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, 2014, 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
87 * See the top comment in cl_object.c for the description of overall locking and
88 * reference-counting design.
90 * See comments below for the description of i/o, page, and dlm-locking
97 * super-class definitions.
99 #include <libcfs/libcfs.h>
100 #include <lu_object.h>
101 #include <linux/atomic.h>
102 #include <linux/mutex.h>
103 #include <linux/radix-tree.h>
104 #include <linux/spinlock.h>
105 #include <linux/wait.h>
106 #include <lustre_dlm.h>
112 struct cl_device_operations;
115 struct cl_object_page_operations;
116 struct cl_object_lock_operations;
119 struct cl_page_slice;
121 struct cl_lock_slice;
123 struct cl_lock_operations;
124 struct cl_page_operations;
133 * Operations for each data device in the client stack.
135 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
137 struct cl_device_operations {
139 * Initialize cl_req. This method is called top-to-bottom on all
140 * devices in the stack to get them a chance to allocate layer-private
141 * data, and to attach them to the cl_req by calling
142 * cl_req_slice_add().
144 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
145 * \see vvp_req_init()
147 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
152 * Device in the client stack.
154 * \see vvp_device, lov_device, lovsub_device, osc_device
158 struct lu_device cd_lu_dev;
159 /** Per-layer operation vector. */
160 const struct cl_device_operations *cd_ops;
163 /** \addtogroup cl_object cl_object
166 * "Data attributes" of cl_object. Data attributes can be updated
167 * independently for a sub-object, and top-object's attributes are calculated
168 * from sub-objects' ones.
171 /** Object size, in bytes */
174 * Known minimal size, in bytes.
176 * This is only valid when at least one DLM lock is held.
179 /** Modification time. Measured in seconds since epoch. */
181 /** Access time. Measured in seconds since epoch. */
183 /** Change time. Measured in seconds since epoch. */
186 * Blocks allocated to this cl_object on the server file system.
188 * \todo XXX An interface for block size is needed.
192 * User identifier for quota purposes.
196 * Group identifier for quota purposes.
200 /* nlink of the directory */
205 * Fields in cl_attr that are being set.
219 * Sub-class of lu_object with methods common for objects on the client
222 * cl_object: represents a regular file system object, both a file and a
223 * stripe. cl_object is based on lu_object: it is identified by a fid,
224 * layered, cached, hashed, and lrued. Important distinction with the server
225 * side, where md_object and dt_object are used, is that cl_object "fans out"
226 * at the lov/sns level: depending on the file layout, single file is
227 * represented as a set of "sub-objects" (stripes). At the implementation
228 * level, struct lov_object contains an array of cl_objects. Each sub-object
229 * is a full-fledged cl_object, having its fid, living in the lru and hash
232 * This leads to the next important difference with the server side: on the
233 * client, it's quite usual to have objects with the different sequence of
234 * layers. For example, typical top-object is composed of the following
240 * whereas its sub-objects are composed of
245 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
246 * track of the object-subobject relationship.
248 * Sub-objects are not cached independently: when top-object is about to
249 * be discarded from the memory, all its sub-objects are torn-down and
252 * \see vvp_object, lov_object, lovsub_object, osc_object
256 struct lu_object co_lu;
257 /** per-object-layer operations */
258 const struct cl_object_operations *co_ops;
259 /** offset of page slice in cl_page buffer */
264 * Description of the client object configuration. This is used for the
265 * creation of a new client object that is identified by a more state than
268 struct cl_object_conf {
270 struct lu_object_conf coc_lu;
273 * Object layout. This is consumed by lov.
275 struct lustre_md *coc_md;
277 * Description of particular stripe location in the
278 * cluster. This is consumed by osc.
280 struct lov_oinfo *coc_oinfo;
283 * VFS inode. This is consumed by vvp.
285 struct inode *coc_inode;
287 * Layout lock handle.
289 struct ldlm_lock *coc_lock;
291 * Operation to handle layout, OBJECT_CONF_XYZ.
297 /** configure layout, set up a new stripe, must be called while
298 * holding layout lock. */
300 /** invalidate the current stripe configuration due to losing
302 OBJECT_CONF_INVALIDATE = 1,
303 /** wait for old layout to go away so that new layout can be
309 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
310 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
314 /** the buffer to return the layout in lov_mds_md format. */
315 struct lu_buf cl_buf;
316 /** size of layout in lov_mds_md format. */
318 /** Layout generation. */
320 /** True if this is a released file.
321 * Temporarily added for released file truncate in ll_setattr_raw().
322 * It will be removed later. -Jinshan */
327 * Operations implemented for each cl object layer.
329 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
331 struct cl_object_operations {
333 * Initialize page slice for this layer. Called top-to-bottom through
334 * every object layer when a new cl_page is instantiated. Layer
335 * keeping private per-page data, or requiring its own page operations
336 * vector should allocate these data here, and attach then to the page
337 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
340 * \retval NULL success.
342 * \retval ERR_PTR(errno) failure code.
344 * \retval valid-pointer pointer to already existing referenced page
345 * to be used instead of newly created.
347 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
348 struct cl_page *page, pgoff_t index);
350 * Initialize lock slice for this layer. Called top-to-bottom through
351 * every object layer when a new cl_lock is instantiated. Layer
352 * keeping private per-lock data, or requiring its own lock operations
353 * vector should allocate these data here, and attach then to the lock
354 * by calling cl_lock_slice_add(). Mandatory.
356 int (*coo_lock_init)(const struct lu_env *env,
357 struct cl_object *obj, struct cl_lock *lock,
358 const struct cl_io *io);
360 * Initialize io state for a given layer.
362 * called top-to-bottom once per io existence to initialize io
363 * state. If layer wants to keep some state for this type of io, it
364 * has to embed struct cl_io_slice in lu_env::le_ses, and register
365 * slice with cl_io_slice_add(). It is guaranteed that all threads
366 * participating in this io share the same session.
368 int (*coo_io_init)(const struct lu_env *env,
369 struct cl_object *obj, struct cl_io *io);
371 * Fill portion of \a attr that this layer controls. This method is
372 * called top-to-bottom through all object layers.
374 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
376 * \return 0: to continue
377 * \return +ve: to stop iterating through layers (but 0 is returned
378 * from enclosing cl_object_attr_get())
379 * \return -ve: to signal error
381 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
382 struct cl_attr *attr);
386 * \a valid is a bitmask composed from enum #cl_attr_valid, and
387 * indicating what attributes are to be set.
389 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
391 * \return the same convention as for
392 * cl_object_operations::coo_attr_get() is used.
394 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
395 const struct cl_attr *attr, unsigned valid);
397 * Update object configuration. Called top-to-bottom to modify object
400 * XXX error conditions and handling.
402 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
403 const struct cl_object_conf *conf);
405 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
406 * object. Layers are supposed to fill parts of \a lvb that will be
407 * shipped to the glimpse originator as a glimpse result.
409 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
410 * \see osc_object_glimpse()
412 int (*coo_glimpse)(const struct lu_env *env,
413 const struct cl_object *obj, struct ost_lvb *lvb);
415 * Object prune method. Called when the layout is going to change on
416 * this object, therefore each layer has to clean up their cache,
417 * mainly pages and locks.
419 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
421 * Object getstripe method.
423 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
424 struct lov_user_md __user *lum);
426 * Find whether there is any callback data (ldlm lock) attached upon
429 int (*coo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
430 ldlm_iterator_t iter, void *data);
432 * Get FIEMAP mapping from the object.
434 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
435 struct ll_fiemap_info_key *fmkey,
436 struct fiemap *fiemap, size_t *buflen);
438 * Get attributes of the object from server. (top->bottom)
440 int (*coo_obd_info_get)(const struct lu_env *env, struct cl_object *obj,
441 struct obd_info *oinfo,
442 struct ptlrpc_request_set *set);
444 * Get data version of the object. (top->bottom)
446 int (*coo_data_version)(const struct lu_env *env, struct cl_object *obj,
447 __u64 *version, int flags);
449 * Get layout and generation of the object.
451 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
452 struct cl_layout *layout);
454 * Get maximum size of the object.
456 loff_t (*coo_maxbytes)(struct cl_object *obj);
460 * Extended header for client object.
462 struct cl_object_header {
463 /** Standard lu_object_header. cl_object::co_lu::lo_header points
465 struct lu_object_header coh_lu;
468 * Parent object. It is assumed that an object has a well-defined
469 * parent, but not a well-defined child (there may be multiple
470 * sub-objects, for the same top-object). cl_object_header::coh_parent
471 * field allows certain code to be written generically, without
472 * limiting possible cl_object layouts unduly.
474 struct cl_object_header *coh_parent;
476 * Protects consistency between cl_attr of parent object and
477 * attributes of sub-objects, that the former is calculated ("merged")
480 * \todo XXX this can be read/write lock if needed.
482 spinlock_t coh_attr_guard;
484 * Size of cl_page + page slices
486 unsigned short coh_page_bufsize;
488 * Number of objects above this one: 0 for a top-object, 1 for its
491 unsigned char coh_nesting;
495 * Helper macro: iterate over all layers of the object \a obj, assigning every
496 * layer top-to-bottom to \a slice.
498 #define cl_object_for_each(slice, obj) \
499 list_for_each_entry((slice), \
500 &(obj)->co_lu.lo_header->loh_layers,\
504 * Helper macro: iterate over all layers of the object \a obj, assigning every
505 * layer bottom-to-top to \a slice.
507 #define cl_object_for_each_reverse(slice, obj) \
508 list_for_each_entry_reverse((slice), \
509 &(obj)->co_lu.lo_header->loh_layers,\
514 #define CL_PAGE_EOF ((pgoff_t)~0ull)
516 /** \addtogroup cl_page cl_page
520 * Layered client page.
522 * cl_page: represents a portion of a file, cached in the memory. All pages
523 * of the given file are of the same size, and are kept in the radix tree
524 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
525 * of the top-level file object are first class cl_objects, they have their
526 * own radix trees of pages and hence page is implemented as a sequence of
527 * struct cl_pages's, linked into double-linked list through
528 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
529 * corresponding radix tree at the corresponding logical offset.
531 * cl_page is associated with VM page of the hosting environment (struct
532 * page in Linux kernel, for example), struct page. It is assumed, that this
533 * association is implemented by one of cl_page layers (top layer in the
534 * current design) that
536 * - intercepts per-VM-page call-backs made by the environment (e.g.,
539 * - translates state (page flag bits) and locking between lustre and
542 * The association between cl_page and struct page is immutable and
543 * established when cl_page is created.
545 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
546 * this io an exclusive access to this page w.r.t. other io attempts and
547 * various events changing page state (such as transfer completion, or
548 * eviction of the page from the memory). Note, that in general cl_io
549 * cannot be identified with a particular thread, and page ownership is not
550 * exactly equal to the current thread holding a lock on the page. Layer
551 * implementing association between cl_page and struct page has to implement
552 * ownership on top of available synchronization mechanisms.
554 * While lustre client maintains the notion of an page ownership by io,
555 * hosting MM/VM usually has its own page concurrency control
556 * mechanisms. For example, in Linux, page access is synchronized by the
557 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
558 * takes care to acquire and release such locks as necessary around the
559 * calls to the file system methods (->readpage(), ->prepare_write(),
560 * ->commit_write(), etc.). This leads to the situation when there are two
561 * different ways to own a page in the client:
563 * - client code explicitly and voluntary owns the page (cl_page_own());
565 * - VM locks a page and then calls the client, that has "to assume"
566 * the ownership from the VM (cl_page_assume()).
568 * Dual methods to release ownership are cl_page_disown() and
569 * cl_page_unassume().
571 * cl_page is reference counted (cl_page::cp_ref). When reference counter
572 * drops to 0, the page is returned to the cache, unless it is in
573 * cl_page_state::CPS_FREEING state, in which case it is immediately
576 * The general logic guaranteeing the absence of "existential races" for
577 * pages is the following:
579 * - there are fixed known ways for a thread to obtain a new reference
582 * - by doing a lookup in the cl_object radix tree, protected by the
585 * - by starting from VM-locked struct page and following some
586 * hosting environment method (e.g., following ->private pointer in
587 * the case of Linux kernel), see cl_vmpage_page();
589 * - when the page enters cl_page_state::CPS_FREEING state, all these
590 * ways are severed with the proper synchronization
591 * (cl_page_delete());
593 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
596 * - no new references to the page in cl_page_state::CPS_FREEING state
597 * are allowed (checked in cl_page_get()).
599 * Together this guarantees that when last reference to a
600 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
601 * page, as neither references to it can be acquired at that point, nor
604 * cl_page is a state machine. States are enumerated in enum
605 * cl_page_state. Possible state transitions are enumerated in
606 * cl_page_state_set(). State transition process (i.e., actual changing of
607 * cl_page::cp_state field) is protected by the lock on the underlying VM
610 * Linux Kernel implementation.
612 * Binding between cl_page and struct page (which is a typedef for
613 * struct page) is implemented in the vvp layer. cl_page is attached to the
614 * ->private pointer of the struct page, together with the setting of
615 * PG_private bit in page->flags, and acquiring additional reference on the
616 * struct page (much like struct buffer_head, or any similar file system
617 * private data structures).
619 * PG_locked lock is used to implement both ownership and transfer
620 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
621 * states. No additional references are acquired for the duration of the
624 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
625 * write-out is "protected" by the special PG_writeback bit.
629 * States of cl_page. cl_page.c assumes particular order here.
631 * The page state machine is rather crude, as it doesn't recognize finer page
632 * states like "dirty" or "up to date". This is because such states are not
633 * always well defined for the whole stack (see, for example, the
634 * implementation of the read-ahead, that hides page up-to-dateness to track
635 * cache hits accurately). Such sub-states are maintained by the layers that
636 * are interested in them.
640 * Page is in the cache, un-owned. Page leaves cached state in the
643 * - [cl_page_state::CPS_OWNED] io comes across the page and
646 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
647 * req-formation engine decides that it wants to include this page
648 * into an cl_req being constructed, and yanks it from the cache;
650 * - [cl_page_state::CPS_FREEING] VM callback is executed to
651 * evict the page form the memory;
653 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
657 * Page is exclusively owned by some cl_io. Page may end up in this
658 * state as a result of
660 * - io creating new page and immediately owning it;
662 * - [cl_page_state::CPS_CACHED] io finding existing cached page
665 * - [cl_page_state::CPS_OWNED] io finding existing owned page
666 * and waiting for owner to release the page;
668 * Page leaves owned state in the following cases:
670 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
671 * the cache, doing nothing;
673 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
676 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
677 * transfer for this page;
679 * - [cl_page_state::CPS_FREEING] io decides to destroy this
680 * page (e.g., as part of truncate or extent lock cancellation).
682 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
686 * Page is being written out, as a part of a transfer. This state is
687 * entered when req-formation logic decided that it wants this page to
688 * be sent through the wire _now_. Specifically, it means that once
689 * this state is achieved, transfer completion handler (with either
690 * success or failure indication) is guaranteed to be executed against
691 * this page independently of any locks and any scheduling decisions
692 * made by the hosting environment (that effectively means that the
693 * page is never put into cl_page_state::CPS_PAGEOUT state "in
694 * advance". This property is mentioned, because it is important when
695 * reasoning about possible dead-locks in the system). The page can
696 * enter this state as a result of
698 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
699 * write-out of this page, or
701 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
702 * that it has enough dirty pages cached to issue a "good"
705 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
706 * is completed---it is moved into cl_page_state::CPS_CACHED state.
708 * Underlying VM page is locked for the duration of transfer.
710 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
714 * Page is being read in, as a part of a transfer. This is quite
715 * similar to the cl_page_state::CPS_PAGEOUT state, except that
716 * read-in is always "immediate"---there is no such thing a sudden
717 * construction of read cl_req from cached, presumably not up to date,
720 * Underlying VM page is locked for the duration of transfer.
722 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
726 * Page is being destroyed. This state is entered when client decides
727 * that page has to be deleted from its host object, as, e.g., a part
730 * Once this state is reached, there is no way to escape it.
732 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
739 /** Host page, the page is from the host inode which the cl_page
743 /** Transient page, the transient cl_page is used to bind a cl_page
744 * to vmpage which is not belonging to the same object of cl_page.
745 * it is used in DirectIO, lockless IO and liblustre. */
750 * Fields are protected by the lock on struct page, except for atomics and
753 * \invariant Data type invariants are in cl_page_invariant(). Basically:
754 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
755 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
756 * cl_page::cp_owner (when set).
759 /** Reference counter. */
761 /** Transfer error. */
763 /** An object this page is a part of. Immutable after creation. */
764 struct cl_object *cp_obj;
766 struct page *cp_vmpage;
767 /** Linkage of pages within group. Pages must be owned */
768 struct list_head cp_batch;
769 /** List of slices. Immutable after creation. */
770 struct list_head cp_layers;
771 /** Linkage of pages within cl_req. */
772 struct list_head cp_flight;
774 * Page state. This field is const to avoid accidental update, it is
775 * modified only internally within cl_page.c. Protected by a VM lock.
777 const enum cl_page_state cp_state;
779 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
782 enum cl_page_type cp_type;
785 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
786 * by sub-io. Protected by a VM lock.
788 struct cl_io *cp_owner;
790 * Owning IO request in cl_page_state::CPS_PAGEOUT and
791 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
792 * the top-level pages. Protected by a VM lock.
794 struct cl_req *cp_req;
795 /** List of references to this page, for debugging. */
796 struct lu_ref cp_reference;
797 /** Link to an object, for debugging. */
798 struct lu_ref_link cp_obj_ref;
799 /** Link to a queue, for debugging. */
800 struct lu_ref_link cp_queue_ref;
801 /** Assigned if doing a sync_io */
802 struct cl_sync_io *cp_sync_io;
806 * Per-layer part of cl_page.
808 * \see vvp_page, lov_page, osc_page
810 struct cl_page_slice {
811 struct cl_page *cpl_page;
814 * Object slice corresponding to this page slice. Immutable after
817 struct cl_object *cpl_obj;
818 const struct cl_page_operations *cpl_ops;
819 /** Linkage into cl_page::cp_layers. Immutable after creation. */
820 struct list_head cpl_linkage;
824 * Lock mode. For the client extent locks.
836 * Requested transfer type.
846 * Per-layer page operations.
848 * Methods taking an \a io argument are for the activity happening in the
849 * context of given \a io. Page is assumed to be owned by that io, except for
850 * the obvious cases (like cl_page_operations::cpo_own()).
852 * \see vvp_page_ops, lov_page_ops, osc_page_ops
854 struct cl_page_operations {
856 * cl_page<->struct page methods. Only one layer in the stack has to
857 * implement these. Current code assumes that this functionality is
858 * provided by the topmost layer, see cl_page_disown0() as an example.
862 * Called when \a io acquires this page into the exclusive
863 * ownership. When this method returns, it is guaranteed that the is
864 * not owned by other io, and no transfer is going on against
868 * \see vvp_page_own(), lov_page_own()
870 int (*cpo_own)(const struct lu_env *env,
871 const struct cl_page_slice *slice,
872 struct cl_io *io, int nonblock);
873 /** Called when ownership it yielded. Optional.
875 * \see cl_page_disown()
876 * \see vvp_page_disown()
878 void (*cpo_disown)(const struct lu_env *env,
879 const struct cl_page_slice *slice, struct cl_io *io);
881 * Called for a page that is already "owned" by \a io from VM point of
884 * \see cl_page_assume()
885 * \see vvp_page_assume(), lov_page_assume()
887 void (*cpo_assume)(const struct lu_env *env,
888 const struct cl_page_slice *slice, struct cl_io *io);
889 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
890 * bottom-to-top when IO releases a page without actually unlocking
893 * \see cl_page_unassume()
894 * \see vvp_page_unassume()
896 void (*cpo_unassume)(const struct lu_env *env,
897 const struct cl_page_slice *slice,
900 * Announces whether the page contains valid data or not by \a uptodate.
902 * \see cl_page_export()
903 * \see vvp_page_export()
905 void (*cpo_export)(const struct lu_env *env,
906 const struct cl_page_slice *slice, int uptodate);
908 * Checks whether underlying VM page is locked (in the suitable
909 * sense). Used for assertions.
911 * \retval -EBUSY: page is protected by a lock of a given mode;
912 * \retval -ENODATA: page is not protected by a lock;
913 * \retval 0: this layer cannot decide. (Should never happen.)
915 int (*cpo_is_vmlocked)(const struct lu_env *env,
916 const struct cl_page_slice *slice);
922 * Called when page is truncated from the object. Optional.
924 * \see cl_page_discard()
925 * \see vvp_page_discard(), osc_page_discard()
927 void (*cpo_discard)(const struct lu_env *env,
928 const struct cl_page_slice *slice,
931 * Called when page is removed from the cache, and is about to being
932 * destroyed. Optional.
934 * \see cl_page_delete()
935 * \see vvp_page_delete(), osc_page_delete()
937 void (*cpo_delete)(const struct lu_env *env,
938 const struct cl_page_slice *slice);
939 /** Destructor. Frees resources and slice itself. */
940 void (*cpo_fini)(const struct lu_env *env,
941 struct cl_page_slice *slice);
943 * Optional debugging helper. Prints given page slice.
945 * \see cl_page_print()
947 int (*cpo_print)(const struct lu_env *env,
948 const struct cl_page_slice *slice,
949 void *cookie, lu_printer_t p);
953 * Transfer methods. See comment on cl_req for a description of
954 * transfer formation and life-cycle.
959 * Request type dependent vector of operations.
961 * Transfer operations depend on transfer mode (cl_req_type). To avoid
962 * passing transfer mode to each and every of these methods, and to
963 * avoid branching on request type inside of the methods, separate
964 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
965 * provided. That is, method invocation usually looks like
967 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
971 * Called when a page is submitted for a transfer as a part of
974 * \return 0 : page is eligible for submission;
975 * \return -EALREADY : skip this page;
976 * \return -ve : error.
978 * \see cl_page_prep()
980 int (*cpo_prep)(const struct lu_env *env,
981 const struct cl_page_slice *slice,
984 * Completion handler. This is guaranteed to be eventually
985 * fired after cl_page_operations::cpo_prep() or
986 * cl_page_operations::cpo_make_ready() call.
988 * This method can be called in a non-blocking context. It is
989 * guaranteed however, that the page involved and its object
990 * are pinned in memory (and, hence, calling cl_page_put() is
993 * \see cl_page_completion()
995 void (*cpo_completion)(const struct lu_env *env,
996 const struct cl_page_slice *slice,
999 * Called when cached page is about to be added to the
1000 * cl_req as a part of req formation.
1002 * \return 0 : proceed with this page;
1003 * \return -EAGAIN : skip this page;
1004 * \return -ve : error.
1006 * \see cl_page_make_ready()
1008 int (*cpo_make_ready)(const struct lu_env *env,
1009 const struct cl_page_slice *slice);
1012 * Tell transfer engine that only [to, from] part of a page should be
1015 * This is used for immediate transfers.
1017 * \todo XXX this is not very good interface. It would be much better
1018 * if all transfer parameters were supplied as arguments to
1019 * cl_io_operations::cio_submit() call, but it is not clear how to do
1020 * this for page queues.
1022 * \see cl_page_clip()
1024 void (*cpo_clip)(const struct lu_env *env,
1025 const struct cl_page_slice *slice,
1028 * \pre the page was queued for transferring.
1029 * \post page is removed from client's pending list, or -EBUSY
1030 * is returned if it has already been in transferring.
1032 * This is one of seldom page operation which is:
1033 * 0. called from top level;
1034 * 1. don't have vmpage locked;
1035 * 2. every layer should synchronize execution of its ->cpo_cancel()
1036 * with completion handlers. Osc uses client obd lock for this
1037 * purpose. Based on there is no vvp_page_cancel and
1038 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1040 * \see osc_page_cancel().
1042 int (*cpo_cancel)(const struct lu_env *env,
1043 const struct cl_page_slice *slice);
1045 * Write out a page by kernel. This is only called by ll_writepage
1048 * \see cl_page_flush()
1050 int (*cpo_flush)(const struct lu_env *env,
1051 const struct cl_page_slice *slice,
1057 * Helper macro, dumping detailed information about \a page into a log.
1059 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1061 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1062 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1063 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1064 CDEBUG(mask, format , ## __VA_ARGS__); \
1069 * Helper macro, dumping shorter information about \a page into a log.
1071 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1073 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1074 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1075 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1076 CDEBUG(mask, format , ## __VA_ARGS__); \
1080 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1082 LASSERT(page->cp_vmpage != NULL);
1083 return page->cp_vmpage;
1087 * Check if a cl_page is in use.
1089 * Client cache holds a refcount, this refcount will be dropped when
1090 * the page is taken out of cache, see vvp_page_delete().
1092 static inline bool __page_in_use(const struct cl_page *page, int refc)
1094 return (atomic_read(&page->cp_ref) > refc + 1);
1098 * Caller itself holds a refcount of cl_page.
1100 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1102 * Caller doesn't hold a refcount.
1104 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1108 /** \addtogroup cl_lock cl_lock
1112 * Extent locking on the client.
1116 * The locking model of the new client code is built around
1120 * data-type representing an extent lock on a regular file. cl_lock is a
1121 * layered object (much like cl_object and cl_page), it consists of a header
1122 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1123 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1125 * Typical cl_lock consists of the two layers:
1127 * - vvp_lock (vvp specific data), and
1128 * - lov_lock (lov specific data).
1130 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1131 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1133 * - lovsub_lock, and
1136 * Each sub-lock is associated with a cl_object (representing stripe
1137 * sub-object or the file to which top-level cl_lock is associated to), and is
1138 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1139 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1140 * is different from cl_page, that doesn't fan out (there is usually exactly
1141 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1142 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1146 * cl_lock is a cacheless data container for the requirements of locks to
1147 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1150 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1151 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1153 * INTERFACE AND USAGE
1155 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1156 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1157 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1158 * consists of multiple sub cl_locks, each sub locks will be enqueued
1159 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1160 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1163 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1164 * method will be called for each layer to release the resource held by this
1165 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1166 * clo_enqueue time, is released.
1168 * LDLM lock can only be canceled if there is no cl_lock using it.
1170 * Overall process of the locking during IO operation is as following:
1172 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1173 * is called on each layer. Responsibility of this method is to add locks,
1174 * needed by a given layer into cl_io.ci_lockset.
1176 * - once locks for all layers were collected, they are sorted to avoid
1177 * dead-locks (cl_io_locks_sort()), and enqueued.
1179 * - when all locks are acquired, IO is performed;
1181 * - locks are released after IO is complete.
1183 * Striping introduces major additional complexity into locking. The
1184 * fundamental problem is that it is generally unsafe to actively use (hold)
1185 * two locks on the different OST servers at the same time, as this introduces
1186 * inter-server dependency and can lead to cascading evictions.
1188 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1189 * that no multi-stripe locks are taken (note that this design abandons POSIX
1190 * read/write semantics). Such pieces ideally can be executed concurrently. At
1191 * the same time, certain types of IO cannot be sub-divived, without
1192 * sacrificing correctness. This includes:
1194 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1197 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1199 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1200 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1201 * has to be held together with the usual lock on [offset, offset + count].
1203 * Interaction with DLM
1205 * In the expected setup, cl_lock is ultimately backed up by a collection of
1206 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1207 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1208 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1209 * description of interaction with DLM.
1215 struct cl_lock_descr {
1216 /** Object this lock is granted for. */
1217 struct cl_object *cld_obj;
1218 /** Index of the first page protected by this lock. */
1220 /** Index of the last page (inclusive) protected by this lock. */
1222 /** Group ID, for group lock */
1225 enum cl_lock_mode cld_mode;
1227 * flags to enqueue lock. A combination of bit-flags from
1228 * enum cl_enq_flags.
1230 __u32 cld_enq_flags;
1233 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1234 #define PDESCR(descr) \
1235 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1236 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1238 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1241 * Layered client lock.
1244 /** List of slices. Immutable after creation. */
1245 struct list_head cll_layers;
1246 /** lock attribute, extent, cl_object, etc. */
1247 struct cl_lock_descr cll_descr;
1251 * Per-layer part of cl_lock
1253 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1255 struct cl_lock_slice {
1256 struct cl_lock *cls_lock;
1257 /** Object slice corresponding to this lock slice. Immutable after
1259 struct cl_object *cls_obj;
1260 const struct cl_lock_operations *cls_ops;
1261 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1262 struct list_head cls_linkage;
1267 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1269 struct cl_lock_operations {
1272 * Attempts to enqueue the lock. Called top-to-bottom.
1274 * \retval 0 this layer has enqueued the lock successfully
1275 * \retval >0 this layer has enqueued the lock, but need to wait on
1276 * @anchor for resources
1277 * \retval -ve failure
1279 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1280 * \see osc_lock_enqueue()
1282 int (*clo_enqueue)(const struct lu_env *env,
1283 const struct cl_lock_slice *slice,
1284 struct cl_io *io, struct cl_sync_io *anchor);
1286 * Cancel a lock, release its DLM lock ref, while does not cancel the
1289 void (*clo_cancel)(const struct lu_env *env,
1290 const struct cl_lock_slice *slice);
1293 * Destructor. Frees resources and the slice.
1295 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1296 * \see osc_lock_fini()
1298 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1300 * Optional debugging helper. Prints given lock slice.
1302 int (*clo_print)(const struct lu_env *env,
1303 void *cookie, lu_printer_t p,
1304 const struct cl_lock_slice *slice);
1307 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1309 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1310 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1311 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1312 CDEBUG(mask, format , ## __VA_ARGS__); \
1316 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1320 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1326 /** \addtogroup cl_page_list cl_page_list
1327 * Page list used to perform collective operations on a group of pages.
1329 * Pages are added to the list one by one. cl_page_list acquires a reference
1330 * for every page in it. Page list is used to perform collective operations on
1333 * - submit pages for an immediate transfer,
1335 * - own pages on behalf of certain io (waiting for each page in turn),
1339 * When list is finalized, it releases references on all pages it still has.
1341 * \todo XXX concurrency control.
1345 struct cl_page_list {
1347 struct list_head pl_pages;
1348 struct task_struct *pl_owner;
1352 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1353 * contains an incoming page list and an outgoing page list.
1356 struct cl_page_list c2_qin;
1357 struct cl_page_list c2_qout;
1360 /** @} cl_page_list */
1362 /** \addtogroup cl_io cl_io
1367 * cl_io represents a high level I/O activity like
1368 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1371 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1372 * important distinction. We want to minimize number of calls to the allocator
1373 * in the fast path, e.g., in the case of read(2) when everything is cached:
1374 * client already owns the lock over region being read, and data are cached
1375 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1376 * per-layer io state is stored in the session, associated with the io, see
1377 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1378 * by using free-lists, see cl_env_get().
1380 * There is a small predefined number of possible io types, enumerated in enum
1383 * cl_io is a state machine, that can be advanced concurrently by the multiple
1384 * threads. It is up to these threads to control the concurrency and,
1385 * specifically, to detect when io is done, and its state can be safely
1388 * For read/write io overall execution plan is as following:
1390 * (0) initialize io state through all layers;
1392 * (1) loop: prepare chunk of work to do
1394 * (2) call all layers to collect locks they need to process current chunk
1396 * (3) sort all locks to avoid dead-locks, and acquire them
1398 * (4) process the chunk: call per-page methods
1399 * cl_io_operations::cio_prepare_write(),
1400 * cl_io_operations::cio_commit_write() for write)
1406 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1407 * address allocation efficiency issues mentioned above), and returns with the
1408 * special error condition from per-page method when current sub-io has to
1409 * block. This causes io loop to be repeated, and lov switches to the next
1410 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1415 /** read system call */
1417 /** write system call */
1419 /** truncate, utime system calls */
1422 * page fault handling
1426 * fsync system call handling
1427 * To write out a range of file
1431 * Miscellaneous io. This is used for occasional io activity that
1432 * doesn't fit into other types. Currently this is used for:
1434 * - cancellation of an extent lock. This io exists as a context
1435 * to write dirty pages from under the lock being canceled back
1438 * - VM induced page write-out. An io context for writing page out
1439 * for memory cleansing;
1441 * - glimpse. An io context to acquire glimpse lock.
1443 * - grouplock. An io context to acquire group lock.
1445 * CIT_MISC io is used simply as a context in which locks and pages
1446 * are manipulated. Such io has no internal "process", that is,
1447 * cl_io_loop() is never called for it.
1454 * States of cl_io state machine
1457 /** Not initialized. */
1461 /** IO iteration started. */
1465 /** Actual IO is in progress. */
1467 /** IO for the current iteration finished. */
1469 /** Locks released. */
1471 /** Iteration completed. */
1473 /** cl_io finalized. */
1478 * IO state private for a layer.
1480 * This is usually embedded into layer session data, rather than allocated
1483 * \see vvp_io, lov_io, osc_io
1485 struct cl_io_slice {
1486 struct cl_io *cis_io;
1487 /** corresponding object slice. Immutable after creation. */
1488 struct cl_object *cis_obj;
1489 /** io operations. Immutable after creation. */
1490 const struct cl_io_operations *cis_iop;
1492 * linkage into a list of all slices for a given cl_io, hanging off
1493 * cl_io::ci_layers. Immutable after creation.
1495 struct list_head cis_linkage;
1498 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1501 struct cl_read_ahead {
1502 /* Maximum page index the readahead window will end.
1503 * This is determined DLM lock coverage, RPC and stripe boundary.
1504 * cra_end is included. */
1506 /* Release routine. If readahead holds resources underneath, this
1507 * function should be called to release it. */
1508 void (*cra_release)(const struct lu_env *env, void *cbdata);
1509 /* Callback data for cra_release routine */
1513 static inline void cl_read_ahead_release(const struct lu_env *env,
1514 struct cl_read_ahead *ra)
1516 if (ra->cra_release != NULL)
1517 ra->cra_release(env, ra->cra_cbdata);
1518 memset(ra, 0, sizeof(*ra));
1523 * Per-layer io operations.
1524 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1526 struct cl_io_operations {
1528 * Vector of io state transition methods for every io type.
1530 * \see cl_page_operations::io
1534 * Prepare io iteration at a given layer.
1536 * Called top-to-bottom at the beginning of each iteration of
1537 * "io loop" (if it makes sense for this type of io). Here
1538 * layer selects what work it will do during this iteration.
1540 * \see cl_io_operations::cio_iter_fini()
1542 int (*cio_iter_init) (const struct lu_env *env,
1543 const struct cl_io_slice *slice);
1545 * Finalize io iteration.
1547 * Called bottom-to-top at the end of each iteration of "io
1548 * loop". Here layers can decide whether IO has to be
1551 * \see cl_io_operations::cio_iter_init()
1553 void (*cio_iter_fini) (const struct lu_env *env,
1554 const struct cl_io_slice *slice);
1556 * Collect locks for the current iteration of io.
1558 * Called top-to-bottom to collect all locks necessary for
1559 * this iteration. This methods shouldn't actually enqueue
1560 * anything, instead it should post a lock through
1561 * cl_io_lock_add(). Once all locks are collected, they are
1562 * sorted and enqueued in the proper order.
1564 int (*cio_lock) (const struct lu_env *env,
1565 const struct cl_io_slice *slice);
1567 * Finalize unlocking.
1569 * Called bottom-to-top to finish layer specific unlocking
1570 * functionality, after generic code released all locks
1571 * acquired by cl_io_operations::cio_lock().
1573 void (*cio_unlock)(const struct lu_env *env,
1574 const struct cl_io_slice *slice);
1576 * Start io iteration.
1578 * Once all locks are acquired, called top-to-bottom to
1579 * commence actual IO. In the current implementation,
1580 * top-level vvp_io_{read,write}_start() does all the work
1581 * synchronously by calling generic_file_*(), so other layers
1582 * are called when everything is done.
1584 int (*cio_start)(const struct lu_env *env,
1585 const struct cl_io_slice *slice);
1587 * Called top-to-bottom at the end of io loop. Here layer
1588 * might wait for an unfinished asynchronous io.
1590 void (*cio_end) (const struct lu_env *env,
1591 const struct cl_io_slice *slice);
1593 * Called bottom-to-top to notify layers that read/write IO
1594 * iteration finished, with \a nob bytes transferred.
1596 void (*cio_advance)(const struct lu_env *env,
1597 const struct cl_io_slice *slice,
1600 * Called once per io, bottom-to-top to release io resources.
1602 void (*cio_fini) (const struct lu_env *env,
1603 const struct cl_io_slice *slice);
1607 * Submit pages from \a queue->c2_qin for IO, and move
1608 * successfully submitted pages into \a queue->c2_qout. Return
1609 * non-zero if failed to submit even the single page. If
1610 * submission failed after some pages were moved into \a
1611 * queue->c2_qout, completion callback with non-zero ioret is
1614 int (*cio_submit)(const struct lu_env *env,
1615 const struct cl_io_slice *slice,
1616 enum cl_req_type crt,
1617 struct cl_2queue *queue);
1619 * Queue async page for write.
1620 * The difference between cio_submit and cio_queue is that
1621 * cio_submit is for urgent request.
1623 int (*cio_commit_async)(const struct lu_env *env,
1624 const struct cl_io_slice *slice,
1625 struct cl_page_list *queue, int from, int to,
1628 * Decide maximum read ahead extent
1630 * \pre io->ci_type == CIT_READ
1632 int (*cio_read_ahead)(const struct lu_env *env,
1633 const struct cl_io_slice *slice,
1634 pgoff_t start, struct cl_read_ahead *ra);
1636 * Optional debugging helper. Print given io slice.
1638 int (*cio_print)(const struct lu_env *env, void *cookie,
1639 lu_printer_t p, const struct cl_io_slice *slice);
1643 * Flags to lock enqueue procedure.
1648 * instruct server to not block, if conflicting lock is found. Instead
1649 * -EWOULDBLOCK is returned immediately.
1651 CEF_NONBLOCK = 0x00000001,
1653 * take lock asynchronously (out of order), as it cannot
1654 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1656 CEF_ASYNC = 0x00000002,
1658 * tell the server to instruct (though a flag in the blocking ast) an
1659 * owner of the conflicting lock, that it can drop dirty pages
1660 * protected by this lock, without sending them to the server.
1662 CEF_DISCARD_DATA = 0x00000004,
1664 * tell the sub layers that it must be a `real' lock. This is used for
1665 * mmapped-buffer locks and glimpse locks that must be never converted
1666 * into lockless mode.
1668 * \see vvp_mmap_locks(), cl_glimpse_lock().
1670 CEF_MUST = 0x00000008,
1672 * tell the sub layers that never request a `real' lock. This flag is
1673 * not used currently.
1675 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1676 * conversion policy: ci_lockreq describes generic information of lock
1677 * requirement for this IO, especially for locks which belong to the
1678 * object doing IO; however, lock itself may have precise requirements
1679 * that are described by the enqueue flags.
1681 CEF_NEVER = 0x00000010,
1683 * for async glimpse lock.
1685 CEF_AGL = 0x00000020,
1687 * enqueue a lock to test DLM lock existence.
1689 CEF_PEEK = 0x00000040,
1691 * mask of enq_flags.
1693 CEF_MASK = 0x0000007f,
1697 * Link between lock and io. Intermediate structure is needed, because the
1698 * same lock can be part of multiple io's simultaneously.
1700 struct cl_io_lock_link {
1701 /** linkage into one of cl_lockset lists. */
1702 struct list_head cill_linkage;
1703 struct cl_lock cill_lock;
1704 /** optional destructor */
1705 void (*cill_fini)(const struct lu_env *env,
1706 struct cl_io_lock_link *link);
1708 #define cill_descr cill_lock.cll_descr
1711 * Lock-set represents a collection of locks, that io needs at a
1712 * time. Generally speaking, client tries to avoid holding multiple locks when
1715 * - holding extent locks over multiple ost's introduces the danger of
1716 * "cascading timeouts";
1718 * - holding multiple locks over the same ost is still dead-lock prone,
1719 * see comment in osc_lock_enqueue(),
1721 * but there are certain situations where this is unavoidable:
1723 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1725 * - truncate has to take [new-size, EOF] lock for correctness;
1727 * - SNS has to take locks across full stripe for correctness;
1729 * - in the case when user level buffer, supplied to {read,write}(file0),
1730 * is a part of a memory mapped lustre file, client has to take a dlm
1731 * locks on file0, and all files that back up the buffer (or a part of
1732 * the buffer, that is being processed in the current chunk, in any
1733 * case, there are situations where at least 2 locks are necessary).
1735 * In such cases we at least try to take locks in the same consistent
1736 * order. To this end, all locks are first collected, then sorted, and then
1740 /** locks to be acquired. */
1741 struct list_head cls_todo;
1742 /** locks acquired. */
1743 struct list_head cls_done;
1747 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1748 * but 'req' is always to be thought as 'request' :-)
1750 enum cl_io_lock_dmd {
1751 /** Always lock data (e.g., O_APPEND). */
1753 /** Layers are free to decide between local and global locking. */
1755 /** Never lock: there is no cache (e.g., liblustre). */
1759 enum cl_fsync_mode {
1760 /** start writeback, do not wait for them to finish */
1762 /** start writeback and wait for them to finish */
1764 /** discard all of dirty pages in a specific file range */
1765 CL_FSYNC_DISCARD = 2,
1766 /** start writeback and make sure they have reached storage before
1767 * return. OST_SYNC RPC must be issued and finished */
1771 struct cl_io_rw_common {
1781 * cl_io is shared by all threads participating in this IO (in current
1782 * implementation only one thread advances IO, but parallel IO design and
1783 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1784 * is up to these threads to serialize their activities, including updates to
1785 * mutable cl_io fields.
1788 /** type of this IO. Immutable after creation. */
1789 enum cl_io_type ci_type;
1790 /** current state of cl_io state machine. */
1791 enum cl_io_state ci_state;
1792 /** main object this io is against. Immutable after creation. */
1793 struct cl_object *ci_obj;
1795 * Upper layer io, of which this io is a part of. Immutable after
1798 struct cl_io *ci_parent;
1799 /** List of slices. Immutable after creation. */
1800 struct list_head ci_layers;
1801 /** list of locks (to be) acquired by this io. */
1802 struct cl_lockset ci_lockset;
1803 /** lock requirements, this is just a help info for sublayers. */
1804 enum cl_io_lock_dmd ci_lockreq;
1807 struct cl_io_rw_common rd;
1810 struct cl_io_rw_common wr;
1814 struct cl_io_rw_common ci_rw;
1815 struct cl_setattr_io {
1816 struct ost_lvb sa_attr;
1817 unsigned int sa_attr_flags;
1818 unsigned int sa_valid;
1819 int sa_stripe_index;
1820 const struct lu_fid *sa_parent_fid;
1821 struct obd_capa *sa_capa;
1823 struct cl_fault_io {
1824 /** page index within file. */
1826 /** bytes valid byte on a faulted page. */
1828 /** writable page? for nopage() only */
1830 /** page of an executable? */
1832 /** page_mkwrite() */
1834 /** resulting page */
1835 struct cl_page *ft_page;
1837 struct cl_fsync_io {
1840 struct obd_capa *fi_capa;
1841 /** file system level fid */
1842 struct lu_fid *fi_fid;
1843 enum cl_fsync_mode fi_mode;
1844 /* how many pages were written/discarded */
1845 unsigned int fi_nr_written;
1848 struct cl_2queue ci_queue;
1851 unsigned int ci_continue:1,
1853 * This io has held grouplock, to inform sublayers that
1854 * don't do lockless i/o.
1858 * The whole IO need to be restarted because layout has been changed
1862 * to not refresh layout - the IO issuer knows that the layout won't
1863 * change(page operations, layout change causes all page to be
1864 * discarded), or it doesn't matter if it changes(sync).
1868 * Check if layout changed after the IO finishes. Mainly for HSM
1869 * requirement. If IO occurs to openning files, it doesn't need to
1870 * verify layout because HSM won't release openning files.
1871 * Right now, only two opertaions need to verify layout: glimpse
1876 * file is released, restore has to to be triggered by vvp layer
1878 ci_restore_needed:1,
1884 * Number of pages owned by this IO. For invariant checking.
1886 unsigned ci_owned_nr;
1891 /** \addtogroup cl_req cl_req
1896 * There are two possible modes of transfer initiation on the client:
1898 * - immediate transfer: this is started when a high level io wants a page
1899 * or a collection of pages to be transferred right away. Examples:
1900 * read-ahead, synchronous read in the case of non-page aligned write,
1901 * page write-out as a part of extent lock cancellation, page write-out
1902 * as a part of memory cleansing. Immediate transfer can be both
1903 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1905 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1906 * when io wants to transfer a page to the server some time later, when
1907 * it can be done efficiently. Example: pages dirtied by the write(2)
1910 * In any case, transfer takes place in the form of a cl_req, which is a
1911 * representation for a network RPC.
1913 * Pages queued for an opportunistic transfer are cached until it is decided
1914 * that efficient RPC can be composed of them. This decision is made by "a
1915 * req-formation engine", currently implemented as a part of osc
1916 * layer. Req-formation depends on many factors: the size of the resulting
1917 * RPC, whether or not multi-object RPCs are supported by the server,
1918 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1920 * For the immediate transfer io submits a cl_page_list, that req-formation
1921 * engine slices into cl_req's, possibly adding cached pages to some of
1922 * the resulting req's.
1924 * Whenever a page from cl_page_list is added to a newly constructed req, its
1925 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1926 * page state is atomically changed from cl_page_state::CPS_OWNED to
1927 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1928 * is zeroed, and cl_page::cp_req is set to the
1929 * req. cl_page_operations::cpo_prep() method at the particular layer might
1930 * return -EALREADY to indicate that it does not need to submit this page
1931 * at all. This is possible, for example, if page, submitted for read,
1932 * became up-to-date in the meantime; and for write, the page don't have
1933 * dirty bit marked. \see cl_io_submit_rw()
1935 * Whenever a cached page is added to a newly constructed req, its
1936 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1937 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1938 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1939 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1940 * might return -EAGAIN to indicate that this page is not eligible for the
1941 * transfer right now.
1945 * Plan is to divide transfers into "priority bands" (indicated when
1946 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1947 * and allow glueing of cached pages to immediate transfers only within single
1948 * band. This would make high priority transfers (like lock cancellation or
1949 * memory pressure induced write-out) really high priority.
1954 * Per-transfer attributes.
1956 struct cl_req_attr {
1957 /** Generic attributes for the server consumption. */
1958 struct obdo *cra_oa;
1960 struct obd_capa *cra_capa;
1962 char cra_jobid[LUSTRE_JOBID_SIZE];
1966 * Transfer request operations definable at every layer.
1968 * Concurrency: transfer formation engine synchronizes calls to all transfer
1971 struct cl_req_operations {
1973 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1974 * complete (all pages are added).
1976 * \see osc_req_prep()
1978 int (*cro_prep)(const struct lu_env *env,
1979 const struct cl_req_slice *slice);
1981 * Called top-to-bottom to fill in \a oa fields. This is called twice
1982 * with different flags, see bug 10150 and osc_build_req().
1984 * \param obj an object from cl_req which attributes are to be set in
1987 * \param oa struct obdo where attributes are placed
1989 * \param flags \a oa fields to be filled.
1991 void (*cro_attr_set)(const struct lu_env *env,
1992 const struct cl_req_slice *slice,
1993 const struct cl_object *obj,
1994 struct cl_req_attr *attr, u64 flags);
1996 * Called top-to-bottom from cl_req_completion() to notify layers that
1997 * transfer completed. Has to free all state allocated by
1998 * cl_device_operations::cdo_req_init().
2000 void (*cro_completion)(const struct lu_env *env,
2001 const struct cl_req_slice *slice, int ioret);
2005 * A per-object state that (potentially multi-object) transfer request keeps.
2008 /** object itself */
2009 struct cl_object *ro_obj;
2010 /** reference to cl_req_obj::ro_obj. For debugging. */
2011 struct lu_ref_link ro_obj_ref;
2012 /* something else? Number of pages for a given object? */
2018 * Transfer requests are not reference counted, because IO sub-system owns
2019 * them exclusively and knows when to free them.
2023 * cl_req is created by cl_req_alloc() that calls
2024 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2025 * state in every layer.
2027 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2028 * contains pages for.
2030 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2031 * called top-to-bottom. At that point layers can modify req, let it pass, or
2032 * deny it completely. This is to support things like SNS that have transfer
2033 * ordering requirements invisible to the individual req-formation engine.
2035 * On transfer completion (or transfer timeout, or failure to initiate the
2036 * transfer of an allocated req), cl_req_operations::cro_completion() method
2037 * is called, after execution of cl_page_operations::cpo_completion() of all
2041 enum cl_req_type crq_type;
2042 /** A list of pages being transferred */
2043 struct list_head crq_pages;
2044 /** Number of pages in cl_req::crq_pages */
2045 unsigned crq_nrpages;
2046 /** An array of objects which pages are in ->crq_pages */
2047 struct cl_req_obj *crq_o;
2048 /** Number of elements in cl_req::crq_objs[] */
2049 unsigned crq_nrobjs;
2050 struct list_head crq_layers;
2054 * Per-layer state for request.
2056 struct cl_req_slice {
2057 struct cl_req *crs_req;
2058 struct cl_device *crs_dev;
2059 struct list_head crs_linkage;
2060 const struct cl_req_operations *crs_ops;
2065 enum cache_stats_item {
2066 /** how many cache lookups were performed */
2068 /** how many times cache lookup resulted in a hit */
2070 /** how many entities are in the cache right now */
2072 /** how many entities in the cache are actively used (and cannot be
2073 * evicted) right now */
2075 /** how many entities were created at all */
2080 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2083 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2085 struct cache_stats {
2086 const char *cs_name;
2087 atomic_t cs_stats[CS_NR];
2090 /** These are not exported so far */
2091 void cache_stats_init (struct cache_stats *cs, const char *name);
2094 * Client-side site. This represents particular client stack. "Global"
2095 * variables should (directly or indirectly) be added here to allow multiple
2096 * clients to co-exist in the single address space.
2099 struct lu_site cs_lu;
2101 * Statistical counters. Atomics do not scale, something better like
2102 * per-cpu counters is needed.
2104 * These are exported as /proc/fs/lustre/llite/.../site
2106 * When interpreting keep in mind that both sub-locks (and sub-pages)
2107 * and top-locks (and top-pages) are accounted here.
2109 struct cache_stats cs_pages;
2110 atomic_t cs_pages_state[CPS_NR];
2113 int cl_site_init(struct cl_site *s, struct cl_device *top);
2114 void cl_site_fini(struct cl_site *s);
2115 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2118 * Output client site statistical counters into a buffer. Suitable for
2119 * ll_rd_*()-style functions.
2121 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2126 * Type conversion and accessory functions.
2130 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2132 return container_of(site, struct cl_site, cs_lu);
2135 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2137 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2138 return container_of0(d, struct cl_device, cd_lu_dev);
2141 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2143 return &d->cd_lu_dev;
2146 static inline struct cl_object *lu2cl(const struct lu_object *o)
2148 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2149 return container_of0(o, struct cl_object, co_lu);
2152 static inline const struct cl_object_conf *
2153 lu2cl_conf(const struct lu_object_conf *conf)
2155 return container_of0(conf, struct cl_object_conf, coc_lu);
2158 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2160 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2163 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2165 return container_of0(h, struct cl_object_header, coh_lu);
2168 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2170 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2174 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2176 return luh2coh(obj->co_lu.lo_header);
2179 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2181 return lu_device_init(&d->cd_lu_dev, t);
2184 static inline void cl_device_fini(struct cl_device *d)
2186 lu_device_fini(&d->cd_lu_dev);
2189 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2190 struct cl_object *obj, pgoff_t index,
2191 const struct cl_page_operations *ops);
2192 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2193 struct cl_object *obj,
2194 const struct cl_lock_operations *ops);
2195 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2196 struct cl_object *obj, const struct cl_io_operations *ops);
2197 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2198 struct cl_device *dev,
2199 const struct cl_req_operations *ops);
2202 /** \defgroup cl_object cl_object
2204 struct cl_object *cl_object_top (struct cl_object *o);
2205 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2206 const struct lu_fid *fid,
2207 const struct cl_object_conf *c);
2209 int cl_object_header_init(struct cl_object_header *h);
2210 void cl_object_header_fini(struct cl_object_header *h);
2211 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2212 void cl_object_get (struct cl_object *o);
2213 void cl_object_attr_lock (struct cl_object *o);
2214 void cl_object_attr_unlock(struct cl_object *o);
2215 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2216 struct cl_attr *attr);
2217 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2218 const struct cl_attr *attr, unsigned valid);
2219 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2220 struct ost_lvb *lvb);
2221 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2222 const struct cl_object_conf *conf);
2223 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2224 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2225 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2226 struct lov_user_md __user *lum);
2227 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2228 ldlm_iterator_t iter, void *data);
2229 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2230 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2232 int cl_object_obd_info_get(const struct lu_env *env, struct cl_object *obj,
2233 struct obd_info *oinfo,
2234 struct ptlrpc_request_set *set);
2235 int cl_object_data_version(const struct lu_env *env, struct cl_object *obj,
2236 __u64 *version, int flags);
2237 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2238 struct cl_layout *cl);
2239 loff_t cl_object_maxbytes(struct cl_object *obj);
2242 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2244 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2246 return cl_object_header(o0) == cl_object_header(o1);
2249 static inline void cl_object_page_init(struct cl_object *clob, int size)
2251 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2252 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2253 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2256 static inline void *cl_object_page_slice(struct cl_object *clob,
2257 struct cl_page *page)
2259 return (void *)((char *)page + clob->co_slice_off);
2263 * Return refcount of cl_object.
2265 static inline int cl_object_refc(struct cl_object *clob)
2267 struct lu_object_header *header = clob->co_lu.lo_header;
2268 return atomic_read(&header->loh_ref);
2273 /** \defgroup cl_page cl_page
2281 /* callback of cl_page_gang_lookup() */
2283 struct cl_page *cl_page_find (const struct lu_env *env,
2284 struct cl_object *obj,
2285 pgoff_t idx, struct page *vmpage,
2286 enum cl_page_type type);
2287 struct cl_page *cl_page_alloc (const struct lu_env *env,
2288 struct cl_object *o, pgoff_t ind,
2289 struct page *vmpage,
2290 enum cl_page_type type);
2291 void cl_page_get (struct cl_page *page);
2292 void cl_page_put (const struct lu_env *env,
2293 struct cl_page *page);
2294 void cl_page_print (const struct lu_env *env, void *cookie,
2295 lu_printer_t printer,
2296 const struct cl_page *pg);
2297 void cl_page_header_print(const struct lu_env *env, void *cookie,
2298 lu_printer_t printer,
2299 const struct cl_page *pg);
2300 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2301 struct cl_page *cl_page_top (struct cl_page *page);
2303 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2304 const struct lu_device_type *dtype);
2309 * Functions dealing with the ownership of page by io.
2313 int cl_page_own (const struct lu_env *env,
2314 struct cl_io *io, struct cl_page *page);
2315 int cl_page_own_try (const struct lu_env *env,
2316 struct cl_io *io, struct cl_page *page);
2317 void cl_page_assume (const struct lu_env *env,
2318 struct cl_io *io, struct cl_page *page);
2319 void cl_page_unassume (const struct lu_env *env,
2320 struct cl_io *io, struct cl_page *pg);
2321 void cl_page_disown (const struct lu_env *env,
2322 struct cl_io *io, struct cl_page *page);
2323 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2330 * Functions dealing with the preparation of a page for a transfer, and
2331 * tracking transfer state.
2334 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2335 struct cl_page *pg, enum cl_req_type crt);
2336 void cl_page_completion (const struct lu_env *env,
2337 struct cl_page *pg, enum cl_req_type crt, int ioret);
2338 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2339 enum cl_req_type crt);
2340 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2341 struct cl_page *pg, enum cl_req_type crt);
2342 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2344 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2345 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2346 struct cl_page *pg);
2352 * \name helper routines
2353 * Functions to discard, delete and export a cl_page.
2356 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2357 struct cl_page *pg);
2358 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2359 int cl_page_is_vmlocked(const struct lu_env *env,
2360 const struct cl_page *pg);
2361 void cl_page_export(const struct lu_env *env,
2362 struct cl_page *pg, int uptodate);
2363 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2364 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2365 size_t cl_page_size(const struct cl_object *obj);
2367 void cl_lock_print(const struct lu_env *env, void *cookie,
2368 lu_printer_t printer, const struct cl_lock *lock);
2369 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2370 lu_printer_t printer,
2371 const struct cl_lock_descr *descr);
2375 * Data structure managing a client's cached pages. A count of
2376 * "unstable" pages is maintained, and an LRU of clean pages is
2377 * maintained. "unstable" pages are pages pinned by the ptlrpc
2378 * layer for recovery purposes.
2380 struct cl_client_cache {
2382 * # of client cache refcount
2383 * # of users (OSCs) + 2 (held by llite and lov)
2387 * # of threads are doing shrinking
2389 unsigned int ccc_lru_shrinkers;
2391 * # of LRU entries available
2393 atomic_long_t ccc_lru_left;
2395 * List of entities(OSCs) for this LRU cache
2397 struct list_head ccc_lru;
2399 * Max # of LRU entries
2401 unsigned long ccc_lru_max;
2403 * Lock to protect ccc_lru list
2405 spinlock_t ccc_lru_lock;
2407 * Set if unstable check is enabled
2409 unsigned int ccc_unstable_check:1;
2411 * # of unstable pages for this mount point
2413 atomic_long_t ccc_unstable_nr;
2415 * Waitq for awaiting unstable pages to reach zero.
2416 * Used at umounting time and signaled on BRW commit
2418 wait_queue_head_t ccc_unstable_waitq;
2421 * cl_cache functions
2423 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2424 void cl_cache_incref(struct cl_client_cache *cache);
2425 void cl_cache_decref(struct cl_client_cache *cache);
2429 /** \defgroup cl_lock cl_lock
2431 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2432 struct cl_lock *lock);
2433 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2434 const struct cl_io *io);
2435 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2436 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2437 const struct lu_device_type *dtype);
2438 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2440 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2441 struct cl_lock *lock, struct cl_sync_io *anchor);
2442 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2446 /** \defgroup cl_io cl_io
2449 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2450 enum cl_io_type iot, struct cl_object *obj);
2451 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2452 enum cl_io_type iot, struct cl_object *obj);
2453 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2454 enum cl_io_type iot, loff_t pos, size_t count);
2455 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2457 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2458 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2459 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2460 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2461 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2462 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2463 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2464 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2465 struct cl_io_lock_link *link);
2466 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2467 struct cl_lock_descr *descr);
2468 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2469 enum cl_req_type iot, struct cl_2queue *queue);
2470 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2471 enum cl_req_type iot, struct cl_2queue *queue,
2473 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2474 struct cl_page_list *queue, int from, int to,
2476 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2477 pgoff_t start, struct cl_read_ahead *ra);
2478 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2480 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2481 struct cl_page_list *queue);
2482 int cl_io_is_going (const struct lu_env *env);
2485 * True, iff \a io is an O_APPEND write(2).
2487 static inline int cl_io_is_append(const struct cl_io *io)
2489 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2492 static inline int cl_io_is_sync_write(const struct cl_io *io)
2494 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2497 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2499 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2503 * True, iff \a io is a truncate(2).
2505 static inline int cl_io_is_trunc(const struct cl_io *io)
2507 return io->ci_type == CIT_SETATTR &&
2508 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2511 struct cl_io *cl_io_top(struct cl_io *io);
2513 void cl_io_print(const struct lu_env *env, void *cookie,
2514 lu_printer_t printer, const struct cl_io *io);
2516 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2518 typeof(foo_io) __foo_io = (foo_io); \
2520 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2521 memset(&__foo_io->base + 1, 0, \
2522 (sizeof *__foo_io) - sizeof __foo_io->base); \
2527 /** \defgroup cl_page_list cl_page_list
2531 * Last page in the page list.
2533 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2535 LASSERT(plist->pl_nr > 0);
2536 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2539 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2541 LASSERT(plist->pl_nr > 0);
2542 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2546 * Iterate over pages in a page list.
2548 #define cl_page_list_for_each(page, list) \
2549 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2552 * Iterate over pages in a page list, taking possible removals into account.
2554 #define cl_page_list_for_each_safe(page, temp, list) \
2555 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2557 void cl_page_list_init (struct cl_page_list *plist);
2558 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2559 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2560 struct cl_page *page);
2561 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2562 struct cl_page *page);
2563 void cl_page_list_splice (struct cl_page_list *list,
2564 struct cl_page_list *head);
2565 void cl_page_list_del (const struct lu_env *env,
2566 struct cl_page_list *plist, struct cl_page *page);
2567 void cl_page_list_disown (const struct lu_env *env,
2568 struct cl_io *io, struct cl_page_list *plist);
2569 int cl_page_list_own (const struct lu_env *env,
2570 struct cl_io *io, struct cl_page_list *plist);
2571 void cl_page_list_assume (const struct lu_env *env,
2572 struct cl_io *io, struct cl_page_list *plist);
2573 void cl_page_list_discard(const struct lu_env *env,
2574 struct cl_io *io, struct cl_page_list *plist);
2575 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2577 void cl_2queue_init (struct cl_2queue *queue);
2578 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2579 void cl_2queue_disown (const struct lu_env *env,
2580 struct cl_io *io, struct cl_2queue *queue);
2581 void cl_2queue_assume (const struct lu_env *env,
2582 struct cl_io *io, struct cl_2queue *queue);
2583 void cl_2queue_discard (const struct lu_env *env,
2584 struct cl_io *io, struct cl_2queue *queue);
2585 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2586 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2588 /** @} cl_page_list */
2590 /** \defgroup cl_req cl_req
2592 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2593 enum cl_req_type crt, int nr_objects);
2595 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2596 struct cl_page *page);
2597 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2598 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2599 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2600 struct cl_req_attr *attr, u64 flags);
2601 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2603 /** \defgroup cl_sync_io cl_sync_io
2607 * Anchor for synchronous transfer. This is allocated on a stack by thread
2608 * doing synchronous transfer, and a pointer to this structure is set up in
2609 * every page submitted for transfer. Transfer completion routine updates
2610 * anchor and wakes up waiting thread when transfer is complete.
2613 /** number of pages yet to be transferred. */
2614 atomic_t csi_sync_nr;
2617 /** barrier of destroy this structure */
2618 atomic_t csi_barrier;
2619 /** completion to be signaled when transfer is complete. */
2620 wait_queue_head_t csi_waitq;
2621 /** callback to invoke when this IO is finished */
2622 void (*csi_end_io)(const struct lu_env *,
2623 struct cl_sync_io *);
2626 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2627 void (*end)(const struct lu_env *, struct cl_sync_io *));
2628 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2630 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2632 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2634 /** @} cl_sync_io */
2638 /** \defgroup cl_env cl_env
2640 * lu_env handling for a client.
2642 * lu_env is an environment within which lustre code executes. Its major part
2643 * is lu_context---a fast memory allocation mechanism that is used to conserve
2644 * precious kernel stack space. Originally lu_env was designed for a server,
2647 * - there is a (mostly) fixed number of threads, and
2649 * - call chains have no non-lustre portions inserted between lustre code.
2651 * On a client both these assumtpion fails, because every user thread can
2652 * potentially execute lustre code as part of a system call, and lustre calls
2653 * into VFS or MM that call back into lustre.
2655 * To deal with that, cl_env wrapper functions implement the following
2658 * - allocation and destruction of environment is amortized by caching no
2659 * longer used environments instead of destroying them;
2661 * - there is a notion of "current" environment, attached to the kernel
2662 * data structure representing current thread Top-level lustre code
2663 * allocates an environment and makes it current, then calls into
2664 * non-lustre code, that in turn calls lustre back. Low-level lustre
2665 * code thus called can fetch environment created by the top-level code
2666 * and reuse it, avoiding additional environment allocation.
2667 * Right now, three interfaces can attach the cl_env to running thread:
2670 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2672 * \see lu_env, lu_context, lu_context_key
2675 struct cl_env_nest {
2680 struct lu_env *cl_env_peek (int *refcheck);
2681 struct lu_env *cl_env_get (int *refcheck);
2682 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2683 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2684 void cl_env_put (struct lu_env *env, int *refcheck);
2685 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2686 void *cl_env_reenter (void);
2687 void cl_env_reexit (void *cookie);
2688 void cl_env_implant (struct lu_env *env, int *refcheck);
2689 void cl_env_unplant (struct lu_env *env, int *refcheck);
2690 unsigned cl_env_cache_purge(unsigned nr);
2691 struct lu_env *cl_env_percpu_get (void);
2692 void cl_env_percpu_put (struct lu_env *env);
2699 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2700 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2702 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2703 struct lu_device_type *ldt,
2704 struct lu_device *next);
2707 int cl_global_init(void);
2708 void cl_global_fini(void);
2710 #endif /* _LINUX_CL_OBJECT_H */