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>
111 struct cl_device_operations;
114 struct cl_object_page_operations;
115 struct cl_object_lock_operations;
118 struct cl_page_slice;
120 struct cl_lock_slice;
122 struct cl_lock_operations;
123 struct cl_page_operations;
132 * Operations for each data device in the client stack.
134 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
136 struct cl_device_operations {
138 * Initialize cl_req. This method is called top-to-bottom on all
139 * devices in the stack to get them a chance to allocate layer-private
140 * data, and to attach them to the cl_req by calling
141 * cl_req_slice_add().
143 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
144 * \see vvp_req_init()
146 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
151 * Device in the client stack.
153 * \see vvp_device, lov_device, lovsub_device, osc_device
157 struct lu_device cd_lu_dev;
158 /** Per-layer operation vector. */
159 const struct cl_device_operations *cd_ops;
162 /** \addtogroup cl_object cl_object
165 * "Data attributes" of cl_object. Data attributes can be updated
166 * independently for a sub-object, and top-object's attributes are calculated
167 * from sub-objects' ones.
170 /** Object size, in bytes */
173 * Known minimal size, in bytes.
175 * This is only valid when at least one DLM lock is held.
178 /** Modification time. Measured in seconds since epoch. */
180 /** Access time. Measured in seconds since epoch. */
182 /** Change time. Measured in seconds since epoch. */
185 * Blocks allocated to this cl_object on the server file system.
187 * \todo XXX An interface for block size is needed.
191 * User identifier for quota purposes.
195 * Group identifier for quota purposes.
199 /* nlink of the directory */
204 * Fields in cl_attr that are being set.
218 * Sub-class of lu_object with methods common for objects on the client
221 * cl_object: represents a regular file system object, both a file and a
222 * stripe. cl_object is based on lu_object: it is identified by a fid,
223 * layered, cached, hashed, and lrued. Important distinction with the server
224 * side, where md_object and dt_object are used, is that cl_object "fans out"
225 * at the lov/sns level: depending on the file layout, single file is
226 * represented as a set of "sub-objects" (stripes). At the implementation
227 * level, struct lov_object contains an array of cl_objects. Each sub-object
228 * is a full-fledged cl_object, having its fid, living in the lru and hash
231 * This leads to the next important difference with the server side: on the
232 * client, it's quite usual to have objects with the different sequence of
233 * layers. For example, typical top-object is composed of the following
239 * whereas its sub-objects are composed of
244 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
245 * track of the object-subobject relationship.
247 * Sub-objects are not cached independently: when top-object is about to
248 * be discarded from the memory, all its sub-objects are torn-down and
251 * \see vvp_object, lov_object, lovsub_object, osc_object
255 struct lu_object co_lu;
256 /** per-object-layer operations */
257 const struct cl_object_operations *co_ops;
258 /** offset of page slice in cl_page buffer */
263 * Description of the client object configuration. This is used for the
264 * creation of a new client object that is identified by a more state than
267 struct cl_object_conf {
269 struct lu_object_conf coc_lu;
272 * Object layout. This is consumed by lov.
274 struct lustre_md *coc_md;
276 * Description of particular stripe location in the
277 * cluster. This is consumed by osc.
279 struct lov_oinfo *coc_oinfo;
282 * VFS inode. This is consumed by vvp.
284 struct inode *coc_inode;
286 * Layout lock handle.
288 struct ldlm_lock *coc_lock;
290 * Operation to handle layout, OBJECT_CONF_XYZ.
296 /** configure layout, set up a new stripe, must be called while
297 * holding layout lock. */
299 /** invalidate the current stripe configuration due to losing
301 OBJECT_CONF_INVALIDATE = 1,
302 /** wait for old layout to go away so that new layout can be
308 * Operations implemented for each cl object layer.
310 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
312 struct cl_object_operations {
314 * Initialize page slice for this layer. Called top-to-bottom through
315 * every object layer when a new cl_page is instantiated. Layer
316 * keeping private per-page data, or requiring its own page operations
317 * vector should allocate these data here, and attach then to the page
318 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
321 * \retval NULL success.
323 * \retval ERR_PTR(errno) failure code.
325 * \retval valid-pointer pointer to already existing referenced page
326 * to be used instead of newly created.
328 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
329 struct cl_page *page, pgoff_t index);
331 * Initialize lock slice for this layer. Called top-to-bottom through
332 * every object layer when a new cl_lock is instantiated. Layer
333 * keeping private per-lock data, or requiring its own lock operations
334 * vector should allocate these data here, and attach then to the lock
335 * by calling cl_lock_slice_add(). Mandatory.
337 int (*coo_lock_init)(const struct lu_env *env,
338 struct cl_object *obj, struct cl_lock *lock,
339 const struct cl_io *io);
341 * Initialize io state for a given layer.
343 * called top-to-bottom once per io existence to initialize io
344 * state. If layer wants to keep some state for this type of io, it
345 * has to embed struct cl_io_slice in lu_env::le_ses, and register
346 * slice with cl_io_slice_add(). It is guaranteed that all threads
347 * participating in this io share the same session.
349 int (*coo_io_init)(const struct lu_env *env,
350 struct cl_object *obj, struct cl_io *io);
352 * Fill portion of \a attr that this layer controls. This method is
353 * called top-to-bottom through all object layers.
355 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
357 * \return 0: to continue
358 * \return +ve: to stop iterating through layers (but 0 is returned
359 * from enclosing cl_object_attr_get())
360 * \return -ve: to signal error
362 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
363 struct cl_attr *attr);
367 * \a valid is a bitmask composed from enum #cl_attr_valid, and
368 * indicating what attributes are to be set.
370 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
372 * \return the same convention as for
373 * cl_object_operations::coo_attr_get() is used.
375 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
376 const struct cl_attr *attr, unsigned valid);
378 * Update object configuration. Called top-to-bottom to modify object
381 * XXX error conditions and handling.
383 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
384 const struct cl_object_conf *conf);
386 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
387 * object. Layers are supposed to fill parts of \a lvb that will be
388 * shipped to the glimpse originator as a glimpse result.
390 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
391 * \see osc_object_glimpse()
393 int (*coo_glimpse)(const struct lu_env *env,
394 const struct cl_object *obj, struct ost_lvb *lvb);
396 * Object prune method. Called when the layout is going to change on
397 * this object, therefore each layer has to clean up their cache,
398 * mainly pages and locks.
400 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
402 * Object getstripe method.
404 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
405 struct lov_user_md __user *lum);
407 * Find whether there is any callback data (ldlm lock) attached upon
410 int (*coo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
411 ldlm_iterator_t iter, void *data);
413 * Get FIEMAP mapping from the object.
415 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
416 struct ll_fiemap_info_key *fmkey,
417 struct fiemap *fiemap, size_t *buflen);
421 * Extended header for client object.
423 struct cl_object_header {
424 /** Standard lu_object_header. cl_object::co_lu::lo_header points
426 struct lu_object_header coh_lu;
429 * Parent object. It is assumed that an object has a well-defined
430 * parent, but not a well-defined child (there may be multiple
431 * sub-objects, for the same top-object). cl_object_header::coh_parent
432 * field allows certain code to be written generically, without
433 * limiting possible cl_object layouts unduly.
435 struct cl_object_header *coh_parent;
437 * Protects consistency between cl_attr of parent object and
438 * attributes of sub-objects, that the former is calculated ("merged")
441 * \todo XXX this can be read/write lock if needed.
443 spinlock_t coh_attr_guard;
445 * Size of cl_page + page slices
447 unsigned short coh_page_bufsize;
449 * Number of objects above this one: 0 for a top-object, 1 for its
452 unsigned char coh_nesting;
456 * Helper macro: iterate over all layers of the object \a obj, assigning every
457 * layer top-to-bottom to \a slice.
459 #define cl_object_for_each(slice, obj) \
460 list_for_each_entry((slice), \
461 &(obj)->co_lu.lo_header->loh_layers,\
465 * Helper macro: iterate over all layers of the object \a obj, assigning every
466 * layer bottom-to-top to \a slice.
468 #define cl_object_for_each_reverse(slice, obj) \
469 list_for_each_entry_reverse((slice), \
470 &(obj)->co_lu.lo_header->loh_layers,\
475 #define CL_PAGE_EOF ((pgoff_t)~0ull)
477 /** \addtogroup cl_page cl_page
481 * Layered client page.
483 * cl_page: represents a portion of a file, cached in the memory. All pages
484 * of the given file are of the same size, and are kept in the radix tree
485 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
486 * of the top-level file object are first class cl_objects, they have their
487 * own radix trees of pages and hence page is implemented as a sequence of
488 * struct cl_pages's, linked into double-linked list through
489 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
490 * corresponding radix tree at the corresponding logical offset.
492 * cl_page is associated with VM page of the hosting environment (struct
493 * page in Linux kernel, for example), struct page. It is assumed, that this
494 * association is implemented by one of cl_page layers (top layer in the
495 * current design) that
497 * - intercepts per-VM-page call-backs made by the environment (e.g.,
500 * - translates state (page flag bits) and locking between lustre and
503 * The association between cl_page and struct page is immutable and
504 * established when cl_page is created.
506 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
507 * this io an exclusive access to this page w.r.t. other io attempts and
508 * various events changing page state (such as transfer completion, or
509 * eviction of the page from the memory). Note, that in general cl_io
510 * cannot be identified with a particular thread, and page ownership is not
511 * exactly equal to the current thread holding a lock on the page. Layer
512 * implementing association between cl_page and struct page has to implement
513 * ownership on top of available synchronization mechanisms.
515 * While lustre client maintains the notion of an page ownership by io,
516 * hosting MM/VM usually has its own page concurrency control
517 * mechanisms. For example, in Linux, page access is synchronized by the
518 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
519 * takes care to acquire and release such locks as necessary around the
520 * calls to the file system methods (->readpage(), ->prepare_write(),
521 * ->commit_write(), etc.). This leads to the situation when there are two
522 * different ways to own a page in the client:
524 * - client code explicitly and voluntary owns the page (cl_page_own());
526 * - VM locks a page and then calls the client, that has "to assume"
527 * the ownership from the VM (cl_page_assume()).
529 * Dual methods to release ownership are cl_page_disown() and
530 * cl_page_unassume().
532 * cl_page is reference counted (cl_page::cp_ref). When reference counter
533 * drops to 0, the page is returned to the cache, unless it is in
534 * cl_page_state::CPS_FREEING state, in which case it is immediately
537 * The general logic guaranteeing the absence of "existential races" for
538 * pages is the following:
540 * - there are fixed known ways for a thread to obtain a new reference
543 * - by doing a lookup in the cl_object radix tree, protected by the
546 * - by starting from VM-locked struct page and following some
547 * hosting environment method (e.g., following ->private pointer in
548 * the case of Linux kernel), see cl_vmpage_page();
550 * - when the page enters cl_page_state::CPS_FREEING state, all these
551 * ways are severed with the proper synchronization
552 * (cl_page_delete());
554 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
557 * - no new references to the page in cl_page_state::CPS_FREEING state
558 * are allowed (checked in cl_page_get()).
560 * Together this guarantees that when last reference to a
561 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
562 * page, as neither references to it can be acquired at that point, nor
565 * cl_page is a state machine. States are enumerated in enum
566 * cl_page_state. Possible state transitions are enumerated in
567 * cl_page_state_set(). State transition process (i.e., actual changing of
568 * cl_page::cp_state field) is protected by the lock on the underlying VM
571 * Linux Kernel implementation.
573 * Binding between cl_page and struct page (which is a typedef for
574 * struct page) is implemented in the vvp layer. cl_page is attached to the
575 * ->private pointer of the struct page, together with the setting of
576 * PG_private bit in page->flags, and acquiring additional reference on the
577 * struct page (much like struct buffer_head, or any similar file system
578 * private data structures).
580 * PG_locked lock is used to implement both ownership and transfer
581 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
582 * states. No additional references are acquired for the duration of the
585 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
586 * write-out is "protected" by the special PG_writeback bit.
590 * States of cl_page. cl_page.c assumes particular order here.
592 * The page state machine is rather crude, as it doesn't recognize finer page
593 * states like "dirty" or "up to date". This is because such states are not
594 * always well defined for the whole stack (see, for example, the
595 * implementation of the read-ahead, that hides page up-to-dateness to track
596 * cache hits accurately). Such sub-states are maintained by the layers that
597 * are interested in them.
601 * Page is in the cache, un-owned. Page leaves cached state in the
604 * - [cl_page_state::CPS_OWNED] io comes across the page and
607 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
608 * req-formation engine decides that it wants to include this page
609 * into an cl_req being constructed, and yanks it from the cache;
611 * - [cl_page_state::CPS_FREEING] VM callback is executed to
612 * evict the page form the memory;
614 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
618 * Page is exclusively owned by some cl_io. Page may end up in this
619 * state as a result of
621 * - io creating new page and immediately owning it;
623 * - [cl_page_state::CPS_CACHED] io finding existing cached page
626 * - [cl_page_state::CPS_OWNED] io finding existing owned page
627 * and waiting for owner to release the page;
629 * Page leaves owned state in the following cases:
631 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
632 * the cache, doing nothing;
634 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
637 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
638 * transfer for this page;
640 * - [cl_page_state::CPS_FREEING] io decides to destroy this
641 * page (e.g., as part of truncate or extent lock cancellation).
643 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
647 * Page is being written out, as a part of a transfer. This state is
648 * entered when req-formation logic decided that it wants this page to
649 * be sent through the wire _now_. Specifically, it means that once
650 * this state is achieved, transfer completion handler (with either
651 * success or failure indication) is guaranteed to be executed against
652 * this page independently of any locks and any scheduling decisions
653 * made by the hosting environment (that effectively means that the
654 * page is never put into cl_page_state::CPS_PAGEOUT state "in
655 * advance". This property is mentioned, because it is important when
656 * reasoning about possible dead-locks in the system). The page can
657 * enter this state as a result of
659 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
660 * write-out of this page, or
662 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
663 * that it has enough dirty pages cached to issue a "good"
666 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
667 * is completed---it is moved into cl_page_state::CPS_CACHED state.
669 * Underlying VM page is locked for the duration of transfer.
671 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
675 * Page is being read in, as a part of a transfer. This is quite
676 * similar to the cl_page_state::CPS_PAGEOUT state, except that
677 * read-in is always "immediate"---there is no such thing a sudden
678 * construction of read cl_req from cached, presumably not up to date,
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 destroyed. This state is entered when client decides
688 * that page has to be deleted from its host object, as, e.g., a part
691 * Once this state is reached, there is no way to escape it.
693 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
700 /** Host page, the page is from the host inode which the cl_page
704 /** Transient page, the transient cl_page is used to bind a cl_page
705 * to vmpage which is not belonging to the same object of cl_page.
706 * it is used in DirectIO, lockless IO and liblustre. */
711 * Fields are protected by the lock on struct page, except for atomics and
714 * \invariant Data type invariants are in cl_page_invariant(). Basically:
715 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
716 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
717 * cl_page::cp_owner (when set).
720 /** Reference counter. */
722 /** Transfer error. */
724 /** An object this page is a part of. Immutable after creation. */
725 struct cl_object *cp_obj;
727 struct page *cp_vmpage;
728 /** Linkage of pages within group. Pages must be owned */
729 struct list_head cp_batch;
730 /** List of slices. Immutable after creation. */
731 struct list_head cp_layers;
732 /** Linkage of pages within cl_req. */
733 struct list_head cp_flight;
735 * Page state. This field is const to avoid accidental update, it is
736 * modified only internally within cl_page.c. Protected by a VM lock.
738 const enum cl_page_state cp_state;
740 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
743 enum cl_page_type cp_type;
746 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
747 * by sub-io. Protected by a VM lock.
749 struct cl_io *cp_owner;
751 * Owning IO request in cl_page_state::CPS_PAGEOUT and
752 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
753 * the top-level pages. Protected by a VM lock.
755 struct cl_req *cp_req;
756 /** List of references to this page, for debugging. */
757 struct lu_ref cp_reference;
758 /** Link to an object, for debugging. */
759 struct lu_ref_link cp_obj_ref;
760 /** Link to a queue, for debugging. */
761 struct lu_ref_link cp_queue_ref;
762 /** Assigned if doing a sync_io */
763 struct cl_sync_io *cp_sync_io;
767 * Per-layer part of cl_page.
769 * \see vvp_page, lov_page, osc_page
771 struct cl_page_slice {
772 struct cl_page *cpl_page;
775 * Object slice corresponding to this page slice. Immutable after
778 struct cl_object *cpl_obj;
779 const struct cl_page_operations *cpl_ops;
780 /** Linkage into cl_page::cp_layers. Immutable after creation. */
781 struct list_head cpl_linkage;
785 * Lock mode. For the client extent locks.
797 * Requested transfer type.
807 * Per-layer page operations.
809 * Methods taking an \a io argument are for the activity happening in the
810 * context of given \a io. Page is assumed to be owned by that io, except for
811 * the obvious cases (like cl_page_operations::cpo_own()).
813 * \see vvp_page_ops, lov_page_ops, osc_page_ops
815 struct cl_page_operations {
817 * cl_page<->struct page methods. Only one layer in the stack has to
818 * implement these. Current code assumes that this functionality is
819 * provided by the topmost layer, see cl_page_disown0() as an example.
823 * Called when \a io acquires this page into the exclusive
824 * ownership. When this method returns, it is guaranteed that the is
825 * not owned by other io, and no transfer is going on against
829 * \see vvp_page_own(), lov_page_own()
831 int (*cpo_own)(const struct lu_env *env,
832 const struct cl_page_slice *slice,
833 struct cl_io *io, int nonblock);
834 /** Called when ownership it yielded. Optional.
836 * \see cl_page_disown()
837 * \see vvp_page_disown()
839 void (*cpo_disown)(const struct lu_env *env,
840 const struct cl_page_slice *slice, struct cl_io *io);
842 * Called for a page that is already "owned" by \a io from VM point of
845 * \see cl_page_assume()
846 * \see vvp_page_assume(), lov_page_assume()
848 void (*cpo_assume)(const struct lu_env *env,
849 const struct cl_page_slice *slice, struct cl_io *io);
850 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
851 * bottom-to-top when IO releases a page without actually unlocking
854 * \see cl_page_unassume()
855 * \see vvp_page_unassume()
857 void (*cpo_unassume)(const struct lu_env *env,
858 const struct cl_page_slice *slice,
861 * Announces whether the page contains valid data or not by \a uptodate.
863 * \see cl_page_export()
864 * \see vvp_page_export()
866 void (*cpo_export)(const struct lu_env *env,
867 const struct cl_page_slice *slice, int uptodate);
869 * Checks whether underlying VM page is locked (in the suitable
870 * sense). Used for assertions.
872 * \retval -EBUSY: page is protected by a lock of a given mode;
873 * \retval -ENODATA: page is not protected by a lock;
874 * \retval 0: this layer cannot decide. (Should never happen.)
876 int (*cpo_is_vmlocked)(const struct lu_env *env,
877 const struct cl_page_slice *slice);
883 * Called when page is truncated from the object. Optional.
885 * \see cl_page_discard()
886 * \see vvp_page_discard(), osc_page_discard()
888 void (*cpo_discard)(const struct lu_env *env,
889 const struct cl_page_slice *slice,
892 * Called when page is removed from the cache, and is about to being
893 * destroyed. Optional.
895 * \see cl_page_delete()
896 * \see vvp_page_delete(), osc_page_delete()
898 void (*cpo_delete)(const struct lu_env *env,
899 const struct cl_page_slice *slice);
900 /** Destructor. Frees resources and slice itself. */
901 void (*cpo_fini)(const struct lu_env *env,
902 struct cl_page_slice *slice);
904 * Optional debugging helper. Prints given page slice.
906 * \see cl_page_print()
908 int (*cpo_print)(const struct lu_env *env,
909 const struct cl_page_slice *slice,
910 void *cookie, lu_printer_t p);
914 * Transfer methods. See comment on cl_req for a description of
915 * transfer formation and life-cycle.
920 * Request type dependent vector of operations.
922 * Transfer operations depend on transfer mode (cl_req_type). To avoid
923 * passing transfer mode to each and every of these methods, and to
924 * avoid branching on request type inside of the methods, separate
925 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
926 * provided. That is, method invocation usually looks like
928 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
932 * Called when a page is submitted for a transfer as a part of
935 * \return 0 : page is eligible for submission;
936 * \return -EALREADY : skip this page;
937 * \return -ve : error.
939 * \see cl_page_prep()
941 int (*cpo_prep)(const struct lu_env *env,
942 const struct cl_page_slice *slice,
945 * Completion handler. This is guaranteed to be eventually
946 * fired after cl_page_operations::cpo_prep() or
947 * cl_page_operations::cpo_make_ready() call.
949 * This method can be called in a non-blocking context. It is
950 * guaranteed however, that the page involved and its object
951 * are pinned in memory (and, hence, calling cl_page_put() is
954 * \see cl_page_completion()
956 void (*cpo_completion)(const struct lu_env *env,
957 const struct cl_page_slice *slice,
960 * Called when cached page is about to be added to the
961 * cl_req as a part of req formation.
963 * \return 0 : proceed with this page;
964 * \return -EAGAIN : skip this page;
965 * \return -ve : error.
967 * \see cl_page_make_ready()
969 int (*cpo_make_ready)(const struct lu_env *env,
970 const struct cl_page_slice *slice);
973 * Tell transfer engine that only [to, from] part of a page should be
976 * This is used for immediate transfers.
978 * \todo XXX this is not very good interface. It would be much better
979 * if all transfer parameters were supplied as arguments to
980 * cl_io_operations::cio_submit() call, but it is not clear how to do
981 * this for page queues.
983 * \see cl_page_clip()
985 void (*cpo_clip)(const struct lu_env *env,
986 const struct cl_page_slice *slice,
989 * \pre the page was queued for transferring.
990 * \post page is removed from client's pending list, or -EBUSY
991 * is returned if it has already been in transferring.
993 * This is one of seldom page operation which is:
994 * 0. called from top level;
995 * 1. don't have vmpage locked;
996 * 2. every layer should synchronize execution of its ->cpo_cancel()
997 * with completion handlers. Osc uses client obd lock for this
998 * purpose. Based on there is no vvp_page_cancel and
999 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1001 * \see osc_page_cancel().
1003 int (*cpo_cancel)(const struct lu_env *env,
1004 const struct cl_page_slice *slice);
1006 * Write out a page by kernel. This is only called by ll_writepage
1009 * \see cl_page_flush()
1011 int (*cpo_flush)(const struct lu_env *env,
1012 const struct cl_page_slice *slice,
1018 * Helper macro, dumping detailed information about \a page into a log.
1020 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1022 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1023 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1024 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1025 CDEBUG(mask, format , ## __VA_ARGS__); \
1030 * Helper macro, dumping shorter information about \a page into a log.
1032 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1034 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1035 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1036 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1037 CDEBUG(mask, format , ## __VA_ARGS__); \
1041 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1043 LASSERT(page->cp_vmpage != NULL);
1044 return page->cp_vmpage;
1048 * Check if a cl_page is in use.
1050 * Client cache holds a refcount, this refcount will be dropped when
1051 * the page is taken out of cache, see vvp_page_delete().
1053 static inline bool __page_in_use(const struct cl_page *page, int refc)
1055 return (atomic_read(&page->cp_ref) > refc + 1);
1059 * Caller itself holds a refcount of cl_page.
1061 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1063 * Caller doesn't hold a refcount.
1065 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1069 /** \addtogroup cl_lock cl_lock
1073 * Extent locking on the client.
1077 * The locking model of the new client code is built around
1081 * data-type representing an extent lock on a regular file. cl_lock is a
1082 * layered object (much like cl_object and cl_page), it consists of a header
1083 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1084 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1086 * Typical cl_lock consists of the two layers:
1088 * - vvp_lock (vvp specific data), and
1089 * - lov_lock (lov specific data).
1091 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1092 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1094 * - lovsub_lock, and
1097 * Each sub-lock is associated with a cl_object (representing stripe
1098 * sub-object or the file to which top-level cl_lock is associated to), and is
1099 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1100 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1101 * is different from cl_page, that doesn't fan out (there is usually exactly
1102 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1103 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1107 * cl_lock is a cacheless data container for the requirements of locks to
1108 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1111 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1112 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1114 * INTERFACE AND USAGE
1116 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1117 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1118 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1119 * consists of multiple sub cl_locks, each sub locks will be enqueued
1120 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1121 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1124 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1125 * method will be called for each layer to release the resource held by this
1126 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1127 * clo_enqueue time, is released.
1129 * LDLM lock can only be canceled if there is no cl_lock using it.
1131 * Overall process of the locking during IO operation is as following:
1133 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1134 * is called on each layer. Responsibility of this method is to add locks,
1135 * needed by a given layer into cl_io.ci_lockset.
1137 * - once locks for all layers were collected, they are sorted to avoid
1138 * dead-locks (cl_io_locks_sort()), and enqueued.
1140 * - when all locks are acquired, IO is performed;
1142 * - locks are released after IO is complete.
1144 * Striping introduces major additional complexity into locking. The
1145 * fundamental problem is that it is generally unsafe to actively use (hold)
1146 * two locks on the different OST servers at the same time, as this introduces
1147 * inter-server dependency and can lead to cascading evictions.
1149 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1150 * that no multi-stripe locks are taken (note that this design abandons POSIX
1151 * read/write semantics). Such pieces ideally can be executed concurrently. At
1152 * the same time, certain types of IO cannot be sub-divived, without
1153 * sacrificing correctness. This includes:
1155 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1158 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1160 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1161 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1162 * has to be held together with the usual lock on [offset, offset + count].
1164 * Interaction with DLM
1166 * In the expected setup, cl_lock is ultimately backed up by a collection of
1167 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1168 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1169 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1170 * description of interaction with DLM.
1176 struct cl_lock_descr {
1177 /** Object this lock is granted for. */
1178 struct cl_object *cld_obj;
1179 /** Index of the first page protected by this lock. */
1181 /** Index of the last page (inclusive) protected by this lock. */
1183 /** Group ID, for group lock */
1186 enum cl_lock_mode cld_mode;
1188 * flags to enqueue lock. A combination of bit-flags from
1189 * enum cl_enq_flags.
1191 __u32 cld_enq_flags;
1194 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1195 #define PDESCR(descr) \
1196 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1197 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1199 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1202 * Layered client lock.
1205 /** List of slices. Immutable after creation. */
1206 struct list_head cll_layers;
1207 /** lock attribute, extent, cl_object, etc. */
1208 struct cl_lock_descr cll_descr;
1212 * Per-layer part of cl_lock
1214 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1216 struct cl_lock_slice {
1217 struct cl_lock *cls_lock;
1218 /** Object slice corresponding to this lock slice. Immutable after
1220 struct cl_object *cls_obj;
1221 const struct cl_lock_operations *cls_ops;
1222 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1223 struct list_head cls_linkage;
1228 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1230 struct cl_lock_operations {
1233 * Attempts to enqueue the lock. Called top-to-bottom.
1235 * \retval 0 this layer has enqueued the lock successfully
1236 * \retval >0 this layer has enqueued the lock, but need to wait on
1237 * @anchor for resources
1238 * \retval -ve failure
1240 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1241 * \see osc_lock_enqueue()
1243 int (*clo_enqueue)(const struct lu_env *env,
1244 const struct cl_lock_slice *slice,
1245 struct cl_io *io, struct cl_sync_io *anchor);
1247 * Cancel a lock, release its DLM lock ref, while does not cancel the
1250 void (*clo_cancel)(const struct lu_env *env,
1251 const struct cl_lock_slice *slice);
1254 * Destructor. Frees resources and the slice.
1256 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1257 * \see osc_lock_fini()
1259 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1261 * Optional debugging helper. Prints given lock slice.
1263 int (*clo_print)(const struct lu_env *env,
1264 void *cookie, lu_printer_t p,
1265 const struct cl_lock_slice *slice);
1268 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1270 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1271 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1272 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1273 CDEBUG(mask, format , ## __VA_ARGS__); \
1277 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1281 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1287 /** \addtogroup cl_page_list cl_page_list
1288 * Page list used to perform collective operations on a group of pages.
1290 * Pages are added to the list one by one. cl_page_list acquires a reference
1291 * for every page in it. Page list is used to perform collective operations on
1294 * - submit pages for an immediate transfer,
1296 * - own pages on behalf of certain io (waiting for each page in turn),
1300 * When list is finalized, it releases references on all pages it still has.
1302 * \todo XXX concurrency control.
1306 struct cl_page_list {
1308 struct list_head pl_pages;
1309 struct task_struct *pl_owner;
1313 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1314 * contains an incoming page list and an outgoing page list.
1317 struct cl_page_list c2_qin;
1318 struct cl_page_list c2_qout;
1321 /** @} cl_page_list */
1323 /** \addtogroup cl_io cl_io
1328 * cl_io represents a high level I/O activity like
1329 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1332 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1333 * important distinction. We want to minimize number of calls to the allocator
1334 * in the fast path, e.g., in the case of read(2) when everything is cached:
1335 * client already owns the lock over region being read, and data are cached
1336 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1337 * per-layer io state is stored in the session, associated with the io, see
1338 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1339 * by using free-lists, see cl_env_get().
1341 * There is a small predefined number of possible io types, enumerated in enum
1344 * cl_io is a state machine, that can be advanced concurrently by the multiple
1345 * threads. It is up to these threads to control the concurrency and,
1346 * specifically, to detect when io is done, and its state can be safely
1349 * For read/write io overall execution plan is as following:
1351 * (0) initialize io state through all layers;
1353 * (1) loop: prepare chunk of work to do
1355 * (2) call all layers to collect locks they need to process current chunk
1357 * (3) sort all locks to avoid dead-locks, and acquire them
1359 * (4) process the chunk: call per-page methods
1360 * cl_io_operations::cio_prepare_write(),
1361 * cl_io_operations::cio_commit_write() for write)
1367 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1368 * address allocation efficiency issues mentioned above), and returns with the
1369 * special error condition from per-page method when current sub-io has to
1370 * block. This causes io loop to be repeated, and lov switches to the next
1371 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1376 /** read system call */
1378 /** write system call */
1380 /** truncate, utime system calls */
1383 * page fault handling
1387 * fsync system call handling
1388 * To write out a range of file
1392 * Miscellaneous io. This is used for occasional io activity that
1393 * doesn't fit into other types. Currently this is used for:
1395 * - cancellation of an extent lock. This io exists as a context
1396 * to write dirty pages from under the lock being canceled back
1399 * - VM induced page write-out. An io context for writing page out
1400 * for memory cleansing;
1402 * - glimpse. An io context to acquire glimpse lock.
1404 * - grouplock. An io context to acquire group lock.
1406 * CIT_MISC io is used simply as a context in which locks and pages
1407 * are manipulated. Such io has no internal "process", that is,
1408 * cl_io_loop() is never called for it.
1415 * States of cl_io state machine
1418 /** Not initialized. */
1422 /** IO iteration started. */
1426 /** Actual IO is in progress. */
1428 /** IO for the current iteration finished. */
1430 /** Locks released. */
1432 /** Iteration completed. */
1434 /** cl_io finalized. */
1439 * IO state private for a layer.
1441 * This is usually embedded into layer session data, rather than allocated
1444 * \see vvp_io, lov_io, osc_io
1446 struct cl_io_slice {
1447 struct cl_io *cis_io;
1448 /** corresponding object slice. Immutable after creation. */
1449 struct cl_object *cis_obj;
1450 /** io operations. Immutable after creation. */
1451 const struct cl_io_operations *cis_iop;
1453 * linkage into a list of all slices for a given cl_io, hanging off
1454 * cl_io::ci_layers. Immutable after creation.
1456 struct list_head cis_linkage;
1459 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1462 struct cl_read_ahead {
1463 /* Maximum page index the readahead window will end.
1464 * This is determined DLM lock coverage, RPC and stripe boundary.
1465 * cra_end is included. */
1467 /* Release routine. If readahead holds resources underneath, this
1468 * function should be called to release it. */
1469 void (*cra_release)(const struct lu_env *env, void *cbdata);
1470 /* Callback data for cra_release routine */
1474 static inline void cl_read_ahead_release(const struct lu_env *env,
1475 struct cl_read_ahead *ra)
1477 if (ra->cra_release != NULL)
1478 ra->cra_release(env, ra->cra_cbdata);
1479 memset(ra, 0, sizeof(*ra));
1484 * Per-layer io operations.
1485 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1487 struct cl_io_operations {
1489 * Vector of io state transition methods for every io type.
1491 * \see cl_page_operations::io
1495 * Prepare io iteration at a given layer.
1497 * Called top-to-bottom at the beginning of each iteration of
1498 * "io loop" (if it makes sense for this type of io). Here
1499 * layer selects what work it will do during this iteration.
1501 * \see cl_io_operations::cio_iter_fini()
1503 int (*cio_iter_init) (const struct lu_env *env,
1504 const struct cl_io_slice *slice);
1506 * Finalize io iteration.
1508 * Called bottom-to-top at the end of each iteration of "io
1509 * loop". Here layers can decide whether IO has to be
1512 * \see cl_io_operations::cio_iter_init()
1514 void (*cio_iter_fini) (const struct lu_env *env,
1515 const struct cl_io_slice *slice);
1517 * Collect locks for the current iteration of io.
1519 * Called top-to-bottom to collect all locks necessary for
1520 * this iteration. This methods shouldn't actually enqueue
1521 * anything, instead it should post a lock through
1522 * cl_io_lock_add(). Once all locks are collected, they are
1523 * sorted and enqueued in the proper order.
1525 int (*cio_lock) (const struct lu_env *env,
1526 const struct cl_io_slice *slice);
1528 * Finalize unlocking.
1530 * Called bottom-to-top to finish layer specific unlocking
1531 * functionality, after generic code released all locks
1532 * acquired by cl_io_operations::cio_lock().
1534 void (*cio_unlock)(const struct lu_env *env,
1535 const struct cl_io_slice *slice);
1537 * Start io iteration.
1539 * Once all locks are acquired, called top-to-bottom to
1540 * commence actual IO. In the current implementation,
1541 * top-level vvp_io_{read,write}_start() does all the work
1542 * synchronously by calling generic_file_*(), so other layers
1543 * are called when everything is done.
1545 int (*cio_start)(const struct lu_env *env,
1546 const struct cl_io_slice *slice);
1548 * Called top-to-bottom at the end of io loop. Here layer
1549 * might wait for an unfinished asynchronous io.
1551 void (*cio_end) (const struct lu_env *env,
1552 const struct cl_io_slice *slice);
1554 * Called bottom-to-top to notify layers that read/write IO
1555 * iteration finished, with \a nob bytes transferred.
1557 void (*cio_advance)(const struct lu_env *env,
1558 const struct cl_io_slice *slice,
1561 * Called once per io, bottom-to-top to release io resources.
1563 void (*cio_fini) (const struct lu_env *env,
1564 const struct cl_io_slice *slice);
1568 * Submit pages from \a queue->c2_qin for IO, and move
1569 * successfully submitted pages into \a queue->c2_qout. Return
1570 * non-zero if failed to submit even the single page. If
1571 * submission failed after some pages were moved into \a
1572 * queue->c2_qout, completion callback with non-zero ioret is
1575 int (*cio_submit)(const struct lu_env *env,
1576 const struct cl_io_slice *slice,
1577 enum cl_req_type crt,
1578 struct cl_2queue *queue);
1580 * Queue async page for write.
1581 * The difference between cio_submit and cio_queue is that
1582 * cio_submit is for urgent request.
1584 int (*cio_commit_async)(const struct lu_env *env,
1585 const struct cl_io_slice *slice,
1586 struct cl_page_list *queue, int from, int to,
1589 * Decide maximum read ahead extent
1591 * \pre io->ci_type == CIT_READ
1593 int (*cio_read_ahead)(const struct lu_env *env,
1594 const struct cl_io_slice *slice,
1595 pgoff_t start, struct cl_read_ahead *ra);
1597 * Optional debugging helper. Print given io slice.
1599 int (*cio_print)(const struct lu_env *env, void *cookie,
1600 lu_printer_t p, const struct cl_io_slice *slice);
1604 * Flags to lock enqueue procedure.
1609 * instruct server to not block, if conflicting lock is found. Instead
1610 * -EWOULDBLOCK is returned immediately.
1612 CEF_NONBLOCK = 0x00000001,
1614 * take lock asynchronously (out of order), as it cannot
1615 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1617 CEF_ASYNC = 0x00000002,
1619 * tell the server to instruct (though a flag in the blocking ast) an
1620 * owner of the conflicting lock, that it can drop dirty pages
1621 * protected by this lock, without sending them to the server.
1623 CEF_DISCARD_DATA = 0x00000004,
1625 * tell the sub layers that it must be a `real' lock. This is used for
1626 * mmapped-buffer locks and glimpse locks that must be never converted
1627 * into lockless mode.
1629 * \see vvp_mmap_locks(), cl_glimpse_lock().
1631 CEF_MUST = 0x00000008,
1633 * tell the sub layers that never request a `real' lock. This flag is
1634 * not used currently.
1636 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1637 * conversion policy: ci_lockreq describes generic information of lock
1638 * requirement for this IO, especially for locks which belong to the
1639 * object doing IO; however, lock itself may have precise requirements
1640 * that are described by the enqueue flags.
1642 CEF_NEVER = 0x00000010,
1644 * for async glimpse lock.
1646 CEF_AGL = 0x00000020,
1648 * enqueue a lock to test DLM lock existence.
1650 CEF_PEEK = 0x00000040,
1652 * mask of enq_flags.
1654 CEF_MASK = 0x0000007f,
1658 * Link between lock and io. Intermediate structure is needed, because the
1659 * same lock can be part of multiple io's simultaneously.
1661 struct cl_io_lock_link {
1662 /** linkage into one of cl_lockset lists. */
1663 struct list_head cill_linkage;
1664 struct cl_lock cill_lock;
1665 /** optional destructor */
1666 void (*cill_fini)(const struct lu_env *env,
1667 struct cl_io_lock_link *link);
1669 #define cill_descr cill_lock.cll_descr
1672 * Lock-set represents a collection of locks, that io needs at a
1673 * time. Generally speaking, client tries to avoid holding multiple locks when
1676 * - holding extent locks over multiple ost's introduces the danger of
1677 * "cascading timeouts";
1679 * - holding multiple locks over the same ost is still dead-lock prone,
1680 * see comment in osc_lock_enqueue(),
1682 * but there are certain situations where this is unavoidable:
1684 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1686 * - truncate has to take [new-size, EOF] lock for correctness;
1688 * - SNS has to take locks across full stripe for correctness;
1690 * - in the case when user level buffer, supplied to {read,write}(file0),
1691 * is a part of a memory mapped lustre file, client has to take a dlm
1692 * locks on file0, and all files that back up the buffer (or a part of
1693 * the buffer, that is being processed in the current chunk, in any
1694 * case, there are situations where at least 2 locks are necessary).
1696 * In such cases we at least try to take locks in the same consistent
1697 * order. To this end, all locks are first collected, then sorted, and then
1701 /** locks to be acquired. */
1702 struct list_head cls_todo;
1703 /** locks acquired. */
1704 struct list_head cls_done;
1708 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1709 * but 'req' is always to be thought as 'request' :-)
1711 enum cl_io_lock_dmd {
1712 /** Always lock data (e.g., O_APPEND). */
1714 /** Layers are free to decide between local and global locking. */
1716 /** Never lock: there is no cache (e.g., liblustre). */
1720 enum cl_fsync_mode {
1721 /** start writeback, do not wait for them to finish */
1723 /** start writeback and wait for them to finish */
1725 /** discard all of dirty pages in a specific file range */
1726 CL_FSYNC_DISCARD = 2,
1727 /** start writeback and make sure they have reached storage before
1728 * return. OST_SYNC RPC must be issued and finished */
1732 struct cl_io_rw_common {
1742 * cl_io is shared by all threads participating in this IO (in current
1743 * implementation only one thread advances IO, but parallel IO design and
1744 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1745 * is up to these threads to serialize their activities, including updates to
1746 * mutable cl_io fields.
1749 /** type of this IO. Immutable after creation. */
1750 enum cl_io_type ci_type;
1751 /** current state of cl_io state machine. */
1752 enum cl_io_state ci_state;
1753 /** main object this io is against. Immutable after creation. */
1754 struct cl_object *ci_obj;
1756 * Upper layer io, of which this io is a part of. Immutable after
1759 struct cl_io *ci_parent;
1760 /** List of slices. Immutable after creation. */
1761 struct list_head ci_layers;
1762 /** list of locks (to be) acquired by this io. */
1763 struct cl_lockset ci_lockset;
1764 /** lock requirements, this is just a help info for sublayers. */
1765 enum cl_io_lock_dmd ci_lockreq;
1768 struct cl_io_rw_common rd;
1771 struct cl_io_rw_common wr;
1775 struct cl_io_rw_common ci_rw;
1776 struct cl_setattr_io {
1777 struct ost_lvb sa_attr;
1778 unsigned int sa_valid;
1779 int sa_stripe_index;
1780 struct lu_fid *sa_parent_fid;
1781 struct obd_capa *sa_capa;
1783 struct cl_fault_io {
1784 /** page index within file. */
1786 /** bytes valid byte on a faulted page. */
1788 /** writable page? for nopage() only */
1790 /** page of an executable? */
1792 /** page_mkwrite() */
1794 /** resulting page */
1795 struct cl_page *ft_page;
1797 struct cl_fsync_io {
1800 struct obd_capa *fi_capa;
1801 /** file system level fid */
1802 struct lu_fid *fi_fid;
1803 enum cl_fsync_mode fi_mode;
1804 /* how many pages were written/discarded */
1805 unsigned int fi_nr_written;
1808 struct cl_2queue ci_queue;
1811 unsigned int ci_continue:1,
1813 * This io has held grouplock, to inform sublayers that
1814 * don't do lockless i/o.
1818 * The whole IO need to be restarted because layout has been changed
1822 * to not refresh layout - the IO issuer knows that the layout won't
1823 * change(page operations, layout change causes all page to be
1824 * discarded), or it doesn't matter if it changes(sync).
1828 * Check if layout changed after the IO finishes. Mainly for HSM
1829 * requirement. If IO occurs to openning files, it doesn't need to
1830 * verify layout because HSM won't release openning files.
1831 * Right now, only two opertaions need to verify layout: glimpse
1836 * file is released, restore has to to be triggered by vvp layer
1838 ci_restore_needed:1,
1844 * Number of pages owned by this IO. For invariant checking.
1846 unsigned ci_owned_nr;
1851 /** \addtogroup cl_req cl_req
1856 * There are two possible modes of transfer initiation on the client:
1858 * - immediate transfer: this is started when a high level io wants a page
1859 * or a collection of pages to be transferred right away. Examples:
1860 * read-ahead, synchronous read in the case of non-page aligned write,
1861 * page write-out as a part of extent lock cancellation, page write-out
1862 * as a part of memory cleansing. Immediate transfer can be both
1863 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1865 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1866 * when io wants to transfer a page to the server some time later, when
1867 * it can be done efficiently. Example: pages dirtied by the write(2)
1870 * In any case, transfer takes place in the form of a cl_req, which is a
1871 * representation for a network RPC.
1873 * Pages queued for an opportunistic transfer are cached until it is decided
1874 * that efficient RPC can be composed of them. This decision is made by "a
1875 * req-formation engine", currently implemented as a part of osc
1876 * layer. Req-formation depends on many factors: the size of the resulting
1877 * RPC, whether or not multi-object RPCs are supported by the server,
1878 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1880 * For the immediate transfer io submits a cl_page_list, that req-formation
1881 * engine slices into cl_req's, possibly adding cached pages to some of
1882 * the resulting req's.
1884 * Whenever a page from cl_page_list is added to a newly constructed req, its
1885 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1886 * page state is atomically changed from cl_page_state::CPS_OWNED to
1887 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1888 * is zeroed, and cl_page::cp_req is set to the
1889 * req. cl_page_operations::cpo_prep() method at the particular layer might
1890 * return -EALREADY to indicate that it does not need to submit this page
1891 * at all. This is possible, for example, if page, submitted for read,
1892 * became up-to-date in the meantime; and for write, the page don't have
1893 * dirty bit marked. \see cl_io_submit_rw()
1895 * Whenever a cached page is added to a newly constructed req, its
1896 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1897 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1898 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1899 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1900 * might return -EAGAIN to indicate that this page is not eligible for the
1901 * transfer right now.
1905 * Plan is to divide transfers into "priority bands" (indicated when
1906 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1907 * and allow glueing of cached pages to immediate transfers only within single
1908 * band. This would make high priority transfers (like lock cancellation or
1909 * memory pressure induced write-out) really high priority.
1914 * Per-transfer attributes.
1916 struct cl_req_attr {
1917 /** Generic attributes for the server consumption. */
1918 struct obdo *cra_oa;
1920 struct obd_capa *cra_capa;
1922 char cra_jobid[LUSTRE_JOBID_SIZE];
1926 * Transfer request operations definable at every layer.
1928 * Concurrency: transfer formation engine synchronizes calls to all transfer
1931 struct cl_req_operations {
1933 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1934 * complete (all pages are added).
1936 * \see osc_req_prep()
1938 int (*cro_prep)(const struct lu_env *env,
1939 const struct cl_req_slice *slice);
1941 * Called top-to-bottom to fill in \a oa fields. This is called twice
1942 * with different flags, see bug 10150 and osc_build_req().
1944 * \param obj an object from cl_req which attributes are to be set in
1947 * \param oa struct obdo where attributes are placed
1949 * \param flags \a oa fields to be filled.
1951 void (*cro_attr_set)(const struct lu_env *env,
1952 const struct cl_req_slice *slice,
1953 const struct cl_object *obj,
1954 struct cl_req_attr *attr, u64 flags);
1956 * Called top-to-bottom from cl_req_completion() to notify layers that
1957 * transfer completed. Has to free all state allocated by
1958 * cl_device_operations::cdo_req_init().
1960 void (*cro_completion)(const struct lu_env *env,
1961 const struct cl_req_slice *slice, int ioret);
1965 * A per-object state that (potentially multi-object) transfer request keeps.
1968 /** object itself */
1969 struct cl_object *ro_obj;
1970 /** reference to cl_req_obj::ro_obj. For debugging. */
1971 struct lu_ref_link ro_obj_ref;
1972 /* something else? Number of pages for a given object? */
1978 * Transfer requests are not reference counted, because IO sub-system owns
1979 * them exclusively and knows when to free them.
1983 * cl_req is created by cl_req_alloc() that calls
1984 * cl_device_operations::cdo_req_init() device methods to allocate per-req
1985 * state in every layer.
1987 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
1988 * contains pages for.
1990 * Once all pages were collected, cl_page_operations::cpo_prep() method is
1991 * called top-to-bottom. At that point layers can modify req, let it pass, or
1992 * deny it completely. This is to support things like SNS that have transfer
1993 * ordering requirements invisible to the individual req-formation engine.
1995 * On transfer completion (or transfer timeout, or failure to initiate the
1996 * transfer of an allocated req), cl_req_operations::cro_completion() method
1997 * is called, after execution of cl_page_operations::cpo_completion() of all
2001 enum cl_req_type crq_type;
2002 /** A list of pages being transfered */
2003 struct list_head crq_pages;
2004 /** Number of pages in cl_req::crq_pages */
2005 unsigned crq_nrpages;
2006 /** An array of objects which pages are in ->crq_pages */
2007 struct cl_req_obj *crq_o;
2008 /** Number of elements in cl_req::crq_objs[] */
2009 unsigned crq_nrobjs;
2010 struct list_head crq_layers;
2014 * Per-layer state for request.
2016 struct cl_req_slice {
2017 struct cl_req *crs_req;
2018 struct cl_device *crs_dev;
2019 struct list_head crs_linkage;
2020 const struct cl_req_operations *crs_ops;
2025 enum cache_stats_item {
2026 /** how many cache lookups were performed */
2028 /** how many times cache lookup resulted in a hit */
2030 /** how many entities are in the cache right now */
2032 /** how many entities in the cache are actively used (and cannot be
2033 * evicted) right now */
2035 /** how many entities were created at all */
2040 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2043 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2045 struct cache_stats {
2046 const char *cs_name;
2047 atomic_t cs_stats[CS_NR];
2050 /** These are not exported so far */
2051 void cache_stats_init (struct cache_stats *cs, const char *name);
2054 * Client-side site. This represents particular client stack. "Global"
2055 * variables should (directly or indirectly) be added here to allow multiple
2056 * clients to co-exist in the single address space.
2059 struct lu_site cs_lu;
2061 * Statistical counters. Atomics do not scale, something better like
2062 * per-cpu counters is needed.
2064 * These are exported as /proc/fs/lustre/llite/.../site
2066 * When interpreting keep in mind that both sub-locks (and sub-pages)
2067 * and top-locks (and top-pages) are accounted here.
2069 struct cache_stats cs_pages;
2070 atomic_t cs_pages_state[CPS_NR];
2073 int cl_site_init(struct cl_site *s, struct cl_device *top);
2074 void cl_site_fini(struct cl_site *s);
2075 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2078 * Output client site statistical counters into a buffer. Suitable for
2079 * ll_rd_*()-style functions.
2081 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2086 * Type conversion and accessory functions.
2090 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2092 return container_of(site, struct cl_site, cs_lu);
2095 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2097 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2098 return container_of0(d, struct cl_device, cd_lu_dev);
2101 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2103 return &d->cd_lu_dev;
2106 static inline struct cl_object *lu2cl(const struct lu_object *o)
2108 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2109 return container_of0(o, struct cl_object, co_lu);
2112 static inline const struct cl_object_conf *
2113 lu2cl_conf(const struct lu_object_conf *conf)
2115 return container_of0(conf, struct cl_object_conf, coc_lu);
2118 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2120 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2123 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2125 return container_of0(h, struct cl_object_header, coh_lu);
2128 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2130 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2134 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2136 return luh2coh(obj->co_lu.lo_header);
2139 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2141 return lu_device_init(&d->cd_lu_dev, t);
2144 static inline void cl_device_fini(struct cl_device *d)
2146 lu_device_fini(&d->cd_lu_dev);
2149 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2150 struct cl_object *obj, pgoff_t index,
2151 const struct cl_page_operations *ops);
2152 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2153 struct cl_object *obj,
2154 const struct cl_lock_operations *ops);
2155 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2156 struct cl_object *obj, const struct cl_io_operations *ops);
2157 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2158 struct cl_device *dev,
2159 const struct cl_req_operations *ops);
2162 /** \defgroup cl_object cl_object
2164 struct cl_object *cl_object_top (struct cl_object *o);
2165 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2166 const struct lu_fid *fid,
2167 const struct cl_object_conf *c);
2169 int cl_object_header_init(struct cl_object_header *h);
2170 void cl_object_header_fini(struct cl_object_header *h);
2171 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2172 void cl_object_get (struct cl_object *o);
2173 void cl_object_attr_lock (struct cl_object *o);
2174 void cl_object_attr_unlock(struct cl_object *o);
2175 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2176 struct cl_attr *attr);
2177 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2178 const struct cl_attr *attr, unsigned valid);
2179 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2180 struct ost_lvb *lvb);
2181 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2182 const struct cl_object_conf *conf);
2183 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2184 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2185 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2186 struct lov_user_md __user *lum);
2187 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2188 ldlm_iterator_t iter, void *data);
2189 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2190 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2194 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2196 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2198 return cl_object_header(o0) == cl_object_header(o1);
2201 static inline void cl_object_page_init(struct cl_object *clob, int size)
2203 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2204 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2205 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2208 static inline void *cl_object_page_slice(struct cl_object *clob,
2209 struct cl_page *page)
2211 return (void *)((char *)page + clob->co_slice_off);
2215 * Return refcount of cl_object.
2217 static inline int cl_object_refc(struct cl_object *clob)
2219 struct lu_object_header *header = clob->co_lu.lo_header;
2220 return atomic_read(&header->loh_ref);
2225 /** \defgroup cl_page cl_page
2233 /* callback of cl_page_gang_lookup() */
2235 struct cl_page *cl_page_find (const struct lu_env *env,
2236 struct cl_object *obj,
2237 pgoff_t idx, struct page *vmpage,
2238 enum cl_page_type type);
2239 struct cl_page *cl_page_alloc (const struct lu_env *env,
2240 struct cl_object *o, pgoff_t ind,
2241 struct page *vmpage,
2242 enum cl_page_type type);
2243 void cl_page_get (struct cl_page *page);
2244 void cl_page_put (const struct lu_env *env,
2245 struct cl_page *page);
2246 void cl_page_print (const struct lu_env *env, void *cookie,
2247 lu_printer_t printer,
2248 const struct cl_page *pg);
2249 void cl_page_header_print(const struct lu_env *env, void *cookie,
2250 lu_printer_t printer,
2251 const struct cl_page *pg);
2252 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2253 struct cl_page *cl_page_top (struct cl_page *page);
2255 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2256 const struct lu_device_type *dtype);
2261 * Functions dealing with the ownership of page by io.
2265 int cl_page_own (const struct lu_env *env,
2266 struct cl_io *io, struct cl_page *page);
2267 int cl_page_own_try (const struct lu_env *env,
2268 struct cl_io *io, struct cl_page *page);
2269 void cl_page_assume (const struct lu_env *env,
2270 struct cl_io *io, struct cl_page *page);
2271 void cl_page_unassume (const struct lu_env *env,
2272 struct cl_io *io, struct cl_page *pg);
2273 void cl_page_disown (const struct lu_env *env,
2274 struct cl_io *io, struct cl_page *page);
2275 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2282 * Functions dealing with the preparation of a page for a transfer, and
2283 * tracking transfer state.
2286 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2287 struct cl_page *pg, enum cl_req_type crt);
2288 void cl_page_completion (const struct lu_env *env,
2289 struct cl_page *pg, enum cl_req_type crt, int ioret);
2290 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2291 enum cl_req_type crt);
2292 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2293 struct cl_page *pg, enum cl_req_type crt);
2294 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2296 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2297 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2298 struct cl_page *pg);
2304 * \name helper routines
2305 * Functions to discard, delete and export a cl_page.
2308 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2309 struct cl_page *pg);
2310 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2311 int cl_page_is_vmlocked(const struct lu_env *env,
2312 const struct cl_page *pg);
2313 void cl_page_export(const struct lu_env *env,
2314 struct cl_page *pg, int uptodate);
2315 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2316 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2317 size_t cl_page_size(const struct cl_object *obj);
2319 void cl_lock_print(const struct lu_env *env, void *cookie,
2320 lu_printer_t printer, const struct cl_lock *lock);
2321 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2322 lu_printer_t printer,
2323 const struct cl_lock_descr *descr);
2327 * Data structure managing a client's cached pages. A count of
2328 * "unstable" pages is maintained, and an LRU of clean pages is
2329 * maintained. "unstable" pages are pages pinned by the ptlrpc
2330 * layer for recovery purposes.
2332 struct cl_client_cache {
2334 * # of client cache refcount
2335 * # of users (OSCs) + 2 (held by llite and lov)
2339 * # of threads are doing shrinking
2341 unsigned int ccc_lru_shrinkers;
2343 * # of LRU entries available
2345 atomic_long_t ccc_lru_left;
2347 * List of entities(OSCs) for this LRU cache
2349 struct list_head ccc_lru;
2351 * Max # of LRU entries
2353 unsigned long ccc_lru_max;
2355 * Lock to protect ccc_lru list
2357 spinlock_t ccc_lru_lock;
2359 * Set if unstable check is enabled
2361 unsigned int ccc_unstable_check:1;
2363 * # of unstable pages for this mount point
2365 atomic_long_t ccc_unstable_nr;
2367 * Waitq for awaiting unstable pages to reach zero.
2368 * Used at umounting time and signaled on BRW commit
2370 wait_queue_head_t ccc_unstable_waitq;
2373 * cl_cache functions
2375 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2376 void cl_cache_incref(struct cl_client_cache *cache);
2377 void cl_cache_decref(struct cl_client_cache *cache);
2381 /** \defgroup cl_lock cl_lock
2383 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2384 struct cl_lock *lock);
2385 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2386 const struct cl_io *io);
2387 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2388 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2389 const struct lu_device_type *dtype);
2390 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2392 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2393 struct cl_lock *lock, struct cl_sync_io *anchor);
2394 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2398 /** \defgroup cl_io cl_io
2401 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2402 enum cl_io_type iot, struct cl_object *obj);
2403 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2404 enum cl_io_type iot, struct cl_object *obj);
2405 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2406 enum cl_io_type iot, loff_t pos, size_t count);
2407 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2409 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2410 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2411 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2412 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2413 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2414 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2415 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2416 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2417 struct cl_io_lock_link *link);
2418 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2419 struct cl_lock_descr *descr);
2420 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2421 enum cl_req_type iot, struct cl_2queue *queue);
2422 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2423 enum cl_req_type iot, struct cl_2queue *queue,
2425 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2426 struct cl_page_list *queue, int from, int to,
2428 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2429 pgoff_t start, struct cl_read_ahead *ra);
2430 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2432 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2433 struct cl_page_list *queue);
2434 int cl_io_is_going (const struct lu_env *env);
2437 * True, iff \a io is an O_APPEND write(2).
2439 static inline int cl_io_is_append(const struct cl_io *io)
2441 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2444 static inline int cl_io_is_sync_write(const struct cl_io *io)
2446 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2449 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2451 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2455 * True, iff \a io is a truncate(2).
2457 static inline int cl_io_is_trunc(const struct cl_io *io)
2459 return io->ci_type == CIT_SETATTR &&
2460 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2463 struct cl_io *cl_io_top(struct cl_io *io);
2465 void cl_io_print(const struct lu_env *env, void *cookie,
2466 lu_printer_t printer, const struct cl_io *io);
2468 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2470 typeof(foo_io) __foo_io = (foo_io); \
2472 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2473 memset(&__foo_io->base + 1, 0, \
2474 (sizeof *__foo_io) - sizeof __foo_io->base); \
2479 /** \defgroup cl_page_list cl_page_list
2483 * Last page in the page list.
2485 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2487 LASSERT(plist->pl_nr > 0);
2488 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2491 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2493 LASSERT(plist->pl_nr > 0);
2494 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2498 * Iterate over pages in a page list.
2500 #define cl_page_list_for_each(page, list) \
2501 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2504 * Iterate over pages in a page list, taking possible removals into account.
2506 #define cl_page_list_for_each_safe(page, temp, list) \
2507 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2509 void cl_page_list_init (struct cl_page_list *plist);
2510 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2511 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2512 struct cl_page *page);
2513 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2514 struct cl_page *page);
2515 void cl_page_list_splice (struct cl_page_list *list,
2516 struct cl_page_list *head);
2517 void cl_page_list_del (const struct lu_env *env,
2518 struct cl_page_list *plist, struct cl_page *page);
2519 void cl_page_list_disown (const struct lu_env *env,
2520 struct cl_io *io, struct cl_page_list *plist);
2521 int cl_page_list_own (const struct lu_env *env,
2522 struct cl_io *io, struct cl_page_list *plist);
2523 void cl_page_list_assume (const struct lu_env *env,
2524 struct cl_io *io, struct cl_page_list *plist);
2525 void cl_page_list_discard(const struct lu_env *env,
2526 struct cl_io *io, struct cl_page_list *plist);
2527 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2529 void cl_2queue_init (struct cl_2queue *queue);
2530 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2531 void cl_2queue_disown (const struct lu_env *env,
2532 struct cl_io *io, struct cl_2queue *queue);
2533 void cl_2queue_assume (const struct lu_env *env,
2534 struct cl_io *io, struct cl_2queue *queue);
2535 void cl_2queue_discard (const struct lu_env *env,
2536 struct cl_io *io, struct cl_2queue *queue);
2537 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2538 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2540 /** @} cl_page_list */
2542 /** \defgroup cl_req cl_req
2544 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2545 enum cl_req_type crt, int nr_objects);
2547 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2548 struct cl_page *page);
2549 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2550 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2551 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2552 struct cl_req_attr *attr, u64 flags);
2553 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2555 /** \defgroup cl_sync_io cl_sync_io
2559 * Anchor for synchronous transfer. This is allocated on a stack by thread
2560 * doing synchronous transfer, and a pointer to this structure is set up in
2561 * every page submitted for transfer. Transfer completion routine updates
2562 * anchor and wakes up waiting thread when transfer is complete.
2565 /** number of pages yet to be transferred. */
2566 atomic_t csi_sync_nr;
2569 /** barrier of destroy this structure */
2570 atomic_t csi_barrier;
2571 /** completion to be signaled when transfer is complete. */
2572 wait_queue_head_t csi_waitq;
2573 /** callback to invoke when this IO is finished */
2574 void (*csi_end_io)(const struct lu_env *,
2575 struct cl_sync_io *);
2578 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2579 void (*end)(const struct lu_env *, struct cl_sync_io *));
2580 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2582 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2584 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2586 /** @} cl_sync_io */
2590 /** \defgroup cl_env cl_env
2592 * lu_env handling for a client.
2594 * lu_env is an environment within which lustre code executes. Its major part
2595 * is lu_context---a fast memory allocation mechanism that is used to conserve
2596 * precious kernel stack space. Originally lu_env was designed for a server,
2599 * - there is a (mostly) fixed number of threads, and
2601 * - call chains have no non-lustre portions inserted between lustre code.
2603 * On a client both these assumtpion fails, because every user thread can
2604 * potentially execute lustre code as part of a system call, and lustre calls
2605 * into VFS or MM that call back into lustre.
2607 * To deal with that, cl_env wrapper functions implement the following
2610 * - allocation and destruction of environment is amortized by caching no
2611 * longer used environments instead of destroying them;
2613 * - there is a notion of "current" environment, attached to the kernel
2614 * data structure representing current thread Top-level lustre code
2615 * allocates an environment and makes it current, then calls into
2616 * non-lustre code, that in turn calls lustre back. Low-level lustre
2617 * code thus called can fetch environment created by the top-level code
2618 * and reuse it, avoiding additional environment allocation.
2619 * Right now, three interfaces can attach the cl_env to running thread:
2622 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2624 * \see lu_env, lu_context, lu_context_key
2627 struct cl_env_nest {
2632 struct lu_env *cl_env_peek (int *refcheck);
2633 struct lu_env *cl_env_get (int *refcheck);
2634 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2635 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2636 void cl_env_put (struct lu_env *env, int *refcheck);
2637 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2638 void *cl_env_reenter (void);
2639 void cl_env_reexit (void *cookie);
2640 void cl_env_implant (struct lu_env *env, int *refcheck);
2641 void cl_env_unplant (struct lu_env *env, int *refcheck);
2642 unsigned cl_env_cache_purge(unsigned nr);
2643 struct lu_env *cl_env_percpu_get (void);
2644 void cl_env_percpu_put (struct lu_env *env);
2651 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2652 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2654 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2655 struct lu_device_type *ldt,
2656 struct lu_device *next);
2659 int cl_global_init(void);
2660 void cl_global_fini(void);
2662 #endif /* _LINUX_CL_OBJECT_H */