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 * Operations implemented for each cl object layer.
311 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
313 struct cl_object_operations {
315 * Initialize page slice for this layer. Called top-to-bottom through
316 * every object layer when a new cl_page is instantiated. Layer
317 * keeping private per-page data, or requiring its own page operations
318 * vector should allocate these data here, and attach then to the page
319 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
322 * \retval NULL success.
324 * \retval ERR_PTR(errno) failure code.
326 * \retval valid-pointer pointer to already existing referenced page
327 * to be used instead of newly created.
329 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
330 struct cl_page *page, pgoff_t index);
332 * Initialize lock slice for this layer. Called top-to-bottom through
333 * every object layer when a new cl_lock is instantiated. Layer
334 * keeping private per-lock data, or requiring its own lock operations
335 * vector should allocate these data here, and attach then to the lock
336 * by calling cl_lock_slice_add(). Mandatory.
338 int (*coo_lock_init)(const struct lu_env *env,
339 struct cl_object *obj, struct cl_lock *lock,
340 const struct cl_io *io);
342 * Initialize io state for a given layer.
344 * called top-to-bottom once per io existence to initialize io
345 * state. If layer wants to keep some state for this type of io, it
346 * has to embed struct cl_io_slice in lu_env::le_ses, and register
347 * slice with cl_io_slice_add(). It is guaranteed that all threads
348 * participating in this io share the same session.
350 int (*coo_io_init)(const struct lu_env *env,
351 struct cl_object *obj, struct cl_io *io);
353 * Fill portion of \a attr that this layer controls. This method is
354 * called top-to-bottom through all object layers.
356 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
358 * \return 0: to continue
359 * \return +ve: to stop iterating through layers (but 0 is returned
360 * from enclosing cl_object_attr_get())
361 * \return -ve: to signal error
363 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
364 struct cl_attr *attr);
368 * \a valid is a bitmask composed from enum #cl_attr_valid, and
369 * indicating what attributes are to be set.
371 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
373 * \return the same convention as for
374 * cl_object_operations::coo_attr_get() is used.
376 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
377 const struct cl_attr *attr, unsigned valid);
379 * Update object configuration. Called top-to-bottom to modify object
382 * XXX error conditions and handling.
384 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
385 const struct cl_object_conf *conf);
387 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
388 * object. Layers are supposed to fill parts of \a lvb that will be
389 * shipped to the glimpse originator as a glimpse result.
391 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
392 * \see osc_object_glimpse()
394 int (*coo_glimpse)(const struct lu_env *env,
395 const struct cl_object *obj, struct ost_lvb *lvb);
397 * Object prune method. Called when the layout is going to change on
398 * this object, therefore each layer has to clean up their cache,
399 * mainly pages and locks.
401 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
403 * Object getstripe method.
405 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
406 struct lov_user_md __user *lum);
408 * Find whether there is any callback data (ldlm lock) attached upon
411 int (*coo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
412 ldlm_iterator_t iter, void *data);
414 * Get FIEMAP mapping from the object.
416 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
417 struct ll_fiemap_info_key *fmkey,
418 struct fiemap *fiemap, size_t *buflen);
420 * Get attributes of the object from server. (top->bottom)
422 int (*coo_obd_info_get)(const struct lu_env *env, struct cl_object *obj,
423 struct obd_info *oinfo,
424 struct ptlrpc_request_set *set);
426 * Get data version of the object. (top->bottom)
428 int (*coo_data_version)(const struct lu_env *env, struct cl_object *obj,
429 __u64 *version, int flags);
433 * Extended header for client object.
435 struct cl_object_header {
436 /** Standard lu_object_header. cl_object::co_lu::lo_header points
438 struct lu_object_header coh_lu;
441 * Parent object. It is assumed that an object has a well-defined
442 * parent, but not a well-defined child (there may be multiple
443 * sub-objects, for the same top-object). cl_object_header::coh_parent
444 * field allows certain code to be written generically, without
445 * limiting possible cl_object layouts unduly.
447 struct cl_object_header *coh_parent;
449 * Protects consistency between cl_attr of parent object and
450 * attributes of sub-objects, that the former is calculated ("merged")
453 * \todo XXX this can be read/write lock if needed.
455 spinlock_t coh_attr_guard;
457 * Size of cl_page + page slices
459 unsigned short coh_page_bufsize;
461 * Number of objects above this one: 0 for a top-object, 1 for its
464 unsigned char coh_nesting;
468 * Helper macro: iterate over all layers of the object \a obj, assigning every
469 * layer top-to-bottom to \a slice.
471 #define cl_object_for_each(slice, obj) \
472 list_for_each_entry((slice), \
473 &(obj)->co_lu.lo_header->loh_layers,\
477 * Helper macro: iterate over all layers of the object \a obj, assigning every
478 * layer bottom-to-top to \a slice.
480 #define cl_object_for_each_reverse(slice, obj) \
481 list_for_each_entry_reverse((slice), \
482 &(obj)->co_lu.lo_header->loh_layers,\
487 #define CL_PAGE_EOF ((pgoff_t)~0ull)
489 /** \addtogroup cl_page cl_page
493 * Layered client page.
495 * cl_page: represents a portion of a file, cached in the memory. All pages
496 * of the given file are of the same size, and are kept in the radix tree
497 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
498 * of the top-level file object are first class cl_objects, they have their
499 * own radix trees of pages and hence page is implemented as a sequence of
500 * struct cl_pages's, linked into double-linked list through
501 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
502 * corresponding radix tree at the corresponding logical offset.
504 * cl_page is associated with VM page of the hosting environment (struct
505 * page in Linux kernel, for example), struct page. It is assumed, that this
506 * association is implemented by one of cl_page layers (top layer in the
507 * current design) that
509 * - intercepts per-VM-page call-backs made by the environment (e.g.,
512 * - translates state (page flag bits) and locking between lustre and
515 * The association between cl_page and struct page is immutable and
516 * established when cl_page is created.
518 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
519 * this io an exclusive access to this page w.r.t. other io attempts and
520 * various events changing page state (such as transfer completion, or
521 * eviction of the page from the memory). Note, that in general cl_io
522 * cannot be identified with a particular thread, and page ownership is not
523 * exactly equal to the current thread holding a lock on the page. Layer
524 * implementing association between cl_page and struct page has to implement
525 * ownership on top of available synchronization mechanisms.
527 * While lustre client maintains the notion of an page ownership by io,
528 * hosting MM/VM usually has its own page concurrency control
529 * mechanisms. For example, in Linux, page access is synchronized by the
530 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
531 * takes care to acquire and release such locks as necessary around the
532 * calls to the file system methods (->readpage(), ->prepare_write(),
533 * ->commit_write(), etc.). This leads to the situation when there are two
534 * different ways to own a page in the client:
536 * - client code explicitly and voluntary owns the page (cl_page_own());
538 * - VM locks a page and then calls the client, that has "to assume"
539 * the ownership from the VM (cl_page_assume()).
541 * Dual methods to release ownership are cl_page_disown() and
542 * cl_page_unassume().
544 * cl_page is reference counted (cl_page::cp_ref). When reference counter
545 * drops to 0, the page is returned to the cache, unless it is in
546 * cl_page_state::CPS_FREEING state, in which case it is immediately
549 * The general logic guaranteeing the absence of "existential races" for
550 * pages is the following:
552 * - there are fixed known ways for a thread to obtain a new reference
555 * - by doing a lookup in the cl_object radix tree, protected by the
558 * - by starting from VM-locked struct page and following some
559 * hosting environment method (e.g., following ->private pointer in
560 * the case of Linux kernel), see cl_vmpage_page();
562 * - when the page enters cl_page_state::CPS_FREEING state, all these
563 * ways are severed with the proper synchronization
564 * (cl_page_delete());
566 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
569 * - no new references to the page in cl_page_state::CPS_FREEING state
570 * are allowed (checked in cl_page_get()).
572 * Together this guarantees that when last reference to a
573 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
574 * page, as neither references to it can be acquired at that point, nor
577 * cl_page is a state machine. States are enumerated in enum
578 * cl_page_state. Possible state transitions are enumerated in
579 * cl_page_state_set(). State transition process (i.e., actual changing of
580 * cl_page::cp_state field) is protected by the lock on the underlying VM
583 * Linux Kernel implementation.
585 * Binding between cl_page and struct page (which is a typedef for
586 * struct page) is implemented in the vvp layer. cl_page is attached to the
587 * ->private pointer of the struct page, together with the setting of
588 * PG_private bit in page->flags, and acquiring additional reference on the
589 * struct page (much like struct buffer_head, or any similar file system
590 * private data structures).
592 * PG_locked lock is used to implement both ownership and transfer
593 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
594 * states. No additional references are acquired for the duration of the
597 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
598 * write-out is "protected" by the special PG_writeback bit.
602 * States of cl_page. cl_page.c assumes particular order here.
604 * The page state machine is rather crude, as it doesn't recognize finer page
605 * states like "dirty" or "up to date". This is because such states are not
606 * always well defined for the whole stack (see, for example, the
607 * implementation of the read-ahead, that hides page up-to-dateness to track
608 * cache hits accurately). Such sub-states are maintained by the layers that
609 * are interested in them.
613 * Page is in the cache, un-owned. Page leaves cached state in the
616 * - [cl_page_state::CPS_OWNED] io comes across the page and
619 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
620 * req-formation engine decides that it wants to include this page
621 * into an cl_req being constructed, and yanks it from the cache;
623 * - [cl_page_state::CPS_FREEING] VM callback is executed to
624 * evict the page form the memory;
626 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
630 * Page is exclusively owned by some cl_io. Page may end up in this
631 * state as a result of
633 * - io creating new page and immediately owning it;
635 * - [cl_page_state::CPS_CACHED] io finding existing cached page
638 * - [cl_page_state::CPS_OWNED] io finding existing owned page
639 * and waiting for owner to release the page;
641 * Page leaves owned state in the following cases:
643 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
644 * the cache, doing nothing;
646 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
649 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
650 * transfer for this page;
652 * - [cl_page_state::CPS_FREEING] io decides to destroy this
653 * page (e.g., as part of truncate or extent lock cancellation).
655 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
659 * Page is being written out, as a part of a transfer. This state is
660 * entered when req-formation logic decided that it wants this page to
661 * be sent through the wire _now_. Specifically, it means that once
662 * this state is achieved, transfer completion handler (with either
663 * success or failure indication) is guaranteed to be executed against
664 * this page independently of any locks and any scheduling decisions
665 * made by the hosting environment (that effectively means that the
666 * page is never put into cl_page_state::CPS_PAGEOUT state "in
667 * advance". This property is mentioned, because it is important when
668 * reasoning about possible dead-locks in the system). The page can
669 * enter this state as a result of
671 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
672 * write-out of this page, or
674 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
675 * that it has enough dirty pages cached to issue a "good"
678 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
679 * is completed---it is moved into cl_page_state::CPS_CACHED state.
681 * Underlying VM page is locked for the duration of transfer.
683 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
687 * Page is being read in, as a part of a transfer. This is quite
688 * similar to the cl_page_state::CPS_PAGEOUT state, except that
689 * read-in is always "immediate"---there is no such thing a sudden
690 * construction of read cl_req from cached, presumably not up to date,
693 * Underlying VM page is locked for the duration of transfer.
695 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
699 * Page is being destroyed. This state is entered when client decides
700 * that page has to be deleted from its host object, as, e.g., a part
703 * Once this state is reached, there is no way to escape it.
705 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
712 /** Host page, the page is from the host inode which the cl_page
716 /** Transient page, the transient cl_page is used to bind a cl_page
717 * to vmpage which is not belonging to the same object of cl_page.
718 * it is used in DirectIO, lockless IO and liblustre. */
723 * Fields are protected by the lock on struct page, except for atomics and
726 * \invariant Data type invariants are in cl_page_invariant(). Basically:
727 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
728 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
729 * cl_page::cp_owner (when set).
732 /** Reference counter. */
734 /** Transfer error. */
736 /** An object this page is a part of. Immutable after creation. */
737 struct cl_object *cp_obj;
739 struct page *cp_vmpage;
740 /** Linkage of pages within group. Pages must be owned */
741 struct list_head cp_batch;
742 /** List of slices. Immutable after creation. */
743 struct list_head cp_layers;
744 /** Linkage of pages within cl_req. */
745 struct list_head cp_flight;
747 * Page state. This field is const to avoid accidental update, it is
748 * modified only internally within cl_page.c. Protected by a VM lock.
750 const enum cl_page_state cp_state;
752 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
755 enum cl_page_type cp_type;
758 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
759 * by sub-io. Protected by a VM lock.
761 struct cl_io *cp_owner;
763 * Owning IO request in cl_page_state::CPS_PAGEOUT and
764 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
765 * the top-level pages. Protected by a VM lock.
767 struct cl_req *cp_req;
768 /** List of references to this page, for debugging. */
769 struct lu_ref cp_reference;
770 /** Link to an object, for debugging. */
771 struct lu_ref_link cp_obj_ref;
772 /** Link to a queue, for debugging. */
773 struct lu_ref_link cp_queue_ref;
774 /** Assigned if doing a sync_io */
775 struct cl_sync_io *cp_sync_io;
779 * Per-layer part of cl_page.
781 * \see vvp_page, lov_page, osc_page
783 struct cl_page_slice {
784 struct cl_page *cpl_page;
787 * Object slice corresponding to this page slice. Immutable after
790 struct cl_object *cpl_obj;
791 const struct cl_page_operations *cpl_ops;
792 /** Linkage into cl_page::cp_layers. Immutable after creation. */
793 struct list_head cpl_linkage;
797 * Lock mode. For the client extent locks.
809 * Requested transfer type.
819 * Per-layer page operations.
821 * Methods taking an \a io argument are for the activity happening in the
822 * context of given \a io. Page is assumed to be owned by that io, except for
823 * the obvious cases (like cl_page_operations::cpo_own()).
825 * \see vvp_page_ops, lov_page_ops, osc_page_ops
827 struct cl_page_operations {
829 * cl_page<->struct page methods. Only one layer in the stack has to
830 * implement these. Current code assumes that this functionality is
831 * provided by the topmost layer, see cl_page_disown0() as an example.
835 * Called when \a io acquires this page into the exclusive
836 * ownership. When this method returns, it is guaranteed that the is
837 * not owned by other io, and no transfer is going on against
841 * \see vvp_page_own(), lov_page_own()
843 int (*cpo_own)(const struct lu_env *env,
844 const struct cl_page_slice *slice,
845 struct cl_io *io, int nonblock);
846 /** Called when ownership it yielded. Optional.
848 * \see cl_page_disown()
849 * \see vvp_page_disown()
851 void (*cpo_disown)(const struct lu_env *env,
852 const struct cl_page_slice *slice, struct cl_io *io);
854 * Called for a page that is already "owned" by \a io from VM point of
857 * \see cl_page_assume()
858 * \see vvp_page_assume(), lov_page_assume()
860 void (*cpo_assume)(const struct lu_env *env,
861 const struct cl_page_slice *slice, struct cl_io *io);
862 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
863 * bottom-to-top when IO releases a page without actually unlocking
866 * \see cl_page_unassume()
867 * \see vvp_page_unassume()
869 void (*cpo_unassume)(const struct lu_env *env,
870 const struct cl_page_slice *slice,
873 * Announces whether the page contains valid data or not by \a uptodate.
875 * \see cl_page_export()
876 * \see vvp_page_export()
878 void (*cpo_export)(const struct lu_env *env,
879 const struct cl_page_slice *slice, int uptodate);
881 * Checks whether underlying VM page is locked (in the suitable
882 * sense). Used for assertions.
884 * \retval -EBUSY: page is protected by a lock of a given mode;
885 * \retval -ENODATA: page is not protected by a lock;
886 * \retval 0: this layer cannot decide. (Should never happen.)
888 int (*cpo_is_vmlocked)(const struct lu_env *env,
889 const struct cl_page_slice *slice);
895 * Called when page is truncated from the object. Optional.
897 * \see cl_page_discard()
898 * \see vvp_page_discard(), osc_page_discard()
900 void (*cpo_discard)(const struct lu_env *env,
901 const struct cl_page_slice *slice,
904 * Called when page is removed from the cache, and is about to being
905 * destroyed. Optional.
907 * \see cl_page_delete()
908 * \see vvp_page_delete(), osc_page_delete()
910 void (*cpo_delete)(const struct lu_env *env,
911 const struct cl_page_slice *slice);
912 /** Destructor. Frees resources and slice itself. */
913 void (*cpo_fini)(const struct lu_env *env,
914 struct cl_page_slice *slice);
916 * Optional debugging helper. Prints given page slice.
918 * \see cl_page_print()
920 int (*cpo_print)(const struct lu_env *env,
921 const struct cl_page_slice *slice,
922 void *cookie, lu_printer_t p);
926 * Transfer methods. See comment on cl_req for a description of
927 * transfer formation and life-cycle.
932 * Request type dependent vector of operations.
934 * Transfer operations depend on transfer mode (cl_req_type). To avoid
935 * passing transfer mode to each and every of these methods, and to
936 * avoid branching on request type inside of the methods, separate
937 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
938 * provided. That is, method invocation usually looks like
940 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
944 * Called when a page is submitted for a transfer as a part of
947 * \return 0 : page is eligible for submission;
948 * \return -EALREADY : skip this page;
949 * \return -ve : error.
951 * \see cl_page_prep()
953 int (*cpo_prep)(const struct lu_env *env,
954 const struct cl_page_slice *slice,
957 * Completion handler. This is guaranteed to be eventually
958 * fired after cl_page_operations::cpo_prep() or
959 * cl_page_operations::cpo_make_ready() call.
961 * This method can be called in a non-blocking context. It is
962 * guaranteed however, that the page involved and its object
963 * are pinned in memory (and, hence, calling cl_page_put() is
966 * \see cl_page_completion()
968 void (*cpo_completion)(const struct lu_env *env,
969 const struct cl_page_slice *slice,
972 * Called when cached page is about to be added to the
973 * cl_req as a part of req formation.
975 * \return 0 : proceed with this page;
976 * \return -EAGAIN : skip this page;
977 * \return -ve : error.
979 * \see cl_page_make_ready()
981 int (*cpo_make_ready)(const struct lu_env *env,
982 const struct cl_page_slice *slice);
985 * Tell transfer engine that only [to, from] part of a page should be
988 * This is used for immediate transfers.
990 * \todo XXX this is not very good interface. It would be much better
991 * if all transfer parameters were supplied as arguments to
992 * cl_io_operations::cio_submit() call, but it is not clear how to do
993 * this for page queues.
995 * \see cl_page_clip()
997 void (*cpo_clip)(const struct lu_env *env,
998 const struct cl_page_slice *slice,
1001 * \pre the page was queued for transferring.
1002 * \post page is removed from client's pending list, or -EBUSY
1003 * is returned if it has already been in transferring.
1005 * This is one of seldom page operation which is:
1006 * 0. called from top level;
1007 * 1. don't have vmpage locked;
1008 * 2. every layer should synchronize execution of its ->cpo_cancel()
1009 * with completion handlers. Osc uses client obd lock for this
1010 * purpose. Based on there is no vvp_page_cancel and
1011 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1013 * \see osc_page_cancel().
1015 int (*cpo_cancel)(const struct lu_env *env,
1016 const struct cl_page_slice *slice);
1018 * Write out a page by kernel. This is only called by ll_writepage
1021 * \see cl_page_flush()
1023 int (*cpo_flush)(const struct lu_env *env,
1024 const struct cl_page_slice *slice,
1030 * Helper macro, dumping detailed information about \a page into a log.
1032 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1034 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1035 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1036 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1037 CDEBUG(mask, format , ## __VA_ARGS__); \
1042 * Helper macro, dumping shorter information about \a page into a log.
1044 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1046 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1047 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1048 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1049 CDEBUG(mask, format , ## __VA_ARGS__); \
1053 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1055 LASSERT(page->cp_vmpage != NULL);
1056 return page->cp_vmpage;
1060 * Check if a cl_page is in use.
1062 * Client cache holds a refcount, this refcount will be dropped when
1063 * the page is taken out of cache, see vvp_page_delete().
1065 static inline bool __page_in_use(const struct cl_page *page, int refc)
1067 return (atomic_read(&page->cp_ref) > refc + 1);
1071 * Caller itself holds a refcount of cl_page.
1073 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1075 * Caller doesn't hold a refcount.
1077 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1081 /** \addtogroup cl_lock cl_lock
1085 * Extent locking on the client.
1089 * The locking model of the new client code is built around
1093 * data-type representing an extent lock on a regular file. cl_lock is a
1094 * layered object (much like cl_object and cl_page), it consists of a header
1095 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1096 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1098 * Typical cl_lock consists of the two layers:
1100 * - vvp_lock (vvp specific data), and
1101 * - lov_lock (lov specific data).
1103 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1104 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1106 * - lovsub_lock, and
1109 * Each sub-lock is associated with a cl_object (representing stripe
1110 * sub-object or the file to which top-level cl_lock is associated to), and is
1111 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1112 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1113 * is different from cl_page, that doesn't fan out (there is usually exactly
1114 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1115 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1119 * cl_lock is a cacheless data container for the requirements of locks to
1120 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1123 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1124 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1126 * INTERFACE AND USAGE
1128 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1129 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1130 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1131 * consists of multiple sub cl_locks, each sub locks will be enqueued
1132 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1133 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1136 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1137 * method will be called for each layer to release the resource held by this
1138 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1139 * clo_enqueue time, is released.
1141 * LDLM lock can only be canceled if there is no cl_lock using it.
1143 * Overall process of the locking during IO operation is as following:
1145 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1146 * is called on each layer. Responsibility of this method is to add locks,
1147 * needed by a given layer into cl_io.ci_lockset.
1149 * - once locks for all layers were collected, they are sorted to avoid
1150 * dead-locks (cl_io_locks_sort()), and enqueued.
1152 * - when all locks are acquired, IO is performed;
1154 * - locks are released after IO is complete.
1156 * Striping introduces major additional complexity into locking. The
1157 * fundamental problem is that it is generally unsafe to actively use (hold)
1158 * two locks on the different OST servers at the same time, as this introduces
1159 * inter-server dependency and can lead to cascading evictions.
1161 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1162 * that no multi-stripe locks are taken (note that this design abandons POSIX
1163 * read/write semantics). Such pieces ideally can be executed concurrently. At
1164 * the same time, certain types of IO cannot be sub-divived, without
1165 * sacrificing correctness. This includes:
1167 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1170 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1172 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1173 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1174 * has to be held together with the usual lock on [offset, offset + count].
1176 * Interaction with DLM
1178 * In the expected setup, cl_lock is ultimately backed up by a collection of
1179 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1180 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1181 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1182 * description of interaction with DLM.
1188 struct cl_lock_descr {
1189 /** Object this lock is granted for. */
1190 struct cl_object *cld_obj;
1191 /** Index of the first page protected by this lock. */
1193 /** Index of the last page (inclusive) protected by this lock. */
1195 /** Group ID, for group lock */
1198 enum cl_lock_mode cld_mode;
1200 * flags to enqueue lock. A combination of bit-flags from
1201 * enum cl_enq_flags.
1203 __u32 cld_enq_flags;
1206 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1207 #define PDESCR(descr) \
1208 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1209 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1211 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1214 * Layered client lock.
1217 /** List of slices. Immutable after creation. */
1218 struct list_head cll_layers;
1219 /** lock attribute, extent, cl_object, etc. */
1220 struct cl_lock_descr cll_descr;
1224 * Per-layer part of cl_lock
1226 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1228 struct cl_lock_slice {
1229 struct cl_lock *cls_lock;
1230 /** Object slice corresponding to this lock slice. Immutable after
1232 struct cl_object *cls_obj;
1233 const struct cl_lock_operations *cls_ops;
1234 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1235 struct list_head cls_linkage;
1240 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1242 struct cl_lock_operations {
1245 * Attempts to enqueue the lock. Called top-to-bottom.
1247 * \retval 0 this layer has enqueued the lock successfully
1248 * \retval >0 this layer has enqueued the lock, but need to wait on
1249 * @anchor for resources
1250 * \retval -ve failure
1252 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1253 * \see osc_lock_enqueue()
1255 int (*clo_enqueue)(const struct lu_env *env,
1256 const struct cl_lock_slice *slice,
1257 struct cl_io *io, struct cl_sync_io *anchor);
1259 * Cancel a lock, release its DLM lock ref, while does not cancel the
1262 void (*clo_cancel)(const struct lu_env *env,
1263 const struct cl_lock_slice *slice);
1266 * Destructor. Frees resources and the slice.
1268 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1269 * \see osc_lock_fini()
1271 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1273 * Optional debugging helper. Prints given lock slice.
1275 int (*clo_print)(const struct lu_env *env,
1276 void *cookie, lu_printer_t p,
1277 const struct cl_lock_slice *slice);
1280 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1282 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1283 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1284 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1285 CDEBUG(mask, format , ## __VA_ARGS__); \
1289 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1293 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1299 /** \addtogroup cl_page_list cl_page_list
1300 * Page list used to perform collective operations on a group of pages.
1302 * Pages are added to the list one by one. cl_page_list acquires a reference
1303 * for every page in it. Page list is used to perform collective operations on
1306 * - submit pages for an immediate transfer,
1308 * - own pages on behalf of certain io (waiting for each page in turn),
1312 * When list is finalized, it releases references on all pages it still has.
1314 * \todo XXX concurrency control.
1318 struct cl_page_list {
1320 struct list_head pl_pages;
1321 struct task_struct *pl_owner;
1325 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1326 * contains an incoming page list and an outgoing page list.
1329 struct cl_page_list c2_qin;
1330 struct cl_page_list c2_qout;
1333 /** @} cl_page_list */
1335 /** \addtogroup cl_io cl_io
1340 * cl_io represents a high level I/O activity like
1341 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1344 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1345 * important distinction. We want to minimize number of calls to the allocator
1346 * in the fast path, e.g., in the case of read(2) when everything is cached:
1347 * client already owns the lock over region being read, and data are cached
1348 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1349 * per-layer io state is stored in the session, associated with the io, see
1350 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1351 * by using free-lists, see cl_env_get().
1353 * There is a small predefined number of possible io types, enumerated in enum
1356 * cl_io is a state machine, that can be advanced concurrently by the multiple
1357 * threads. It is up to these threads to control the concurrency and,
1358 * specifically, to detect when io is done, and its state can be safely
1361 * For read/write io overall execution plan is as following:
1363 * (0) initialize io state through all layers;
1365 * (1) loop: prepare chunk of work to do
1367 * (2) call all layers to collect locks they need to process current chunk
1369 * (3) sort all locks to avoid dead-locks, and acquire them
1371 * (4) process the chunk: call per-page methods
1372 * cl_io_operations::cio_prepare_write(),
1373 * cl_io_operations::cio_commit_write() for write)
1379 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1380 * address allocation efficiency issues mentioned above), and returns with the
1381 * special error condition from per-page method when current sub-io has to
1382 * block. This causes io loop to be repeated, and lov switches to the next
1383 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1388 /** read system call */
1390 /** write system call */
1392 /** truncate, utime system calls */
1395 * page fault handling
1399 * fsync system call handling
1400 * To write out a range of file
1404 * Miscellaneous io. This is used for occasional io activity that
1405 * doesn't fit into other types. Currently this is used for:
1407 * - cancellation of an extent lock. This io exists as a context
1408 * to write dirty pages from under the lock being canceled back
1411 * - VM induced page write-out. An io context for writing page out
1412 * for memory cleansing;
1414 * - glimpse. An io context to acquire glimpse lock.
1416 * - grouplock. An io context to acquire group lock.
1418 * CIT_MISC io is used simply as a context in which locks and pages
1419 * are manipulated. Such io has no internal "process", that is,
1420 * cl_io_loop() is never called for it.
1427 * States of cl_io state machine
1430 /** Not initialized. */
1434 /** IO iteration started. */
1438 /** Actual IO is in progress. */
1440 /** IO for the current iteration finished. */
1442 /** Locks released. */
1444 /** Iteration completed. */
1446 /** cl_io finalized. */
1451 * IO state private for a layer.
1453 * This is usually embedded into layer session data, rather than allocated
1456 * \see vvp_io, lov_io, osc_io
1458 struct cl_io_slice {
1459 struct cl_io *cis_io;
1460 /** corresponding object slice. Immutable after creation. */
1461 struct cl_object *cis_obj;
1462 /** io operations. Immutable after creation. */
1463 const struct cl_io_operations *cis_iop;
1465 * linkage into a list of all slices for a given cl_io, hanging off
1466 * cl_io::ci_layers. Immutable after creation.
1468 struct list_head cis_linkage;
1471 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1474 struct cl_read_ahead {
1475 /* Maximum page index the readahead window will end.
1476 * This is determined DLM lock coverage, RPC and stripe boundary.
1477 * cra_end is included. */
1479 /* Release routine. If readahead holds resources underneath, this
1480 * function should be called to release it. */
1481 void (*cra_release)(const struct lu_env *env, void *cbdata);
1482 /* Callback data for cra_release routine */
1486 static inline void cl_read_ahead_release(const struct lu_env *env,
1487 struct cl_read_ahead *ra)
1489 if (ra->cra_release != NULL)
1490 ra->cra_release(env, ra->cra_cbdata);
1491 memset(ra, 0, sizeof(*ra));
1496 * Per-layer io operations.
1497 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1499 struct cl_io_operations {
1501 * Vector of io state transition methods for every io type.
1503 * \see cl_page_operations::io
1507 * Prepare io iteration at a given layer.
1509 * Called top-to-bottom at the beginning of each iteration of
1510 * "io loop" (if it makes sense for this type of io). Here
1511 * layer selects what work it will do during this iteration.
1513 * \see cl_io_operations::cio_iter_fini()
1515 int (*cio_iter_init) (const struct lu_env *env,
1516 const struct cl_io_slice *slice);
1518 * Finalize io iteration.
1520 * Called bottom-to-top at the end of each iteration of "io
1521 * loop". Here layers can decide whether IO has to be
1524 * \see cl_io_operations::cio_iter_init()
1526 void (*cio_iter_fini) (const struct lu_env *env,
1527 const struct cl_io_slice *slice);
1529 * Collect locks for the current iteration of io.
1531 * Called top-to-bottom to collect all locks necessary for
1532 * this iteration. This methods shouldn't actually enqueue
1533 * anything, instead it should post a lock through
1534 * cl_io_lock_add(). Once all locks are collected, they are
1535 * sorted and enqueued in the proper order.
1537 int (*cio_lock) (const struct lu_env *env,
1538 const struct cl_io_slice *slice);
1540 * Finalize unlocking.
1542 * Called bottom-to-top to finish layer specific unlocking
1543 * functionality, after generic code released all locks
1544 * acquired by cl_io_operations::cio_lock().
1546 void (*cio_unlock)(const struct lu_env *env,
1547 const struct cl_io_slice *slice);
1549 * Start io iteration.
1551 * Once all locks are acquired, called top-to-bottom to
1552 * commence actual IO. In the current implementation,
1553 * top-level vvp_io_{read,write}_start() does all the work
1554 * synchronously by calling generic_file_*(), so other layers
1555 * are called when everything is done.
1557 int (*cio_start)(const struct lu_env *env,
1558 const struct cl_io_slice *slice);
1560 * Called top-to-bottom at the end of io loop. Here layer
1561 * might wait for an unfinished asynchronous io.
1563 void (*cio_end) (const struct lu_env *env,
1564 const struct cl_io_slice *slice);
1566 * Called bottom-to-top to notify layers that read/write IO
1567 * iteration finished, with \a nob bytes transferred.
1569 void (*cio_advance)(const struct lu_env *env,
1570 const struct cl_io_slice *slice,
1573 * Called once per io, bottom-to-top to release io resources.
1575 void (*cio_fini) (const struct lu_env *env,
1576 const struct cl_io_slice *slice);
1580 * Submit pages from \a queue->c2_qin for IO, and move
1581 * successfully submitted pages into \a queue->c2_qout. Return
1582 * non-zero if failed to submit even the single page. If
1583 * submission failed after some pages were moved into \a
1584 * queue->c2_qout, completion callback with non-zero ioret is
1587 int (*cio_submit)(const struct lu_env *env,
1588 const struct cl_io_slice *slice,
1589 enum cl_req_type crt,
1590 struct cl_2queue *queue);
1592 * Queue async page for write.
1593 * The difference between cio_submit and cio_queue is that
1594 * cio_submit is for urgent request.
1596 int (*cio_commit_async)(const struct lu_env *env,
1597 const struct cl_io_slice *slice,
1598 struct cl_page_list *queue, int from, int to,
1601 * Decide maximum read ahead extent
1603 * \pre io->ci_type == CIT_READ
1605 int (*cio_read_ahead)(const struct lu_env *env,
1606 const struct cl_io_slice *slice,
1607 pgoff_t start, struct cl_read_ahead *ra);
1609 * Optional debugging helper. Print given io slice.
1611 int (*cio_print)(const struct lu_env *env, void *cookie,
1612 lu_printer_t p, const struct cl_io_slice *slice);
1616 * Flags to lock enqueue procedure.
1621 * instruct server to not block, if conflicting lock is found. Instead
1622 * -EWOULDBLOCK is returned immediately.
1624 CEF_NONBLOCK = 0x00000001,
1626 * take lock asynchronously (out of order), as it cannot
1627 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1629 CEF_ASYNC = 0x00000002,
1631 * tell the server to instruct (though a flag in the blocking ast) an
1632 * owner of the conflicting lock, that it can drop dirty pages
1633 * protected by this lock, without sending them to the server.
1635 CEF_DISCARD_DATA = 0x00000004,
1637 * tell the sub layers that it must be a `real' lock. This is used for
1638 * mmapped-buffer locks and glimpse locks that must be never converted
1639 * into lockless mode.
1641 * \see vvp_mmap_locks(), cl_glimpse_lock().
1643 CEF_MUST = 0x00000008,
1645 * tell the sub layers that never request a `real' lock. This flag is
1646 * not used currently.
1648 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1649 * conversion policy: ci_lockreq describes generic information of lock
1650 * requirement for this IO, especially for locks which belong to the
1651 * object doing IO; however, lock itself may have precise requirements
1652 * that are described by the enqueue flags.
1654 CEF_NEVER = 0x00000010,
1656 * for async glimpse lock.
1658 CEF_AGL = 0x00000020,
1660 * enqueue a lock to test DLM lock existence.
1662 CEF_PEEK = 0x00000040,
1664 * mask of enq_flags.
1666 CEF_MASK = 0x0000007f,
1670 * Link between lock and io. Intermediate structure is needed, because the
1671 * same lock can be part of multiple io's simultaneously.
1673 struct cl_io_lock_link {
1674 /** linkage into one of cl_lockset lists. */
1675 struct list_head cill_linkage;
1676 struct cl_lock cill_lock;
1677 /** optional destructor */
1678 void (*cill_fini)(const struct lu_env *env,
1679 struct cl_io_lock_link *link);
1681 #define cill_descr cill_lock.cll_descr
1684 * Lock-set represents a collection of locks, that io needs at a
1685 * time. Generally speaking, client tries to avoid holding multiple locks when
1688 * - holding extent locks over multiple ost's introduces the danger of
1689 * "cascading timeouts";
1691 * - holding multiple locks over the same ost is still dead-lock prone,
1692 * see comment in osc_lock_enqueue(),
1694 * but there are certain situations where this is unavoidable:
1696 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1698 * - truncate has to take [new-size, EOF] lock for correctness;
1700 * - SNS has to take locks across full stripe for correctness;
1702 * - in the case when user level buffer, supplied to {read,write}(file0),
1703 * is a part of a memory mapped lustre file, client has to take a dlm
1704 * locks on file0, and all files that back up the buffer (or a part of
1705 * the buffer, that is being processed in the current chunk, in any
1706 * case, there are situations where at least 2 locks are necessary).
1708 * In such cases we at least try to take locks in the same consistent
1709 * order. To this end, all locks are first collected, then sorted, and then
1713 /** locks to be acquired. */
1714 struct list_head cls_todo;
1715 /** locks acquired. */
1716 struct list_head cls_done;
1720 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1721 * but 'req' is always to be thought as 'request' :-)
1723 enum cl_io_lock_dmd {
1724 /** Always lock data (e.g., O_APPEND). */
1726 /** Layers are free to decide between local and global locking. */
1728 /** Never lock: there is no cache (e.g., liblustre). */
1732 enum cl_fsync_mode {
1733 /** start writeback, do not wait for them to finish */
1735 /** start writeback and wait for them to finish */
1737 /** discard all of dirty pages in a specific file range */
1738 CL_FSYNC_DISCARD = 2,
1739 /** start writeback and make sure they have reached storage before
1740 * return. OST_SYNC RPC must be issued and finished */
1744 struct cl_io_rw_common {
1754 * cl_io is shared by all threads participating in this IO (in current
1755 * implementation only one thread advances IO, but parallel IO design and
1756 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1757 * is up to these threads to serialize their activities, including updates to
1758 * mutable cl_io fields.
1761 /** type of this IO. Immutable after creation. */
1762 enum cl_io_type ci_type;
1763 /** current state of cl_io state machine. */
1764 enum cl_io_state ci_state;
1765 /** main object this io is against. Immutable after creation. */
1766 struct cl_object *ci_obj;
1768 * Upper layer io, of which this io is a part of. Immutable after
1771 struct cl_io *ci_parent;
1772 /** List of slices. Immutable after creation. */
1773 struct list_head ci_layers;
1774 /** list of locks (to be) acquired by this io. */
1775 struct cl_lockset ci_lockset;
1776 /** lock requirements, this is just a help info for sublayers. */
1777 enum cl_io_lock_dmd ci_lockreq;
1780 struct cl_io_rw_common rd;
1783 struct cl_io_rw_common wr;
1787 struct cl_io_rw_common ci_rw;
1788 struct cl_setattr_io {
1789 struct ost_lvb sa_attr;
1790 unsigned int sa_attr_flags;
1791 unsigned int sa_valid;
1792 int sa_stripe_index;
1793 const struct lu_fid *sa_parent_fid;
1794 struct obd_capa *sa_capa;
1796 struct cl_fault_io {
1797 /** page index within file. */
1799 /** bytes valid byte on a faulted page. */
1801 /** writable page? for nopage() only */
1803 /** page of an executable? */
1805 /** page_mkwrite() */
1807 /** resulting page */
1808 struct cl_page *ft_page;
1810 struct cl_fsync_io {
1813 struct obd_capa *fi_capa;
1814 /** file system level fid */
1815 struct lu_fid *fi_fid;
1816 enum cl_fsync_mode fi_mode;
1817 /* how many pages were written/discarded */
1818 unsigned int fi_nr_written;
1821 struct cl_2queue ci_queue;
1824 unsigned int ci_continue:1,
1826 * This io has held grouplock, to inform sublayers that
1827 * don't do lockless i/o.
1831 * The whole IO need to be restarted because layout has been changed
1835 * to not refresh layout - the IO issuer knows that the layout won't
1836 * change(page operations, layout change causes all page to be
1837 * discarded), or it doesn't matter if it changes(sync).
1841 * Check if layout changed after the IO finishes. Mainly for HSM
1842 * requirement. If IO occurs to openning files, it doesn't need to
1843 * verify layout because HSM won't release openning files.
1844 * Right now, only two opertaions need to verify layout: glimpse
1849 * file is released, restore has to to be triggered by vvp layer
1851 ci_restore_needed:1,
1857 * Number of pages owned by this IO. For invariant checking.
1859 unsigned ci_owned_nr;
1864 /** \addtogroup cl_req cl_req
1869 * There are two possible modes of transfer initiation on the client:
1871 * - immediate transfer: this is started when a high level io wants a page
1872 * or a collection of pages to be transferred right away. Examples:
1873 * read-ahead, synchronous read in the case of non-page aligned write,
1874 * page write-out as a part of extent lock cancellation, page write-out
1875 * as a part of memory cleansing. Immediate transfer can be both
1876 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1878 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1879 * when io wants to transfer a page to the server some time later, when
1880 * it can be done efficiently. Example: pages dirtied by the write(2)
1883 * In any case, transfer takes place in the form of a cl_req, which is a
1884 * representation for a network RPC.
1886 * Pages queued for an opportunistic transfer are cached until it is decided
1887 * that efficient RPC can be composed of them. This decision is made by "a
1888 * req-formation engine", currently implemented as a part of osc
1889 * layer. Req-formation depends on many factors: the size of the resulting
1890 * RPC, whether or not multi-object RPCs are supported by the server,
1891 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1893 * For the immediate transfer io submits a cl_page_list, that req-formation
1894 * engine slices into cl_req's, possibly adding cached pages to some of
1895 * the resulting req's.
1897 * Whenever a page from cl_page_list is added to a newly constructed req, its
1898 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1899 * page state is atomically changed from cl_page_state::CPS_OWNED to
1900 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1901 * is zeroed, and cl_page::cp_req is set to the
1902 * req. cl_page_operations::cpo_prep() method at the particular layer might
1903 * return -EALREADY to indicate that it does not need to submit this page
1904 * at all. This is possible, for example, if page, submitted for read,
1905 * became up-to-date in the meantime; and for write, the page don't have
1906 * dirty bit marked. \see cl_io_submit_rw()
1908 * Whenever a cached page is added to a newly constructed req, its
1909 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1910 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1911 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1912 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1913 * might return -EAGAIN to indicate that this page is not eligible for the
1914 * transfer right now.
1918 * Plan is to divide transfers into "priority bands" (indicated when
1919 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1920 * and allow glueing of cached pages to immediate transfers only within single
1921 * band. This would make high priority transfers (like lock cancellation or
1922 * memory pressure induced write-out) really high priority.
1927 * Per-transfer attributes.
1929 struct cl_req_attr {
1930 /** Generic attributes for the server consumption. */
1931 struct obdo *cra_oa;
1933 struct obd_capa *cra_capa;
1935 char cra_jobid[LUSTRE_JOBID_SIZE];
1939 * Transfer request operations definable at every layer.
1941 * Concurrency: transfer formation engine synchronizes calls to all transfer
1944 struct cl_req_operations {
1946 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1947 * complete (all pages are added).
1949 * \see osc_req_prep()
1951 int (*cro_prep)(const struct lu_env *env,
1952 const struct cl_req_slice *slice);
1954 * Called top-to-bottom to fill in \a oa fields. This is called twice
1955 * with different flags, see bug 10150 and osc_build_req().
1957 * \param obj an object from cl_req which attributes are to be set in
1960 * \param oa struct obdo where attributes are placed
1962 * \param flags \a oa fields to be filled.
1964 void (*cro_attr_set)(const struct lu_env *env,
1965 const struct cl_req_slice *slice,
1966 const struct cl_object *obj,
1967 struct cl_req_attr *attr, u64 flags);
1969 * Called top-to-bottom from cl_req_completion() to notify layers that
1970 * transfer completed. Has to free all state allocated by
1971 * cl_device_operations::cdo_req_init().
1973 void (*cro_completion)(const struct lu_env *env,
1974 const struct cl_req_slice *slice, int ioret);
1978 * A per-object state that (potentially multi-object) transfer request keeps.
1981 /** object itself */
1982 struct cl_object *ro_obj;
1983 /** reference to cl_req_obj::ro_obj. For debugging. */
1984 struct lu_ref_link ro_obj_ref;
1985 /* something else? Number of pages for a given object? */
1991 * Transfer requests are not reference counted, because IO sub-system owns
1992 * them exclusively and knows when to free them.
1996 * cl_req is created by cl_req_alloc() that calls
1997 * cl_device_operations::cdo_req_init() device methods to allocate per-req
1998 * state in every layer.
2000 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2001 * contains pages for.
2003 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2004 * called top-to-bottom. At that point layers can modify req, let it pass, or
2005 * deny it completely. This is to support things like SNS that have transfer
2006 * ordering requirements invisible to the individual req-formation engine.
2008 * On transfer completion (or transfer timeout, or failure to initiate the
2009 * transfer of an allocated req), cl_req_operations::cro_completion() method
2010 * is called, after execution of cl_page_operations::cpo_completion() of all
2014 enum cl_req_type crq_type;
2015 /** A list of pages being transfered */
2016 struct list_head crq_pages;
2017 /** Number of pages in cl_req::crq_pages */
2018 unsigned crq_nrpages;
2019 /** An array of objects which pages are in ->crq_pages */
2020 struct cl_req_obj *crq_o;
2021 /** Number of elements in cl_req::crq_objs[] */
2022 unsigned crq_nrobjs;
2023 struct list_head crq_layers;
2027 * Per-layer state for request.
2029 struct cl_req_slice {
2030 struct cl_req *crs_req;
2031 struct cl_device *crs_dev;
2032 struct list_head crs_linkage;
2033 const struct cl_req_operations *crs_ops;
2038 enum cache_stats_item {
2039 /** how many cache lookups were performed */
2041 /** how many times cache lookup resulted in a hit */
2043 /** how many entities are in the cache right now */
2045 /** how many entities in the cache are actively used (and cannot be
2046 * evicted) right now */
2048 /** how many entities were created at all */
2053 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2056 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2058 struct cache_stats {
2059 const char *cs_name;
2060 atomic_t cs_stats[CS_NR];
2063 /** These are not exported so far */
2064 void cache_stats_init (struct cache_stats *cs, const char *name);
2067 * Client-side site. This represents particular client stack. "Global"
2068 * variables should (directly or indirectly) be added here to allow multiple
2069 * clients to co-exist in the single address space.
2072 struct lu_site cs_lu;
2074 * Statistical counters. Atomics do not scale, something better like
2075 * per-cpu counters is needed.
2077 * These are exported as /proc/fs/lustre/llite/.../site
2079 * When interpreting keep in mind that both sub-locks (and sub-pages)
2080 * and top-locks (and top-pages) are accounted here.
2082 struct cache_stats cs_pages;
2083 atomic_t cs_pages_state[CPS_NR];
2086 int cl_site_init(struct cl_site *s, struct cl_device *top);
2087 void cl_site_fini(struct cl_site *s);
2088 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2091 * Output client site statistical counters into a buffer. Suitable for
2092 * ll_rd_*()-style functions.
2094 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2099 * Type conversion and accessory functions.
2103 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2105 return container_of(site, struct cl_site, cs_lu);
2108 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2110 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2111 return container_of0(d, struct cl_device, cd_lu_dev);
2114 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2116 return &d->cd_lu_dev;
2119 static inline struct cl_object *lu2cl(const struct lu_object *o)
2121 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2122 return container_of0(o, struct cl_object, co_lu);
2125 static inline const struct cl_object_conf *
2126 lu2cl_conf(const struct lu_object_conf *conf)
2128 return container_of0(conf, struct cl_object_conf, coc_lu);
2131 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2133 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2136 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2138 return container_of0(h, struct cl_object_header, coh_lu);
2141 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2143 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2147 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2149 return luh2coh(obj->co_lu.lo_header);
2152 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2154 return lu_device_init(&d->cd_lu_dev, t);
2157 static inline void cl_device_fini(struct cl_device *d)
2159 lu_device_fini(&d->cd_lu_dev);
2162 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2163 struct cl_object *obj, pgoff_t index,
2164 const struct cl_page_operations *ops);
2165 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2166 struct cl_object *obj,
2167 const struct cl_lock_operations *ops);
2168 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2169 struct cl_object *obj, const struct cl_io_operations *ops);
2170 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2171 struct cl_device *dev,
2172 const struct cl_req_operations *ops);
2175 /** \defgroup cl_object cl_object
2177 struct cl_object *cl_object_top (struct cl_object *o);
2178 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2179 const struct lu_fid *fid,
2180 const struct cl_object_conf *c);
2182 int cl_object_header_init(struct cl_object_header *h);
2183 void cl_object_header_fini(struct cl_object_header *h);
2184 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2185 void cl_object_get (struct cl_object *o);
2186 void cl_object_attr_lock (struct cl_object *o);
2187 void cl_object_attr_unlock(struct cl_object *o);
2188 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2189 struct cl_attr *attr);
2190 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2191 const struct cl_attr *attr, unsigned valid);
2192 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2193 struct ost_lvb *lvb);
2194 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2195 const struct cl_object_conf *conf);
2196 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2197 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2198 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2199 struct lov_user_md __user *lum);
2200 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2201 ldlm_iterator_t iter, void *data);
2202 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2203 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2205 int cl_object_obd_info_get(const struct lu_env *env, struct cl_object *obj,
2206 struct obd_info *oinfo,
2207 struct ptlrpc_request_set *set);
2208 int cl_object_data_version(const struct lu_env *env, struct cl_object *obj,
2209 __u64 *version, int flags);
2212 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2214 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2216 return cl_object_header(o0) == cl_object_header(o1);
2219 static inline void cl_object_page_init(struct cl_object *clob, int size)
2221 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2222 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2223 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2226 static inline void *cl_object_page_slice(struct cl_object *clob,
2227 struct cl_page *page)
2229 return (void *)((char *)page + clob->co_slice_off);
2233 * Return refcount of cl_object.
2235 static inline int cl_object_refc(struct cl_object *clob)
2237 struct lu_object_header *header = clob->co_lu.lo_header;
2238 return atomic_read(&header->loh_ref);
2243 /** \defgroup cl_page cl_page
2251 /* callback of cl_page_gang_lookup() */
2253 struct cl_page *cl_page_find (const struct lu_env *env,
2254 struct cl_object *obj,
2255 pgoff_t idx, struct page *vmpage,
2256 enum cl_page_type type);
2257 struct cl_page *cl_page_alloc (const struct lu_env *env,
2258 struct cl_object *o, pgoff_t ind,
2259 struct page *vmpage,
2260 enum cl_page_type type);
2261 void cl_page_get (struct cl_page *page);
2262 void cl_page_put (const struct lu_env *env,
2263 struct cl_page *page);
2264 void cl_page_print (const struct lu_env *env, void *cookie,
2265 lu_printer_t printer,
2266 const struct cl_page *pg);
2267 void cl_page_header_print(const struct lu_env *env, void *cookie,
2268 lu_printer_t printer,
2269 const struct cl_page *pg);
2270 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2271 struct cl_page *cl_page_top (struct cl_page *page);
2273 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2274 const struct lu_device_type *dtype);
2279 * Functions dealing with the ownership of page by io.
2283 int cl_page_own (const struct lu_env *env,
2284 struct cl_io *io, struct cl_page *page);
2285 int cl_page_own_try (const struct lu_env *env,
2286 struct cl_io *io, struct cl_page *page);
2287 void cl_page_assume (const struct lu_env *env,
2288 struct cl_io *io, struct cl_page *page);
2289 void cl_page_unassume (const struct lu_env *env,
2290 struct cl_io *io, struct cl_page *pg);
2291 void cl_page_disown (const struct lu_env *env,
2292 struct cl_io *io, struct cl_page *page);
2293 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2300 * Functions dealing with the preparation of a page for a transfer, and
2301 * tracking transfer state.
2304 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2305 struct cl_page *pg, enum cl_req_type crt);
2306 void cl_page_completion (const struct lu_env *env,
2307 struct cl_page *pg, enum cl_req_type crt, int ioret);
2308 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2309 enum cl_req_type crt);
2310 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2311 struct cl_page *pg, enum cl_req_type crt);
2312 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2314 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2315 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2316 struct cl_page *pg);
2322 * \name helper routines
2323 * Functions to discard, delete and export a cl_page.
2326 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2327 struct cl_page *pg);
2328 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2329 int cl_page_is_vmlocked(const struct lu_env *env,
2330 const struct cl_page *pg);
2331 void cl_page_export(const struct lu_env *env,
2332 struct cl_page *pg, int uptodate);
2333 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2334 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2335 size_t cl_page_size(const struct cl_object *obj);
2337 void cl_lock_print(const struct lu_env *env, void *cookie,
2338 lu_printer_t printer, const struct cl_lock *lock);
2339 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2340 lu_printer_t printer,
2341 const struct cl_lock_descr *descr);
2345 * Data structure managing a client's cached pages. A count of
2346 * "unstable" pages is maintained, and an LRU of clean pages is
2347 * maintained. "unstable" pages are pages pinned by the ptlrpc
2348 * layer for recovery purposes.
2350 struct cl_client_cache {
2352 * # of client cache refcount
2353 * # of users (OSCs) + 2 (held by llite and lov)
2357 * # of threads are doing shrinking
2359 unsigned int ccc_lru_shrinkers;
2361 * # of LRU entries available
2363 atomic_long_t ccc_lru_left;
2365 * List of entities(OSCs) for this LRU cache
2367 struct list_head ccc_lru;
2369 * Max # of LRU entries
2371 unsigned long ccc_lru_max;
2373 * Lock to protect ccc_lru list
2375 spinlock_t ccc_lru_lock;
2377 * Set if unstable check is enabled
2379 unsigned int ccc_unstable_check:1;
2381 * # of unstable pages for this mount point
2383 atomic_long_t ccc_unstable_nr;
2385 * Waitq for awaiting unstable pages to reach zero.
2386 * Used at umounting time and signaled on BRW commit
2388 wait_queue_head_t ccc_unstable_waitq;
2391 * cl_cache functions
2393 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2394 void cl_cache_incref(struct cl_client_cache *cache);
2395 void cl_cache_decref(struct cl_client_cache *cache);
2399 /** \defgroup cl_lock cl_lock
2401 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2402 struct cl_lock *lock);
2403 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2404 const struct cl_io *io);
2405 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2406 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2407 const struct lu_device_type *dtype);
2408 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2410 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2411 struct cl_lock *lock, struct cl_sync_io *anchor);
2412 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2416 /** \defgroup cl_io cl_io
2419 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2420 enum cl_io_type iot, struct cl_object *obj);
2421 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2422 enum cl_io_type iot, struct cl_object *obj);
2423 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2424 enum cl_io_type iot, loff_t pos, size_t count);
2425 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2427 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2428 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2429 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2430 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2431 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2432 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2433 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2434 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2435 struct cl_io_lock_link *link);
2436 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2437 struct cl_lock_descr *descr);
2438 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2439 enum cl_req_type iot, struct cl_2queue *queue);
2440 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2441 enum cl_req_type iot, struct cl_2queue *queue,
2443 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2444 struct cl_page_list *queue, int from, int to,
2446 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2447 pgoff_t start, struct cl_read_ahead *ra);
2448 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2450 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2451 struct cl_page_list *queue);
2452 int cl_io_is_going (const struct lu_env *env);
2455 * True, iff \a io is an O_APPEND write(2).
2457 static inline int cl_io_is_append(const struct cl_io *io)
2459 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2462 static inline int cl_io_is_sync_write(const struct cl_io *io)
2464 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2467 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2469 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2473 * True, iff \a io is a truncate(2).
2475 static inline int cl_io_is_trunc(const struct cl_io *io)
2477 return io->ci_type == CIT_SETATTR &&
2478 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2481 struct cl_io *cl_io_top(struct cl_io *io);
2483 void cl_io_print(const struct lu_env *env, void *cookie,
2484 lu_printer_t printer, const struct cl_io *io);
2486 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2488 typeof(foo_io) __foo_io = (foo_io); \
2490 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2491 memset(&__foo_io->base + 1, 0, \
2492 (sizeof *__foo_io) - sizeof __foo_io->base); \
2497 /** \defgroup cl_page_list cl_page_list
2501 * Last page in the page list.
2503 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2505 LASSERT(plist->pl_nr > 0);
2506 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2509 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2511 LASSERT(plist->pl_nr > 0);
2512 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2516 * Iterate over pages in a page list.
2518 #define cl_page_list_for_each(page, list) \
2519 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2522 * Iterate over pages in a page list, taking possible removals into account.
2524 #define cl_page_list_for_each_safe(page, temp, list) \
2525 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2527 void cl_page_list_init (struct cl_page_list *plist);
2528 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2529 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2530 struct cl_page *page);
2531 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2532 struct cl_page *page);
2533 void cl_page_list_splice (struct cl_page_list *list,
2534 struct cl_page_list *head);
2535 void cl_page_list_del (const struct lu_env *env,
2536 struct cl_page_list *plist, struct cl_page *page);
2537 void cl_page_list_disown (const struct lu_env *env,
2538 struct cl_io *io, struct cl_page_list *plist);
2539 int cl_page_list_own (const struct lu_env *env,
2540 struct cl_io *io, struct cl_page_list *plist);
2541 void cl_page_list_assume (const struct lu_env *env,
2542 struct cl_io *io, struct cl_page_list *plist);
2543 void cl_page_list_discard(const struct lu_env *env,
2544 struct cl_io *io, struct cl_page_list *plist);
2545 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2547 void cl_2queue_init (struct cl_2queue *queue);
2548 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2549 void cl_2queue_disown (const struct lu_env *env,
2550 struct cl_io *io, struct cl_2queue *queue);
2551 void cl_2queue_assume (const struct lu_env *env,
2552 struct cl_io *io, struct cl_2queue *queue);
2553 void cl_2queue_discard (const struct lu_env *env,
2554 struct cl_io *io, struct cl_2queue *queue);
2555 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2556 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2558 /** @} cl_page_list */
2560 /** \defgroup cl_req cl_req
2562 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2563 enum cl_req_type crt, int nr_objects);
2565 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2566 struct cl_page *page);
2567 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2568 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2569 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2570 struct cl_req_attr *attr, u64 flags);
2571 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2573 /** \defgroup cl_sync_io cl_sync_io
2577 * Anchor for synchronous transfer. This is allocated on a stack by thread
2578 * doing synchronous transfer, and a pointer to this structure is set up in
2579 * every page submitted for transfer. Transfer completion routine updates
2580 * anchor and wakes up waiting thread when transfer is complete.
2583 /** number of pages yet to be transferred. */
2584 atomic_t csi_sync_nr;
2587 /** barrier of destroy this structure */
2588 atomic_t csi_barrier;
2589 /** completion to be signaled when transfer is complete. */
2590 wait_queue_head_t csi_waitq;
2591 /** callback to invoke when this IO is finished */
2592 void (*csi_end_io)(const struct lu_env *,
2593 struct cl_sync_io *);
2596 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2597 void (*end)(const struct lu_env *, struct cl_sync_io *));
2598 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2600 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2602 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2604 /** @} cl_sync_io */
2608 /** \defgroup cl_env cl_env
2610 * lu_env handling for a client.
2612 * lu_env is an environment within which lustre code executes. Its major part
2613 * is lu_context---a fast memory allocation mechanism that is used to conserve
2614 * precious kernel stack space. Originally lu_env was designed for a server,
2617 * - there is a (mostly) fixed number of threads, and
2619 * - call chains have no non-lustre portions inserted between lustre code.
2621 * On a client both these assumtpion fails, because every user thread can
2622 * potentially execute lustre code as part of a system call, and lustre calls
2623 * into VFS or MM that call back into lustre.
2625 * To deal with that, cl_env wrapper functions implement the following
2628 * - allocation and destruction of environment is amortized by caching no
2629 * longer used environments instead of destroying them;
2631 * - there is a notion of "current" environment, attached to the kernel
2632 * data structure representing current thread Top-level lustre code
2633 * allocates an environment and makes it current, then calls into
2634 * non-lustre code, that in turn calls lustre back. Low-level lustre
2635 * code thus called can fetch environment created by the top-level code
2636 * and reuse it, avoiding additional environment allocation.
2637 * Right now, three interfaces can attach the cl_env to running thread:
2640 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2642 * \see lu_env, lu_context, lu_context_key
2645 struct cl_env_nest {
2650 struct lu_env *cl_env_peek (int *refcheck);
2651 struct lu_env *cl_env_get (int *refcheck);
2652 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2653 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2654 void cl_env_put (struct lu_env *env, int *refcheck);
2655 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2656 void *cl_env_reenter (void);
2657 void cl_env_reexit (void *cookie);
2658 void cl_env_implant (struct lu_env *env, int *refcheck);
2659 void cl_env_unplant (struct lu_env *env, int *refcheck);
2660 unsigned cl_env_cache_purge(unsigned nr);
2661 struct lu_env *cl_env_percpu_get (void);
2662 void cl_env_percpu_put (struct lu_env *env);
2669 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2670 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2672 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2673 struct lu_device_type *ldt,
2674 struct lu_device *next);
2677 int cl_global_init(void);
2678 void cl_global_fini(void);
2680 #endif /* _LINUX_CL_OBJECT_H */