy a lock on hold will be
1210 * recorded, but only honored when last hold on a lock is released;
1212 * - placing a "user" on a lock guarantees that lock will not leave
1213 * cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING,
1214 * cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of
1215 * states, once it enters this set. That is, if a user is added onto a
1216 * lock in a state not from this set, it doesn't immediately enforce
1217 * lock to move to this set, but once lock enters this set it will
1218 * remain there until all users are removed. Lock users are counted in
1219 * cl_lock::cll_users.
1221 * User is used to assure that lock is not canceled or destroyed while
1222 * it is being enqueued, or actively used by some IO.
1224 * Currently, a user always comes with a hold (cl_lock_invariant()
1225 * checks that a number of holds is not less than a number of users).
1229 * This is how lock state-machine operates. struct cl_lock contains a mutex
1230 * cl_lock::cll_guard that protects struct fields.
1232 * - mutex is taken, and cl_lock::cll_state is examined.
1234 * - for every state there are possible target states where lock can move
1235 * into. They are tried in order. Attempts to move into next state are
1236 * done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try().
1238 * - if the transition can be performed immediately, state is changed,
1239 * and mutex is released.
1241 * - if the transition requires blocking, _try() function returns
1242 * cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to
1243 * sleep, waiting for possibility of lock state change. It is woken
1244 * up when some event occurs, that makes lock state change possible
1245 * (e.g., the reception of the reply from the server), and repeats
1248 * Top-lock and sub-lock has separate mutexes and the latter has to be taken
1249 * first to avoid dead-lock.
1251 * To see an example of interaction of all these issues, take a look at the
1252 * lov_cl.c:lov_lock_enqueue() function. It is called as a part of
1253 * cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by
1254 * advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note
1255 * also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It
1256 * also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be
1257 * done in parallel, rather than one after another (this is used for glimpse
1258 * locks, that cannot dead-lock).
1260 * INTERFACE AND USAGE
1262 * struct cl_lock_operations provide a number of call-backs that are invoked
1263 * when events of interest occurs. Layers can intercept and handle glimpse,
1264 * blocking, cancel ASTs and a reception of the reply from the server.
1266 * One important difference with the old client locking model is that new
1267 * client has a representation for the top-lock, whereas in the old code only
1268 * sub-locks existed as real data structures and file-level locks are
1269 * represented by "request sets" that are created and destroyed on each and
1270 * every lock creation.
1272 * Top-locks are cached, and can be found in the cache by the system calls. It
1273 * is possible that top-lock is in cache, but some of its sub-locks were
1274 * canceled and destroyed. In that case top-lock has to be enqueued again
1275 * before it can be used.
1277 * Overall process of the locking during IO operation is as following:
1279 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1280 * is called on each layer. Responsibility of this method is to add locks,
1281 * needed by a given layer into cl_io.ci_lockset.
1283 * - once locks for all layers were collected, they are sorted to avoid
1284 * dead-locks (cl_io_locks_sort()), and enqueued.
1286 * - when all locks are acquired, IO is performed;
1288 * - locks are released into cache.
1290 * Striping introduces major additional complexity into locking. The
1291 * fundamental problem is that it is generally unsafe to actively use (hold)
1292 * two locks on the different OST servers at the same time, as this introduces
1293 * inter-server dependency and can lead to cascading evictions.
1295 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1296 * that no multi-stripe locks are taken (note that this design abandons POSIX
1297 * read/write semantics). Such pieces ideally can be executed concurrently. At
1298 * the same time, certain types of IO cannot be sub-divived, without
1299 * sacrificing correctness. This includes:
1301 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1304 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1306 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1307 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1308 * has to be held together with the usual lock on [offset, offset + count].
1310 * As multi-stripe locks have to be allowed, it makes sense to cache them, so
1311 * that, for example, a sequence of O_APPEND writes can proceed quickly
1312 * without going down to the individual stripes to do lock matching. On the
1313 * other hand, multi-stripe locks shouldn't be used by normal read/write
1314 * calls. To achieve this, every layer can implement ->clo_fits_into() method,
1315 * that is called by lock matching code (cl_lock_lookup()), and that can be
1316 * used to selectively disable matching of certain locks for certain IOs. For
1317 * exmaple, lov layer implements lov_lock_fits_into() that allow multi-stripe
1318 * locks to be matched only for truncates and O_APPEND writes.
1320 * Interaction with DLM
1322 * In the expected setup, cl_lock is ultimately backed up by a collection of
1323 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1324 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1325 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1326 * description of interaction with DLM.
1332 struct cl_lock_descr {
1333 /** Object this lock is granted for. */
1334 struct cl_object *cld_obj;
1335 /** Index of the first page protected by this lock. */
1337 /** Index of the last page (inclusive) protected by this lock. */
1339 /** Group ID, for group lock */
1342 enum cl_lock_mode cld_mode;
1344 * flags to enqueue lock. A combination of bit-flags from
1345 * enum cl_enq_flags.
1347 __u32 cld_enq_flags;
1350 #define DDESCR "%s(%d):[%lu, %lu]"
1351 #define PDESCR(descr) \
1352 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1353 (descr)->cld_start, (descr)->cld_end
1355 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1358 * Lock state-machine states.
1363 * Possible state transitions:
1365 * +------------------>NEW
1367 * | | cl_enqueue_try()
1369 * | cl_unuse_try() V
1370 * | +--------------QUEUING (*)
1372 * | | | cl_enqueue_try()
1374 * | | cl_unuse_try() V
1375 * sub-lock | +-------------ENQUEUED (*)
1377 * | | | cl_wait_try()
1382 * | | HELD<---------+
1384 * | | | | cl_use_try()
1385 * | | cl_unuse_try() | |
1388 * | +------------>INTRANSIT (D) <--+
1390 * | cl_unuse_try() | | cached lock found
1391 * | | | cl_use_try()
1394 * +------------------CACHED---------+
1403 * In states marked with (*) transition to the same state (i.e., a loop
1404 * in the diagram) is possible.
1406 * (R) is the point where Receive call-back is invoked: it allows layers
1407 * to handle arrival of lock reply.
1409 * (C) is the point where Cancellation call-back is invoked.
1411 * (D) is the transit state which means the lock is changing.
1413 * Transition to FREEING state is possible from any other state in the
1414 * diagram in case of unrecoverable error.
1418 * These states are for individual cl_lock object. Top-lock and its sub-locks
1419 * can be in the different states. Another way to say this is that we have
1420 * nested state-machines.
1422 * Separate QUEUING and ENQUEUED states are needed to support non-blocking
1423 * operation for locks with multiple sub-locks. Imagine lock on a file F, that
1424 * intersects 3 stripes S0, S1, and S2. To enqueue F client has to send
1425 * enqueue to S0, wait for its completion, then send enqueue for S1, wait for
1426 * its completion and at last enqueue lock for S2, and wait for its
1427 * completion. In that case, top-lock is in QUEUING state while S0, S1 are
1428 * handled, and is in ENQUEUED state after enqueue to S2 has been sent (note
1429 * that in this case, sub-locks move from state to state, and top-lock remains
1430 * in the same state).
1432 enum cl_lock_state {
1434 * Lock that wasn't yet enqueued
1438 * Enqueue is in progress, blocking for some intermediate interaction
1439 * with the other side.
1443 * Lock is fully enqueued, waiting for server to reply when it is
1448 * Lock granted, actively used by some IO.
1452 * This state is used to mark the lock is being used, or unused.
1453 * We need this state because the lock may have several sublocks,
1454 * so it's impossible to have an atomic way to bring all sublocks
1455 * into CLS_HELD state at use case, or all sublocks to CLS_CACHED
1457 * If a thread is referring to a lock, and it sees the lock is in this
1458 * state, it must wait for the lock.
1459 * See state diagram for details.
1463 * Lock granted, not used.
1467 * Lock is being destroyed.
1473 enum cl_lock_flags {
1475 * lock has been cancelled. This flag is never cleared once set (by
1476 * cl_lock_cancel0()).
1478 CLF_CANCELLED = 1 << 0,
1479 /** cancellation is pending for this lock. */
1480 CLF_CANCELPEND = 1 << 1,
1481 /** destruction is pending for this lock. */
1482 CLF_DOOMED = 1 << 2,
1483 /** from enqueue RPC reply upcall. */
1484 CLF_FROM_UPCALL= 1 << 3,
1490 * Lock closure is a collection of locks (both top-locks and sub-locks) that
1491 * might be updated in a result of an operation on a certain lock (which lock
1492 * this is a closure of).
1494 * Closures are needed to guarantee dead-lock freedom in the presence of
1496 * - nested state-machines (top-lock state-machine composed of sub-lock
1497 * state-machines), and
1499 * - shared sub-locks.
1501 * Specifically, many operations, such as lock enqueue, wait, unlock,
1502 * etc. start from a top-lock, and then operate on a sub-locks of this
1503 * top-lock, holding a top-lock mutex. When sub-lock state changes as a result
1504 * of such operation, this change has to be propagated to all top-locks that
1505 * share this sub-lock. Obviously, no natural lock ordering (e.g.,
1506 * top-to-bottom or bottom-to-top) captures this scenario, so try-locking has
1507 * to be used. Lock closure systematizes this try-and-repeat logic.
1509 struct cl_lock_closure {
1511 * Lock that is mutexed when closure construction is started. When
1512 * closure in is `wait' mode (cl_lock_closure::clc_wait), mutex on
1513 * origin is released before waiting.
1515 struct cl_lock *clc_origin;
1517 * List of enclosed locks, so far. Locks are linked here through
1518 * cl_lock::cll_inclosure.
1520 struct list_head clc_list;
1522 * True iff closure is in a `wait' mode. This determines what
1523 * cl_lock_enclosure() does when a lock L to be added to the closure
1524 * is currently mutexed by some other thread.
1526 * If cl_lock_closure::clc_wait is not set, then closure construction
1527 * fails with CLO_REPEAT immediately.
1529 * In wait mode, cl_lock_enclosure() waits until next attempt to build
1530 * a closure might succeed. To this end it releases an origin mutex
1531 * (cl_lock_closure::clc_origin), that has to be the only lock mutex
1532 * owned by the current thread, and then waits on L mutex (by grabbing
1533 * it and immediately releasing), before returning CLO_REPEAT to the
1537 /** Number of locks in the closure. */
1542 * Layered client lock.
1545 /** Reference counter. */
1547 /** List of slices. Immutable after creation. */
1548 struct list_head cll_layers;
1550 * Linkage into cl_lock::cll_descr::cld_obj::coh_locks list. Protected
1551 * by cl_lock::cll_descr::cld_obj::coh_lock_guard.
1553 struct list_head cll_linkage;
1555 * Parameters of this lock. Protected by
1556 * cl_lock::cll_descr::cld_obj::coh_lock_guard nested within
1557 * cl_lock::cll_guard. Modified only on lock creation and in
1560 struct cl_lock_descr cll_descr;
1561 /** Protected by cl_lock::cll_guard. */
1562 enum cl_lock_state cll_state;
1563 /** signals state changes. */
1564 wait_queue_head_t cll_wq;
1566 * Recursive lock, most fields in cl_lock{} are protected by this.
1568 * Locking rules: this mutex is never held across network
1569 * communication, except when lock is being canceled.
1571 * Lock ordering: a mutex of a sub-lock is taken first, then a mutex
1572 * on a top-lock. Other direction is implemented through a
1573 * try-lock-repeat loop. Mutices of unrelated locks can be taken only
1576 * \see osc_lock_enqueue_wait(), lov_lock_cancel(), lov_sublock_wait().
1578 struct mutex cll_guard;
1579 struct task_struct *cll_guarder;
1583 * the owner for INTRANSIT state
1585 struct task_struct *cll_intransit_owner;
1588 * Number of holds on a lock. A hold prevents a lock from being
1589 * canceled and destroyed. Protected by cl_lock::cll_guard.
1591 * \see cl_lock_hold(), cl_lock_unhold(), cl_lock_release()
1595 * Number of lock users. Valid in cl_lock_state::CLS_HELD state
1596 * only. Lock user pins lock in CLS_HELD state. Protected by
1597 * cl_lock::cll_guard.
1599 * \see cl_wait(), cl_unuse().
1603 * Flag bit-mask. Values from enum cl_lock_flags. Updates are
1604 * protected by cl_lock::cll_guard.
1606 unsigned long cll_flags;
1608 * A linkage into a list of locks in a closure.
1610 * \see cl_lock_closure
1612 struct list_head cll_inclosure;
1614 * Confict lock at queuing time.
1616 struct cl_lock *cll_conflict;
1618 * A list of references to this lock, for debugging.
1620 struct lu_ref cll_reference;
1622 * A list of holds on this lock, for debugging.
1624 struct lu_ref cll_holders;
1626 * A reference for cl_lock::cll_descr::cld_obj. For debugging.
1628 struct lu_ref_link cll_obj_ref;
1629 #ifdef CONFIG_LOCKDEP
1630 /* "dep_map" name is assumed by lockdep.h macros. */
1631 struct lockdep_map dep_map;
1636 * Per-layer part of cl_lock
1638 * \see ccc_lock, lov_lock, lovsub_lock, osc_lock
1640 struct cl_lock_slice {
1641 struct cl_lock *cls_lock;
1642 /** Object slice corresponding to this lock slice. Immutable after
1644 struct cl_object *cls_obj;
1645 const struct cl_lock_operations *cls_ops;
1646 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1647 struct list_head cls_linkage;
1651 * Possible (non-error) return values of ->clo_{enqueue,wait,unlock}().
1653 * NOTE: lov_subresult() depends on ordering here.
1655 enum cl_lock_transition {
1656 /** operation cannot be completed immediately. Wait for state change. */
1658 /** operation had to release lock mutex, restart. */
1660 /** lower layer re-enqueued. */
1666 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1668 struct cl_lock_operations {
1670 * \name statemachine
1672 * State machine transitions. These 3 methods are called to transfer
1673 * lock from one state to another, as described in the commentary
1674 * above enum #cl_lock_state.
1676 * \retval 0 this layer has nothing more to do to before
1677 * transition to the target state happens;
1679 * \retval CLO_REPEAT method had to release and re-acquire cl_lock
1680 * mutex, repeat invocation of transition method
1681 * across all layers;
1683 * \retval CLO_WAIT this layer cannot move to the target state
1684 * immediately, as it has to wait for certain event
1685 * (e.g., the communication with the server). It
1686 * is guaranteed, that when the state transfer
1687 * becomes possible, cl_lock::cll_wq wait-queue
1688 * is signaled. Caller can wait for this event by
1689 * calling cl_lock_state_wait();
1691 * \retval -ve failure, abort state transition, move the lock
1692 * into cl_lock_state::CLS_FREEING state, and set
1693 * cl_lock::cll_error.
1695 * Once all layers voted to agree to transition (by returning 0), lock
1696 * is moved into corresponding target state. All state transition
1697 * methods are optional.
1701 * Attempts to enqueue the lock. Called top-to-bottom.
1703 * \see ccc_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1704 * \see osc_lock_enqueue()
1706 int (*clo_enqueue)(const struct lu_env *env,
1707 const struct cl_lock_slice *slice,
1708 struct cl_io *io, __u32 enqflags);
1710 * Attempts to wait for enqueue result. Called top-to-bottom.
1712 * \see ccc_lock_wait(), lov_lock_wait(), osc_lock_wait()
1714 int (*clo_wait)(const struct lu_env *env,
1715 const struct cl_lock_slice *slice);
1717 * Attempts to unlock the lock. Called bottom-to-top. In addition to
1718 * usual return values of lock state-machine methods, this can return
1719 * -ESTALE to indicate that lock cannot be returned to the cache, and
1720 * has to be re-initialized.
1721 * unuse is a one-shot operation, so it must NOT return CLO_WAIT.
1723 * \see ccc_lock_unuse(), lov_lock_unuse(), osc_lock_unuse()
1725 int (*clo_unuse)(const struct lu_env *env,
1726 const struct cl_lock_slice *slice);
1728 * Notifies layer that cached lock is started being used.
1730 * \pre lock->cll_state == CLS_CACHED
1732 * \see lov_lock_use(), osc_lock_use()
1734 int (*clo_use)(const struct lu_env *env,
1735 const struct cl_lock_slice *slice);
1736 /** @} statemachine */
1738 * A method invoked when lock state is changed (as a result of state
1739 * transition). This is used, for example, to track when the state of
1740 * a sub-lock changes, to propagate this change to the corresponding
1741 * top-lock. Optional
1743 * \see lovsub_lock_state()
1745 void (*clo_state)(const struct lu_env *env,
1746 const struct cl_lock_slice *slice,
1747 enum cl_lock_state st);
1749 * Returns true, iff given lock is suitable for the given io, idea
1750 * being, that there are certain "unsafe" locks, e.g., ones acquired
1751 * for O_APPEND writes, that we don't want to re-use for a normal
1752 * write, to avoid the danger of cascading evictions. Optional. Runs
1753 * under cl_object_header::coh_lock_guard.
1755 * XXX this should take more information about lock needed by
1756 * io. Probably lock description or something similar.
1758 * \see lov_fits_into()
1760 int (*clo_fits_into)(const struct lu_env *env,
1761 const struct cl_lock_slice *slice,
1762 const struct cl_lock_descr *need,
1763 const struct cl_io *io);
1766 * Asynchronous System Traps. All of then are optional, all are
1767 * executed bottom-to-top.
1772 * Cancellation callback. Cancel a lock voluntarily, or under
1773 * the request of server.
1775 void (*clo_cancel)(const struct lu_env *env,
1776 const struct cl_lock_slice *slice);
1778 * Lock weighting ast. Executed to estimate how precious this lock
1779 * is. The sum of results across all layers is used to determine
1780 * whether lock worth keeping in cache given present memory usage.
1782 * \see osc_lock_weigh(), vvp_lock_weigh(), lovsub_lock_weigh().
1784 unsigned long (*clo_weigh)(const struct lu_env *env,
1785 const struct cl_lock_slice *slice);
1789 * \see lovsub_lock_closure()
1791 int (*clo_closure)(const struct lu_env *env,
1792 const struct cl_lock_slice *slice,
1793 struct cl_lock_closure *closure);
1795 * Executed bottom-to-top when lock description changes (e.g., as a
1796 * result of server granting more generous lock than was requested).
1798 * \see lovsub_lock_modify()
1800 int (*clo_modify)(const struct lu_env *env,
1801 const struct cl_lock_slice *slice,
1802 const struct cl_lock_descr *updated);
1804 * Notifies layers (bottom-to-top) that lock is going to be
1805 * destroyed. Responsibility of layers is to prevent new references on
1806 * this lock from being acquired once this method returns.
1808 * This can be called multiple times due to the races.
1810 * \see cl_lock_delete()
1811 * \see osc_lock_delete(), lovsub_lock_delete()
1813 void (*clo_delete)(const struct lu_env *env,
1814 const struct cl_lock_slice *slice);
1816 * Destructor. Frees resources and the slice.
1818 * \see ccc_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1819 * \see osc_lock_fini()
1821 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1823 * Optional debugging helper. Prints given lock slice.
1825 int (*clo_print)(const struct lu_env *env,
1826 void *cookie, lu_printer_t p,
1827 const struct cl_lock_slice *slice);
1830 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1832 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1834 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1835 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1836 CDEBUG(mask, format , ## __VA_ARGS__); \
1840 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1844 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1850 /** \addtogroup cl_page_list cl_page_list
1851 * Page list used to perform collective operations on a group of pages.
1853 * Pages are added to the list one by one. cl_page_list acquires a reference
1854 * for every page in it. Page list is used to perform collective operations on
1857 * - submit pages for an immediate transfer,
1859 * - own pages on behalf of certain io (waiting for each page in turn),
1863 * When list is finalized, it releases references on all pages it still has.
1865 * \todo XXX concurrency control.
1869 struct cl_page_list {
1871 struct list_head pl_pages;
1872 struct task_struct *pl_owner;
1876 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1877 * contains an incoming page list and an outgoing page list.
1880 struct cl_page_list c2_qin;
1881 struct cl_page_list c2_qout;
1884 /** @} cl_page_list */
1886 /** \addtogroup cl_io cl_io
1891 * cl_io represents a high level I/O activity like
1892 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1895 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1896 * important distinction. We want to minimize number of calls to the allocator
1897 * in the fast path, e.g., in the case of read(2) when everything is cached:
1898 * client already owns the lock over region being read, and data are cached
1899 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1900 * per-layer io state is stored in the session, associated with the io, see
1901 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1902 * by using free-lists, see cl_env_get().
1904 * There is a small predefined number of possible io types, enumerated in enum
1907 * cl_io is a state machine, that can be advanced concurrently by the multiple
1908 * threads. It is up to these threads to control the concurrency and,
1909 * specifically, to detect when io is done, and its state can be safely
1912 * For read/write io overall execution plan is as following:
1914 * (0) initialize io state through all layers;
1916 * (1) loop: prepare chunk of work to do
1918 * (2) call all layers to collect locks they need to process current chunk
1920 * (3) sort all locks to avoid dead-locks, and acquire them
1922 * (4) process the chunk: call per-page methods
1923 * (cl_io_operations::cio_read_page() for read,
1924 * cl_io_operations::cio_prepare_write(),
1925 * cl_io_operations::cio_commit_write() for write)
1931 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1932 * address allocation efficiency issues mentioned above), and returns with the
1933 * special error condition from per-page method when current sub-io has to
1934 * block. This causes io loop to be repeated, and lov switches to the next
1935 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1940 /** read system call */
1942 /** write system call */
1944 /** truncate, utime system calls */
1947 * page fault handling
1951 * fsync system call handling
1952 * To write out a range of file
1956 * Miscellaneous io. This is used for occasional io activity that
1957 * doesn't fit into other types. Currently this is used for:
1959 * - cancellation of an extent lock. This io exists as a context
1960 * to write dirty pages from under the lock being canceled back
1963 * - VM induced page write-out. An io context for writing page out
1964 * for memory cleansing;
1966 * - glimpse. An io context to acquire glimpse lock.
1968 * - grouplock. An io context to acquire group lock.
1970 * CIT_MISC io is used simply as a context in which locks and pages
1971 * are manipulated. Such io has no internal "process", that is,
1972 * cl_io_loop() is never called for it.
1979 * States of cl_io state machine
1982 /** Not initialized. */
1986 /** IO iteration started. */
1990 /** Actual IO is in progress. */
1992 /** IO for the current iteration finished. */
1994 /** Locks released. */
1996 /** Iteration completed. */
1998 /** cl_io finalized. */
2003 * IO state private for a layer.
2005 * This is usually embedded into layer session data, rather than allocated
2008 * \see vvp_io, lov_io, osc_io, ccc_io
2010 struct cl_io_slice {
2011 struct cl_io *cis_io;
2012 /** corresponding object slice. Immutable after creation. */
2013 struct cl_object *cis_obj;
2014 /** io operations. Immutable after creation. */
2015 const struct cl_io_operations *cis_iop;
2017 * linkage into a list of all slices for a given cl_io, hanging off
2018 * cl_io::ci_layers. Immutable after creation.
2020 struct list_head cis_linkage;
2025 * Per-layer io operations.
2026 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
2028 struct cl_io_operations {
2030 * Vector of io state transition methods for every io type.
2032 * \see cl_page_operations::io
2036 * Prepare io iteration at a given layer.
2038 * Called top-to-bottom at the beginning of each iteration of
2039 * "io loop" (if it makes sense for this type of io). Here
2040 * layer selects what work it will do during this iteration.
2042 * \see cl_io_operations::cio_iter_fini()
2044 int (*cio_iter_init) (const struct lu_env *env,
2045 const struct cl_io_slice *slice);
2047 * Finalize io iteration.
2049 * Called bottom-to-top at the end of each iteration of "io
2050 * loop". Here layers can decide whether IO has to be
2053 * \see cl_io_operations::cio_iter_init()
2055 void (*cio_iter_fini) (const struct lu_env *env,
2056 const struct cl_io_slice *slice);
2058 * Collect locks for the current iteration of io.
2060 * Called top-to-bottom to collect all locks necessary for
2061 * this iteration. This methods shouldn't actually enqueue
2062 * anything, instead it should post a lock through
2063 * cl_io_lock_add(). Once all locks are collected, they are
2064 * sorted and enqueued in the proper order.
2066 int (*cio_lock) (const struct lu_env *env,
2067 const struct cl_io_slice *slice);
2069 * Finalize unlocking.
2071 * Called bottom-to-top to finish layer specific unlocking
2072 * functionality, after generic code released all locks
2073 * acquired by cl_io_operations::cio_lock().
2075 void (*cio_unlock)(const struct lu_env *env,
2076 const struct cl_io_slice *slice);
2078 * Start io iteration.
2080 * Once all locks are acquired, called top-to-bottom to
2081 * commence actual IO. In the current implementation,
2082 * top-level vvp_io_{read,write}_start() does all the work
2083 * synchronously by calling generic_file_*(), so other layers
2084 * are called when everything is done.
2086 int (*cio_start)(const struct lu_env *env,
2087 const struct cl_io_slice *slice);
2089 * Called top-to-bottom at the end of io loop. Here layer
2090 * might wait for an unfinished asynchronous io.
2092 void (*cio_end) (const struct lu_env *env,
2093 const struct cl_io_slice *slice);
2095 * Called bottom-to-top to notify layers that read/write IO
2096 * iteration finished, with \a nob bytes transferred.
2098 void (*cio_advance)(const struct lu_env *env,
2099 const struct cl_io_slice *slice,
2102 * Called once per io, bottom-to-top to release io resources.
2104 void (*cio_fini) (const struct lu_env *env,
2105 const struct cl_io_slice *slice);
2109 * Submit pages from \a queue->c2_qin for IO, and move
2110 * successfully submitted pages into \a queue->c2_qout. Return
2111 * non-zero if failed to submit even the single page. If
2112 * submission failed after some pages were moved into \a
2113 * queue->c2_qout, completion callback with non-zero ioret is
2116 int (*cio_submit)(const struct lu_env *env,
2117 const struct cl_io_slice *slice,
2118 enum cl_req_type crt,
2119 struct cl_2queue *queue);
2122 * Read missing page.
2124 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
2125 * method, when it hits not-up-to-date page in the range. Optional.
2127 * \pre io->ci_type == CIT_READ
2129 int (*cio_read_page)(const struct lu_env *env,
2130 const struct cl_io_slice *slice,
2131 const struct cl_page_slice *page);
2133 * Prepare write of a \a page. Called bottom-to-top by a top-level
2134 * cl_io_operations::op[CIT_WRITE]::cio_start() to prepare page for
2135 * get data from user-level buffer.
2137 * \pre io->ci_type == CIT_WRITE
2139 * \see vvp_io_prepare_write(), lov_io_prepare_write(),
2140 * osc_io_prepare_write().
2142 int (*cio_prepare_write)(const struct lu_env *env,
2143 const struct cl_io_slice *slice,
2144 const struct cl_page_slice *page,
2145 unsigned from, unsigned to);
2148 * \pre io->ci_type == CIT_WRITE
2150 * \see vvp_io_commit_write(), lov_io_commit_write(),
2151 * osc_io_commit_write().
2153 int (*cio_commit_write)(const struct lu_env *env,
2154 const struct cl_io_slice *slice,
2155 const struct cl_page_slice *page,
2156 unsigned from, unsigned to);
2158 * Optional debugging helper. Print given io slice.
2160 int (*cio_print)(const struct lu_env *env, void *cookie,
2161 lu_printer_t p, const struct cl_io_slice *slice);
2165 * Flags to lock enqueue procedure.
2170 * instruct server to not block, if conflicting lock is found. Instead
2171 * -EWOULDBLOCK is returned immediately.
2173 CEF_NONBLOCK = 0x00000001,
2175 * take lock asynchronously (out of order), as it cannot
2176 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
2178 CEF_ASYNC = 0x00000002,
2180 * tell the server to instruct (though a flag in the blocking ast) an
2181 * owner of the conflicting lock, that it can drop dirty pages
2182 * protected by this lock, without sending them to the server.
2184 CEF_DISCARD_DATA = 0x00000004,
2186 * tell the sub layers that it must be a `real' lock. This is used for
2187 * mmapped-buffer locks and glimpse locks that must be never converted
2188 * into lockless mode.
2190 * \see vvp_mmap_locks(), cl_glimpse_lock().
2192 CEF_MUST = 0x00000008,
2194 * tell the sub layers that never request a `real' lock. This flag is
2195 * not used currently.
2197 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
2198 * conversion policy: ci_lockreq describes generic information of lock
2199 * requirement for this IO, especially for locks which belong to the
2200 * object doing IO; however, lock itself may have precise requirements
2201 * that are described by the enqueue flags.
2203 CEF_NEVER = 0x00000010,
2205 * for async glimpse lock.
2207 CEF_AGL = 0x00000020,
2209 * mask of enq_flags.
2211 CEF_MASK = 0x0000003f,
2215 * Link between lock and io. Intermediate structure is needed, because the
2216 * same lock can be part of multiple io's simultaneously.
2218 struct cl_io_lock_link {
2219 /** linkage into one of cl_lockset lists. */
2220 struct list_head cill_linkage;
2221 struct cl_lock_descr cill_descr;
2222 struct cl_lock *cill_lock;
2223 /** optional destructor */
2224 void (*cill_fini)(const struct lu_env *env,
2225 struct cl_io_lock_link *link);
2229 * Lock-set represents a collection of locks, that io needs at a
2230 * time. Generally speaking, client tries to avoid holding multiple locks when
2233 * - holding extent locks over multiple ost's introduces the danger of
2234 * "cascading timeouts";
2236 * - holding multiple locks over the same ost is still dead-lock prone,
2237 * see comment in osc_lock_enqueue(),
2239 * but there are certain situations where this is unavoidable:
2241 * - O_APPEND writes have to take [0, EOF] lock for correctness;
2243 * - truncate has to take [new-size, EOF] lock for correctness;
2245 * - SNS has to take locks across full stripe for correctness;
2247 * - in the case when user level buffer, supplied to {read,write}(file0),
2248 * is a part of a memory mapped lustre file, client has to take a dlm
2249 * locks on file0, and all files that back up the buffer (or a part of
2250 * the buffer, that is being processed in the current chunk, in any
2251 * case, there are situations where at least 2 locks are necessary).
2253 * In such cases we at least try to take locks in the same consistent
2254 * order. To this end, all locks are first collected, then sorted, and then
2258 /** locks to be acquired. */
2259 struct list_head cls_todo;
2260 /** locks currently being processed. */
2261 struct list_head cls_curr;
2262 /** locks acquired. */
2263 struct list_head cls_done;
2267 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
2268 * but 'req' is always to be thought as 'request' :-)
2270 enum cl_io_lock_dmd {
2271 /** Always lock data (e.g., O_APPEND). */
2273 /** Layers are free to decide between local and global locking. */
2275 /** Never lock: there is no cache (e.g., liblustre). */
2279 enum cl_fsync_mode {
2280 /** start writeback, do not wait for them to finish */
2282 /** start writeback and wait for them to finish */
2284 /** discard all of dirty pages in a specific file range */
2285 CL_FSYNC_DISCARD = 2,
2286 /** start writeback and make sure they have reached storage before
2287 * return. OST_SYNC RPC must be issued and finished */
2291 struct cl_io_rw_common {
2301 * cl_io is shared by all threads participating in this IO (in current
2302 * implementation only one thread advances IO, but parallel IO design and
2303 * concurrent copy_*_user() require multiple threads acting on the same IO. It
2304 * is up to these threads to serialize their activities, including updates to
2305 * mutable cl_io fields.
2308 /** type of this IO. Immutable after creation. */
2309 enum cl_io_type ci_type;
2310 /** current state of cl_io state machine. */
2311 enum cl_io_state ci_state;
2312 /** main object this io is against. Immutable after creation. */
2313 struct cl_object *ci_obj;
2315 * Upper layer io, of which this io is a part of. Immutable after
2318 struct cl_io *ci_parent;
2319 /** List of slices. Immutable after creation. */
2320 struct list_head ci_layers;
2321 /** list of locks (to be) acquired by this io. */
2322 struct cl_lockset ci_lockset;
2323 /** lock requirements, this is just a help info for sublayers. */
2324 enum cl_io_lock_dmd ci_lockreq;
2327 struct cl_io_rw_common rd;
2330 struct cl_io_rw_common wr;
2334 struct cl_io_rw_common ci_rw;
2335 struct cl_setattr_io {
2336 struct ost_lvb sa_attr;
2337 unsigned int sa_valid;
2338 struct obd_capa *sa_capa;
2340 struct cl_fault_io {
2341 /** page index within file. */
2343 /** bytes valid byte on a faulted page. */
2345 /** writable page? for nopage() only */
2347 /** page of an executable? */
2349 /** page_mkwrite() */
2351 /** resulting page */
2352 struct cl_page *ft_page;
2354 struct cl_fsync_io {
2357 struct obd_capa *fi_capa;
2358 /** file system level fid */
2359 struct lu_fid *fi_fid;
2360 enum cl_fsync_mode fi_mode;
2361 /* how many pages were written/discarded */
2362 unsigned int fi_nr_written;
2365 struct cl_2queue ci_queue;
2368 unsigned int ci_continue:1,
2370 * This io has held grouplock, to inform sublayers that
2371 * don't do lockless i/o.
2375 * The whole IO need to be restarted because layout has been changed
2379 * to not refresh layout - the IO issuer knows that the layout won't
2380 * change(page operations, layout change causes all page to be
2381 * discarded), or it doesn't matter if it changes(sync).
2385 * Check if layout changed after the IO finishes. Mainly for HSM
2386 * requirement. If IO occurs to openning files, it doesn't need to
2387 * verify layout because HSM won't release openning files.
2388 * Right now, only two opertaions need to verify layout: glimpse
2393 * file is released, restore has to to be triggered by vvp layer
2395 ci_restore_needed:1;
2397 * Number of pages owned by this IO. For invariant checking.
2399 unsigned ci_owned_nr;
2404 /** \addtogroup cl_req cl_req
2409 * There are two possible modes of transfer initiation on the client:
2411 * - immediate transfer: this is started when a high level io wants a page
2412 * or a collection of pages to be transferred right away. Examples:
2413 * read-ahead, synchronous read in the case of non-page aligned write,
2414 * page write-out as a part of extent lock cancellation, page write-out
2415 * as a part of memory cleansing. Immediate transfer can be both
2416 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
2418 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
2419 * when io wants to transfer a page to the server some time later, when
2420 * it can be done efficiently. Example: pages dirtied by the write(2)
2423 * In any case, transfer takes place in the form of a cl_req, which is a
2424 * representation for a network RPC.
2426 * Pages queued for an opportunistic transfer are cached until it is decided
2427 * that efficient RPC can be composed of them. This decision is made by "a
2428 * req-formation engine", currently implemented as a part of osc
2429 * layer. Req-formation depends on many factors: the size of the resulting
2430 * RPC, whether or not multi-object RPCs are supported by the server,
2431 * max-rpc-in-flight limitations, size of the dirty cache, etc.
2433 * For the immediate transfer io submits a cl_page_list, that req-formation
2434 * engine slices into cl_req's, possibly adding cached pages to some of
2435 * the resulting req's.
2437 * Whenever a page from cl_page_list is added to a newly constructed req, its
2438 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
2439 * page state is atomically changed from cl_page_state::CPS_OWNED to
2440 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
2441 * is zeroed, and cl_page::cp_req is set to the
2442 * req. cl_page_operations::cpo_prep() method at the particular layer might
2443 * return -EALREADY to indicate that it does not need to submit this page
2444 * at all. This is possible, for example, if page, submitted for read,
2445 * became up-to-date in the meantime; and for write, the page don't have
2446 * dirty bit marked. \see cl_io_submit_rw()
2448 * Whenever a cached page is added to a newly constructed req, its
2449 * cl_page_operations::cpo_make_ready() layer methods are called. At that
2450 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
2451 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
2452 * req. cl_page_operations::cpo_make_ready() method at the particular layer
2453 * might return -EAGAIN to indicate that this page is not eligible for the
2454 * transfer right now.
2458 * Plan is to divide transfers into "priority bands" (indicated when
2459 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
2460 * and allow glueing of cached pages to immediate transfers only within single
2461 * band. This would make high priority transfers (like lock cancellation or
2462 * memory pressure induced write-out) really high priority.
2467 * Per-transfer attributes.
2469 struct cl_req_attr {
2470 /** Generic attributes for the server consumption. */
2471 struct obdo *cra_oa;
2473 struct obd_capa *cra_capa;
2475 char cra_jobid[JOBSTATS_JOBID_SIZE];
2479 * Transfer request operations definable at every layer.
2481 * Concurrency: transfer formation engine synchronizes calls to all transfer
2484 struct cl_req_operations {
2486 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
2487 * complete (all pages are added).
2489 * \see osc_req_prep()
2491 int (*cro_prep)(const struct lu_env *env,
2492 const struct cl_req_slice *slice);
2494 * Called top-to-bottom to fill in \a oa fields. This is called twice
2495 * with different flags, see bug 10150 and osc_build_req().
2497 * \param obj an object from cl_req which attributes are to be set in
2500 * \param oa struct obdo where attributes are placed
2502 * \param flags \a oa fields to be filled.
2504 void (*cro_attr_set)(const struct lu_env *env,
2505 const struct cl_req_slice *slice,
2506 const struct cl_object *obj,
2507 struct cl_req_attr *attr, obd_valid flags);
2509 * Called top-to-bottom from cl_req_completion() to notify layers that
2510 * transfer completed. Has to free all state allocated by
2511 * cl_device_operations::cdo_req_init().
2513 void (*cro_completion)(const struct lu_env *env,
2514 const struct cl_req_slice *slice, int ioret);
2518 * A per-object state that (potentially multi-object) transfer request keeps.
2521 /** object itself */
2522 struct cl_object *ro_obj;
2523 /** reference to cl_req_obj::ro_obj. For debugging. */
2524 struct lu_ref_link ro_obj_ref;
2525 /* something else? Number of pages for a given object? */
2531 * Transfer requests are not reference counted, because IO sub-system owns
2532 * them exclusively and knows when to free them.
2536 * cl_req is created by cl_req_alloc() that calls
2537 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2538 * state in every layer.
2540 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2541 * contains pages for.
2543 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2544 * called top-to-bottom. At that point layers can modify req, let it pass, or
2545 * deny it completely. This is to support things like SNS that have transfer
2546 * ordering requirements invisible to the individual req-formation engine.
2548 * On transfer completion (or transfer timeout, or failure to initiate the
2549 * transfer of an allocated req), cl_req_operations::cro_completion() method
2550 * is called, after execution of cl_page_operations::cpo_completion() of all
2554 enum cl_req_type crq_type;
2555 /** A list of pages being transfered */
2556 struct list_head crq_pages;
2557 /** Number of pages in cl_req::crq_pages */
2558 unsigned crq_nrpages;
2559 /** An array of objects which pages are in ->crq_pages */
2560 struct cl_req_obj *crq_o;
2561 /** Number of elements in cl_req::crq_objs[] */
2562 unsigned crq_nrobjs;
2563 struct list_head crq_layers;
2567 * Per-layer state for request.
2569 struct cl_req_slice {
2570 struct cl_req *crs_req;
2571 struct cl_device *crs_dev;
2572 struct list_head crs_linkage;
2573 const struct cl_req_operations *crs_ops;
2578 enum cache_stats_item {
2579 /** how many cache lookups were performed */
2581 /** how many times cache lookup resulted in a hit */
2583 /** how many entities are in the cache right now */
2585 /** how many entities in the cache are actively used (and cannot be
2586 * evicted) right now */
2588 /** how many entities were created at all */
2593 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2596 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2598 struct cache_stats {
2599 const char *cs_name;
2600 atomic_t cs_stats[CS_NR];
2603 /** These are not exported so far */
2604 void cache_stats_init (struct cache_stats *cs, const char *name);
2607 * Client-side site. This represents particular client stack. "Global"
2608 * variables should (directly or indirectly) be added here to allow multiple
2609 * clients to co-exist in the single address space.
2612 struct lu_site cs_lu;
2614 * Statistical counters. Atomics do not scale, something better like
2615 * per-cpu counters is needed.
2617 * These are exported as /proc/fs/lustre/llite/.../site
2619 * When interpreting keep in mind that both sub-locks (and sub-pages)
2620 * and top-locks (and top-pages) are accounted here.
2622 struct cache_stats cs_pages;
2623 struct cache_stats cs_locks;
2624 atomic_t cs_pages_state[CPS_NR];
2625 atomic_t cs_locks_state[CLS_NR];
2628 int cl_site_init (struct cl_site *s, struct cl_device *top);
2629 void cl_site_fini (struct cl_site *s);
2630 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2633 * Output client site statistical counters into a buffer. Suitable for
2634 * ll_rd_*()-style functions.
2636 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2641 * Type conversion and accessory functions.
2645 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2647 return container_of(site, struct cl_site, cs_lu);
2650 static inline int lu_device_is_cl(const struct lu_device *d)
2652 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2655 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2657 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2658 return container_of0(d, struct cl_device, cd_lu_dev);
2661 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2663 return &d->cd_lu_dev;
2666 static inline struct cl_object *lu2cl(const struct lu_object *o)
2668 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2669 return container_of0(o, struct cl_object, co_lu);
2672 static inline const struct cl_object_conf *
2673 lu2cl_conf(const struct lu_object_conf *conf)
2675 return container_of0(conf, struct cl_object_conf, coc_lu);
2678 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2680 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2683 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2685 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2686 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2689 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2691 return container_of0(h, struct cl_object_header, coh_lu);
2694 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2696 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2700 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2702 return luh2coh(obj->co_lu.lo_header);
2705 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2707 return lu_device_init(&d->cd_lu_dev, t);
2710 static inline void cl_device_fini(struct cl_device *d)
2712 lu_device_fini(&d->cd_lu_dev);
2715 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2716 struct cl_object *obj,
2717 const struct cl_page_operations *ops);
2718 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2719 struct cl_object *obj,
2720 const struct cl_lock_operations *ops);
2721 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2722 struct cl_object *obj, const struct cl_io_operations *ops);
2723 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2724 struct cl_device *dev,
2725 const struct cl_req_operations *ops);
2728 /** \defgroup cl_object cl_object
2730 struct cl_object *cl_object_top (struct cl_object *o);
2731 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2732 const struct lu_fid *fid,
2733 const struct cl_object_conf *c);
2735 int cl_object_header_init(struct cl_object_header *h);
2736 void cl_object_header_fini(struct cl_object_header *h);
2737 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2738 void cl_object_get (struct cl_object *o);
2739 void cl_object_attr_lock (struct cl_object *o);
2740 void cl_object_attr_unlock(struct cl_object *o);
2741 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2742 struct cl_attr *attr);
2743 int cl_object_attr_set (const struct lu_env *env, struct cl_object *obj,
2744 const struct cl_attr *attr, unsigned valid);
2745 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2746 struct ost_lvb *lvb);
2747 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2748 const struct cl_object_conf *conf);
2749 void cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2750 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2751 int cl_object_has_locks (struct cl_object *obj);
2754 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2756 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2758 return cl_object_header(o0) == cl_object_header(o1);
2761 static inline void cl_object_page_init(struct cl_object *clob, int size)
2763 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2764 cl_object_header(clob)->coh_page_bufsize += ALIGN(size, 8);
2767 static inline void *cl_object_page_slice(struct cl_object *clob,
2768 struct cl_page *page)
2770 return (void *)((char *)page + clob->co_slice_off);
2775 /** \defgroup cl_page cl_page
2784 /* callback of cl_page_gang_lookup() */
2785 typedef int (*cl_page_gang_cb_t) (const struct lu_env *, struct cl_io *,
2786 struct cl_page *, void *);
2787 int cl_page_gang_lookup (const struct lu_env *env,
2788 struct cl_object *obj,
2790 pgoff_t start, pgoff_t end,
2791 cl_page_gang_cb_t cb, void *cbdata);
2792 struct cl_page *cl_page_lookup (struct cl_object_header *hdr,
2794 struct cl_page *cl_page_find (const struct lu_env *env,
2795 struct cl_object *obj,
2796 pgoff_t idx, struct page *vmpage,
2797 enum cl_page_type type);
2798 struct cl_page *cl_page_find_sub (const struct lu_env *env,
2799 struct cl_object *obj,
2800 pgoff_t idx, struct page *vmpage,
2801 struct cl_page *parent);
2802 void cl_page_get (struct cl_page *page);
2803 void cl_page_put (const struct lu_env *env,
2804 struct cl_page *page);
2805 void cl_page_print (const struct lu_env *env, void *cookie,
2806 lu_printer_t printer,
2807 const struct cl_page *pg);
2808 void cl_page_header_print(const struct lu_env *env, void *cookie,
2809 lu_printer_t printer,
2810 const struct cl_page *pg);
2811 struct page *cl_page_vmpage (const struct lu_env *env,
2812 struct cl_page *page);
2813 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2814 struct cl_page *cl_page_top (struct cl_page *page);
2816 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2817 const struct lu_device_type *dtype);
2822 * Functions dealing with the ownership of page by io.
2826 int cl_page_own (const struct lu_env *env,
2827 struct cl_io *io, struct cl_page *page);
2828 int cl_page_own_try (const struct lu_env *env,
2829 struct cl_io *io, struct cl_page *page);
2830 void cl_page_assume (const struct lu_env *env,
2831 struct cl_io *io, struct cl_page *page);
2832 void cl_page_unassume (const struct lu_env *env,
2833 struct cl_io *io, struct cl_page *pg);
2834 void cl_page_disown (const struct lu_env *env,
2835 struct cl_io *io, struct cl_page *page);
2836 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2843 * Functions dealing with the preparation of a page for a transfer, and
2844 * tracking transfer state.
2847 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2848 struct cl_page *pg, enum cl_req_type crt);
2849 void cl_page_completion (const struct lu_env *env,
2850 struct cl_page *pg, enum cl_req_type crt, int ioret);
2851 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2852 enum cl_req_type crt);
2853 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2854 struct cl_page *pg, enum cl_req_type crt);
2855 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2857 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2858 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2859 struct cl_page *pg);
2865 * \name helper routines
2866 * Functions to discard, delete and export a cl_page.
2869 void cl_page_discard (const struct lu_env *env, struct cl_io *io,
2870 struct cl_page *pg);
2871 void cl_page_delete (const struct lu_env *env, struct cl_page *pg);
2872 int cl_page_unmap (const struct lu_env *env, struct cl_io *io,
2873 struct cl_page *pg);
2874 int cl_page_is_vmlocked (const struct lu_env *env,
2875 const struct cl_page *pg);
2876 void cl_page_export (const struct lu_env *env,
2877 struct cl_page *pg, int uptodate);
2878 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2879 struct cl_page *page);
2880 loff_t cl_offset (const struct cl_object *obj, pgoff_t idx);
2881 pgoff_t cl_index (const struct cl_object *obj, loff_t offset);
2882 int cl_page_size (const struct cl_object *obj);
2883 int cl_pages_prune (const struct lu_env *env, struct cl_object *obj);
2885 void cl_lock_print (const struct lu_env *env, void *cookie,
2886 lu_printer_t printer, const struct cl_lock *lock);
2887 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2888 lu_printer_t printer,
2889 const struct cl_lock_descr *descr);
2894 /** \defgroup cl_lock cl_lock
2897 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2898 const struct cl_lock_descr *need,
2899 const char *scope, const void *source);
2900 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
2901 const struct cl_lock_descr *need,
2902 const char *scope, const void *source);
2903 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2904 const struct cl_lock_descr *need,
2905 const char *scope, const void *source);
2906 struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env,
2907 struct cl_object *obj, pgoff_t index,
2908 struct cl_lock *except, int pending,
2910 static inline struct cl_lock *cl_lock_at_page(const struct lu_env *env,
2911 struct cl_object *obj,
2912 struct cl_page *page,
2913 struct cl_lock *except,
2914 int pending, int canceld)
2916 LASSERT(cl_object_header(obj) == cl_object_header(page->cp_obj));
2917 return cl_lock_at_pgoff(env, obj, page->cp_index, except,
2921 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2922 const struct lu_device_type *dtype);
2924 void cl_lock_get (struct cl_lock *lock);
2925 void cl_lock_get_trust (struct cl_lock *lock);
2926 void cl_lock_put (const struct lu_env *env, struct cl_lock *lock);
2927 void cl_lock_hold_add (const struct lu_env *env, struct cl_lock *lock,
2928 const char *scope, const void *source);
2929 void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
2930 const char *scope, const void *source);
2931 void cl_lock_unhold (const struct lu_env *env, struct cl_lock *lock,
2932 const char *scope, const void *source);
2933 void cl_lock_release (const struct lu_env *env, struct cl_lock *lock,
2934 const char *scope, const void *source);
2935 void cl_lock_user_add (const struct lu_env *env, struct cl_lock *lock);
2936 void cl_lock_user_del (const struct lu_env *env, struct cl_lock *lock);
2938 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
2939 struct cl_lock *lock);
2940 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
2941 enum cl_lock_state state);
2942 int cl_lock_is_intransit(struct cl_lock *lock);
2944 int cl_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock,
2947 /** \name statemachine statemachine
2948 * Interface to lock state machine consists of 3 parts:
2950 * - "try" functions that attempt to effect a state transition. If state
2951 * transition is not possible right now (e.g., if it has to wait for some
2952 * asynchronous event to occur), these functions return
2953 * cl_lock_transition::CLO_WAIT.
2955 * - "non-try" functions that implement synchronous blocking interface on
2956 * top of non-blocking "try" functions. These functions repeatedly call
2957 * corresponding "try" versions, and if state transition is not possible
2958 * immediately, wait for lock state change.
2960 * - methods from cl_lock_operations, called by "try" functions. Lock can
2961 * be advanced to the target state only when all layers voted that they
2962 * are ready for this transition. "Try" functions call methods under lock
2963 * mutex. If a layer had to release a mutex, it re-acquires it and returns
2964 * cl_lock_transition::CLO_REPEAT, causing "try" function to call all
2967 * TRY NON-TRY METHOD FINAL STATE
2969 * cl_enqueue_try() cl_enqueue() cl_lock_operations::clo_enqueue() CLS_ENQUEUED
2971 * cl_wait_try() cl_wait() cl_lock_operations::clo_wait() CLS_HELD
2973 * cl_unuse_try() cl_unuse() cl_lock_operations::clo_unuse() CLS_CACHED
2975 * cl_use_try() NONE cl_lock_operations::clo_use() CLS_HELD
2979 int cl_enqueue (const struct lu_env *env, struct cl_lock *lock,
2980 struct cl_io *io, __u32 flags);
2981 int cl_wait (const struct lu_env *env, struct cl_lock *lock);
2982 void cl_unuse (const struct lu_env *env, struct cl_lock *lock);
2983 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
2984 struct cl_io *io, __u32 flags);
2985 int cl_unuse_try (const struct lu_env *env, struct cl_lock *lock);
2986 int cl_wait_try (const struct lu_env *env, struct cl_lock *lock);
2987 int cl_use_try (const struct lu_env *env, struct cl_lock *lock, int atomic);
2989 /** @} statemachine */
2991 void cl_lock_signal (const struct lu_env *env, struct cl_lock *lock);
2992 int cl_lock_state_wait (const struct lu_env *env, struct cl_lock *lock);
2993 void cl_lock_state_set (const struct lu_env *env, struct cl_lock *lock,
2994 enum cl_lock_state state);
2995 int cl_queue_match (const struct list_head *queue,
2996 const struct cl_lock_descr *need);
2998 void cl_lock_mutex_get (const struct lu_env *env, struct cl_lock *lock);
2999 int cl_lock_mutex_try (const struct lu_env *env, struct cl_lock *lock);
3000 void cl_lock_mutex_put (const struct lu_env *env, struct cl_lock *lock);
3001 int cl_lock_is_mutexed (struct cl_lock *lock);
3002 int cl_lock_nr_mutexed (const struct lu_env *env);
3003 int cl_lock_discard_pages(const struct lu_env *env, struct cl_lock *lock);
3004 int cl_lock_ext_match (const struct cl_lock_descr *has,
3005 const struct cl_lock_descr *need);
3006 int cl_lock_descr_match(const struct cl_lock_descr *has,
3007 const struct cl_lock_descr *need);
3008 int cl_lock_mode_match (enum cl_lock_mode has, enum cl_lock_mode need);
3009 int cl_lock_modify (const struct lu_env *env, struct cl_lock *lock,
3010 const struct cl_lock_descr *desc);
3012 void cl_lock_closure_init (const struct lu_env *env,
3013 struct cl_lock_closure *closure,
3014 struct cl_lock *origin, int wait);
3015 void cl_lock_closure_fini (struct cl_lock_closure *closure);
3016 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
3017 struct cl_lock_closure *closure);
3018 void cl_lock_disclosure (const struct lu_env *env,
3019 struct cl_lock_closure *closure);
3020 int cl_lock_enclosure (const struct lu_env *env, struct cl_lock *lock,
3021 struct cl_lock_closure *closure);
3023 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
3024 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock);
3025 void cl_lock_error (const struct lu_env *env, struct cl_lock *lock, int error);
3026 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int wait);
3028 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock);
3032 /** \defgroup cl_io cl_io
3035 int cl_io_init (const struct lu_env *env, struct cl_io *io,
3036 enum cl_io_type iot, struct cl_object *obj);
3037 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
3038 enum cl_io_type iot, struct cl_object *obj);
3039 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
3040 enum cl_io_type iot, loff_t pos, size_t count);
3041 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
3043 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
3044 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
3045 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
3046 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
3047 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
3048 int cl_io_start (const struct lu_env *env, struct cl_io *io);
3049 void cl_io_end (const struct lu_env *env, struct cl_io *io);
3050 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
3051 struct cl_io_lock_link *link);
3052 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
3053 struct cl_lock_descr *descr);
3054 int cl_io_read_page (const struct lu_env *env, struct cl_io *io,
3055 struct cl_page *page);
3056 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
3057 struct cl_page *page, unsigned from, unsigned to);
3058 int cl_io_commit_write (const struct lu_env *env, struct cl_io *io,
3059 struct cl_page *page, unsigned from, unsigned to);
3060 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
3061 enum cl_req_type iot, struct cl_2queue *queue);
3062 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
3063 enum cl_req_type iot, struct cl_2queue *queue,
3065 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
3067 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
3068 struct cl_page_list *queue);
3069 int cl_io_is_going (const struct lu_env *env);
3072 * True, iff \a io is an O_APPEND write(2).
3074 static inline int cl_io_is_append(const struct cl_io *io)
3076 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
3079 static inline int cl_io_is_sync_write(const struct cl_io *io)
3081 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
3084 static inline int cl_io_is_mkwrite(const struct cl_io *io)
3086 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
3090 * True, iff \a io is a truncate(2).
3092 static inline int cl_io_is_trunc(const struct cl_io *io)
3094 return io->ci_type == CIT_SETATTR &&
3095 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
3098 struct cl_io *cl_io_top(struct cl_io *io);
3100 void cl_io_print(const struct lu_env *env, void *cookie,
3101 lu_printer_t printer, const struct cl_io *io);
3103 #define CL_IO_SLICE_CLEAN(foo_io, base) \
3105 typeof(foo_io) __foo_io = (foo_io); \
3107 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
3108 memset(&__foo_io->base + 1, 0, \
3109 sizeof(*__foo_io) - sizeof(__foo_io->base)); \
3114 /** \defgroup cl_page_list cl_page_list
3118 * Last page in the page list.
3120 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
3122 LASSERT(plist->pl_nr > 0);
3123 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
3127 * Iterate over pages in a page list.
3129 #define cl_page_list_for_each(page, list) \
3130 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
3133 * Iterate over pages in a page list, taking possible removals into account.
3135 #define cl_page_list_for_each_safe(page, temp, list) \
3136 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
3138 void cl_page_list_init (struct cl_page_list *plist);
3139 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
3140 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
3141 struct cl_page *page);
3142 void cl_page_list_splice (struct cl_page_list *list,
3143 struct cl_page_list *head);
3144 void cl_page_list_del (const struct lu_env *env,
3145 struct cl_page_list *plist, struct cl_page *page);
3146 void cl_page_list_disown (const struct lu_env *env,
3147 struct cl_io *io, struct cl_page_list *plist);
3148 int cl_page_list_own (const struct lu_env *env,
3149 struct cl_io *io, struct cl_page_list *plist);
3150 void cl_page_list_assume (const struct lu_env *env,
3151 struct cl_io *io, struct cl_page_list *plist);
3152 void cl_page_list_discard(const struct lu_env *env,
3153 struct cl_io *io, struct cl_page_list *plist);
3154 int cl_page_list_unmap (const struct lu_env *env,
3155 struct cl_io *io, struct cl_page_list *plist);
3156 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
3158 void cl_2queue_init (struct cl_2queue *queue);
3159 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
3160 void cl_2queue_disown (const struct lu_env *env,
3161 struct cl_io *io, struct cl_2queue *queue);
3162 void cl_2queue_assume (const struct lu_env *env,
3163 struct cl_io *io, struct cl_2queue *queue);
3164 void cl_2queue_discard (const struct lu_env *env,
3165 struct cl_io *io, struct cl_2queue *queue);
3166 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
3167 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
3169 /** @} cl_page_list */
3171 /** \defgroup cl_req cl_req
3173 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
3174 enum cl_req_type crt, int nr_objects);
3176 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
3177 struct cl_page *page);
3178 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
3179 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
3180 void cl_req_attr_set (const struct lu_env *env, struct cl_req *req,
3181 struct cl_req_attr *attr, obd_valid flags);
3182 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
3184 /** \defgroup cl_sync_io cl_sync_io
3188 * Anchor for synchronous transfer. This is allocated on a stack by thread
3189 * doing synchronous transfer, and a pointer to this structure is set up in
3190 * every page submitted for transfer. Transfer completion routine updates
3191 * anchor and wakes up waiting thread when transfer is complete.
3194 /** number of pages yet to be transferred. */
3195 atomic_t csi_sync_nr;
3198 /** barrier of destroy this structure */
3199 atomic_t csi_barrier;
3200 /** completion to be signaled when transfer is complete. */
3201 wait_queue_head_t csi_waitq;
3204 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages);
3205 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
3206 struct cl_page_list *queue, struct cl_sync_io *anchor,
3208 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret);
3210 /** @} cl_sync_io */
3214 /** \defgroup cl_env cl_env
3216 * lu_env handling for a client.
3218 * lu_env is an environment within which lustre code executes. Its major part
3219 * is lu_context---a fast memory allocation mechanism that is used to conserve
3220 * precious kernel stack space. Originally lu_env was designed for a server,
3223 * - there is a (mostly) fixed number of threads, and
3225 * - call chains have no non-lustre portions inserted between lustre code.
3227 * On a client both these assumtpion fails, because every user thread can
3228 * potentially execute lustre code as part of a system call, and lustre calls
3229 * into VFS or MM that call back into lustre.
3231 * To deal with that, cl_env wrapper functions implement the following
3234 * - allocation and destruction of environment is amortized by caching no
3235 * longer used environments instead of destroying them;
3237 * - there is a notion of "current" environment, attached to the kernel
3238 * data structure representing current thread Top-level lustre code
3239 * allocates an environment and makes it current, then calls into
3240 * non-lustre code, that in turn calls lustre back. Low-level lustre
3241 * code thus called can fetch environment created by the top-level code
3242 * and reuse it, avoiding additional environment allocation.
3243 * Right now, three interfaces can attach the cl_env to running thread:
3246 * - cl_env_reexit(cl_env_reenter had to be called priorly)
3248 * \see lu_env, lu_context, lu_context_key
3251 struct cl_env_nest {
3256 struct lu_env *cl_env_peek (int *refcheck);
3257 struct lu_env *cl_env_get (int *refcheck);
3258 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
3259 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
3260 void cl_env_put (struct lu_env *env, int *refcheck);
3261 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
3262 void *cl_env_reenter (void);
3263 void cl_env_reexit (void *cookie);
3264 void cl_env_implant (struct lu_env *env, int *refcheck);
3265 void cl_env_unplant (struct lu_env *env, int *refcheck);
3272 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
3273 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
3275 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
3276 struct lu_device_type *ldt,
3277 struct lu_device *next);
3280 int cl_global_init(void);
3281 void cl_global_fini(void);
3283 #endif /* _LINUX_CL_OBJECT_H */