1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 #ifndef _LUSTRE_CL_OBJECT_H
40 #define _LUSTRE_CL_OBJECT_H
42 /** \defgroup clio clio
44 * Client objects implement io operations and cache pages.
46 * Examples: lov and osc are implementations of cl interface.
48 * Big Theory Statement.
52 * Client implementation is based on the following data-types:
58 * - cl_lock represents an extent lock on an object.
60 * - cl_io represents high-level i/o activity such as whole read/write
61 * system call, or write-out of pages from under the lock being
62 * canceled. cl_io has sub-ios that can be stopped and resumed
63 * independently, thus achieving high degree of transfer
64 * parallelism. Single cl_io can be advanced forward by
65 * the multiple threads (although in the most usual case of
66 * read/write system call it is associated with the single user
67 * thread, that issued the system call).
69 * - cl_req represents a collection of pages for a transfer. cl_req is
70 * constructed by req-forming engine that tries to saturate
71 * transport with large and continuous transfers.
75 * - to avoid confusion high-level I/O operation like read or write system
76 * call is referred to as "an io", whereas low-level I/O operation, like
77 * RPC, is referred to as "a transfer"
79 * - "generic code" means generic (not file system specific) code in the
80 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
81 * is not layer specific.
87 * - cl_object_header::coh_page_guard
88 * - cl_object_header::coh_lock_guard
91 * See the top comment in cl_object.c for the description of overall locking and
92 * reference-counting design.
94 * See comments below for the description of i/o, page, and dlm-locking
101 * super-class definitions.
103 #include <lu_object.h>
106 # include <linux/mutex.h>
107 # include <linux/radix-tree.h>
113 struct cl_device_operations;
116 struct cl_object_page_operations;
117 struct cl_object_lock_operations;
120 struct cl_page_slice;
122 struct cl_lock_slice;
124 struct cl_lock_operations;
125 struct cl_page_operations;
134 * Operations for each data device in the client stack.
136 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
138 struct cl_device_operations {
140 * Initialize cl_req. This method is called top-to-bottom on all
141 * devices in the stack to get them a chance to allocate layer-private
142 * data, and to attach them to the cl_req by calling
143 * cl_req_slice_add().
145 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
146 * \see ccc_req_init()
148 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
153 * Device in the client stack.
155 * \see ccc_device, lov_device, lovsub_device, osc_device
159 struct lu_device cd_lu_dev;
160 /** Per-layer operation vector. */
161 const struct cl_device_operations *cd_ops;
164 /** \addtogroup cl_object cl_object
167 * "Data attributes" of cl_object. Data attributes can be updated
168 * independently for a sub-object, and top-object's attributes are calculated
169 * from sub-objects' ones.
172 /** Object size, in bytes */
175 * Known minimal size, in bytes.
177 * This is only valid when at least one DLM lock is held.
180 /** Modification time. Measured in seconds since epoch. */
182 /** Access time. Measured in seconds since epoch. */
184 /** Change time. Measured in seconds since epoch. */
187 * Blocks allocated to this cl_object on the server file system.
189 * \todo XXX An interface for block size is needed.
193 * User identifier for quota purposes.
197 * Group identifier for quota purposes.
203 * Fields in cl_attr that are being set.
217 * Sub-class of lu_object with methods common for objects on the client
220 * cl_object: represents a regular file system object, both a file and a
221 * stripe. cl_object is based on lu_object: it is identified by a fid,
222 * layered, cached, hashed, and lrued. Important distinction with the server
223 * side, where md_object and dt_object are used, is that cl_object "fans out"
224 * at the lov/sns level: depending on the file layout, single file is
225 * represented as a set of "sub-objects" (stripes). At the implementation
226 * level, struct lov_object contains an array of cl_objects. Each sub-object
227 * is a full-fledged cl_object, having its fid, living in the lru and hash
230 * This leads to the next important difference with the server side: on the
231 * client, it's quite usual to have objects with the different sequence of
232 * layers. For example, typical top-object is composed of the following
238 * whereas its sub-objects are composed of
243 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
244 * track of the object-subobject relationship.
246 * Sub-objects are not cached independently: when top-object is about to
247 * be discarded from the memory, all its sub-objects are torn-down and
250 * \see ccc_object, lov_object, lovsub_object, osc_object
254 struct lu_object co_lu;
255 /** per-object-layer operations */
256 const struct cl_object_operations *co_ops;
260 * Description of the client object configuration. This is used for the
261 * creation of a new client object that is identified by a more state than
264 struct cl_object_conf {
266 struct lu_object_conf coc_lu;
269 * Object layout. This is consumed by lov.
271 struct lustre_md *coc_md;
273 * Description of particular stripe location in the
274 * cluster. This is consumed by osc.
276 struct lov_oinfo *coc_oinfo;
279 * VFS inode. This is consumed by vvp.
281 struct inode *coc_inode;
285 * Operations implemented for each cl object layer.
287 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
289 struct cl_object_operations {
291 * Initialize page slice for this layer. Called top-to-bottom through
292 * every object layer when a new cl_page is instantiated. Layer
293 * keeping private per-page data, or requiring its own page operations
294 * vector should allocate these data here, and attach then to the page
295 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
298 * \retval NULL success.
300 * \retval ERR_PTR(errno) failure code.
302 * \retval valid-pointer pointer to already existing referenced page
303 * to be used instead of newly created.
305 struct cl_page *(*coo_page_init)(const struct lu_env *env,
306 struct cl_object *obj,
307 struct cl_page *page,
310 * Initialize lock slice for this layer. Called top-to-bottom through
311 * every object layer when a new cl_lock is instantiated. Layer
312 * keeping private per-lock data, or requiring its own lock operations
313 * vector should allocate these data here, and attach then to the lock
314 * by calling cl_lock_slice_add(). Mandatory.
316 int (*coo_lock_init)(const struct lu_env *env,
317 struct cl_object *obj, struct cl_lock *lock,
318 const struct cl_io *io);
320 * Initialize io state for a given layer.
322 * called top-to-bottom once per io existence to initialize io
323 * state. If layer wants to keep some state for this type of io, it
324 * has to embed struct cl_io_slice in lu_env::le_ses, and register
325 * slice with cl_io_slice_add(). It is guaranteed that all threads
326 * participating in this io share the same session.
328 int (*coo_io_init)(const struct lu_env *env,
329 struct cl_object *obj, struct cl_io *io);
331 * Fill portion of \a attr that this layer controls. This method is
332 * called top-to-bottom through all object layers.
334 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
336 * \return 0: to continue
337 * \return +ve: to stop iterating through layers (but 0 is returned
338 * from enclosing cl_object_attr_get())
339 * \return -ve: to signal error
341 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
342 struct cl_attr *attr);
346 * \a valid is a bitmask composed from enum #cl_attr_valid, and
347 * indicating what attributes are to be set.
349 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
351 * \return the same convention as for
352 * cl_object_operations::coo_attr_get() is used.
354 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj,
355 const struct cl_attr *attr, unsigned valid);
357 * Update object configuration. Called top-to-bottom to modify object
360 * XXX error conditions and handling.
362 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
363 const struct cl_object_conf *conf);
365 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
366 * object. Layers are supposed to fill parts of \a lvb that will be
367 * shipped to the glimpse originator as a glimpse result.
369 * \see ccc_object_glimpse(), lovsub_object_glimpse(),
370 * \see osc_object_glimpse()
372 int (*coo_glimpse)(const struct lu_env *env,
373 const struct cl_object *obj, struct ost_lvb *lvb);
377 * Extended header for client object.
379 struct cl_object_header {
380 /** Standard lu_object_header. cl_object::co_lu::lo_header points
382 struct lu_object_header coh_lu;
384 * \todo XXX move locks below to the separate cache-lines, they are
385 * mostly useless otherwise.
388 /** Lock protecting page tree. */
389 cfs_spinlock_t coh_page_guard;
390 /** Lock protecting lock list. */
391 cfs_spinlock_t coh_lock_guard;
393 /** Radix tree of cl_page's, cached for this object. */
394 struct radix_tree_root coh_tree;
395 /** # of pages in radix tree. */
396 unsigned long coh_pages;
397 /** List of cl_lock's granted for this object. */
398 cfs_list_t coh_locks;
401 * Parent object. It is assumed that an object has a well-defined
402 * parent, but not a well-defined child (there may be multiple
403 * sub-objects, for the same top-object). cl_object_header::coh_parent
404 * field allows certain code to be written generically, without
405 * limiting possible cl_object layouts unduly.
407 struct cl_object_header *coh_parent;
409 * Protects consistency between cl_attr of parent object and
410 * attributes of sub-objects, that the former is calculated ("merged")
413 * \todo XXX this can be read/write lock if needed.
415 cfs_spinlock_t coh_attr_guard;
417 * Number of objects above this one: 0 for a top-object, 1 for its
420 unsigned coh_nesting;
424 * Helper macro: iterate over all layers of the object \a obj, assigning every
425 * layer top-to-bottom to \a slice.
427 #define cl_object_for_each(slice, obj) \
428 cfs_list_for_each_entry((slice), \
429 &(obj)->co_lu.lo_header->loh_layers, \
432 * Helper macro: iterate over all layers of the object \a obj, assigning every
433 * layer bottom-to-top to \a slice.
435 #define cl_object_for_each_reverse(slice, obj) \
436 cfs_list_for_each_entry_reverse((slice), \
437 &(obj)->co_lu.lo_header->loh_layers, \
442 #define pgoff_t unsigned long
445 #define CL_PAGE_EOF ((pgoff_t)~0ull)
447 /** \addtogroup cl_page cl_page
451 * Layered client page.
453 * cl_page: represents a portion of a file, cached in the memory. All pages
454 * of the given file are of the same size, and are kept in the radix tree
455 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
456 * of the top-level file object are first class cl_objects, they have their
457 * own radix trees of pages and hence page is implemented as a sequence of
458 * struct cl_pages's, linked into double-linked list through
459 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
460 * corresponding radix tree at the corresponding logical offset.
462 * cl_page is associated with VM page of the hosting environment (struct
463 * page in Linux kernel, for example), cfs_page_t. It is assumed, that this
464 * association is implemented by one of cl_page layers (top layer in the
465 * current design) that
467 * - intercepts per-VM-page call-backs made by the environment (e.g.,
470 * - translates state (page flag bits) and locking between lustre and
473 * The association between cl_page and cfs_page_t is immutable and
474 * established when cl_page is created.
476 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
477 * this io an exclusive access to this page w.r.t. other io attempts and
478 * various events changing page state (such as transfer completion, or
479 * eviction of the page from the memory). Note, that in general cl_io
480 * cannot be identified with a particular thread, and page ownership is not
481 * exactly equal to the current thread holding a lock on the page. Layer
482 * implementing association between cl_page and cfs_page_t has to implement
483 * ownership on top of available synchronization mechanisms.
485 * While lustre client maintains the notion of an page ownership by io,
486 * hosting MM/VM usually has its own page concurrency control
487 * mechanisms. For example, in Linux, page access is synchronized by the
488 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
489 * takes care to acquire and release such locks as necessary around the
490 * calls to the file system methods (->readpage(), ->prepare_write(),
491 * ->commit_write(), etc.). This leads to the situation when there are two
492 * different ways to own a page in the client:
494 * - client code explicitly and voluntary owns the page (cl_page_own());
496 * - VM locks a page and then calls the client, that has "to assume"
497 * the ownership from the VM (cl_page_assume()).
499 * Dual methods to release ownership are cl_page_disown() and
500 * cl_page_unassume().
502 * cl_page is reference counted (cl_page::cp_ref). When reference counter
503 * drops to 0, the page is returned to the cache, unless it is in
504 * cl_page_state::CPS_FREEING state, in which case it is immediately
507 * The general logic guaranteeing the absence of "existential races" for
508 * pages is the following:
510 * - there are fixed known ways for a thread to obtain a new reference
513 * - by doing a lookup in the cl_object radix tree, protected by the
516 * - by starting from VM-locked cfs_page_t and following some
517 * hosting environment method (e.g., following ->private pointer in
518 * the case of Linux kernel), see cl_vmpage_page();
520 * - when the page enters cl_page_state::CPS_FREEING state, all these
521 * ways are severed with the proper synchronization
522 * (cl_page_delete());
524 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
527 * - no new references to the page in cl_page_state::CPS_FREEING state
528 * are allowed (checked in cl_page_get()).
530 * Together this guarantees that when last reference to a
531 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
532 * page, as neither references to it can be acquired at that point, nor
535 * cl_page is a state machine. States are enumerated in enum
536 * cl_page_state. Possible state transitions are enumerated in
537 * cl_page_state_set(). State transition process (i.e., actual changing of
538 * cl_page::cp_state field) is protected by the lock on the underlying VM
541 * Linux Kernel implementation.
543 * Binding between cl_page and cfs_page_t (which is a typedef for
544 * struct page) is implemented in the vvp layer. cl_page is attached to the
545 * ->private pointer of the struct page, together with the setting of
546 * PG_private bit in page->flags, and acquiring additional reference on the
547 * struct page (much like struct buffer_head, or any similar file system
548 * private data structures).
550 * PG_locked lock is used to implement both ownership and transfer
551 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
552 * states. No additional references are acquired for the duration of the
555 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
556 * write-out is "protected" by the special PG_writeback bit.
560 * States of cl_page. cl_page.c assumes particular order here.
562 * The page state machine is rather crude, as it doesn't recognize finer page
563 * states like "dirty" or "up to date". This is because such states are not
564 * always well defined for the whole stack (see, for example, the
565 * implementation of the read-ahead, that hides page up-to-dateness to track
566 * cache hits accurately). Such sub-states are maintained by the layers that
567 * are interested in them.
571 * Page is in the cache, un-owned. Page leaves cached state in the
574 * - [cl_page_state::CPS_OWNED] io comes across the page and
577 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
578 * req-formation engine decides that it wants to include this page
579 * into an cl_req being constructed, and yanks it from the cache;
581 * - [cl_page_state::CPS_FREEING] VM callback is executed to
582 * evict the page form the memory;
584 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
588 * Page is exclusively owned by some cl_io. Page may end up in this
589 * state as a result of
591 * - io creating new page and immediately owning it;
593 * - [cl_page_state::CPS_CACHED] io finding existing cached page
596 * - [cl_page_state::CPS_OWNED] io finding existing owned page
597 * and waiting for owner to release the page;
599 * Page leaves owned state in the following cases:
601 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
602 * the cache, doing nothing;
604 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
607 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
608 * transfer for this page;
610 * - [cl_page_state::CPS_FREEING] io decides to destroy this
611 * page (e.g., as part of truncate or extent lock cancellation).
613 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
617 * Page is being written out, as a part of a transfer. This state is
618 * entered when req-formation logic decided that it wants this page to
619 * be sent through the wire _now_. Specifically, it means that once
620 * this state is achieved, transfer completion handler (with either
621 * success or failure indication) is guaranteed to be executed against
622 * this page independently of any locks and any scheduling decisions
623 * made by the hosting environment (that effectively means that the
624 * page is never put into cl_page_state::CPS_PAGEOUT state "in
625 * advance". This property is mentioned, because it is important when
626 * reasoning about possible dead-locks in the system). The page can
627 * enter this state as a result of
629 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
630 * write-out of this page, or
632 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
633 * that it has enough dirty pages cached to issue a "good"
636 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
637 * is completed---it is moved into cl_page_state::CPS_CACHED state.
639 * Underlying VM page is locked for the duration of transfer.
641 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
645 * Page is being read in, as a part of a transfer. This is quite
646 * similar to the cl_page_state::CPS_PAGEOUT state, except that
647 * read-in is always "immediate"---there is no such thing a sudden
648 * construction of read cl_req from cached, presumably not up to date,
651 * Underlying VM page is locked for the duration of transfer.
653 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
657 * Page is being destroyed. This state is entered when client decides
658 * that page has to be deleted from its host object, as, e.g., a part
661 * Once this state is reached, there is no way to escape it.
663 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
670 /** Host page, the page is from the host inode which the cl_page
674 /** Transient page, the transient cl_page is used to bind a cl_page
675 * to vmpage which is not belonging to the same object of cl_page.
676 * it is used in DirectIO, lockless IO and liblustre. */
681 * Flags maintained for every cl_page.
685 * Set when pagein completes. Used for debugging (read completes at
686 * most once for a page).
688 CPF_READ_COMPLETED = 1 << 0
692 * Fields are protected by the lock on cfs_page_t, except for atomics and
695 * \invariant Data type invariants are in cl_page_invariant(). Basically:
696 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
697 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
698 * cl_page::cp_owner (when set).
701 /** Reference counter. */
703 /** An object this page is a part of. Immutable after creation. */
704 struct cl_object *cp_obj;
705 /** Logical page index within the object. Immutable after creation. */
707 /** List of slices. Immutable after creation. */
708 cfs_list_t cp_layers;
709 /** Parent page, NULL for top-level page. Immutable after creation. */
710 struct cl_page *cp_parent;
711 /** Lower-layer page. NULL for bottommost page. Immutable after
713 struct cl_page *cp_child;
715 * Page state. This field is const to avoid accidental update, it is
716 * modified only internally within cl_page.c. Protected by a VM lock.
718 const enum cl_page_state cp_state;
720 * Linkage of pages within some group. Protected by
721 * cl_page::cp_mutex. */
723 /** Mutex serializing membership of a page in a batch. */
724 cfs_mutex_t cp_mutex;
725 /** Linkage of pages within cl_req. */
726 cfs_list_t cp_flight;
727 /** Transfer error. */
731 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
734 enum cl_page_type cp_type;
737 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
738 * by sub-io. Protected by a VM lock.
740 struct cl_io *cp_owner;
742 * Debug information, the task is owning the page.
746 * Owning IO request in cl_page_state::CPS_PAGEOUT and
747 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
748 * the top-level pages. Protected by a VM lock.
750 struct cl_req *cp_req;
751 /** List of references to this page, for debugging. */
752 struct lu_ref cp_reference;
753 /** Link to an object, for debugging. */
754 struct lu_ref_link *cp_obj_ref;
755 /** Link to a queue, for debugging. */
756 struct lu_ref_link *cp_queue_ref;
757 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */
759 /** Assigned if doing a sync_io */
760 struct cl_sync_io *cp_sync_io;
764 * Per-layer part of cl_page.
766 * \see ccc_page, lov_page, osc_page
768 struct cl_page_slice {
769 struct cl_page *cpl_page;
771 * Object slice corresponding to this page slice. Immutable after
774 struct cl_object *cpl_obj;
775 const struct cl_page_operations *cpl_ops;
776 /** Linkage into cl_page::cp_layers. Immutable after creation. */
777 cfs_list_t cpl_linkage;
781 * Lock mode. For the client extent locks.
783 * \warning: cl_lock_mode_match() assumes particular ordering here.
788 * Mode of a lock that protects no data, and exists only as a
789 * placeholder. This is used for `glimpse' requests. A phantom lock
790 * might get promoted to real lock at some point.
799 * Requested transfer type.
809 * Per-layer page operations.
811 * Methods taking an \a io argument are for the activity happening in the
812 * context of given \a io. Page is assumed to be owned by that io, except for
813 * the obvious cases (like cl_page_operations::cpo_own()).
815 * \see vvp_page_ops, lov_page_ops, osc_page_ops
817 struct cl_page_operations {
819 * cl_page<->cfs_page_t methods. Only one layer in the stack has to
820 * implement these. Current code assumes that this functionality is
821 * provided by the topmost layer, see cl_page_disown0() as an example.
825 * \return the underlying VM page. Optional.
827 cfs_page_t *(*cpo_vmpage)(const struct lu_env *env,
828 const struct cl_page_slice *slice);
830 * Called when \a io acquires this page into the exclusive
831 * ownership. When this method returns, it is guaranteed that the is
832 * not owned by other io, and no transfer is going on against
836 * \see vvp_page_own(), lov_page_own()
838 int (*cpo_own)(const struct lu_env *env,
839 const struct cl_page_slice *slice,
840 struct cl_io *io, int nonblock);
841 /** Called when ownership it yielded. Optional.
843 * \see cl_page_disown()
844 * \see vvp_page_disown()
846 void (*cpo_disown)(const struct lu_env *env,
847 const struct cl_page_slice *slice, struct cl_io *io);
849 * Called for a page that is already "owned" by \a io from VM point of
852 * \see cl_page_assume()
853 * \see vvp_page_assume(), lov_page_assume()
855 void (*cpo_assume)(const struct lu_env *env,
856 const struct cl_page_slice *slice, struct cl_io *io);
857 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
858 * bottom-to-top when IO releases a page without actually unlocking
861 * \see cl_page_unassume()
862 * \see vvp_page_unassume()
864 void (*cpo_unassume)(const struct lu_env *env,
865 const struct cl_page_slice *slice,
868 * Announces whether the page contains valid data or not by \a uptodate.
870 * \see cl_page_export()
871 * \see vvp_page_export()
873 void (*cpo_export)(const struct lu_env *env,
874 const struct cl_page_slice *slice, int uptodate);
876 * Unmaps page from the user space (if it is mapped).
878 * \see cl_page_unmap()
879 * \see vvp_page_unmap()
881 int (*cpo_unmap)(const struct lu_env *env,
882 const struct cl_page_slice *slice, struct cl_io *io);
884 * Checks whether underlying VM page is locked (in the suitable
885 * sense). Used for assertions.
887 * \retval -EBUSY: page is protected by a lock of a given mode;
888 * \retval -ENODATA: page is not protected by a lock;
889 * \retval 0: this layer cannot decide. (Should never happen.)
891 int (*cpo_is_vmlocked)(const struct lu_env *env,
892 const struct cl_page_slice *slice);
898 * Called when page is truncated from the object. Optional.
900 * \see cl_page_discard()
901 * \see vvp_page_discard(), osc_page_discard()
903 void (*cpo_discard)(const struct lu_env *env,
904 const struct cl_page_slice *slice,
907 * Called when page is removed from the cache, and is about to being
908 * destroyed. Optional.
910 * \see cl_page_delete()
911 * \see vvp_page_delete(), osc_page_delete()
913 void (*cpo_delete)(const struct lu_env *env,
914 const struct cl_page_slice *slice);
915 /** Destructor. Frees resources and slice itself. */
916 void (*cpo_fini)(const struct lu_env *env,
917 struct cl_page_slice *slice);
920 * Checks whether the page is protected by a cl_lock. This is a
921 * per-layer method, because certain layers have ways to check for the
922 * lock much more efficiently than through the generic locks scan, or
923 * implement locking mechanisms separate from cl_lock, e.g.,
924 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
925 * being canceled, or scheduled for cancellation as soon as the last
926 * user goes away, too.
928 * \retval -EBUSY: page is protected by a lock of a given mode;
929 * \retval -ENODATA: page is not protected by a lock;
930 * \retval 0: this layer cannot decide.
932 * \see cl_page_is_under_lock()
934 int (*cpo_is_under_lock)(const struct lu_env *env,
935 const struct cl_page_slice *slice,
939 * Optional debugging helper. Prints given page slice.
941 * \see cl_page_print()
943 int (*cpo_print)(const struct lu_env *env,
944 const struct cl_page_slice *slice,
945 void *cookie, lu_printer_t p);
949 * Transfer methods. See comment on cl_req for a description of
950 * transfer formation and life-cycle.
955 * Request type dependent vector of operations.
957 * Transfer operations depend on transfer mode (cl_req_type). To avoid
958 * passing transfer mode to each and every of these methods, and to
959 * avoid branching on request type inside of the methods, separate
960 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
961 * provided. That is, method invocation usually looks like
963 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
967 * Called when a page is submitted for a transfer as a part of
970 * \return 0 : page is eligible for submission;
971 * \return -EALREADY : skip this page;
972 * \return -ve : error.
974 * \see cl_page_prep()
976 int (*cpo_prep)(const struct lu_env *env,
977 const struct cl_page_slice *slice,
980 * Completion handler. This is guaranteed to be eventually
981 * fired after cl_page_operations::cpo_prep() or
982 * cl_page_operations::cpo_make_ready() call.
984 * This method can be called in a non-blocking context. It is
985 * guaranteed however, that the page involved and its object
986 * are pinned in memory (and, hence, calling cl_page_put() is
989 * \see cl_page_completion()
991 void (*cpo_completion)(const struct lu_env *env,
992 const struct cl_page_slice *slice,
995 * Called when cached page is about to be added to the
996 * cl_req as a part of req formation.
998 * \return 0 : proceed with this page;
999 * \return -EAGAIN : skip this page;
1000 * \return -ve : error.
1002 * \see cl_page_make_ready()
1004 int (*cpo_make_ready)(const struct lu_env *env,
1005 const struct cl_page_slice *slice);
1007 * Announce that this page is to be written out
1008 * opportunistically, that is, page is dirty, it is not
1009 * necessary to start write-out transfer right now, but
1010 * eventually page has to be written out.
1012 * Main caller of this is the write path (see
1013 * vvp_io_commit_write()), using this method to build a
1014 * "transfer cache" from which large transfers are then
1015 * constructed by the req-formation engine.
1017 * \todo XXX it would make sense to add page-age tracking
1018 * semantics here, and to oblige the req-formation engine to
1019 * send the page out not later than it is too old.
1021 * \see cl_page_cache_add()
1023 int (*cpo_cache_add)(const struct lu_env *env,
1024 const struct cl_page_slice *slice,
1028 * Tell transfer engine that only [to, from] part of a page should be
1031 * This is used for immediate transfers.
1033 * \todo XXX this is not very good interface. It would be much better
1034 * if all transfer parameters were supplied as arguments to
1035 * cl_io_operations::cio_submit() call, but it is not clear how to do
1036 * this for page queues.
1038 * \see cl_page_clip()
1040 void (*cpo_clip)(const struct lu_env *env,
1041 const struct cl_page_slice *slice,
1044 * \pre the page was queued for transferring.
1045 * \post page is removed from client's pending list, or -EBUSY
1046 * is returned if it has already been in transferring.
1048 * This is one of seldom page operation which is:
1049 * 0. called from top level;
1050 * 1. don't have vmpage locked;
1051 * 2. every layer should synchronize execution of its ->cpo_cancel()
1052 * with completion handlers. Osc uses client obd lock for this
1053 * purpose. Based on there is no vvp_page_cancel and
1054 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1056 * \see osc_page_cancel().
1058 int (*cpo_cancel)(const struct lu_env *env,
1059 const struct cl_page_slice *slice);
1064 * Helper macro, dumping detailed information about \a page into a log.
1066 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1068 static DECLARE_LU_CDEBUG_PRINT_INFO(__info, mask); \
1070 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1071 cl_page_print(env, &__info, lu_cdebug_printer, page); \
1072 CDEBUG(mask, format , ## __VA_ARGS__); \
1077 * Helper macro, dumping shorter information about \a page into a log.
1079 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1081 static DECLARE_LU_CDEBUG_PRINT_INFO(__info, mask); \
1083 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1084 cl_page_header_print(env, &__info, lu_cdebug_printer, page); \
1085 CDEBUG(mask, format , ## __VA_ARGS__); \
1091 /** \addtogroup cl_lock cl_lock
1095 * Extent locking on the client.
1099 * The locking model of the new client code is built around
1103 * data-type representing an extent lock on a regular file. cl_lock is a
1104 * layered object (much like cl_object and cl_page), it consists of a header
1105 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1106 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1108 * All locks for a given object are linked into cl_object_header::coh_locks
1109 * list (protected by cl_object_header::coh_lock_guard spin-lock) through
1110 * cl_lock::cll_linkage. Currently this list is not sorted in any way. We can
1111 * sort it in starting lock offset, or use altogether different data structure
1114 * Typical cl_lock consists of the two layers:
1116 * - vvp_lock (vvp specific data), and
1117 * - lov_lock (lov specific data).
1119 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1120 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1122 * - lovsub_lock, and
1125 * Each sub-lock is associated with a cl_object (representing stripe
1126 * sub-object or the file to which top-level cl_lock is associated to), and is
1127 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1128 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1129 * is different from cl_page, that doesn't fan out (there is usually exactly
1130 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1131 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1135 * cl_lock is reference counted. When reference counter drops to 0, lock is
1136 * placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING
1137 * lock is destroyed when last reference is released. Referencing between
1138 * top-lock and its sub-locks is described in the lov documentation module.
1142 * Also, cl_lock is a state machine. This requires some clarification. One of
1143 * the goals of client IO re-write was to make IO path non-blocking, or at
1144 * least to make it easier to make it non-blocking in the future. Here
1145 * `non-blocking' means that when a system call (read, write, truncate)
1146 * reaches a situation where it has to wait for a communication with the
1147 * server, it should --instead of waiting-- remember its current state and
1148 * switch to some other work. E.g,. instead of waiting for a lock enqueue,
1149 * client should proceed doing IO on the next stripe, etc. Obviously this is
1150 * rather radical redesign, and it is not planned to be fully implemented at
1151 * this time, instead we are putting some infrastructure in place, that would
1152 * make it easier to do asynchronous non-blocking IO easier in the
1153 * future. Specifically, where old locking code goes to sleep (waiting for
1154 * enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When
1155 * enqueue reply comes, its completion handler signals that lock state-machine
1156 * is ready to transit to the next state. There is some generic code in
1157 * cl_lock.c that sleeps, waiting for these signals. As a result, for users of
1158 * this cl_lock.c code, it looks like locking is done in normal blocking
1159 * fashion, and it the same time it is possible to switch to the non-blocking
1160 * locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c
1163 * For a description of state machine states and transitions see enum
1166 * There are two ways to restrict a set of states which lock might move to:
1168 * - placing a "hold" on a lock guarantees that lock will not be moved
1169 * into cl_lock_state::CLS_FREEING state until hold is released. Hold
1170 * can be only acquired on a lock that is not in
1171 * cl_lock_state::CLS_FREEING. All holds on a lock are counted in
1172 * cl_lock::cll_holds. Hold protects lock from cancellation and
1173 * destruction. Requests to cancel and destroy a lock on hold will be
1174 * recorded, but only honored when last hold on a lock is released;
1176 * - placing a "user" on a lock guarantees that lock will not leave
1177 * cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING,
1178 * cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of
1179 * states, once it enters this set. That is, if a user is added onto a
1180 * lock in a state not from this set, it doesn't immediately enforce
1181 * lock to move to this set, but once lock enters this set it will
1182 * remain there until all users are removed. Lock users are counted in
1183 * cl_lock::cll_users.
1185 * User is used to assure that lock is not canceled or destroyed while
1186 * it is being enqueued, or actively used by some IO.
1188 * Currently, a user always comes with a hold (cl_lock_invariant()
1189 * checks that a number of holds is not less than a number of users).
1193 * This is how lock state-machine operates. struct cl_lock contains a mutex
1194 * cl_lock::cll_guard that protects struct fields.
1196 * - mutex is taken, and cl_lock::cll_state is examined.
1198 * - for every state there are possible target states where lock can move
1199 * into. They are tried in order. Attempts to move into next state are
1200 * done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try().
1202 * - if the transition can be performed immediately, state is changed,
1203 * and mutex is released.
1205 * - if the transition requires blocking, _try() function returns
1206 * cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to
1207 * sleep, waiting for possibility of lock state change. It is woken
1208 * up when some event occurs, that makes lock state change possible
1209 * (e.g., the reception of the reply from the server), and repeats
1212 * Top-lock and sub-lock has separate mutexes and the latter has to be taken
1213 * first to avoid dead-lock.
1215 * To see an example of interaction of all these issues, take a look at the
1216 * lov_cl.c:lov_lock_enqueue() function. It is called as a part of
1217 * cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by
1218 * advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note
1219 * also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It
1220 * also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be
1221 * done in parallel, rather than one after another (this is used for glimpse
1222 * locks, that cannot dead-lock).
1224 * INTERFACE AND USAGE
1226 * struct cl_lock_operations provide a number of call-backs that are invoked
1227 * when events of interest occurs. Layers can intercept and handle glimpse,
1228 * blocking, cancel ASTs and a reception of the reply from the server.
1230 * One important difference with the old client locking model is that new
1231 * client has a representation for the top-lock, whereas in the old code only
1232 * sub-locks existed as real data structures and file-level locks are
1233 * represented by "request sets" that are created and destroyed on each and
1234 * every lock creation.
1236 * Top-locks are cached, and can be found in the cache by the system calls. It
1237 * is possible that top-lock is in cache, but some of its sub-locks were
1238 * canceled and destroyed. In that case top-lock has to be enqueued again
1239 * before it can be used.
1241 * Overall process of the locking during IO operation is as following:
1243 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1244 * is called on each layer. Responsibility of this method is to add locks,
1245 * needed by a given layer into cl_io.ci_lockset.
1247 * - once locks for all layers were collected, they are sorted to avoid
1248 * dead-locks (cl_io_locks_sort()), and enqueued.
1250 * - when all locks are acquired, IO is performed;
1252 * - locks are released into cache.
1254 * Striping introduces major additional complexity into locking. The
1255 * fundamental problem is that it is generally unsafe to actively use (hold)
1256 * two locks on the different OST servers at the same time, as this introduces
1257 * inter-server dependency and can lead to cascading evictions.
1259 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1260 * that no multi-stripe locks are taken (note that this design abandons POSIX
1261 * read/write semantics). Such pieces ideally can be executed concurrently. At
1262 * the same time, certain types of IO cannot be sub-divived, without
1263 * sacrificing correctness. This includes:
1265 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1268 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1270 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1271 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1272 * has to be held together with the usual lock on [offset, offset + count].
1274 * As multi-stripe locks have to be allowed, it makes sense to cache them, so
1275 * that, for example, a sequence of O_APPEND writes can proceed quickly
1276 * without going down to the individual stripes to do lock matching. On the
1277 * other hand, multi-stripe locks shouldn't be used by normal read/write
1278 * calls. To achieve this, every layer can implement ->clo_fits_into() method,
1279 * that is called by lock matching code (cl_lock_lookup()), and that can be
1280 * used to selectively disable matching of certain locks for certain IOs. For
1281 * exmaple, lov layer implements lov_lock_fits_into() that allow multi-stripe
1282 * locks to be matched only for truncates and O_APPEND writes.
1284 * Interaction with DLM
1286 * In the expected setup, cl_lock is ultimately backed up by a collection of
1287 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1288 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1289 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1290 * description of interaction with DLM.
1296 struct cl_lock_descr {
1297 /** Object this lock is granted for. */
1298 struct cl_object *cld_obj;
1299 /** Index of the first page protected by this lock. */
1301 /** Index of the last page (inclusive) protected by this lock. */
1303 /** Group ID, for group lock */
1306 enum cl_lock_mode cld_mode;
1308 * flags to enqueue lock. A combination of bit-flags from
1309 * enum cl_enq_flags.
1311 __u32 cld_enq_flags;
1314 #define DDESCR "%s(%d):[%lu, %lu]"
1315 #define PDESCR(descr) \
1316 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1317 (descr)->cld_start, (descr)->cld_end
1319 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1322 * Lock state-machine states.
1327 * Possible state transitions:
1329 * +------------------>NEW
1331 * | | cl_enqueue_try()
1333 * | cl_unuse_try() V
1334 * | +--------------QUEUING (*)
1336 * | | | cl_enqueue_try()
1338 * | | cl_unuse_try() V
1339 * sub-lock | +-------------ENQUEUED (*)
1341 * | | | cl_wait_try()
1346 * | | HELD<---------+
1348 * | | | | cl_use_try()
1349 * | | cl_unuse_try() | |
1352 * | +------------>INTRANSIT (D) <--+
1354 * | cl_unuse_try() | | cached lock found
1355 * | | | cl_use_try()
1358 * +------------------CACHED---------+
1367 * In states marked with (*) transition to the same state (i.e., a loop
1368 * in the diagram) is possible.
1370 * (R) is the point where Receive call-back is invoked: it allows layers
1371 * to handle arrival of lock reply.
1373 * (C) is the point where Cancellation call-back is invoked.
1375 * (D) is the transit state which means the lock is changing.
1377 * Transition to FREEING state is possible from any other state in the
1378 * diagram in case of unrecoverable error.
1382 * These states are for individual cl_lock object. Top-lock and its sub-locks
1383 * can be in the different states. Another way to say this is that we have
1384 * nested state-machines.
1386 * Separate QUEUING and ENQUEUED states are needed to support non-blocking
1387 * operation for locks with multiple sub-locks. Imagine lock on a file F, that
1388 * intersects 3 stripes S0, S1, and S2. To enqueue F client has to send
1389 * enqueue to S0, wait for its completion, then send enqueue for S1, wait for
1390 * its completion and at last enqueue lock for S2, and wait for its
1391 * completion. In that case, top-lock is in QUEUING state while S0, S1 are
1392 * handled, and is in ENQUEUED state after enqueue to S2 has been sent (note
1393 * that in this case, sub-locks move from state to state, and top-lock remains
1394 * in the same state).
1396 enum cl_lock_state {
1398 * Lock that wasn't yet enqueued
1402 * Enqueue is in progress, blocking for some intermediate interaction
1403 * with the other side.
1407 * Lock is fully enqueued, waiting for server to reply when it is
1412 * Lock granted, actively used by some IO.
1416 * This state is used to mark the lock is being used, or unused.
1417 * We need this state because the lock may have several sublocks,
1418 * so it's impossible to have an atomic way to bring all sublocks
1419 * into CLS_HELD state at use case, or all sublocks to CLS_CACHED
1421 * If a thread is referring to a lock, and it sees the lock is in this
1422 * state, it must wait for the lock.
1423 * See state diagram for details.
1427 * Lock granted, not used.
1431 * Lock is being destroyed.
1437 enum cl_lock_flags {
1439 * lock has been cancelled. This flag is never cleared once set (by
1440 * cl_lock_cancel0()).
1442 CLF_CANCELLED = 1 << 0,
1443 /** cancellation is pending for this lock. */
1444 CLF_CANCELPEND = 1 << 1,
1445 /** destruction is pending for this lock. */
1452 * Lock closure is a collection of locks (both top-locks and sub-locks) that
1453 * might be updated in a result of an operation on a certain lock (which lock
1454 * this is a closure of).
1456 * Closures are needed to guarantee dead-lock freedom in the presence of
1458 * - nested state-machines (top-lock state-machine composed of sub-lock
1459 * state-machines), and
1461 * - shared sub-locks.
1463 * Specifically, many operations, such as lock enqueue, wait, unlock,
1464 * etc. start from a top-lock, and then operate on a sub-locks of this
1465 * top-lock, holding a top-lock mutex. When sub-lock state changes as a result
1466 * of such operation, this change has to be propagated to all top-locks that
1467 * share this sub-lock. Obviously, no natural lock ordering (e.g.,
1468 * top-to-bottom or bottom-to-top) captures this scenario, so try-locking has
1469 * to be used. Lock closure systematizes this try-and-repeat logic.
1471 struct cl_lock_closure {
1473 * Lock that is mutexed when closure construction is started. When
1474 * closure in is `wait' mode (cl_lock_closure::clc_wait), mutex on
1475 * origin is released before waiting.
1477 struct cl_lock *clc_origin;
1479 * List of enclosed locks, so far. Locks are linked here through
1480 * cl_lock::cll_inclosure.
1482 cfs_list_t clc_list;
1484 * True iff closure is in a `wait' mode. This determines what
1485 * cl_lock_enclosure() does when a lock L to be added to the closure
1486 * is currently mutexed by some other thread.
1488 * If cl_lock_closure::clc_wait is not set, then closure construction
1489 * fails with CLO_REPEAT immediately.
1491 * In wait mode, cl_lock_enclosure() waits until next attempt to build
1492 * a closure might succeed. To this end it releases an origin mutex
1493 * (cl_lock_closure::clc_origin), that has to be the only lock mutex
1494 * owned by the current thread, and then waits on L mutex (by grabbing
1495 * it and immediately releasing), before returning CLO_REPEAT to the
1499 /** Number of locks in the closure. */
1504 * Layered client lock.
1507 /** Reference counter. */
1508 cfs_atomic_t cll_ref;
1509 /** List of slices. Immutable after creation. */
1510 cfs_list_t cll_layers;
1512 * Linkage into cl_lock::cll_descr::cld_obj::coh_locks list. Protected
1513 * by cl_lock::cll_descr::cld_obj::coh_lock_guard.
1515 cfs_list_t cll_linkage;
1517 * Parameters of this lock. Protected by
1518 * cl_lock::cll_descr::cld_obj::coh_lock_guard nested within
1519 * cl_lock::cll_guard. Modified only on lock creation and in
1522 struct cl_lock_descr cll_descr;
1523 /** Protected by cl_lock::cll_guard. */
1524 enum cl_lock_state cll_state;
1525 /** signals state changes. */
1528 * Recursive lock, most fields in cl_lock{} are protected by this.
1530 * Locking rules: this mutex is never held across network
1531 * communication, except when lock is being canceled.
1533 * Lock ordering: a mutex of a sub-lock is taken first, then a mutex
1534 * on a top-lock. Other direction is implemented through a
1535 * try-lock-repeat loop. Mutices of unrelated locks can be taken only
1538 * \see osc_lock_enqueue_wait(), lov_lock_cancel(), lov_sublock_wait().
1540 cfs_mutex_t cll_guard;
1541 cfs_task_t *cll_guarder;
1545 * the owner for INTRANSIT state
1547 cfs_task_t *cll_intransit_owner;
1550 * Number of holds on a lock. A hold prevents a lock from being
1551 * canceled and destroyed. Protected by cl_lock::cll_guard.
1553 * \see cl_lock_hold(), cl_lock_unhold(), cl_lock_release()
1557 * Number of lock users. Valid in cl_lock_state::CLS_HELD state
1558 * only. Lock user pins lock in CLS_HELD state. Protected by
1559 * cl_lock::cll_guard.
1561 * \see cl_wait(), cl_unuse().
1565 * Flag bit-mask. Values from enum cl_lock_flags. Updates are
1566 * protected by cl_lock::cll_guard.
1568 unsigned long cll_flags;
1570 * A linkage into a list of locks in a closure.
1572 * \see cl_lock_closure
1574 cfs_list_t cll_inclosure;
1576 * Confict lock at queuing time.
1578 struct cl_lock *cll_conflict;
1580 * A list of references to this lock, for debugging.
1582 struct lu_ref cll_reference;
1584 * A list of holds on this lock, for debugging.
1586 struct lu_ref cll_holders;
1588 * A reference for cl_lock::cll_descr::cld_obj. For debugging.
1590 struct lu_ref_link *cll_obj_ref;
1591 #ifdef CONFIG_LOCKDEP
1592 /* "dep_map" name is assumed by lockdep.h macros. */
1593 struct lockdep_map dep_map;
1598 * Per-layer part of cl_lock
1600 * \see ccc_lock, lov_lock, lovsub_lock, osc_lock
1602 struct cl_lock_slice {
1603 struct cl_lock *cls_lock;
1604 /** Object slice corresponding to this lock slice. Immutable after
1606 struct cl_object *cls_obj;
1607 const struct cl_lock_operations *cls_ops;
1608 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1609 cfs_list_t cls_linkage;
1613 * Possible (non-error) return values of ->clo_{enqueue,wait,unlock}().
1615 * NOTE: lov_subresult() depends on ordering here.
1617 enum cl_lock_transition {
1618 /** operation cannot be completed immediately. Wait for state change. */
1620 /** operation had to release lock mutex, restart. */
1626 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1628 struct cl_lock_operations {
1630 * \name statemachine
1632 * State machine transitions. These 3 methods are called to transfer
1633 * lock from one state to another, as described in the commentary
1634 * above enum #cl_lock_state.
1636 * \retval 0 this layer has nothing more to do to before
1637 * transition to the target state happens;
1639 * \retval CLO_REPEAT method had to release and re-acquire cl_lock
1640 * mutex, repeat invocation of transition method
1641 * across all layers;
1643 * \retval CLO_WAIT this layer cannot move to the target state
1644 * immediately, as it has to wait for certain event
1645 * (e.g., the communication with the server). It
1646 * is guaranteed, that when the state transfer
1647 * becomes possible, cl_lock::cll_wq wait-queue
1648 * is signaled. Caller can wait for this event by
1649 * calling cl_lock_state_wait();
1651 * \retval -ve failure, abort state transition, move the lock
1652 * into cl_lock_state::CLS_FREEING state, and set
1653 * cl_lock::cll_error.
1655 * Once all layers voted to agree to transition (by returning 0), lock
1656 * is moved into corresponding target state. All state transition
1657 * methods are optional.
1661 * Attempts to enqueue the lock. Called top-to-bottom.
1663 * \see ccc_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1664 * \see osc_lock_enqueue()
1666 int (*clo_enqueue)(const struct lu_env *env,
1667 const struct cl_lock_slice *slice,
1668 struct cl_io *io, __u32 enqflags);
1670 * Attempts to wait for enqueue result. Called top-to-bottom.
1672 * \see ccc_lock_wait(), lov_lock_wait(), osc_lock_wait()
1674 int (*clo_wait)(const struct lu_env *env,
1675 const struct cl_lock_slice *slice);
1677 * Attempts to unlock the lock. Called bottom-to-top. In addition to
1678 * usual return values of lock state-machine methods, this can return
1679 * -ESTALE to indicate that lock cannot be returned to the cache, and
1680 * has to be re-initialized.
1681 * unuse is a one-shot operation, so it must NOT return CLO_WAIT.
1683 * \see ccc_lock_unuse(), lov_lock_unuse(), osc_lock_unuse()
1685 int (*clo_unuse)(const struct lu_env *env,
1686 const struct cl_lock_slice *slice);
1688 * Notifies layer that cached lock is started being used.
1690 * \pre lock->cll_state == CLS_CACHED
1692 * \see lov_lock_use(), osc_lock_use()
1694 int (*clo_use)(const struct lu_env *env,
1695 const struct cl_lock_slice *slice);
1696 /** @} statemachine */
1698 * A method invoked when lock state is changed (as a result of state
1699 * transition). This is used, for example, to track when the state of
1700 * a sub-lock changes, to propagate this change to the corresponding
1701 * top-lock. Optional
1703 * \see lovsub_lock_state()
1705 void (*clo_state)(const struct lu_env *env,
1706 const struct cl_lock_slice *slice,
1707 enum cl_lock_state st);
1709 * Returns true, iff given lock is suitable for the given io, idea
1710 * being, that there are certain "unsafe" locks, e.g., ones acquired
1711 * for O_APPEND writes, that we don't want to re-use for a normal
1712 * write, to avoid the danger of cascading evictions. Optional. Runs
1713 * under cl_object_header::coh_lock_guard.
1715 * XXX this should take more information about lock needed by
1716 * io. Probably lock description or something similar.
1718 * \see lov_fits_into()
1720 int (*clo_fits_into)(const struct lu_env *env,
1721 const struct cl_lock_slice *slice,
1722 const struct cl_lock_descr *need,
1723 const struct cl_io *io);
1726 * Asynchronous System Traps. All of then are optional, all are
1727 * executed bottom-to-top.
1732 * Cancellation callback. Cancel a lock voluntarily, or under
1733 * the request of server.
1735 void (*clo_cancel)(const struct lu_env *env,
1736 const struct cl_lock_slice *slice);
1738 * Lock weighting ast. Executed to estimate how precious this lock
1739 * is. The sum of results across all layers is used to determine
1740 * whether lock worth keeping in cache given present memory usage.
1742 * \see osc_lock_weigh(), vvp_lock_weigh(), lovsub_lock_weigh().
1744 unsigned long (*clo_weigh)(const struct lu_env *env,
1745 const struct cl_lock_slice *slice);
1749 * \see lovsub_lock_closure()
1751 int (*clo_closure)(const struct lu_env *env,
1752 const struct cl_lock_slice *slice,
1753 struct cl_lock_closure *closure);
1755 * Executed bottom-to-top when lock description changes (e.g., as a
1756 * result of server granting more generous lock than was requested).
1758 * \see lovsub_lock_modify()
1760 int (*clo_modify)(const struct lu_env *env,
1761 const struct cl_lock_slice *slice,
1762 const struct cl_lock_descr *updated);
1764 * Notifies layers (bottom-to-top) that lock is going to be
1765 * destroyed. Responsibility of layers is to prevent new references on
1766 * this lock from being acquired once this method returns.
1768 * This can be called multiple times due to the races.
1770 * \see cl_lock_delete()
1771 * \see osc_lock_delete(), lovsub_lock_delete()
1773 void (*clo_delete)(const struct lu_env *env,
1774 const struct cl_lock_slice *slice);
1776 * Destructor. Frees resources and the slice.
1778 * \see ccc_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1779 * \see osc_lock_fini()
1781 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1783 * Optional debugging helper. Prints given lock slice.
1785 int (*clo_print)(const struct lu_env *env,
1786 void *cookie, lu_printer_t p,
1787 const struct cl_lock_slice *slice);
1790 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1792 static DECLARE_LU_CDEBUG_PRINT_INFO(__info, mask); \
1794 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1795 cl_lock_print(env, &__info, lu_cdebug_printer, lock); \
1796 CDEBUG(mask, format , ## __VA_ARGS__); \
1802 /** \addtogroup cl_page_list cl_page_list
1803 * Page list used to perform collective operations on a group of pages.
1805 * Pages are added to the list one by one. cl_page_list acquires a reference
1806 * for every page in it. Page list is used to perform collective operations on
1809 * - submit pages for an immediate transfer,
1811 * - own pages on behalf of certain io (waiting for each page in turn),
1815 * When list is finalized, it releases references on all pages it still has.
1817 * \todo XXX concurrency control.
1821 struct cl_page_list {
1823 cfs_list_t pl_pages;
1824 cfs_task_t *pl_owner;
1828 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1829 * contains an incoming page list and an outgoing page list.
1832 struct cl_page_list c2_qin;
1833 struct cl_page_list c2_qout;
1836 /** @} cl_page_list */
1838 /** \addtogroup cl_io cl_io
1843 * cl_io represents a high level I/O activity like
1844 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1847 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1848 * important distinction. We want to minimize number of calls to the allocator
1849 * in the fast path, e.g., in the case of read(2) when everything is cached:
1850 * client already owns the lock over region being read, and data are cached
1851 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1852 * per-layer io state is stored in the session, associated with the io, see
1853 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1854 * by using free-lists, see cl_env_get().
1856 * There is a small predefined number of possible io types, enumerated in enum
1859 * cl_io is a state machine, that can be advanced concurrently by the multiple
1860 * threads. It is up to these threads to control the concurrency and,
1861 * specifically, to detect when io is done, and its state can be safely
1864 * For read/write io overall execution plan is as following:
1866 * (0) initialize io state through all layers;
1868 * (1) loop: prepare chunk of work to do
1870 * (2) call all layers to collect locks they need to process current chunk
1872 * (3) sort all locks to avoid dead-locks, and acquire them
1874 * (4) process the chunk: call per-page methods
1875 * (cl_io_operations::cio_read_page() for read,
1876 * cl_io_operations::cio_prepare_write(),
1877 * cl_io_operations::cio_commit_write() for write)
1883 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1884 * address allocation efficiency issues mentioned above), and returns with the
1885 * special error condition from per-page method when current sub-io has to
1886 * block. This causes io loop to be repeated, and lov switches to the next
1887 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1892 /** read system call */
1894 /** write system call */
1896 /** truncate, utime system calls */
1899 * page fault handling
1903 * Miscellaneous io. This is used for occasional io activity that
1904 * doesn't fit into other types. Currently this is used for:
1906 * - cancellation of an extent lock. This io exists as a context
1907 * to write dirty pages from under the lock being canceled back
1910 * - VM induced page write-out. An io context for writing page out
1911 * for memory cleansing;
1913 * - glimpse. An io context to acquire glimpse lock.
1915 * - grouplock. An io context to acquire group lock.
1917 * CIT_MISC io is used simply as a context in which locks and pages
1918 * are manipulated. Such io has no internal "process", that is,
1919 * cl_io_loop() is never called for it.
1926 * States of cl_io state machine
1929 /** Not initialized. */
1933 /** IO iteration started. */
1937 /** Actual IO is in progress. */
1939 /** IO for the current iteration finished. */
1941 /** Locks released. */
1943 /** Iteration completed. */
1945 /** cl_io finalized. */
1949 enum cl_req_priority {
1955 * IO state private for a layer.
1957 * This is usually embedded into layer session data, rather than allocated
1960 * \see vvp_io, lov_io, osc_io, ccc_io
1962 struct cl_io_slice {
1963 struct cl_io *cis_io;
1964 /** corresponding object slice. Immutable after creation. */
1965 struct cl_object *cis_obj;
1966 /** io operations. Immutable after creation. */
1967 const struct cl_io_operations *cis_iop;
1969 * linkage into a list of all slices for a given cl_io, hanging off
1970 * cl_io::ci_layers. Immutable after creation.
1972 cfs_list_t cis_linkage;
1977 * Per-layer io operations.
1978 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1980 struct cl_io_operations {
1982 * Vector of io state transition methods for every io type.
1984 * \see cl_page_operations::io
1988 * Prepare io iteration at a given layer.
1990 * Called top-to-bottom at the beginning of each iteration of
1991 * "io loop" (if it makes sense for this type of io). Here
1992 * layer selects what work it will do during this iteration.
1994 * \see cl_io_operations::cio_iter_fini()
1996 int (*cio_iter_init) (const struct lu_env *env,
1997 const struct cl_io_slice *slice);
1999 * Finalize io iteration.
2001 * Called bottom-to-top at the end of each iteration of "io
2002 * loop". Here layers can decide whether IO has to be
2005 * \see cl_io_operations::cio_iter_init()
2007 void (*cio_iter_fini) (const struct lu_env *env,
2008 const struct cl_io_slice *slice);
2010 * Collect locks for the current iteration of io.
2012 * Called top-to-bottom to collect all locks necessary for
2013 * this iteration. This methods shouldn't actually enqueue
2014 * anything, instead it should post a lock through
2015 * cl_io_lock_add(). Once all locks are collected, they are
2016 * sorted and enqueued in the proper order.
2018 int (*cio_lock) (const struct lu_env *env,
2019 const struct cl_io_slice *slice);
2021 * Finalize unlocking.
2023 * Called bottom-to-top to finish layer specific unlocking
2024 * functionality, after generic code released all locks
2025 * acquired by cl_io_operations::cio_lock().
2027 void (*cio_unlock)(const struct lu_env *env,
2028 const struct cl_io_slice *slice);
2030 * Start io iteration.
2032 * Once all locks are acquired, called top-to-bottom to
2033 * commence actual IO. In the current implementation,
2034 * top-level vvp_io_{read,write}_start() does all the work
2035 * synchronously by calling generic_file_*(), so other layers
2036 * are called when everything is done.
2038 int (*cio_start)(const struct lu_env *env,
2039 const struct cl_io_slice *slice);
2041 * Called top-to-bottom at the end of io loop. Here layer
2042 * might wait for an unfinished asynchronous io.
2044 void (*cio_end) (const struct lu_env *env,
2045 const struct cl_io_slice *slice);
2047 * Called bottom-to-top to notify layers that read/write IO
2048 * iteration finished, with \a nob bytes transferred.
2050 void (*cio_advance)(const struct lu_env *env,
2051 const struct cl_io_slice *slice,
2054 * Called once per io, bottom-to-top to release io resources.
2056 void (*cio_fini) (const struct lu_env *env,
2057 const struct cl_io_slice *slice);
2061 * Submit pages from \a queue->c2_qin for IO, and move
2062 * successfully submitted pages into \a queue->c2_qout. Return
2063 * non-zero if failed to submit even the single page. If
2064 * submission failed after some pages were moved into \a
2065 * queue->c2_qout, completion callback with non-zero ioret is
2068 int (*cio_submit)(const struct lu_env *env,
2069 const struct cl_io_slice *slice,
2070 enum cl_req_type crt,
2071 struct cl_2queue *queue,
2072 enum cl_req_priority priority);
2075 * Read missing page.
2077 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
2078 * method, when it hits not-up-to-date page in the range. Optional.
2080 * \pre io->ci_type == CIT_READ
2082 int (*cio_read_page)(const struct lu_env *env,
2083 const struct cl_io_slice *slice,
2084 const struct cl_page_slice *page);
2086 * Prepare write of a \a page. Called bottom-to-top by a top-level
2087 * cl_io_operations::op[CIT_WRITE]::cio_start() to prepare page for
2088 * get data from user-level buffer.
2090 * \pre io->ci_type == CIT_WRITE
2092 * \see vvp_io_prepare_write(), lov_io_prepare_write(),
2093 * osc_io_prepare_write().
2095 int (*cio_prepare_write)(const struct lu_env *env,
2096 const struct cl_io_slice *slice,
2097 const struct cl_page_slice *page,
2098 unsigned from, unsigned to);
2101 * \pre io->ci_type == CIT_WRITE
2103 * \see vvp_io_commit_write(), lov_io_commit_write(),
2104 * osc_io_commit_write().
2106 int (*cio_commit_write)(const struct lu_env *env,
2107 const struct cl_io_slice *slice,
2108 const struct cl_page_slice *page,
2109 unsigned from, unsigned to);
2111 * Optional debugging helper. Print given io slice.
2113 int (*cio_print)(const struct lu_env *env, void *cookie,
2114 lu_printer_t p, const struct cl_io_slice *slice);
2118 * Flags to lock enqueue procedure.
2123 * instruct server to not block, if conflicting lock is found. Instead
2124 * -EWOULDBLOCK is returned immediately.
2126 CEF_NONBLOCK = 0x00000001,
2128 * take lock asynchronously (out of order), as it cannot
2129 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
2131 CEF_ASYNC = 0x00000002,
2133 * tell the server to instruct (though a flag in the blocking ast) an
2134 * owner of the conflicting lock, that it can drop dirty pages
2135 * protected by this lock, without sending them to the server.
2137 CEF_DISCARD_DATA = 0x00000004,
2139 * tell the sub layers that it must be a `real' lock. This is used for
2140 * mmapped-buffer locks and glimpse locks that must be never converted
2141 * into lockless mode.
2143 * \see vvp_mmap_locks(), cl_glimpse_lock().
2145 CEF_MUST = 0x00000008,
2147 * tell the sub layers that never request a `real' lock. This flag is
2148 * not used currently.
2150 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
2151 * conversion policy: ci_lockreq describes generic information of lock
2152 * requirement for this IO, especially for locks which belong to the
2153 * object doing IO; however, lock itself may have precise requirements
2154 * that are described by the enqueue flags.
2156 CEF_NEVER = 0x00000010,
2158 * mask of enq_flags.
2160 CEF_MASK = 0x0000001f
2164 * Link between lock and io. Intermediate structure is needed, because the
2165 * same lock can be part of multiple io's simultaneously.
2167 struct cl_io_lock_link {
2168 /** linkage into one of cl_lockset lists. */
2169 cfs_list_t cill_linkage;
2170 struct cl_lock_descr cill_descr;
2171 struct cl_lock *cill_lock;
2172 /** optional destructor */
2173 void (*cill_fini)(const struct lu_env *env,
2174 struct cl_io_lock_link *link);
2178 * Lock-set represents a collection of locks, that io needs at a
2179 * time. Generally speaking, client tries to avoid holding multiple locks when
2182 * - holding extent locks over multiple ost's introduces the danger of
2183 * "cascading timeouts";
2185 * - holding multiple locks over the same ost is still dead-lock prone,
2186 * see comment in osc_lock_enqueue(),
2188 * but there are certain situations where this is unavoidable:
2190 * - O_APPEND writes have to take [0, EOF] lock for correctness;
2192 * - truncate has to take [new-size, EOF] lock for correctness;
2194 * - SNS has to take locks across full stripe for correctness;
2196 * - in the case when user level buffer, supplied to {read,write}(file0),
2197 * is a part of a memory mapped lustre file, client has to take a dlm
2198 * locks on file0, and all files that back up the buffer (or a part of
2199 * the buffer, that is being processed in the current chunk, in any
2200 * case, there are situations where at least 2 locks are necessary).
2202 * In such cases we at least try to take locks in the same consistent
2203 * order. To this end, all locks are first collected, then sorted, and then
2207 /** locks to be acquired. */
2208 cfs_list_t cls_todo;
2209 /** locks currently being processed. */
2210 cfs_list_t cls_curr;
2211 /** locks acquired. */
2212 cfs_list_t cls_done;
2216 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
2217 * but 'req' is always to be thought as 'request' :-)
2219 enum cl_io_lock_dmd {
2220 /** Always lock data (e.g., O_APPEND). */
2222 /** Layers are free to decide between local and global locking. */
2224 /** Never lock: there is no cache (e.g., liblustre). */
2228 struct cl_io_rw_common {
2238 * cl_io is shared by all threads participating in this IO (in current
2239 * implementation only one thread advances IO, but parallel IO design and
2240 * concurrent copy_*_user() require multiple threads acting on the same IO. It
2241 * is up to these threads to serialize their activities, including updates to
2242 * mutable cl_io fields.
2245 /** type of this IO. Immutable after creation. */
2246 enum cl_io_type ci_type;
2247 /** current state of cl_io state machine. */
2248 enum cl_io_state ci_state;
2249 /** main object this io is against. Immutable after creation. */
2250 struct cl_object *ci_obj;
2252 * Upper layer io, of which this io is a part of. Immutable after
2255 struct cl_io *ci_parent;
2256 /** List of slices. Immutable after creation. */
2257 cfs_list_t ci_layers;
2258 /** list of locks (to be) acquired by this io. */
2259 struct cl_lockset ci_lockset;
2260 /** lock requirements, this is just a help info for sublayers. */
2261 enum cl_io_lock_dmd ci_lockreq;
2263 * This io has held grouplock, to inform sublayers that
2264 * don't do lockless i/o.
2269 struct cl_io_rw_common rd;
2272 struct cl_io_rw_common wr;
2275 struct cl_io_rw_common ci_rw;
2276 struct cl_setattr_io {
2277 struct ost_lvb sa_attr;
2278 unsigned int sa_valid;
2279 struct obd_capa *sa_capa;
2281 struct cl_fault_io {
2282 /** page index within file. */
2284 /** bytes valid byte on a faulted page. */
2286 /** writable page? */
2288 /** page of an executable? */
2290 /** resulting page */
2291 struct cl_page *ft_page;
2294 struct cl_2queue ci_queue;
2299 * Number of pages owned by this IO. For invariant checking.
2301 unsigned ci_owned_nr;
2306 /** \addtogroup cl_req cl_req
2311 * There are two possible modes of transfer initiation on the client:
2313 * - immediate transfer: this is started when a high level io wants a page
2314 * or a collection of pages to be transferred right away. Examples:
2315 * read-ahead, synchronous read in the case of non-page aligned write,
2316 * page write-out as a part of extent lock cancellation, page write-out
2317 * as a part of memory cleansing. Immediate transfer can be both
2318 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
2320 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
2321 * when io wants to transfer a page to the server some time later, when
2322 * it can be done efficiently. Example: pages dirtied by the write(2)
2325 * In any case, transfer takes place in the form of a cl_req, which is a
2326 * representation for a network RPC.
2328 * Pages queued for an opportunistic transfer are cached until it is decided
2329 * that efficient RPC can be composed of them. This decision is made by "a
2330 * req-formation engine", currently implemented as a part of osc
2331 * layer. Req-formation depends on many factors: the size of the resulting
2332 * RPC, whether or not multi-object RPCs are supported by the server,
2333 * max-rpc-in-flight limitations, size of the dirty cache, etc.
2335 * For the immediate transfer io submits a cl_page_list, that req-formation
2336 * engine slices into cl_req's, possibly adding cached pages to some of
2337 * the resulting req's.
2339 * Whenever a page from cl_page_list is added to a newly constructed req, its
2340 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
2341 * page state is atomically changed from cl_page_state::CPS_OWNED to
2342 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
2343 * is zeroed, and cl_page::cp_req is set to the
2344 * req. cl_page_operations::cpo_prep() method at the particular layer might
2345 * return -EALREADY to indicate that it does not need to submit this page
2346 * at all. This is possible, for example, if page, submitted for read,
2347 * became up-to-date in the meantime; and for write, the page don't have
2348 * dirty bit marked. \see cl_io_submit_rw()
2350 * Whenever a cached page is added to a newly constructed req, its
2351 * cl_page_operations::cpo_make_ready() layer methods are called. At that
2352 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
2353 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
2354 * req. cl_page_operations::cpo_make_ready() method at the particular layer
2355 * might return -EAGAIN to indicate that this page is not eligible for the
2356 * transfer right now.
2360 * Plan is to divide transfers into "priority bands" (indicated when
2361 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
2362 * and allow glueing of cached pages to immediate transfers only within single
2363 * band. This would make high priority transfers (like lock cancellation or
2364 * memory pressure induced write-out) really high priority.
2369 * Per-transfer attributes.
2371 struct cl_req_attr {
2372 /** Generic attributes for the server consumption. */
2373 struct obdo *cra_oa;
2375 struct obd_capa *cra_capa;
2379 * Transfer request operations definable at every layer.
2381 * Concurrency: transfer formation engine synchronizes calls to all transfer
2384 struct cl_req_operations {
2386 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
2387 * complete (all pages are added).
2389 * \see osc_req_prep()
2391 int (*cro_prep)(const struct lu_env *env,
2392 const struct cl_req_slice *slice);
2394 * Called top-to-bottom to fill in \a oa fields. This is called twice
2395 * with different flags, see bug 10150 and osc_build_req().
2397 * \param obj an object from cl_req which attributes are to be set in
2400 * \param oa struct obdo where attributes are placed
2402 * \param flags \a oa fields to be filled.
2404 void (*cro_attr_set)(const struct lu_env *env,
2405 const struct cl_req_slice *slice,
2406 const struct cl_object *obj,
2407 struct cl_req_attr *attr, obd_valid flags);
2409 * Called top-to-bottom from cl_req_completion() to notify layers that
2410 * transfer completed. Has to free all state allocated by
2411 * cl_device_operations::cdo_req_init().
2413 void (*cro_completion)(const struct lu_env *env,
2414 const struct cl_req_slice *slice, int ioret);
2418 * A per-object state that (potentially multi-object) transfer request keeps.
2421 /** object itself */
2422 struct cl_object *ro_obj;
2423 /** reference to cl_req_obj::ro_obj. For debugging. */
2424 struct lu_ref_link *ro_obj_ref;
2425 /* something else? Number of pages for a given object? */
2431 * Transfer requests are not reference counted, because IO sub-system owns
2432 * them exclusively and knows when to free them.
2436 * cl_req is created by cl_req_alloc() that calls
2437 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2438 * state in every layer.
2440 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2441 * contains pages for.
2443 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2444 * called top-to-bottom. At that point layers can modify req, let it pass, or
2445 * deny it completely. This is to support things like SNS that have transfer
2446 * ordering requirements invisible to the individual req-formation engine.
2448 * On transfer completion (or transfer timeout, or failure to initiate the
2449 * transfer of an allocated req), cl_req_operations::cro_completion() method
2450 * is called, after execution of cl_page_operations::cpo_completion() of all
2454 enum cl_req_type crq_type;
2455 /** A list of pages being transfered */
2456 cfs_list_t crq_pages;
2457 /** Number of pages in cl_req::crq_pages */
2458 unsigned crq_nrpages;
2459 /** An array of objects which pages are in ->crq_pages */
2460 struct cl_req_obj *crq_o;
2461 /** Number of elements in cl_req::crq_objs[] */
2462 unsigned crq_nrobjs;
2463 cfs_list_t crq_layers;
2467 * Per-layer state for request.
2469 struct cl_req_slice {
2470 struct cl_req *crs_req;
2471 struct cl_device *crs_dev;
2472 cfs_list_t crs_linkage;
2473 const struct cl_req_operations *crs_ops;
2479 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2481 struct cache_stats {
2482 const char *cs_name;
2483 /** how many entities were created at all */
2484 cfs_atomic_t cs_created;
2485 /** how many cache lookups were performed */
2486 cfs_atomic_t cs_lookup;
2487 /** how many times cache lookup resulted in a hit */
2488 cfs_atomic_t cs_hit;
2489 /** how many entities are in the cache right now */
2490 cfs_atomic_t cs_total;
2491 /** how many entities in the cache are actively used (and cannot be
2492 * evicted) right now */
2493 cfs_atomic_t cs_busy;
2496 /** These are not exported so far */
2497 void cache_stats_init (struct cache_stats *cs, const char *name);
2498 int cache_stats_print(const struct cache_stats *cs,
2499 char *page, int count, int header);
2502 * Client-side site. This represents particular client stack. "Global"
2503 * variables should (directly or indirectly) be added here to allow multiple
2504 * clients to co-exist in the single address space.
2507 struct lu_site cs_lu;
2509 * Statistical counters. Atomics do not scale, something better like
2510 * per-cpu counters is needed.
2512 * These are exported as /proc/fs/lustre/llite/.../site
2514 * When interpreting keep in mind that both sub-locks (and sub-pages)
2515 * and top-locks (and top-pages) are accounted here.
2517 struct cache_stats cs_pages;
2518 struct cache_stats cs_locks;
2519 cfs_atomic_t cs_pages_state[CPS_NR];
2520 cfs_atomic_t cs_locks_state[CLS_NR];
2523 int cl_site_init (struct cl_site *s, struct cl_device *top);
2524 void cl_site_fini (struct cl_site *s);
2525 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2528 * Output client site statistical counters into a buffer. Suitable for
2529 * ll_rd_*()-style functions.
2531 int cl_site_stats_print(const struct cl_site *s, char *page, int count);
2536 * Type conversion and accessory functions.
2540 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2542 return container_of(site, struct cl_site, cs_lu);
2545 static inline int lu_device_is_cl(const struct lu_device *d)
2547 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2550 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2552 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2553 return container_of0(d, struct cl_device, cd_lu_dev);
2556 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2558 return &d->cd_lu_dev;
2561 static inline struct cl_object *lu2cl(const struct lu_object *o)
2563 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2564 return container_of0(o, struct cl_object, co_lu);
2567 static inline const struct cl_object_conf *
2568 lu2cl_conf(const struct lu_object_conf *conf)
2570 return container_of0(conf, struct cl_object_conf, coc_lu);
2573 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2575 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2578 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2580 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2581 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2584 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2586 return container_of0(h, struct cl_object_header, coh_lu);
2589 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2591 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2595 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2597 return luh2coh(obj->co_lu.lo_header);
2600 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2602 return lu_device_init(&d->cd_lu_dev, t);
2605 static inline void cl_device_fini(struct cl_device *d)
2607 lu_device_fini(&d->cd_lu_dev);
2610 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2611 struct cl_object *obj,
2612 const struct cl_page_operations *ops);
2613 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2614 struct cl_object *obj,
2615 const struct cl_lock_operations *ops);
2616 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2617 struct cl_object *obj, const struct cl_io_operations *ops);
2618 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2619 struct cl_device *dev,
2620 const struct cl_req_operations *ops);
2623 /** \defgroup cl_object cl_object
2625 struct cl_object *cl_object_top (struct cl_object *o);
2626 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2627 const struct lu_fid *fid,
2628 const struct cl_object_conf *c);
2630 int cl_object_header_init(struct cl_object_header *h);
2631 void cl_object_header_fini(struct cl_object_header *h);
2632 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2633 void cl_object_get (struct cl_object *o);
2634 void cl_object_attr_lock (struct cl_object *o);
2635 void cl_object_attr_unlock(struct cl_object *o);
2636 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2637 struct cl_attr *attr);
2638 int cl_object_attr_set (const struct lu_env *env, struct cl_object *obj,
2639 const struct cl_attr *attr, unsigned valid);
2640 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2641 struct ost_lvb *lvb);
2642 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2643 const struct cl_object_conf *conf);
2644 void cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2645 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2646 int cl_object_has_locks (struct cl_object *obj);
2649 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2651 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2653 return cl_object_header(o0) == cl_object_header(o1);
2658 /** \defgroup cl_page cl_page
2666 int cl_page_gang_lookup (const struct lu_env *env,
2667 struct cl_object *obj,
2669 pgoff_t start, pgoff_t end,
2670 struct cl_page_list *plist);
2671 struct cl_page *cl_page_lookup (struct cl_object_header *hdr,
2673 struct cl_page *cl_page_find (const struct lu_env *env,
2674 struct cl_object *obj,
2675 pgoff_t idx, struct page *vmpage,
2676 enum cl_page_type type);
2677 struct cl_page *cl_page_find_sub (const struct lu_env *env,
2678 struct cl_object *obj,
2679 pgoff_t idx, struct page *vmpage,
2680 struct cl_page *parent);
2681 void cl_page_get (struct cl_page *page);
2682 void cl_page_put (const struct lu_env *env,
2683 struct cl_page *page);
2684 void cl_page_print (const struct lu_env *env, void *cookie,
2685 lu_printer_t printer,
2686 const struct cl_page *pg);
2687 void cl_page_header_print(const struct lu_env *env, void *cookie,
2688 lu_printer_t printer,
2689 const struct cl_page *pg);
2690 cfs_page_t *cl_page_vmpage (const struct lu_env *env,
2691 struct cl_page *page);
2692 struct cl_page *cl_vmpage_page (cfs_page_t *vmpage, struct cl_object *obj);
2693 struct cl_page *cl_page_top (struct cl_page *page);
2695 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2696 const struct lu_device_type *dtype);
2701 * Functions dealing with the ownership of page by io.
2705 int cl_page_own (const struct lu_env *env,
2706 struct cl_io *io, struct cl_page *page);
2707 int cl_page_own_try (const struct lu_env *env,
2708 struct cl_io *io, struct cl_page *page);
2709 void cl_page_assume (const struct lu_env *env,
2710 struct cl_io *io, struct cl_page *page);
2711 void cl_page_unassume (const struct lu_env *env,
2712 struct cl_io *io, struct cl_page *pg);
2713 void cl_page_disown (const struct lu_env *env,
2714 struct cl_io *io, struct cl_page *page);
2715 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2722 * Functions dealing with the preparation of a page for a transfer, and
2723 * tracking transfer state.
2726 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2727 struct cl_page *pg, enum cl_req_type crt);
2728 void cl_page_completion (const struct lu_env *env,
2729 struct cl_page *pg, enum cl_req_type crt, int ioret);
2730 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2731 enum cl_req_type crt);
2732 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2733 struct cl_page *pg, enum cl_req_type crt);
2734 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2736 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2742 * \name helper routines
2743 * Functions to discard, delete and export a cl_page.
2746 void cl_page_discard (const struct lu_env *env, struct cl_io *io,
2747 struct cl_page *pg);
2748 void cl_page_delete (const struct lu_env *env, struct cl_page *pg);
2749 int cl_page_unmap (const struct lu_env *env, struct cl_io *io,
2750 struct cl_page *pg);
2751 int cl_page_is_vmlocked (const struct lu_env *env,
2752 const struct cl_page *pg);
2753 void cl_page_export (const struct lu_env *env,
2754 struct cl_page *pg, int uptodate);
2755 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2756 struct cl_page *page);
2757 loff_t cl_offset (const struct cl_object *obj, pgoff_t idx);
2758 pgoff_t cl_index (const struct cl_object *obj, loff_t offset);
2759 int cl_page_size (const struct cl_object *obj);
2760 int cl_pages_prune (const struct lu_env *env, struct cl_object *obj);
2762 void cl_lock_print (const struct lu_env *env, void *cookie,
2763 lu_printer_t printer, const struct cl_lock *lock);
2764 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2765 lu_printer_t printer,
2766 const struct cl_lock_descr *descr);
2771 /** \defgroup cl_lock cl_lock
2774 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2775 const struct cl_lock_descr *need,
2776 const char *scope, const void *source);
2777 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
2778 const struct cl_lock_descr *need,
2779 const char *scope, const void *source);
2780 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2781 const struct cl_lock_descr *need,
2782 const char *scope, const void *source);
2783 struct cl_lock *cl_lock_at_page(const struct lu_env *env, struct cl_object *obj,
2784 struct cl_page *page, struct cl_lock *except,
2785 int pending, int canceld);
2787 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2788 const struct lu_device_type *dtype);
2790 void cl_lock_get (struct cl_lock *lock);
2791 void cl_lock_get_trust (struct cl_lock *lock);
2792 void cl_lock_put (const struct lu_env *env, struct cl_lock *lock);
2793 void cl_lock_hold_add (const struct lu_env *env, struct cl_lock *lock,
2794 const char *scope, const void *source);
2795 void cl_lock_unhold (const struct lu_env *env, struct cl_lock *lock,
2796 const char *scope, const void *source);
2797 void cl_lock_release (const struct lu_env *env, struct cl_lock *lock,
2798 const char *scope, const void *source);
2799 void cl_lock_user_add (const struct lu_env *env, struct cl_lock *lock);
2800 int cl_lock_user_del (const struct lu_env *env, struct cl_lock *lock);
2802 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
2803 struct cl_lock *lock);
2804 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
2805 enum cl_lock_state state);
2806 int cl_lock_is_intransit(struct cl_lock *lock);
2808 int cl_lock_enqueue_wait(const struct lu_env *env, struct cl_lock *lock,
2811 /** \name statemachine statemachine
2812 * Interface to lock state machine consists of 3 parts:
2814 * - "try" functions that attempt to effect a state transition. If state
2815 * transition is not possible right now (e.g., if it has to wait for some
2816 * asynchronous event to occur), these functions return
2817 * cl_lock_transition::CLO_WAIT.
2819 * - "non-try" functions that implement synchronous blocking interface on
2820 * top of non-blocking "try" functions. These functions repeatedly call
2821 * corresponding "try" versions, and if state transition is not possible
2822 * immediately, wait for lock state change.
2824 * - methods from cl_lock_operations, called by "try" functions. Lock can
2825 * be advanced to the target state only when all layers voted that they
2826 * are ready for this transition. "Try" functions call methods under lock
2827 * mutex. If a layer had to release a mutex, it re-acquires it and returns
2828 * cl_lock_transition::CLO_REPEAT, causing "try" function to call all
2831 * TRY NON-TRY METHOD FINAL STATE
2833 * cl_enqueue_try() cl_enqueue() cl_lock_operations::clo_enqueue() CLS_ENQUEUED
2835 * cl_wait_try() cl_wait() cl_lock_operations::clo_wait() CLS_HELD
2837 * cl_unuse_try() cl_unuse() cl_lock_operations::clo_unuse() CLS_CACHED
2839 * cl_use_try() NONE cl_lock_operations::clo_use() CLS_HELD
2843 int cl_enqueue (const struct lu_env *env, struct cl_lock *lock,
2844 struct cl_io *io, __u32 flags);
2845 int cl_wait (const struct lu_env *env, struct cl_lock *lock);
2846 void cl_unuse (const struct lu_env *env, struct cl_lock *lock);
2847 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
2848 struct cl_io *io, __u32 flags);
2849 int cl_unuse_try (const struct lu_env *env, struct cl_lock *lock);
2850 int cl_wait_try (const struct lu_env *env, struct cl_lock *lock);
2851 int cl_use_try (const struct lu_env *env, struct cl_lock *lock, int atomic);
2853 /** @} statemachine */
2855 void cl_lock_signal (const struct lu_env *env, struct cl_lock *lock);
2856 int cl_lock_state_wait (const struct lu_env *env, struct cl_lock *lock);
2857 void cl_lock_state_set (const struct lu_env *env, struct cl_lock *lock,
2858 enum cl_lock_state state);
2859 int cl_queue_match (const cfs_list_t *queue,
2860 const struct cl_lock_descr *need);
2862 void cl_lock_mutex_get (const struct lu_env *env, struct cl_lock *lock);
2863 int cl_lock_mutex_try (const struct lu_env *env, struct cl_lock *lock);
2864 void cl_lock_mutex_put (const struct lu_env *env, struct cl_lock *lock);
2865 int cl_lock_is_mutexed (struct cl_lock *lock);
2866 int cl_lock_nr_mutexed (const struct lu_env *env);
2867 int cl_lock_page_out (const struct lu_env *env, struct cl_lock *lock,
2869 int cl_lock_ext_match (const struct cl_lock_descr *has,
2870 const struct cl_lock_descr *need);
2871 int cl_lock_descr_match(const struct cl_lock_descr *has,
2872 const struct cl_lock_descr *need);
2873 int cl_lock_mode_match (enum cl_lock_mode has, enum cl_lock_mode need);
2874 int cl_lock_modify (const struct lu_env *env, struct cl_lock *lock,
2875 const struct cl_lock_descr *desc);
2877 void cl_lock_closure_init (const struct lu_env *env,
2878 struct cl_lock_closure *closure,
2879 struct cl_lock *origin, int wait);
2880 void cl_lock_closure_fini (struct cl_lock_closure *closure);
2881 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
2882 struct cl_lock_closure *closure);
2883 void cl_lock_disclosure (const struct lu_env *env,
2884 struct cl_lock_closure *closure);
2885 int cl_lock_enclosure (const struct lu_env *env, struct cl_lock *lock,
2886 struct cl_lock_closure *closure);
2888 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2889 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock);
2890 void cl_lock_error (const struct lu_env *env, struct cl_lock *lock, int error);
2891 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int wait);
2893 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock);
2897 /** \defgroup cl_io cl_io
2900 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2901 enum cl_io_type iot, struct cl_object *obj);
2902 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2903 enum cl_io_type iot, struct cl_object *obj);
2904 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2905 enum cl_io_type iot, loff_t pos, size_t count);
2906 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2908 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2909 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2910 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2911 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2912 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2913 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2914 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2915 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2916 struct cl_io_lock_link *link);
2917 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2918 struct cl_lock_descr *descr);
2919 int cl_io_read_page (const struct lu_env *env, struct cl_io *io,
2920 struct cl_page *page);
2921 int cl_io_prepare_write(const struct lu_env *env, struct cl_io *io,
2922 struct cl_page *page, unsigned from, unsigned to);
2923 int cl_io_commit_write (const struct lu_env *env, struct cl_io *io,
2924 struct cl_page *page, unsigned from, unsigned to);
2925 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2926 enum cl_req_type iot, struct cl_2queue *queue,
2927 enum cl_req_priority priority);
2928 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2929 enum cl_req_type iot, struct cl_2queue *queue,
2930 enum cl_req_priority priority, long timeout);
2931 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2933 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2934 struct cl_page_list *queue);
2935 int cl_io_is_going (const struct lu_env *env);
2938 * True, iff \a io is an O_APPEND write(2).
2940 static inline int cl_io_is_append(const struct cl_io *io)
2942 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2946 * True, iff \a io is a truncate(2).
2948 static inline int cl_io_is_trunc(const struct cl_io *io)
2950 return io->ci_type == CIT_SETATTR &&
2951 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2954 struct cl_io *cl_io_top(struct cl_io *io);
2956 void cl_io_print(const struct lu_env *env, void *cookie,
2957 lu_printer_t printer, const struct cl_io *io);
2959 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2961 typeof(foo_io) __foo_io = (foo_io); \
2963 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2964 memset(&__foo_io->base + 1, 0, \
2965 (sizeof *__foo_io) - sizeof __foo_io->base); \
2970 /** \defgroup cl_page_list cl_page_list
2974 * Last page in the page list.
2976 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2978 LASSERT(plist->pl_nr > 0);
2979 return cfs_list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2983 * Iterate over pages in a page list.
2985 #define cl_page_list_for_each(page, list) \
2986 cfs_list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2989 * Iterate over pages in a page list, taking possible removals into account.
2991 #define cl_page_list_for_each_safe(page, temp, list) \
2992 cfs_list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2994 void cl_page_list_init (struct cl_page_list *plist);
2995 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2996 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2997 struct cl_page *page);
2998 void cl_page_list_splice (struct cl_page_list *list,
2999 struct cl_page_list *head);
3000 void cl_page_list_del (const struct lu_env *env,
3001 struct cl_page_list *plist, struct cl_page *page);
3002 void cl_page_list_disown (const struct lu_env *env,
3003 struct cl_io *io, struct cl_page_list *plist);
3004 int cl_page_list_own (const struct lu_env *env,
3005 struct cl_io *io, struct cl_page_list *plist);
3006 void cl_page_list_assume (const struct lu_env *env,
3007 struct cl_io *io, struct cl_page_list *plist);
3008 void cl_page_list_discard(const struct lu_env *env,
3009 struct cl_io *io, struct cl_page_list *plist);
3010 int cl_page_list_unmap (const struct lu_env *env,
3011 struct cl_io *io, struct cl_page_list *plist);
3012 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
3014 void cl_2queue_init (struct cl_2queue *queue);
3015 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
3016 void cl_2queue_disown (const struct lu_env *env,
3017 struct cl_io *io, struct cl_2queue *queue);
3018 void cl_2queue_assume (const struct lu_env *env,
3019 struct cl_io *io, struct cl_2queue *queue);
3020 void cl_2queue_discard (const struct lu_env *env,
3021 struct cl_io *io, struct cl_2queue *queue);
3022 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
3023 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
3025 /** @} cl_page_list */
3027 /** \defgroup cl_req cl_req
3029 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
3030 enum cl_req_type crt, int nr_objects);
3032 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
3033 struct cl_page *page);
3034 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
3035 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
3036 void cl_req_attr_set (const struct lu_env *env, struct cl_req *req,
3037 struct cl_req_attr *attr, obd_valid flags);
3038 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
3040 /** \defgroup cl_sync_io cl_sync_io
3044 * Anchor for synchronous transfer. This is allocated on a stack by thread
3045 * doing synchronous transfer, and a pointer to this structure is set up in
3046 * every page submitted for transfer. Transfer completion routine updates
3047 * anchor and wakes up waiting thread when transfer is complete.
3050 /** number of pages yet to be transferred. */
3051 cfs_atomic_t csi_sync_nr;
3052 /** completion to be signaled when transfer is complete. */
3053 cfs_waitq_t csi_waitq;
3058 void cl_sync_io_init(struct cl_sync_io *anchor, int nrpages);
3059 int cl_sync_io_wait(const struct lu_env *env, struct cl_io *io,
3060 struct cl_page_list *queue, struct cl_sync_io *anchor,
3062 void cl_sync_io_note(struct cl_sync_io *anchor, int ioret);
3064 /** @} cl_sync_io */
3068 /** \defgroup cl_env cl_env
3070 * lu_env handling for a client.
3072 * lu_env is an environment within which lustre code executes. Its major part
3073 * is lu_context---a fast memory allocation mechanism that is used to conserve
3074 * precious kernel stack space. Originally lu_env was designed for a server,
3077 * - there is a (mostly) fixed number of threads, and
3079 * - call chains have no non-lustre portions inserted between lustre code.
3081 * On a client both these assumtpion fails, because every user thread can
3082 * potentially execute lustre code as part of a system call, and lustre calls
3083 * into VFS or MM that call back into lustre.
3085 * To deal with that, cl_env wrapper functions implement the following
3088 * - allocation and destruction of environment is amortized by caching no
3089 * longer used environments instead of destroying them;
3091 * - there is a notion of "current" environment, attached to the kernel
3092 * data structure representing current thread Top-level lustre code
3093 * allocates an environment and makes it current, then calls into
3094 * non-lustre code, that in turn calls lustre back. Low-level lustre
3095 * code thus called can fetch environment created by the top-level code
3096 * and reuse it, avoiding additional environment allocation.
3097 * Right now, three interfaces can attach the cl_env to running thread:
3100 * - cl_env_reexit(cl_env_reenter had to be called priorly)
3102 * \see lu_env, lu_context, lu_context_key
3105 struct cl_env_nest {
3110 struct lu_env *cl_env_peek (int *refcheck);
3111 struct lu_env *cl_env_get (int *refcheck);
3112 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
3113 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
3114 void cl_env_put (struct lu_env *env, int *refcheck);
3115 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
3116 void *cl_env_reenter (void);
3117 void cl_env_reexit (void *cookie);
3118 void cl_env_implant (struct lu_env *env, int *refcheck);
3119 void cl_env_unplant (struct lu_env *env, int *refcheck);
3120 unsigned cl_env_cache_purge(unsigned nr);
3127 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
3128 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
3130 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
3131 struct lu_device_type *ldt,
3132 struct lu_device *next);
3135 #endif /* _LINUX_CL_OBJECT_H */