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);
456 * Extended header for client object.
458 struct cl_object_header {
459 /** Standard lu_object_header. cl_object::co_lu::lo_header points
461 struct lu_object_header coh_lu;
464 * Parent object. It is assumed that an object has a well-defined
465 * parent, but not a well-defined child (there may be multiple
466 * sub-objects, for the same top-object). cl_object_header::coh_parent
467 * field allows certain code to be written generically, without
468 * limiting possible cl_object layouts unduly.
470 struct cl_object_header *coh_parent;
472 * Protects consistency between cl_attr of parent object and
473 * attributes of sub-objects, that the former is calculated ("merged")
476 * \todo XXX this can be read/write lock if needed.
478 spinlock_t coh_attr_guard;
480 * Size of cl_page + page slices
482 unsigned short coh_page_bufsize;
484 * Number of objects above this one: 0 for a top-object, 1 for its
487 unsigned char coh_nesting;
491 * Helper macro: iterate over all layers of the object \a obj, assigning every
492 * layer top-to-bottom to \a slice.
494 #define cl_object_for_each(slice, obj) \
495 list_for_each_entry((slice), \
496 &(obj)->co_lu.lo_header->loh_layers,\
500 * Helper macro: iterate over all layers of the object \a obj, assigning every
501 * layer bottom-to-top to \a slice.
503 #define cl_object_for_each_reverse(slice, obj) \
504 list_for_each_entry_reverse((slice), \
505 &(obj)->co_lu.lo_header->loh_layers,\
510 #define CL_PAGE_EOF ((pgoff_t)~0ull)
512 /** \addtogroup cl_page cl_page
516 * Layered client page.
518 * cl_page: represents a portion of a file, cached in the memory. All pages
519 * of the given file are of the same size, and are kept in the radix tree
520 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
521 * of the top-level file object are first class cl_objects, they have their
522 * own radix trees of pages and hence page is implemented as a sequence of
523 * struct cl_pages's, linked into double-linked list through
524 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
525 * corresponding radix tree at the corresponding logical offset.
527 * cl_page is associated with VM page of the hosting environment (struct
528 * page in Linux kernel, for example), struct page. It is assumed, that this
529 * association is implemented by one of cl_page layers (top layer in the
530 * current design) that
532 * - intercepts per-VM-page call-backs made by the environment (e.g.,
535 * - translates state (page flag bits) and locking between lustre and
538 * The association between cl_page and struct page is immutable and
539 * established when cl_page is created.
541 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
542 * this io an exclusive access to this page w.r.t. other io attempts and
543 * various events changing page state (such as transfer completion, or
544 * eviction of the page from the memory). Note, that in general cl_io
545 * cannot be identified with a particular thread, and page ownership is not
546 * exactly equal to the current thread holding a lock on the page. Layer
547 * implementing association between cl_page and struct page has to implement
548 * ownership on top of available synchronization mechanisms.
550 * While lustre client maintains the notion of an page ownership by io,
551 * hosting MM/VM usually has its own page concurrency control
552 * mechanisms. For example, in Linux, page access is synchronized by the
553 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
554 * takes care to acquire and release such locks as necessary around the
555 * calls to the file system methods (->readpage(), ->prepare_write(),
556 * ->commit_write(), etc.). This leads to the situation when there are two
557 * different ways to own a page in the client:
559 * - client code explicitly and voluntary owns the page (cl_page_own());
561 * - VM locks a page and then calls the client, that has "to assume"
562 * the ownership from the VM (cl_page_assume()).
564 * Dual methods to release ownership are cl_page_disown() and
565 * cl_page_unassume().
567 * cl_page is reference counted (cl_page::cp_ref). When reference counter
568 * drops to 0, the page is returned to the cache, unless it is in
569 * cl_page_state::CPS_FREEING state, in which case it is immediately
572 * The general logic guaranteeing the absence of "existential races" for
573 * pages is the following:
575 * - there are fixed known ways for a thread to obtain a new reference
578 * - by doing a lookup in the cl_object radix tree, protected by the
581 * - by starting from VM-locked struct page and following some
582 * hosting environment method (e.g., following ->private pointer in
583 * the case of Linux kernel), see cl_vmpage_page();
585 * - when the page enters cl_page_state::CPS_FREEING state, all these
586 * ways are severed with the proper synchronization
587 * (cl_page_delete());
589 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
592 * - no new references to the page in cl_page_state::CPS_FREEING state
593 * are allowed (checked in cl_page_get()).
595 * Together this guarantees that when last reference to a
596 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
597 * page, as neither references to it can be acquired at that point, nor
600 * cl_page is a state machine. States are enumerated in enum
601 * cl_page_state. Possible state transitions are enumerated in
602 * cl_page_state_set(). State transition process (i.e., actual changing of
603 * cl_page::cp_state field) is protected by the lock on the underlying VM
606 * Linux Kernel implementation.
608 * Binding between cl_page and struct page (which is a typedef for
609 * struct page) is implemented in the vvp layer. cl_page is attached to the
610 * ->private pointer of the struct page, together with the setting of
611 * PG_private bit in page->flags, and acquiring additional reference on the
612 * struct page (much like struct buffer_head, or any similar file system
613 * private data structures).
615 * PG_locked lock is used to implement both ownership and transfer
616 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
617 * states. No additional references are acquired for the duration of the
620 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
621 * write-out is "protected" by the special PG_writeback bit.
625 * States of cl_page. cl_page.c assumes particular order here.
627 * The page state machine is rather crude, as it doesn't recognize finer page
628 * states like "dirty" or "up to date". This is because such states are not
629 * always well defined for the whole stack (see, for example, the
630 * implementation of the read-ahead, that hides page up-to-dateness to track
631 * cache hits accurately). Such sub-states are maintained by the layers that
632 * are interested in them.
636 * Page is in the cache, un-owned. Page leaves cached state in the
639 * - [cl_page_state::CPS_OWNED] io comes across the page and
642 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
643 * req-formation engine decides that it wants to include this page
644 * into an cl_req being constructed, and yanks it from the cache;
646 * - [cl_page_state::CPS_FREEING] VM callback is executed to
647 * evict the page form the memory;
649 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
653 * Page is exclusively owned by some cl_io. Page may end up in this
654 * state as a result of
656 * - io creating new page and immediately owning it;
658 * - [cl_page_state::CPS_CACHED] io finding existing cached page
661 * - [cl_page_state::CPS_OWNED] io finding existing owned page
662 * and waiting for owner to release the page;
664 * Page leaves owned state in the following cases:
666 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
667 * the cache, doing nothing;
669 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
672 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
673 * transfer for this page;
675 * - [cl_page_state::CPS_FREEING] io decides to destroy this
676 * page (e.g., as part of truncate or extent lock cancellation).
678 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
682 * Page is being written out, as a part of a transfer. This state is
683 * entered when req-formation logic decided that it wants this page to
684 * be sent through the wire _now_. Specifically, it means that once
685 * this state is achieved, transfer completion handler (with either
686 * success or failure indication) is guaranteed to be executed against
687 * this page independently of any locks and any scheduling decisions
688 * made by the hosting environment (that effectively means that the
689 * page is never put into cl_page_state::CPS_PAGEOUT state "in
690 * advance". This property is mentioned, because it is important when
691 * reasoning about possible dead-locks in the system). The page can
692 * enter this state as a result of
694 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
695 * write-out of this page, or
697 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
698 * that it has enough dirty pages cached to issue a "good"
701 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
702 * is completed---it is moved into cl_page_state::CPS_CACHED state.
704 * Underlying VM page is locked for the duration of transfer.
706 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
710 * Page is being read in, as a part of a transfer. This is quite
711 * similar to the cl_page_state::CPS_PAGEOUT state, except that
712 * read-in is always "immediate"---there is no such thing a sudden
713 * construction of read cl_req from cached, presumably not up to date,
716 * Underlying VM page is locked for the duration of transfer.
718 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
722 * Page is being destroyed. This state is entered when client decides
723 * that page has to be deleted from its host object, as, e.g., a part
726 * Once this state is reached, there is no way to escape it.
728 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
735 /** Host page, the page is from the host inode which the cl_page
739 /** Transient page, the transient cl_page is used to bind a cl_page
740 * to vmpage which is not belonging to the same object of cl_page.
741 * it is used in DirectIO, lockless IO and liblustre. */
746 * Fields are protected by the lock on struct page, except for atomics and
749 * \invariant Data type invariants are in cl_page_invariant(). Basically:
750 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
751 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
752 * cl_page::cp_owner (when set).
755 /** Reference counter. */
757 /** Transfer error. */
759 /** An object this page is a part of. Immutable after creation. */
760 struct cl_object *cp_obj;
762 struct page *cp_vmpage;
763 /** Linkage of pages within group. Pages must be owned */
764 struct list_head cp_batch;
765 /** List of slices. Immutable after creation. */
766 struct list_head cp_layers;
767 /** Linkage of pages within cl_req. */
768 struct list_head cp_flight;
770 * Page state. This field is const to avoid accidental update, it is
771 * modified only internally within cl_page.c. Protected by a VM lock.
773 const enum cl_page_state cp_state;
775 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
778 enum cl_page_type cp_type;
781 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
782 * by sub-io. Protected by a VM lock.
784 struct cl_io *cp_owner;
786 * Owning IO request in cl_page_state::CPS_PAGEOUT and
787 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
788 * the top-level pages. Protected by a VM lock.
790 struct cl_req *cp_req;
791 /** List of references to this page, for debugging. */
792 struct lu_ref cp_reference;
793 /** Link to an object, for debugging. */
794 struct lu_ref_link cp_obj_ref;
795 /** Link to a queue, for debugging. */
796 struct lu_ref_link cp_queue_ref;
797 /** Assigned if doing a sync_io */
798 struct cl_sync_io *cp_sync_io;
802 * Per-layer part of cl_page.
804 * \see vvp_page, lov_page, osc_page
806 struct cl_page_slice {
807 struct cl_page *cpl_page;
810 * Object slice corresponding to this page slice. Immutable after
813 struct cl_object *cpl_obj;
814 const struct cl_page_operations *cpl_ops;
815 /** Linkage into cl_page::cp_layers. Immutable after creation. */
816 struct list_head cpl_linkage;
820 * Lock mode. For the client extent locks.
832 * Requested transfer type.
842 * Per-layer page operations.
844 * Methods taking an \a io argument are for the activity happening in the
845 * context of given \a io. Page is assumed to be owned by that io, except for
846 * the obvious cases (like cl_page_operations::cpo_own()).
848 * \see vvp_page_ops, lov_page_ops, osc_page_ops
850 struct cl_page_operations {
852 * cl_page<->struct page methods. Only one layer in the stack has to
853 * implement these. Current code assumes that this functionality is
854 * provided by the topmost layer, see cl_page_disown0() as an example.
858 * Called when \a io acquires this page into the exclusive
859 * ownership. When this method returns, it is guaranteed that the is
860 * not owned by other io, and no transfer is going on against
864 * \see vvp_page_own(), lov_page_own()
866 int (*cpo_own)(const struct lu_env *env,
867 const struct cl_page_slice *slice,
868 struct cl_io *io, int nonblock);
869 /** Called when ownership it yielded. Optional.
871 * \see cl_page_disown()
872 * \see vvp_page_disown()
874 void (*cpo_disown)(const struct lu_env *env,
875 const struct cl_page_slice *slice, struct cl_io *io);
877 * Called for a page that is already "owned" by \a io from VM point of
880 * \see cl_page_assume()
881 * \see vvp_page_assume(), lov_page_assume()
883 void (*cpo_assume)(const struct lu_env *env,
884 const struct cl_page_slice *slice, struct cl_io *io);
885 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
886 * bottom-to-top when IO releases a page without actually unlocking
889 * \see cl_page_unassume()
890 * \see vvp_page_unassume()
892 void (*cpo_unassume)(const struct lu_env *env,
893 const struct cl_page_slice *slice,
896 * Announces whether the page contains valid data or not by \a uptodate.
898 * \see cl_page_export()
899 * \see vvp_page_export()
901 void (*cpo_export)(const struct lu_env *env,
902 const struct cl_page_slice *slice, int uptodate);
904 * Checks whether underlying VM page is locked (in the suitable
905 * sense). Used for assertions.
907 * \retval -EBUSY: page is protected by a lock of a given mode;
908 * \retval -ENODATA: page is not protected by a lock;
909 * \retval 0: this layer cannot decide. (Should never happen.)
911 int (*cpo_is_vmlocked)(const struct lu_env *env,
912 const struct cl_page_slice *slice);
918 * Called when page is truncated from the object. Optional.
920 * \see cl_page_discard()
921 * \see vvp_page_discard(), osc_page_discard()
923 void (*cpo_discard)(const struct lu_env *env,
924 const struct cl_page_slice *slice,
927 * Called when page is removed from the cache, and is about to being
928 * destroyed. Optional.
930 * \see cl_page_delete()
931 * \see vvp_page_delete(), osc_page_delete()
933 void (*cpo_delete)(const struct lu_env *env,
934 const struct cl_page_slice *slice);
935 /** Destructor. Frees resources and slice itself. */
936 void (*cpo_fini)(const struct lu_env *env,
937 struct cl_page_slice *slice);
939 * Optional debugging helper. Prints given page slice.
941 * \see cl_page_print()
943 int (*cpo_print)(const struct lu_env *env,
944 const struct cl_page_slice *slice,
945 void *cookie, lu_printer_t p);
949 * Transfer methods. See comment on cl_req for a description of
950 * transfer formation and life-cycle.
955 * Request type dependent vector of operations.
957 * Transfer operations depend on transfer mode (cl_req_type). To avoid
958 * passing transfer mode to each and every of these methods, and to
959 * avoid branching on request type inside of the methods, separate
960 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
961 * provided. That is, method invocation usually looks like
963 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
967 * Called when a page is submitted for a transfer as a part of
970 * \return 0 : page is eligible for submission;
971 * \return -EALREADY : skip this page;
972 * \return -ve : error.
974 * \see cl_page_prep()
976 int (*cpo_prep)(const struct lu_env *env,
977 const struct cl_page_slice *slice,
980 * Completion handler. This is guaranteed to be eventually
981 * fired after cl_page_operations::cpo_prep() or
982 * cl_page_operations::cpo_make_ready() call.
984 * This method can be called in a non-blocking context. It is
985 * guaranteed however, that the page involved and its object
986 * are pinned in memory (and, hence, calling cl_page_put() is
989 * \see cl_page_completion()
991 void (*cpo_completion)(const struct lu_env *env,
992 const struct cl_page_slice *slice,
995 * Called when cached page is about to be added to the
996 * cl_req as a part of req formation.
998 * \return 0 : proceed with this page;
999 * \return -EAGAIN : skip this page;
1000 * \return -ve : error.
1002 * \see cl_page_make_ready()
1004 int (*cpo_make_ready)(const struct lu_env *env,
1005 const struct cl_page_slice *slice);
1008 * Tell transfer engine that only [to, from] part of a page should be
1011 * This is used for immediate transfers.
1013 * \todo XXX this is not very good interface. It would be much better
1014 * if all transfer parameters were supplied as arguments to
1015 * cl_io_operations::cio_submit() call, but it is not clear how to do
1016 * this for page queues.
1018 * \see cl_page_clip()
1020 void (*cpo_clip)(const struct lu_env *env,
1021 const struct cl_page_slice *slice,
1024 * \pre the page was queued for transferring.
1025 * \post page is removed from client's pending list, or -EBUSY
1026 * is returned if it has already been in transferring.
1028 * This is one of seldom page operation which is:
1029 * 0. called from top level;
1030 * 1. don't have vmpage locked;
1031 * 2. every layer should synchronize execution of its ->cpo_cancel()
1032 * with completion handlers. Osc uses client obd lock for this
1033 * purpose. Based on there is no vvp_page_cancel and
1034 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1036 * \see osc_page_cancel().
1038 int (*cpo_cancel)(const struct lu_env *env,
1039 const struct cl_page_slice *slice);
1041 * Write out a page by kernel. This is only called by ll_writepage
1044 * \see cl_page_flush()
1046 int (*cpo_flush)(const struct lu_env *env,
1047 const struct cl_page_slice *slice,
1053 * Helper macro, dumping detailed information about \a page into a log.
1055 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1057 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1058 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1059 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1060 CDEBUG(mask, format , ## __VA_ARGS__); \
1065 * Helper macro, dumping shorter information about \a page into a log.
1067 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1069 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1070 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1071 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1072 CDEBUG(mask, format , ## __VA_ARGS__); \
1076 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1078 LASSERT(page->cp_vmpage != NULL);
1079 return page->cp_vmpage;
1083 * Check if a cl_page is in use.
1085 * Client cache holds a refcount, this refcount will be dropped when
1086 * the page is taken out of cache, see vvp_page_delete().
1088 static inline bool __page_in_use(const struct cl_page *page, int refc)
1090 return (atomic_read(&page->cp_ref) > refc + 1);
1094 * Caller itself holds a refcount of cl_page.
1096 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1098 * Caller doesn't hold a refcount.
1100 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1104 /** \addtogroup cl_lock cl_lock
1108 * Extent locking on the client.
1112 * The locking model of the new client code is built around
1116 * data-type representing an extent lock on a regular file. cl_lock is a
1117 * layered object (much like cl_object and cl_page), it consists of a header
1118 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1119 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1121 * Typical cl_lock consists of the two layers:
1123 * - vvp_lock (vvp specific data), and
1124 * - lov_lock (lov specific data).
1126 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1127 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1129 * - lovsub_lock, and
1132 * Each sub-lock is associated with a cl_object (representing stripe
1133 * sub-object or the file to which top-level cl_lock is associated to), and is
1134 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1135 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1136 * is different from cl_page, that doesn't fan out (there is usually exactly
1137 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1138 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1142 * cl_lock is a cacheless data container for the requirements of locks to
1143 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1146 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1147 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1149 * INTERFACE AND USAGE
1151 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1152 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1153 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1154 * consists of multiple sub cl_locks, each sub locks will be enqueued
1155 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1156 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1159 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1160 * method will be called for each layer to release the resource held by this
1161 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1162 * clo_enqueue time, is released.
1164 * LDLM lock can only be canceled if there is no cl_lock using it.
1166 * Overall process of the locking during IO operation is as following:
1168 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1169 * is called on each layer. Responsibility of this method is to add locks,
1170 * needed by a given layer into cl_io.ci_lockset.
1172 * - once locks for all layers were collected, they are sorted to avoid
1173 * dead-locks (cl_io_locks_sort()), and enqueued.
1175 * - when all locks are acquired, IO is performed;
1177 * - locks are released after IO is complete.
1179 * Striping introduces major additional complexity into locking. The
1180 * fundamental problem is that it is generally unsafe to actively use (hold)
1181 * two locks on the different OST servers at the same time, as this introduces
1182 * inter-server dependency and can lead to cascading evictions.
1184 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1185 * that no multi-stripe locks are taken (note that this design abandons POSIX
1186 * read/write semantics). Such pieces ideally can be executed concurrently. At
1187 * the same time, certain types of IO cannot be sub-divived, without
1188 * sacrificing correctness. This includes:
1190 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1193 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1195 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1196 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1197 * has to be held together with the usual lock on [offset, offset + count].
1199 * Interaction with DLM
1201 * In the expected setup, cl_lock is ultimately backed up by a collection of
1202 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1203 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1204 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1205 * description of interaction with DLM.
1211 struct cl_lock_descr {
1212 /** Object this lock is granted for. */
1213 struct cl_object *cld_obj;
1214 /** Index of the first page protected by this lock. */
1216 /** Index of the last page (inclusive) protected by this lock. */
1218 /** Group ID, for group lock */
1221 enum cl_lock_mode cld_mode;
1223 * flags to enqueue lock. A combination of bit-flags from
1224 * enum cl_enq_flags.
1226 __u32 cld_enq_flags;
1229 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1230 #define PDESCR(descr) \
1231 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1232 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1234 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1237 * Layered client lock.
1240 /** List of slices. Immutable after creation. */
1241 struct list_head cll_layers;
1242 /** lock attribute, extent, cl_object, etc. */
1243 struct cl_lock_descr cll_descr;
1247 * Per-layer part of cl_lock
1249 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1251 struct cl_lock_slice {
1252 struct cl_lock *cls_lock;
1253 /** Object slice corresponding to this lock slice. Immutable after
1255 struct cl_object *cls_obj;
1256 const struct cl_lock_operations *cls_ops;
1257 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1258 struct list_head cls_linkage;
1263 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1265 struct cl_lock_operations {
1268 * Attempts to enqueue the lock. Called top-to-bottom.
1270 * \retval 0 this layer has enqueued the lock successfully
1271 * \retval >0 this layer has enqueued the lock, but need to wait on
1272 * @anchor for resources
1273 * \retval -ve failure
1275 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1276 * \see osc_lock_enqueue()
1278 int (*clo_enqueue)(const struct lu_env *env,
1279 const struct cl_lock_slice *slice,
1280 struct cl_io *io, struct cl_sync_io *anchor);
1282 * Cancel a lock, release its DLM lock ref, while does not cancel the
1285 void (*clo_cancel)(const struct lu_env *env,
1286 const struct cl_lock_slice *slice);
1289 * Destructor. Frees resources and the slice.
1291 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1292 * \see osc_lock_fini()
1294 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1296 * Optional debugging helper. Prints given lock slice.
1298 int (*clo_print)(const struct lu_env *env,
1299 void *cookie, lu_printer_t p,
1300 const struct cl_lock_slice *slice);
1303 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1305 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1306 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1307 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1308 CDEBUG(mask, format , ## __VA_ARGS__); \
1312 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1316 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1322 /** \addtogroup cl_page_list cl_page_list
1323 * Page list used to perform collective operations on a group of pages.
1325 * Pages are added to the list one by one. cl_page_list acquires a reference
1326 * for every page in it. Page list is used to perform collective operations on
1329 * - submit pages for an immediate transfer,
1331 * - own pages on behalf of certain io (waiting for each page in turn),
1335 * When list is finalized, it releases references on all pages it still has.
1337 * \todo XXX concurrency control.
1341 struct cl_page_list {
1343 struct list_head pl_pages;
1344 struct task_struct *pl_owner;
1348 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1349 * contains an incoming page list and an outgoing page list.
1352 struct cl_page_list c2_qin;
1353 struct cl_page_list c2_qout;
1356 /** @} cl_page_list */
1358 /** \addtogroup cl_io cl_io
1363 * cl_io represents a high level I/O activity like
1364 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1367 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1368 * important distinction. We want to minimize number of calls to the allocator
1369 * in the fast path, e.g., in the case of read(2) when everything is cached:
1370 * client already owns the lock over region being read, and data are cached
1371 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1372 * per-layer io state is stored in the session, associated with the io, see
1373 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1374 * by using free-lists, see cl_env_get().
1376 * There is a small predefined number of possible io types, enumerated in enum
1379 * cl_io is a state machine, that can be advanced concurrently by the multiple
1380 * threads. It is up to these threads to control the concurrency and,
1381 * specifically, to detect when io is done, and its state can be safely
1384 * For read/write io overall execution plan is as following:
1386 * (0) initialize io state through all layers;
1388 * (1) loop: prepare chunk of work to do
1390 * (2) call all layers to collect locks they need to process current chunk
1392 * (3) sort all locks to avoid dead-locks, and acquire them
1394 * (4) process the chunk: call per-page methods
1395 * cl_io_operations::cio_prepare_write(),
1396 * cl_io_operations::cio_commit_write() for write)
1402 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1403 * address allocation efficiency issues mentioned above), and returns with the
1404 * special error condition from per-page method when current sub-io has to
1405 * block. This causes io loop to be repeated, and lov switches to the next
1406 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1411 /** read system call */
1413 /** write system call */
1415 /** truncate, utime system calls */
1418 * page fault handling
1422 * fsync system call handling
1423 * To write out a range of file
1427 * Miscellaneous io. This is used for occasional io activity that
1428 * doesn't fit into other types. Currently this is used for:
1430 * - cancellation of an extent lock. This io exists as a context
1431 * to write dirty pages from under the lock being canceled back
1434 * - VM induced page write-out. An io context for writing page out
1435 * for memory cleansing;
1437 * - glimpse. An io context to acquire glimpse lock.
1439 * - grouplock. An io context to acquire group lock.
1441 * CIT_MISC io is used simply as a context in which locks and pages
1442 * are manipulated. Such io has no internal "process", that is,
1443 * cl_io_loop() is never called for it.
1450 * States of cl_io state machine
1453 /** Not initialized. */
1457 /** IO iteration started. */
1461 /** Actual IO is in progress. */
1463 /** IO for the current iteration finished. */
1465 /** Locks released. */
1467 /** Iteration completed. */
1469 /** cl_io finalized. */
1474 * IO state private for a layer.
1476 * This is usually embedded into layer session data, rather than allocated
1479 * \see vvp_io, lov_io, osc_io
1481 struct cl_io_slice {
1482 struct cl_io *cis_io;
1483 /** corresponding object slice. Immutable after creation. */
1484 struct cl_object *cis_obj;
1485 /** io operations. Immutable after creation. */
1486 const struct cl_io_operations *cis_iop;
1488 * linkage into a list of all slices for a given cl_io, hanging off
1489 * cl_io::ci_layers. Immutable after creation.
1491 struct list_head cis_linkage;
1494 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1497 struct cl_read_ahead {
1498 /* Maximum page index the readahead window will end.
1499 * This is determined DLM lock coverage, RPC and stripe boundary.
1500 * cra_end is included. */
1502 /* Release routine. If readahead holds resources underneath, this
1503 * function should be called to release it. */
1504 void (*cra_release)(const struct lu_env *env, void *cbdata);
1505 /* Callback data for cra_release routine */
1509 static inline void cl_read_ahead_release(const struct lu_env *env,
1510 struct cl_read_ahead *ra)
1512 if (ra->cra_release != NULL)
1513 ra->cra_release(env, ra->cra_cbdata);
1514 memset(ra, 0, sizeof(*ra));
1519 * Per-layer io operations.
1520 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1522 struct cl_io_operations {
1524 * Vector of io state transition methods for every io type.
1526 * \see cl_page_operations::io
1530 * Prepare io iteration at a given layer.
1532 * Called top-to-bottom at the beginning of each iteration of
1533 * "io loop" (if it makes sense for this type of io). Here
1534 * layer selects what work it will do during this iteration.
1536 * \see cl_io_operations::cio_iter_fini()
1538 int (*cio_iter_init) (const struct lu_env *env,
1539 const struct cl_io_slice *slice);
1541 * Finalize io iteration.
1543 * Called bottom-to-top at the end of each iteration of "io
1544 * loop". Here layers can decide whether IO has to be
1547 * \see cl_io_operations::cio_iter_init()
1549 void (*cio_iter_fini) (const struct lu_env *env,
1550 const struct cl_io_slice *slice);
1552 * Collect locks for the current iteration of io.
1554 * Called top-to-bottom to collect all locks necessary for
1555 * this iteration. This methods shouldn't actually enqueue
1556 * anything, instead it should post a lock through
1557 * cl_io_lock_add(). Once all locks are collected, they are
1558 * sorted and enqueued in the proper order.
1560 int (*cio_lock) (const struct lu_env *env,
1561 const struct cl_io_slice *slice);
1563 * Finalize unlocking.
1565 * Called bottom-to-top to finish layer specific unlocking
1566 * functionality, after generic code released all locks
1567 * acquired by cl_io_operations::cio_lock().
1569 void (*cio_unlock)(const struct lu_env *env,
1570 const struct cl_io_slice *slice);
1572 * Start io iteration.
1574 * Once all locks are acquired, called top-to-bottom to
1575 * commence actual IO. In the current implementation,
1576 * top-level vvp_io_{read,write}_start() does all the work
1577 * synchronously by calling generic_file_*(), so other layers
1578 * are called when everything is done.
1580 int (*cio_start)(const struct lu_env *env,
1581 const struct cl_io_slice *slice);
1583 * Called top-to-bottom at the end of io loop. Here layer
1584 * might wait for an unfinished asynchronous io.
1586 void (*cio_end) (const struct lu_env *env,
1587 const struct cl_io_slice *slice);
1589 * Called bottom-to-top to notify layers that read/write IO
1590 * iteration finished, with \a nob bytes transferred.
1592 void (*cio_advance)(const struct lu_env *env,
1593 const struct cl_io_slice *slice,
1596 * Called once per io, bottom-to-top to release io resources.
1598 void (*cio_fini) (const struct lu_env *env,
1599 const struct cl_io_slice *slice);
1603 * Submit pages from \a queue->c2_qin for IO, and move
1604 * successfully submitted pages into \a queue->c2_qout. Return
1605 * non-zero if failed to submit even the single page. If
1606 * submission failed after some pages were moved into \a
1607 * queue->c2_qout, completion callback with non-zero ioret is
1610 int (*cio_submit)(const struct lu_env *env,
1611 const struct cl_io_slice *slice,
1612 enum cl_req_type crt,
1613 struct cl_2queue *queue);
1615 * Queue async page for write.
1616 * The difference between cio_submit and cio_queue is that
1617 * cio_submit is for urgent request.
1619 int (*cio_commit_async)(const struct lu_env *env,
1620 const struct cl_io_slice *slice,
1621 struct cl_page_list *queue, int from, int to,
1624 * Decide maximum read ahead extent
1626 * \pre io->ci_type == CIT_READ
1628 int (*cio_read_ahead)(const struct lu_env *env,
1629 const struct cl_io_slice *slice,
1630 pgoff_t start, struct cl_read_ahead *ra);
1632 * Optional debugging helper. Print given io slice.
1634 int (*cio_print)(const struct lu_env *env, void *cookie,
1635 lu_printer_t p, const struct cl_io_slice *slice);
1639 * Flags to lock enqueue procedure.
1644 * instruct server to not block, if conflicting lock is found. Instead
1645 * -EWOULDBLOCK is returned immediately.
1647 CEF_NONBLOCK = 0x00000001,
1649 * take lock asynchronously (out of order), as it cannot
1650 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1652 CEF_ASYNC = 0x00000002,
1654 * tell the server to instruct (though a flag in the blocking ast) an
1655 * owner of the conflicting lock, that it can drop dirty pages
1656 * protected by this lock, without sending them to the server.
1658 CEF_DISCARD_DATA = 0x00000004,
1660 * tell the sub layers that it must be a `real' lock. This is used for
1661 * mmapped-buffer locks and glimpse locks that must be never converted
1662 * into lockless mode.
1664 * \see vvp_mmap_locks(), cl_glimpse_lock().
1666 CEF_MUST = 0x00000008,
1668 * tell the sub layers that never request a `real' lock. This flag is
1669 * not used currently.
1671 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1672 * conversion policy: ci_lockreq describes generic information of lock
1673 * requirement for this IO, especially for locks which belong to the
1674 * object doing IO; however, lock itself may have precise requirements
1675 * that are described by the enqueue flags.
1677 CEF_NEVER = 0x00000010,
1679 * for async glimpse lock.
1681 CEF_AGL = 0x00000020,
1683 * enqueue a lock to test DLM lock existence.
1685 CEF_PEEK = 0x00000040,
1687 * mask of enq_flags.
1689 CEF_MASK = 0x0000007f,
1693 * Link between lock and io. Intermediate structure is needed, because the
1694 * same lock can be part of multiple io's simultaneously.
1696 struct cl_io_lock_link {
1697 /** linkage into one of cl_lockset lists. */
1698 struct list_head cill_linkage;
1699 struct cl_lock cill_lock;
1700 /** optional destructor */
1701 void (*cill_fini)(const struct lu_env *env,
1702 struct cl_io_lock_link *link);
1704 #define cill_descr cill_lock.cll_descr
1707 * Lock-set represents a collection of locks, that io needs at a
1708 * time. Generally speaking, client tries to avoid holding multiple locks when
1711 * - holding extent locks over multiple ost's introduces the danger of
1712 * "cascading timeouts";
1714 * - holding multiple locks over the same ost is still dead-lock prone,
1715 * see comment in osc_lock_enqueue(),
1717 * but there are certain situations where this is unavoidable:
1719 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1721 * - truncate has to take [new-size, EOF] lock for correctness;
1723 * - SNS has to take locks across full stripe for correctness;
1725 * - in the case when user level buffer, supplied to {read,write}(file0),
1726 * is a part of a memory mapped lustre file, client has to take a dlm
1727 * locks on file0, and all files that back up the buffer (or a part of
1728 * the buffer, that is being processed in the current chunk, in any
1729 * case, there are situations where at least 2 locks are necessary).
1731 * In such cases we at least try to take locks in the same consistent
1732 * order. To this end, all locks are first collected, then sorted, and then
1736 /** locks to be acquired. */
1737 struct list_head cls_todo;
1738 /** locks acquired. */
1739 struct list_head cls_done;
1743 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1744 * but 'req' is always to be thought as 'request' :-)
1746 enum cl_io_lock_dmd {
1747 /** Always lock data (e.g., O_APPEND). */
1749 /** Layers are free to decide between local and global locking. */
1751 /** Never lock: there is no cache (e.g., liblustre). */
1755 enum cl_fsync_mode {
1756 /** start writeback, do not wait for them to finish */
1758 /** start writeback and wait for them to finish */
1760 /** discard all of dirty pages in a specific file range */
1761 CL_FSYNC_DISCARD = 2,
1762 /** start writeback and make sure they have reached storage before
1763 * return. OST_SYNC RPC must be issued and finished */
1767 struct cl_io_rw_common {
1777 * cl_io is shared by all threads participating in this IO (in current
1778 * implementation only one thread advances IO, but parallel IO design and
1779 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1780 * is up to these threads to serialize their activities, including updates to
1781 * mutable cl_io fields.
1784 /** type of this IO. Immutable after creation. */
1785 enum cl_io_type ci_type;
1786 /** current state of cl_io state machine. */
1787 enum cl_io_state ci_state;
1788 /** main object this io is against. Immutable after creation. */
1789 struct cl_object *ci_obj;
1791 * Upper layer io, of which this io is a part of. Immutable after
1794 struct cl_io *ci_parent;
1795 /** List of slices. Immutable after creation. */
1796 struct list_head ci_layers;
1797 /** list of locks (to be) acquired by this io. */
1798 struct cl_lockset ci_lockset;
1799 /** lock requirements, this is just a help info for sublayers. */
1800 enum cl_io_lock_dmd ci_lockreq;
1803 struct cl_io_rw_common rd;
1806 struct cl_io_rw_common wr;
1810 struct cl_io_rw_common ci_rw;
1811 struct cl_setattr_io {
1812 struct ost_lvb sa_attr;
1813 unsigned int sa_attr_flags;
1814 unsigned int sa_valid;
1815 int sa_stripe_index;
1816 const struct lu_fid *sa_parent_fid;
1817 struct obd_capa *sa_capa;
1819 struct cl_fault_io {
1820 /** page index within file. */
1822 /** bytes valid byte on a faulted page. */
1824 /** writable page? for nopage() only */
1826 /** page of an executable? */
1828 /** page_mkwrite() */
1830 /** resulting page */
1831 struct cl_page *ft_page;
1833 struct cl_fsync_io {
1836 struct obd_capa *fi_capa;
1837 /** file system level fid */
1838 struct lu_fid *fi_fid;
1839 enum cl_fsync_mode fi_mode;
1840 /* how many pages were written/discarded */
1841 unsigned int fi_nr_written;
1844 struct cl_2queue ci_queue;
1847 unsigned int ci_continue:1,
1849 * This io has held grouplock, to inform sublayers that
1850 * don't do lockless i/o.
1854 * The whole IO need to be restarted because layout has been changed
1858 * to not refresh layout - the IO issuer knows that the layout won't
1859 * change(page operations, layout change causes all page to be
1860 * discarded), or it doesn't matter if it changes(sync).
1864 * Check if layout changed after the IO finishes. Mainly for HSM
1865 * requirement. If IO occurs to openning files, it doesn't need to
1866 * verify layout because HSM won't release openning files.
1867 * Right now, only two opertaions need to verify layout: glimpse
1872 * file is released, restore has to to be triggered by vvp layer
1874 ci_restore_needed:1,
1880 * Number of pages owned by this IO. For invariant checking.
1882 unsigned ci_owned_nr;
1887 /** \addtogroup cl_req cl_req
1892 * There are two possible modes of transfer initiation on the client:
1894 * - immediate transfer: this is started when a high level io wants a page
1895 * or a collection of pages to be transferred right away. Examples:
1896 * read-ahead, synchronous read in the case of non-page aligned write,
1897 * page write-out as a part of extent lock cancellation, page write-out
1898 * as a part of memory cleansing. Immediate transfer can be both
1899 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1901 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1902 * when io wants to transfer a page to the server some time later, when
1903 * it can be done efficiently. Example: pages dirtied by the write(2)
1906 * In any case, transfer takes place in the form of a cl_req, which is a
1907 * representation for a network RPC.
1909 * Pages queued for an opportunistic transfer are cached until it is decided
1910 * that efficient RPC can be composed of them. This decision is made by "a
1911 * req-formation engine", currently implemented as a part of osc
1912 * layer. Req-formation depends on many factors: the size of the resulting
1913 * RPC, whether or not multi-object RPCs are supported by the server,
1914 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1916 * For the immediate transfer io submits a cl_page_list, that req-formation
1917 * engine slices into cl_req's, possibly adding cached pages to some of
1918 * the resulting req's.
1920 * Whenever a page from cl_page_list is added to a newly constructed req, its
1921 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1922 * page state is atomically changed from cl_page_state::CPS_OWNED to
1923 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1924 * is zeroed, and cl_page::cp_req is set to the
1925 * req. cl_page_operations::cpo_prep() method at the particular layer might
1926 * return -EALREADY to indicate that it does not need to submit this page
1927 * at all. This is possible, for example, if page, submitted for read,
1928 * became up-to-date in the meantime; and for write, the page don't have
1929 * dirty bit marked. \see cl_io_submit_rw()
1931 * Whenever a cached page is added to a newly constructed req, its
1932 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1933 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1934 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1935 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1936 * might return -EAGAIN to indicate that this page is not eligible for the
1937 * transfer right now.
1941 * Plan is to divide transfers into "priority bands" (indicated when
1942 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1943 * and allow glueing of cached pages to immediate transfers only within single
1944 * band. This would make high priority transfers (like lock cancellation or
1945 * memory pressure induced write-out) really high priority.
1950 * Per-transfer attributes.
1952 struct cl_req_attr {
1953 /** Generic attributes for the server consumption. */
1954 struct obdo *cra_oa;
1956 struct obd_capa *cra_capa;
1958 char cra_jobid[LUSTRE_JOBID_SIZE];
1962 * Transfer request operations definable at every layer.
1964 * Concurrency: transfer formation engine synchronizes calls to all transfer
1967 struct cl_req_operations {
1969 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1970 * complete (all pages are added).
1972 * \see osc_req_prep()
1974 int (*cro_prep)(const struct lu_env *env,
1975 const struct cl_req_slice *slice);
1977 * Called top-to-bottom to fill in \a oa fields. This is called twice
1978 * with different flags, see bug 10150 and osc_build_req().
1980 * \param obj an object from cl_req which attributes are to be set in
1983 * \param oa struct obdo where attributes are placed
1985 * \param flags \a oa fields to be filled.
1987 void (*cro_attr_set)(const struct lu_env *env,
1988 const struct cl_req_slice *slice,
1989 const struct cl_object *obj,
1990 struct cl_req_attr *attr, u64 flags);
1992 * Called top-to-bottom from cl_req_completion() to notify layers that
1993 * transfer completed. Has to free all state allocated by
1994 * cl_device_operations::cdo_req_init().
1996 void (*cro_completion)(const struct lu_env *env,
1997 const struct cl_req_slice *slice, int ioret);
2001 * A per-object state that (potentially multi-object) transfer request keeps.
2004 /** object itself */
2005 struct cl_object *ro_obj;
2006 /** reference to cl_req_obj::ro_obj. For debugging. */
2007 struct lu_ref_link ro_obj_ref;
2008 /* something else? Number of pages for a given object? */
2014 * Transfer requests are not reference counted, because IO sub-system owns
2015 * them exclusively and knows when to free them.
2019 * cl_req is created by cl_req_alloc() that calls
2020 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2021 * state in every layer.
2023 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2024 * contains pages for.
2026 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2027 * called top-to-bottom. At that point layers can modify req, let it pass, or
2028 * deny it completely. This is to support things like SNS that have transfer
2029 * ordering requirements invisible to the individual req-formation engine.
2031 * On transfer completion (or transfer timeout, or failure to initiate the
2032 * transfer of an allocated req), cl_req_operations::cro_completion() method
2033 * is called, after execution of cl_page_operations::cpo_completion() of all
2037 enum cl_req_type crq_type;
2038 /** A list of pages being transferred */
2039 struct list_head crq_pages;
2040 /** Number of pages in cl_req::crq_pages */
2041 unsigned crq_nrpages;
2042 /** An array of objects which pages are in ->crq_pages */
2043 struct cl_req_obj *crq_o;
2044 /** Number of elements in cl_req::crq_objs[] */
2045 unsigned crq_nrobjs;
2046 struct list_head crq_layers;
2050 * Per-layer state for request.
2052 struct cl_req_slice {
2053 struct cl_req *crs_req;
2054 struct cl_device *crs_dev;
2055 struct list_head crs_linkage;
2056 const struct cl_req_operations *crs_ops;
2061 enum cache_stats_item {
2062 /** how many cache lookups were performed */
2064 /** how many times cache lookup resulted in a hit */
2066 /** how many entities are in the cache right now */
2068 /** how many entities in the cache are actively used (and cannot be
2069 * evicted) right now */
2071 /** how many entities were created at all */
2076 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2079 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2081 struct cache_stats {
2082 const char *cs_name;
2083 atomic_t cs_stats[CS_NR];
2086 /** These are not exported so far */
2087 void cache_stats_init (struct cache_stats *cs, const char *name);
2090 * Client-side site. This represents particular client stack. "Global"
2091 * variables should (directly or indirectly) be added here to allow multiple
2092 * clients to co-exist in the single address space.
2095 struct lu_site cs_lu;
2097 * Statistical counters. Atomics do not scale, something better like
2098 * per-cpu counters is needed.
2100 * These are exported as /proc/fs/lustre/llite/.../site
2102 * When interpreting keep in mind that both sub-locks (and sub-pages)
2103 * and top-locks (and top-pages) are accounted here.
2105 struct cache_stats cs_pages;
2106 atomic_t cs_pages_state[CPS_NR];
2109 int cl_site_init(struct cl_site *s, struct cl_device *top);
2110 void cl_site_fini(struct cl_site *s);
2111 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2114 * Output client site statistical counters into a buffer. Suitable for
2115 * ll_rd_*()-style functions.
2117 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2122 * Type conversion and accessory functions.
2126 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2128 return container_of(site, struct cl_site, cs_lu);
2131 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2133 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2134 return container_of0(d, struct cl_device, cd_lu_dev);
2137 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2139 return &d->cd_lu_dev;
2142 static inline struct cl_object *lu2cl(const struct lu_object *o)
2144 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2145 return container_of0(o, struct cl_object, co_lu);
2148 static inline const struct cl_object_conf *
2149 lu2cl_conf(const struct lu_object_conf *conf)
2151 return container_of0(conf, struct cl_object_conf, coc_lu);
2154 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2156 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2159 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2161 return container_of0(h, struct cl_object_header, coh_lu);
2164 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2166 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2170 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2172 return luh2coh(obj->co_lu.lo_header);
2175 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2177 return lu_device_init(&d->cd_lu_dev, t);
2180 static inline void cl_device_fini(struct cl_device *d)
2182 lu_device_fini(&d->cd_lu_dev);
2185 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2186 struct cl_object *obj, pgoff_t index,
2187 const struct cl_page_operations *ops);
2188 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2189 struct cl_object *obj,
2190 const struct cl_lock_operations *ops);
2191 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2192 struct cl_object *obj, const struct cl_io_operations *ops);
2193 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2194 struct cl_device *dev,
2195 const struct cl_req_operations *ops);
2198 /** \defgroup cl_object cl_object
2200 struct cl_object *cl_object_top (struct cl_object *o);
2201 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2202 const struct lu_fid *fid,
2203 const struct cl_object_conf *c);
2205 int cl_object_header_init(struct cl_object_header *h);
2206 void cl_object_header_fini(struct cl_object_header *h);
2207 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2208 void cl_object_get (struct cl_object *o);
2209 void cl_object_attr_lock (struct cl_object *o);
2210 void cl_object_attr_unlock(struct cl_object *o);
2211 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2212 struct cl_attr *attr);
2213 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2214 const struct cl_attr *attr, unsigned valid);
2215 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2216 struct ost_lvb *lvb);
2217 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2218 const struct cl_object_conf *conf);
2219 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2220 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2221 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2222 struct lov_user_md __user *lum);
2223 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2224 ldlm_iterator_t iter, void *data);
2225 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2226 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2228 int cl_object_obd_info_get(const struct lu_env *env, struct cl_object *obj,
2229 struct obd_info *oinfo,
2230 struct ptlrpc_request_set *set);
2231 int cl_object_data_version(const struct lu_env *env, struct cl_object *obj,
2232 __u64 *version, int flags);
2233 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2234 struct cl_layout *cl);
2237 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2239 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2241 return cl_object_header(o0) == cl_object_header(o1);
2244 static inline void cl_object_page_init(struct cl_object *clob, int size)
2246 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2247 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2248 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2251 static inline void *cl_object_page_slice(struct cl_object *clob,
2252 struct cl_page *page)
2254 return (void *)((char *)page + clob->co_slice_off);
2258 * Return refcount of cl_object.
2260 static inline int cl_object_refc(struct cl_object *clob)
2262 struct lu_object_header *header = clob->co_lu.lo_header;
2263 return atomic_read(&header->loh_ref);
2268 /** \defgroup cl_page cl_page
2276 /* callback of cl_page_gang_lookup() */
2278 struct cl_page *cl_page_find (const struct lu_env *env,
2279 struct cl_object *obj,
2280 pgoff_t idx, struct page *vmpage,
2281 enum cl_page_type type);
2282 struct cl_page *cl_page_alloc (const struct lu_env *env,
2283 struct cl_object *o, pgoff_t ind,
2284 struct page *vmpage,
2285 enum cl_page_type type);
2286 void cl_page_get (struct cl_page *page);
2287 void cl_page_put (const struct lu_env *env,
2288 struct cl_page *page);
2289 void cl_page_print (const struct lu_env *env, void *cookie,
2290 lu_printer_t printer,
2291 const struct cl_page *pg);
2292 void cl_page_header_print(const struct lu_env *env, void *cookie,
2293 lu_printer_t printer,
2294 const struct cl_page *pg);
2295 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2296 struct cl_page *cl_page_top (struct cl_page *page);
2298 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2299 const struct lu_device_type *dtype);
2304 * Functions dealing with the ownership of page by io.
2308 int cl_page_own (const struct lu_env *env,
2309 struct cl_io *io, struct cl_page *page);
2310 int cl_page_own_try (const struct lu_env *env,
2311 struct cl_io *io, struct cl_page *page);
2312 void cl_page_assume (const struct lu_env *env,
2313 struct cl_io *io, struct cl_page *page);
2314 void cl_page_unassume (const struct lu_env *env,
2315 struct cl_io *io, struct cl_page *pg);
2316 void cl_page_disown (const struct lu_env *env,
2317 struct cl_io *io, struct cl_page *page);
2318 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2325 * Functions dealing with the preparation of a page for a transfer, and
2326 * tracking transfer state.
2329 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2330 struct cl_page *pg, enum cl_req_type crt);
2331 void cl_page_completion (const struct lu_env *env,
2332 struct cl_page *pg, enum cl_req_type crt, int ioret);
2333 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2334 enum cl_req_type crt);
2335 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2336 struct cl_page *pg, enum cl_req_type crt);
2337 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2339 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2340 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2341 struct cl_page *pg);
2347 * \name helper routines
2348 * Functions to discard, delete and export a cl_page.
2351 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2352 struct cl_page *pg);
2353 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2354 int cl_page_is_vmlocked(const struct lu_env *env,
2355 const struct cl_page *pg);
2356 void cl_page_export(const struct lu_env *env,
2357 struct cl_page *pg, int uptodate);
2358 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2359 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2360 size_t cl_page_size(const struct cl_object *obj);
2362 void cl_lock_print(const struct lu_env *env, void *cookie,
2363 lu_printer_t printer, const struct cl_lock *lock);
2364 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2365 lu_printer_t printer,
2366 const struct cl_lock_descr *descr);
2370 * Data structure managing a client's cached pages. A count of
2371 * "unstable" pages is maintained, and an LRU of clean pages is
2372 * maintained. "unstable" pages are pages pinned by the ptlrpc
2373 * layer for recovery purposes.
2375 struct cl_client_cache {
2377 * # of client cache refcount
2378 * # of users (OSCs) + 2 (held by llite and lov)
2382 * # of threads are doing shrinking
2384 unsigned int ccc_lru_shrinkers;
2386 * # of LRU entries available
2388 atomic_long_t ccc_lru_left;
2390 * List of entities(OSCs) for this LRU cache
2392 struct list_head ccc_lru;
2394 * Max # of LRU entries
2396 unsigned long ccc_lru_max;
2398 * Lock to protect ccc_lru list
2400 spinlock_t ccc_lru_lock;
2402 * Set if unstable check is enabled
2404 unsigned int ccc_unstable_check:1;
2406 * # of unstable pages for this mount point
2408 atomic_long_t ccc_unstable_nr;
2410 * Waitq for awaiting unstable pages to reach zero.
2411 * Used at umounting time and signaled on BRW commit
2413 wait_queue_head_t ccc_unstable_waitq;
2416 * cl_cache functions
2418 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2419 void cl_cache_incref(struct cl_client_cache *cache);
2420 void cl_cache_decref(struct cl_client_cache *cache);
2424 /** \defgroup cl_lock cl_lock
2426 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2427 struct cl_lock *lock);
2428 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2429 const struct cl_io *io);
2430 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2431 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2432 const struct lu_device_type *dtype);
2433 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2435 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2436 struct cl_lock *lock, struct cl_sync_io *anchor);
2437 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2441 /** \defgroup cl_io cl_io
2444 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2445 enum cl_io_type iot, struct cl_object *obj);
2446 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2447 enum cl_io_type iot, struct cl_object *obj);
2448 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2449 enum cl_io_type iot, loff_t pos, size_t count);
2450 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2452 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2453 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2454 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2455 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2456 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2457 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2458 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2459 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2460 struct cl_io_lock_link *link);
2461 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2462 struct cl_lock_descr *descr);
2463 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2464 enum cl_req_type iot, struct cl_2queue *queue);
2465 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2466 enum cl_req_type iot, struct cl_2queue *queue,
2468 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2469 struct cl_page_list *queue, int from, int to,
2471 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2472 pgoff_t start, struct cl_read_ahead *ra);
2473 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2475 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2476 struct cl_page_list *queue);
2477 int cl_io_is_going (const struct lu_env *env);
2480 * True, iff \a io is an O_APPEND write(2).
2482 static inline int cl_io_is_append(const struct cl_io *io)
2484 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2487 static inline int cl_io_is_sync_write(const struct cl_io *io)
2489 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2492 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2494 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2498 * True, iff \a io is a truncate(2).
2500 static inline int cl_io_is_trunc(const struct cl_io *io)
2502 return io->ci_type == CIT_SETATTR &&
2503 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2506 struct cl_io *cl_io_top(struct cl_io *io);
2508 void cl_io_print(const struct lu_env *env, void *cookie,
2509 lu_printer_t printer, const struct cl_io *io);
2511 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2513 typeof(foo_io) __foo_io = (foo_io); \
2515 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2516 memset(&__foo_io->base + 1, 0, \
2517 (sizeof *__foo_io) - sizeof __foo_io->base); \
2522 /** \defgroup cl_page_list cl_page_list
2526 * Last page in the page list.
2528 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2530 LASSERT(plist->pl_nr > 0);
2531 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2534 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2536 LASSERT(plist->pl_nr > 0);
2537 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2541 * Iterate over pages in a page list.
2543 #define cl_page_list_for_each(page, list) \
2544 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2547 * Iterate over pages in a page list, taking possible removals into account.
2549 #define cl_page_list_for_each_safe(page, temp, list) \
2550 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2552 void cl_page_list_init (struct cl_page_list *plist);
2553 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2554 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2555 struct cl_page *page);
2556 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2557 struct cl_page *page);
2558 void cl_page_list_splice (struct cl_page_list *list,
2559 struct cl_page_list *head);
2560 void cl_page_list_del (const struct lu_env *env,
2561 struct cl_page_list *plist, struct cl_page *page);
2562 void cl_page_list_disown (const struct lu_env *env,
2563 struct cl_io *io, struct cl_page_list *plist);
2564 int cl_page_list_own (const struct lu_env *env,
2565 struct cl_io *io, struct cl_page_list *plist);
2566 void cl_page_list_assume (const struct lu_env *env,
2567 struct cl_io *io, struct cl_page_list *plist);
2568 void cl_page_list_discard(const struct lu_env *env,
2569 struct cl_io *io, struct cl_page_list *plist);
2570 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2572 void cl_2queue_init (struct cl_2queue *queue);
2573 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2574 void cl_2queue_disown (const struct lu_env *env,
2575 struct cl_io *io, struct cl_2queue *queue);
2576 void cl_2queue_assume (const struct lu_env *env,
2577 struct cl_io *io, struct cl_2queue *queue);
2578 void cl_2queue_discard (const struct lu_env *env,
2579 struct cl_io *io, struct cl_2queue *queue);
2580 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2581 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2583 /** @} cl_page_list */
2585 /** \defgroup cl_req cl_req
2587 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2588 enum cl_req_type crt, int nr_objects);
2590 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2591 struct cl_page *page);
2592 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2593 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2594 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2595 struct cl_req_attr *attr, u64 flags);
2596 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2598 /** \defgroup cl_sync_io cl_sync_io
2602 * Anchor for synchronous transfer. This is allocated on a stack by thread
2603 * doing synchronous transfer, and a pointer to this structure is set up in
2604 * every page submitted for transfer. Transfer completion routine updates
2605 * anchor and wakes up waiting thread when transfer is complete.
2608 /** number of pages yet to be transferred. */
2609 atomic_t csi_sync_nr;
2612 /** barrier of destroy this structure */
2613 atomic_t csi_barrier;
2614 /** completion to be signaled when transfer is complete. */
2615 wait_queue_head_t csi_waitq;
2616 /** callback to invoke when this IO is finished */
2617 void (*csi_end_io)(const struct lu_env *,
2618 struct cl_sync_io *);
2621 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2622 void (*end)(const struct lu_env *, struct cl_sync_io *));
2623 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2625 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2627 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2629 /** @} cl_sync_io */
2633 /** \defgroup cl_env cl_env
2635 * lu_env handling for a client.
2637 * lu_env is an environment within which lustre code executes. Its major part
2638 * is lu_context---a fast memory allocation mechanism that is used to conserve
2639 * precious kernel stack space. Originally lu_env was designed for a server,
2642 * - there is a (mostly) fixed number of threads, and
2644 * - call chains have no non-lustre portions inserted between lustre code.
2646 * On a client both these assumtpion fails, because every user thread can
2647 * potentially execute lustre code as part of a system call, and lustre calls
2648 * into VFS or MM that call back into lustre.
2650 * To deal with that, cl_env wrapper functions implement the following
2653 * - allocation and destruction of environment is amortized by caching no
2654 * longer used environments instead of destroying them;
2656 * - there is a notion of "current" environment, attached to the kernel
2657 * data structure representing current thread Top-level lustre code
2658 * allocates an environment and makes it current, then calls into
2659 * non-lustre code, that in turn calls lustre back. Low-level lustre
2660 * code thus called can fetch environment created by the top-level code
2661 * and reuse it, avoiding additional environment allocation.
2662 * Right now, three interfaces can attach the cl_env to running thread:
2665 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2667 * \see lu_env, lu_context, lu_context_key
2670 struct cl_env_nest {
2675 struct lu_env *cl_env_peek (int *refcheck);
2676 struct lu_env *cl_env_get (int *refcheck);
2677 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2678 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2679 void cl_env_put (struct lu_env *env, int *refcheck);
2680 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2681 void *cl_env_reenter (void);
2682 void cl_env_reexit (void *cookie);
2683 void cl_env_implant (struct lu_env *env, int *refcheck);
2684 void cl_env_unplant (struct lu_env *env, int *refcheck);
2685 unsigned cl_env_cache_purge(unsigned nr);
2686 struct lu_env *cl_env_percpu_get (void);
2687 void cl_env_percpu_put (struct lu_env *env);
2694 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2695 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2697 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2698 struct lu_device_type *ldt,
2699 struct lu_device *next);
2702 int cl_global_init(void);
2703 void cl_global_fini(void);
2705 #endif /* _LINUX_CL_OBJECT_H */