4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 #ifndef _LUSTRE_CL_OBJECT_H
37 #define _LUSTRE_CL_OBJECT_H
39 /** \defgroup clio clio
41 * Client objects implement io operations and cache pages.
43 * Examples: lov and osc are implementations of cl interface.
45 * Big Theory Statement.
49 * Client implementation is based on the following data-types:
55 * - cl_lock represents an extent lock on an object.
57 * - cl_io represents high-level i/o activity such as whole read/write
58 * system call, or write-out of pages from under the lock being
59 * canceled. cl_io has sub-ios that can be stopped and resumed
60 * independently, thus achieving high degree of transfer
61 * parallelism. Single cl_io can be advanced forward by
62 * the multiple threads (although in the most usual case of
63 * read/write system call it is associated with the single user
64 * thread, that issued the system call).
66 * - cl_req represents a collection of pages for a transfer. cl_req is
67 * constructed by req-forming engine that tries to saturate
68 * transport with large and continuous transfers.
72 * - to avoid confusion high-level I/O operation like read or write system
73 * call is referred to as "an io", whereas low-level I/O operation, like
74 * RPC, is referred to as "a transfer"
76 * - "generic code" means generic (not file system specific) code in the
77 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
78 * is not layer specific.
84 * - cl_object_header::coh_page_guard
87 * See the top comment in cl_object.c for the description of overall locking and
88 * reference-counting design.
90 * See comments below for the description of i/o, page, and dlm-locking
97 * super-class definitions.
99 #include <libcfs/libcfs.h>
100 #include <lu_object.h>
101 #include <linux/atomic.h>
102 #include <linux/mutex.h>
103 #include <linux/radix-tree.h>
104 #include <linux/spinlock.h>
105 #include <linux/wait.h>
110 struct cl_device_operations;
113 struct cl_object_page_operations;
114 struct cl_object_lock_operations;
117 struct cl_page_slice;
119 struct cl_lock_slice;
121 struct cl_lock_operations;
122 struct cl_page_operations;
131 * Operations for each data device in the client stack.
133 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
135 struct cl_device_operations {
137 * Initialize cl_req. This method is called top-to-bottom on all
138 * devices in the stack to get them a chance to allocate layer-private
139 * data, and to attach them to the cl_req by calling
140 * cl_req_slice_add().
142 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
143 * \see ccc_req_init()
145 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
150 * Device in the client stack.
152 * \see vvp_device, lov_device, lovsub_device, osc_device
156 struct lu_device cd_lu_dev;
157 /** Per-layer operation vector. */
158 const struct cl_device_operations *cd_ops;
161 /** \addtogroup cl_object cl_object
164 * "Data attributes" of cl_object. Data attributes can be updated
165 * independently for a sub-object, and top-object's attributes are calculated
166 * from sub-objects' ones.
169 /** Object size, in bytes */
172 * Known minimal size, in bytes.
174 * This is only valid when at least one DLM lock is held.
177 /** Modification time. Measured in seconds since epoch. */
179 /** Access time. Measured in seconds since epoch. */
181 /** Change time. Measured in seconds since epoch. */
184 * Blocks allocated to this cl_object on the server file system.
186 * \todo XXX An interface for block size is needed.
190 * User identifier for quota purposes.
194 * Group identifier for quota purposes.
198 /* nlink of the directory */
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 vvp_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;
257 /** offset of page slice in cl_page buffer */
262 * Description of the client object configuration. This is used for the
263 * creation of a new client object that is identified by a more state than
266 struct cl_object_conf {
268 struct lu_object_conf coc_lu;
271 * Object layout. This is consumed by lov.
273 struct lustre_md *coc_md;
275 * Description of particular stripe location in the
276 * cluster. This is consumed by osc.
278 struct lov_oinfo *coc_oinfo;
281 * VFS inode. This is consumed by vvp.
283 struct inode *coc_inode;
285 * Layout lock handle.
287 struct ldlm_lock *coc_lock;
289 * Operation to handle layout, OBJECT_CONF_XYZ.
295 /** configure layout, set up a new stripe, must be called while
296 * holding layout lock. */
298 /** invalidate the current stripe configuration due to losing
300 OBJECT_CONF_INVALIDATE = 1,
301 /** wait for old layout to go away so that new layout can be
307 * Operations implemented for each cl object layer.
309 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
311 struct cl_object_operations {
313 * Initialize page slice for this layer. Called top-to-bottom through
314 * every object layer when a new cl_page is instantiated. Layer
315 * keeping private per-page data, or requiring its own page operations
316 * vector should allocate these data here, and attach then to the page
317 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
320 * \retval NULL success.
322 * \retval ERR_PTR(errno) failure code.
324 * \retval valid-pointer pointer to already existing referenced page
325 * to be used instead of newly created.
327 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
328 struct cl_page *page, pgoff_t index);
330 * Initialize lock slice for this layer. Called top-to-bottom through
331 * every object layer when a new cl_lock is instantiated. Layer
332 * keeping private per-lock data, or requiring its own lock operations
333 * vector should allocate these data here, and attach then to the lock
334 * by calling cl_lock_slice_add(). Mandatory.
336 int (*coo_lock_init)(const struct lu_env *env,
337 struct cl_object *obj, struct cl_lock *lock,
338 const struct cl_io *io);
340 * Initialize io state for a given layer.
342 * called top-to-bottom once per io existence to initialize io
343 * state. If layer wants to keep some state for this type of io, it
344 * has to embed struct cl_io_slice in lu_env::le_ses, and register
345 * slice with cl_io_slice_add(). It is guaranteed that all threads
346 * participating in this io share the same session.
348 int (*coo_io_init)(const struct lu_env *env,
349 struct cl_object *obj, struct cl_io *io);
351 * Fill portion of \a attr that this layer controls. This method is
352 * called top-to-bottom through all object layers.
354 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
356 * \return 0: to continue
357 * \return +ve: to stop iterating through layers (but 0 is returned
358 * from enclosing cl_object_attr_get())
359 * \return -ve: to signal error
361 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
362 struct cl_attr *attr);
366 * \a valid is a bitmask composed from enum #cl_attr_valid, and
367 * indicating what attributes are to be set.
369 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
371 * \return the same convention as for
372 * cl_object_operations::coo_attr_get() is used.
374 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
375 const struct cl_attr *attr, unsigned valid);
377 * Update object configuration. Called top-to-bottom to modify object
380 * XXX error conditions and handling.
382 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
383 const struct cl_object_conf *conf);
385 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
386 * object. Layers are supposed to fill parts of \a lvb that will be
387 * shipped to the glimpse originator as a glimpse result.
389 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
390 * \see osc_object_glimpse()
392 int (*coo_glimpse)(const struct lu_env *env,
393 const struct cl_object *obj, struct ost_lvb *lvb);
395 * Object prune method. Called when the layout is going to change on
396 * this object, therefore each layer has to clean up their cache,
397 * mainly pages and locks.
399 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
401 * Object getstripe method.
403 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
404 struct lov_user_md __user *lum);
408 * Extended header for client object.
410 struct cl_object_header {
411 /** Standard lu_object_header. cl_object::co_lu::lo_header points
413 struct lu_object_header coh_lu;
416 * Parent object. It is assumed that an object has a well-defined
417 * parent, but not a well-defined child (there may be multiple
418 * sub-objects, for the same top-object). cl_object_header::coh_parent
419 * field allows certain code to be written generically, without
420 * limiting possible cl_object layouts unduly.
422 struct cl_object_header *coh_parent;
424 * Protects consistency between cl_attr of parent object and
425 * attributes of sub-objects, that the former is calculated ("merged")
428 * \todo XXX this can be read/write lock if needed.
430 spinlock_t coh_attr_guard;
432 * Size of cl_page + page slices
434 unsigned short coh_page_bufsize;
436 * Number of objects above this one: 0 for a top-object, 1 for its
439 unsigned char coh_nesting;
443 * Helper macro: iterate over all layers of the object \a obj, assigning every
444 * layer top-to-bottom to \a slice.
446 #define cl_object_for_each(slice, obj) \
447 list_for_each_entry((slice), \
448 &(obj)->co_lu.lo_header->loh_layers,\
452 * Helper macro: iterate over all layers of the object \a obj, assigning every
453 * layer bottom-to-top to \a slice.
455 #define cl_object_for_each_reverse(slice, obj) \
456 list_for_each_entry_reverse((slice), \
457 &(obj)->co_lu.lo_header->loh_layers,\
462 #define CL_PAGE_EOF ((pgoff_t)~0ull)
464 /** \addtogroup cl_page cl_page
468 * Layered client page.
470 * cl_page: represents a portion of a file, cached in the memory. All pages
471 * of the given file are of the same size, and are kept in the radix tree
472 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
473 * of the top-level file object are first class cl_objects, they have their
474 * own radix trees of pages and hence page is implemented as a sequence of
475 * struct cl_pages's, linked into double-linked list through
476 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
477 * corresponding radix tree at the corresponding logical offset.
479 * cl_page is associated with VM page of the hosting environment (struct
480 * page in Linux kernel, for example), struct page. It is assumed, that this
481 * association is implemented by one of cl_page layers (top layer in the
482 * current design) that
484 * - intercepts per-VM-page call-backs made by the environment (e.g.,
487 * - translates state (page flag bits) and locking between lustre and
490 * The association between cl_page and struct page is immutable and
491 * established when cl_page is created.
493 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
494 * this io an exclusive access to this page w.r.t. other io attempts and
495 * various events changing page state (such as transfer completion, or
496 * eviction of the page from the memory). Note, that in general cl_io
497 * cannot be identified with a particular thread, and page ownership is not
498 * exactly equal to the current thread holding a lock on the page. Layer
499 * implementing association between cl_page and struct page has to implement
500 * ownership on top of available synchronization mechanisms.
502 * While lustre client maintains the notion of an page ownership by io,
503 * hosting MM/VM usually has its own page concurrency control
504 * mechanisms. For example, in Linux, page access is synchronized by the
505 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
506 * takes care to acquire and release such locks as necessary around the
507 * calls to the file system methods (->readpage(), ->prepare_write(),
508 * ->commit_write(), etc.). This leads to the situation when there are two
509 * different ways to own a page in the client:
511 * - client code explicitly and voluntary owns the page (cl_page_own());
513 * - VM locks a page and then calls the client, that has "to assume"
514 * the ownership from the VM (cl_page_assume()).
516 * Dual methods to release ownership are cl_page_disown() and
517 * cl_page_unassume().
519 * cl_page is reference counted (cl_page::cp_ref). When reference counter
520 * drops to 0, the page is returned to the cache, unless it is in
521 * cl_page_state::CPS_FREEING state, in which case it is immediately
524 * The general logic guaranteeing the absence of "existential races" for
525 * pages is the following:
527 * - there are fixed known ways for a thread to obtain a new reference
530 * - by doing a lookup in the cl_object radix tree, protected by the
533 * - by starting from VM-locked struct page and following some
534 * hosting environment method (e.g., following ->private pointer in
535 * the case of Linux kernel), see cl_vmpage_page();
537 * - when the page enters cl_page_state::CPS_FREEING state, all these
538 * ways are severed with the proper synchronization
539 * (cl_page_delete());
541 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
544 * - no new references to the page in cl_page_state::CPS_FREEING state
545 * are allowed (checked in cl_page_get()).
547 * Together this guarantees that when last reference to a
548 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
549 * page, as neither references to it can be acquired at that point, nor
552 * cl_page is a state machine. States are enumerated in enum
553 * cl_page_state. Possible state transitions are enumerated in
554 * cl_page_state_set(). State transition process (i.e., actual changing of
555 * cl_page::cp_state field) is protected by the lock on the underlying VM
558 * Linux Kernel implementation.
560 * Binding between cl_page and struct page (which is a typedef for
561 * struct page) is implemented in the vvp layer. cl_page is attached to the
562 * ->private pointer of the struct page, together with the setting of
563 * PG_private bit in page->flags, and acquiring additional reference on the
564 * struct page (much like struct buffer_head, or any similar file system
565 * private data structures).
567 * PG_locked lock is used to implement both ownership and transfer
568 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
569 * states. No additional references are acquired for the duration of the
572 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
573 * write-out is "protected" by the special PG_writeback bit.
577 * States of cl_page. cl_page.c assumes particular order here.
579 * The page state machine is rather crude, as it doesn't recognize finer page
580 * states like "dirty" or "up to date". This is because such states are not
581 * always well defined for the whole stack (see, for example, the
582 * implementation of the read-ahead, that hides page up-to-dateness to track
583 * cache hits accurately). Such sub-states are maintained by the layers that
584 * are interested in them.
588 * Page is in the cache, un-owned. Page leaves cached state in the
591 * - [cl_page_state::CPS_OWNED] io comes across the page and
594 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
595 * req-formation engine decides that it wants to include this page
596 * into an cl_req being constructed, and yanks it from the cache;
598 * - [cl_page_state::CPS_FREEING] VM callback is executed to
599 * evict the page form the memory;
601 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
605 * Page is exclusively owned by some cl_io. Page may end up in this
606 * state as a result of
608 * - io creating new page and immediately owning it;
610 * - [cl_page_state::CPS_CACHED] io finding existing cached page
613 * - [cl_page_state::CPS_OWNED] io finding existing owned page
614 * and waiting for owner to release the page;
616 * Page leaves owned state in the following cases:
618 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
619 * the cache, doing nothing;
621 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
624 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
625 * transfer for this page;
627 * - [cl_page_state::CPS_FREEING] io decides to destroy this
628 * page (e.g., as part of truncate or extent lock cancellation).
630 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
634 * Page is being written out, as a part of a transfer. This state is
635 * entered when req-formation logic decided that it wants this page to
636 * be sent through the wire _now_. Specifically, it means that once
637 * this state is achieved, transfer completion handler (with either
638 * success or failure indication) is guaranteed to be executed against
639 * this page independently of any locks and any scheduling decisions
640 * made by the hosting environment (that effectively means that the
641 * page is never put into cl_page_state::CPS_PAGEOUT state "in
642 * advance". This property is mentioned, because it is important when
643 * reasoning about possible dead-locks in the system). The page can
644 * enter this state as a result of
646 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
647 * write-out of this page, or
649 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
650 * that it has enough dirty pages cached to issue a "good"
653 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
654 * is completed---it is moved into cl_page_state::CPS_CACHED state.
656 * Underlying VM page is locked for the duration of transfer.
658 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
662 * Page is being read in, as a part of a transfer. This is quite
663 * similar to the cl_page_state::CPS_PAGEOUT state, except that
664 * read-in is always "immediate"---there is no such thing a sudden
665 * construction of read cl_req from cached, presumably not up to date,
668 * Underlying VM page is locked for the duration of transfer.
670 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
674 * Page is being destroyed. This state is entered when client decides
675 * that page has to be deleted from its host object, as, e.g., a part
678 * Once this state is reached, there is no way to escape it.
680 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
687 /** Host page, the page is from the host inode which the cl_page
691 /** Transient page, the transient cl_page is used to bind a cl_page
692 * to vmpage which is not belonging to the same object of cl_page.
693 * it is used in DirectIO, lockless IO and liblustre. */
698 * Fields are protected by the lock on struct page, except for atomics and
701 * \invariant Data type invariants are in cl_page_invariant(). Basically:
702 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
703 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
704 * cl_page::cp_owner (when set).
707 /** Reference counter. */
709 /** Transfer error. */
711 /** An object this page is a part of. Immutable after creation. */
712 struct cl_object *cp_obj;
714 struct page *cp_vmpage;
715 /** Linkage of pages within group. Pages must be owned */
716 struct list_head cp_batch;
717 /** List of slices. Immutable after creation. */
718 struct list_head cp_layers;
719 /** Linkage of pages within cl_req. */
720 struct list_head cp_flight;
722 * Page state. This field is const to avoid accidental update, it is
723 * modified only internally within cl_page.c. Protected by a VM lock.
725 const enum cl_page_state cp_state;
727 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
730 enum cl_page_type cp_type;
733 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
734 * by sub-io. Protected by a VM lock.
736 struct cl_io *cp_owner;
738 * Owning IO request in cl_page_state::CPS_PAGEOUT and
739 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
740 * the top-level pages. Protected by a VM lock.
742 struct cl_req *cp_req;
743 /** List of references to this page, for debugging. */
744 struct lu_ref cp_reference;
745 /** Link to an object, for debugging. */
746 struct lu_ref_link cp_obj_ref;
747 /** Link to a queue, for debugging. */
748 struct lu_ref_link cp_queue_ref;
749 /** Assigned if doing a sync_io */
750 struct cl_sync_io *cp_sync_io;
754 * Per-layer part of cl_page.
756 * \see vvp_page, lov_page, osc_page
758 struct cl_page_slice {
759 struct cl_page *cpl_page;
762 * Object slice corresponding to this page slice. Immutable after
765 struct cl_object *cpl_obj;
766 const struct cl_page_operations *cpl_ops;
767 /** Linkage into cl_page::cp_layers. Immutable after creation. */
768 struct list_head cpl_linkage;
772 * Lock mode. For the client extent locks.
784 * Requested transfer type.
794 * Per-layer page operations.
796 * Methods taking an \a io argument are for the activity happening in the
797 * context of given \a io. Page is assumed to be owned by that io, except for
798 * the obvious cases (like cl_page_operations::cpo_own()).
800 * \see vvp_page_ops, lov_page_ops, osc_page_ops
802 struct cl_page_operations {
804 * cl_page<->struct page methods. Only one layer in the stack has to
805 * implement these. Current code assumes that this functionality is
806 * provided by the topmost layer, see cl_page_disown0() as an example.
810 * Called when \a io acquires this page into the exclusive
811 * ownership. When this method returns, it is guaranteed that the is
812 * not owned by other io, and no transfer is going on against
816 * \see vvp_page_own(), lov_page_own()
818 int (*cpo_own)(const struct lu_env *env,
819 const struct cl_page_slice *slice,
820 struct cl_io *io, int nonblock);
821 /** Called when ownership it yielded. Optional.
823 * \see cl_page_disown()
824 * \see vvp_page_disown()
826 void (*cpo_disown)(const struct lu_env *env,
827 const struct cl_page_slice *slice, struct cl_io *io);
829 * Called for a page that is already "owned" by \a io from VM point of
832 * \see cl_page_assume()
833 * \see vvp_page_assume(), lov_page_assume()
835 void (*cpo_assume)(const struct lu_env *env,
836 const struct cl_page_slice *slice, struct cl_io *io);
837 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
838 * bottom-to-top when IO releases a page without actually unlocking
841 * \see cl_page_unassume()
842 * \see vvp_page_unassume()
844 void (*cpo_unassume)(const struct lu_env *env,
845 const struct cl_page_slice *slice,
848 * Announces whether the page contains valid data or not by \a uptodate.
850 * \see cl_page_export()
851 * \see vvp_page_export()
853 void (*cpo_export)(const struct lu_env *env,
854 const struct cl_page_slice *slice, int uptodate);
856 * Checks whether underlying VM page is locked (in the suitable
857 * sense). Used for assertions.
859 * \retval -EBUSY: page is protected by a lock of a given mode;
860 * \retval -ENODATA: page is not protected by a lock;
861 * \retval 0: this layer cannot decide. (Should never happen.)
863 int (*cpo_is_vmlocked)(const struct lu_env *env,
864 const struct cl_page_slice *slice);
870 * Called when page is truncated from the object. Optional.
872 * \see cl_page_discard()
873 * \see vvp_page_discard(), osc_page_discard()
875 void (*cpo_discard)(const struct lu_env *env,
876 const struct cl_page_slice *slice,
879 * Called when page is removed from the cache, and is about to being
880 * destroyed. Optional.
882 * \see cl_page_delete()
883 * \see vvp_page_delete(), osc_page_delete()
885 void (*cpo_delete)(const struct lu_env *env,
886 const struct cl_page_slice *slice);
887 /** Destructor. Frees resources and slice itself. */
888 void (*cpo_fini)(const struct lu_env *env,
889 struct cl_page_slice *slice);
892 * Checks whether the page is protected by a cl_lock. This is a
893 * per-layer method, because certain layers have ways to check for the
894 * lock much more efficiently than through the generic locks scan, or
895 * implement locking mechanisms separate from cl_lock, e.g.,
896 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
897 * being canceled, or scheduled for cancellation as soon as the last
898 * user goes away, too.
900 * \retval -EBUSY: page is protected by a lock of a given mode;
901 * \retval -ENODATA: page is not protected by a lock;
902 * \retval 0: this layer cannot decide.
904 * \see cl_page_is_under_lock()
906 int (*cpo_is_under_lock)(const struct lu_env *env,
907 const struct cl_page_slice *slice,
908 struct cl_io *io, pgoff_t *max);
911 * Optional debugging helper. Prints given page slice.
913 * \see cl_page_print()
915 int (*cpo_print)(const struct lu_env *env,
916 const struct cl_page_slice *slice,
917 void *cookie, lu_printer_t p);
921 * Transfer methods. See comment on cl_req for a description of
922 * transfer formation and life-cycle.
927 * Request type dependent vector of operations.
929 * Transfer operations depend on transfer mode (cl_req_type). To avoid
930 * passing transfer mode to each and every of these methods, and to
931 * avoid branching on request type inside of the methods, separate
932 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
933 * provided. That is, method invocation usually looks like
935 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
939 * Called when a page is submitted for a transfer as a part of
942 * \return 0 : page is eligible for submission;
943 * \return -EALREADY : skip this page;
944 * \return -ve : error.
946 * \see cl_page_prep()
948 int (*cpo_prep)(const struct lu_env *env,
949 const struct cl_page_slice *slice,
952 * Completion handler. This is guaranteed to be eventually
953 * fired after cl_page_operations::cpo_prep() or
954 * cl_page_operations::cpo_make_ready() call.
956 * This method can be called in a non-blocking context. It is
957 * guaranteed however, that the page involved and its object
958 * are pinned in memory (and, hence, calling cl_page_put() is
961 * \see cl_page_completion()
963 void (*cpo_completion)(const struct lu_env *env,
964 const struct cl_page_slice *slice,
967 * Called when cached page is about to be added to the
968 * cl_req as a part of req formation.
970 * \return 0 : proceed with this page;
971 * \return -EAGAIN : skip this page;
972 * \return -ve : error.
974 * \see cl_page_make_ready()
976 int (*cpo_make_ready)(const struct lu_env *env,
977 const struct cl_page_slice *slice);
980 * Tell transfer engine that only [to, from] part of a page should be
983 * This is used for immediate transfers.
985 * \todo XXX this is not very good interface. It would be much better
986 * if all transfer parameters were supplied as arguments to
987 * cl_io_operations::cio_submit() call, but it is not clear how to do
988 * this for page queues.
990 * \see cl_page_clip()
992 void (*cpo_clip)(const struct lu_env *env,
993 const struct cl_page_slice *slice,
996 * \pre the page was queued for transferring.
997 * \post page is removed from client's pending list, or -EBUSY
998 * is returned if it has already been in transferring.
1000 * This is one of seldom page operation which is:
1001 * 0. called from top level;
1002 * 1. don't have vmpage locked;
1003 * 2. every layer should synchronize execution of its ->cpo_cancel()
1004 * with completion handlers. Osc uses client obd lock for this
1005 * purpose. Based on there is no vvp_page_cancel and
1006 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1008 * \see osc_page_cancel().
1010 int (*cpo_cancel)(const struct lu_env *env,
1011 const struct cl_page_slice *slice);
1013 * Write out a page by kernel. This is only called by ll_writepage
1016 * \see cl_page_flush()
1018 int (*cpo_flush)(const struct lu_env *env,
1019 const struct cl_page_slice *slice,
1025 * Helper macro, dumping detailed information about \a page into a log.
1027 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1029 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1030 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1031 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1032 CDEBUG(mask, format , ## __VA_ARGS__); \
1037 * Helper macro, dumping shorter information about \a page into a log.
1039 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1041 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1042 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1043 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1044 CDEBUG(mask, format , ## __VA_ARGS__); \
1048 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1050 LASSERT(page->cp_vmpage != NULL);
1051 return page->cp_vmpage;
1055 * Check if a cl_page is in use.
1057 * Client cache holds a refcount, this refcount will be dropped when
1058 * the page is taken out of cache, see vvp_page_delete().
1060 static inline bool __page_in_use(const struct cl_page *page, int refc)
1062 return (atomic_read(&page->cp_ref) > refc + 1);
1066 * Caller itself holds a refcount of cl_page.
1068 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1070 * Caller doesn't hold a refcount.
1072 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1076 /** \addtogroup cl_lock cl_lock
1080 * Extent locking on the client.
1084 * The locking model of the new client code is built around
1088 * data-type representing an extent lock on a regular file. cl_lock is a
1089 * layered object (much like cl_object and cl_page), it consists of a header
1090 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1091 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1093 * Typical cl_lock consists of the two layers:
1095 * - vvp_lock (vvp specific data), and
1096 * - lov_lock (lov specific data).
1098 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1099 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1101 * - lovsub_lock, and
1104 * Each sub-lock is associated with a cl_object (representing stripe
1105 * sub-object or the file to which top-level cl_lock is associated to), and is
1106 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1107 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1108 * is different from cl_page, that doesn't fan out (there is usually exactly
1109 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1110 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1114 * cl_lock is a cacheless data container for the requirements of locks to
1115 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1118 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1119 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1121 * INTERFACE AND USAGE
1123 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1124 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1125 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1126 * consists of multiple sub cl_locks, each sub locks will be enqueued
1127 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1128 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1131 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1132 * method will be called for each layer to release the resource held by this
1133 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1134 * clo_enqueue time, is released.
1136 * LDLM lock can only be canceled if there is no cl_lock using it.
1138 * Overall process of the locking during IO operation is as following:
1140 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1141 * is called on each layer. Responsibility of this method is to add locks,
1142 * needed by a given layer into cl_io.ci_lockset.
1144 * - once locks for all layers were collected, they are sorted to avoid
1145 * dead-locks (cl_io_locks_sort()), and enqueued.
1147 * - when all locks are acquired, IO is performed;
1149 * - locks are released after IO is complete.
1151 * Striping introduces major additional complexity into locking. The
1152 * fundamental problem is that it is generally unsafe to actively use (hold)
1153 * two locks on the different OST servers at the same time, as this introduces
1154 * inter-server dependency and can lead to cascading evictions.
1156 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1157 * that no multi-stripe locks are taken (note that this design abandons POSIX
1158 * read/write semantics). Such pieces ideally can be executed concurrently. At
1159 * the same time, certain types of IO cannot be sub-divived, without
1160 * sacrificing correctness. This includes:
1162 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1165 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1167 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1168 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1169 * has to be held together with the usual lock on [offset, offset + count].
1171 * Interaction with DLM
1173 * In the expected setup, cl_lock is ultimately backed up by a collection of
1174 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1175 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1176 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1177 * description of interaction with DLM.
1183 struct cl_lock_descr {
1184 /** Object this lock is granted for. */
1185 struct cl_object *cld_obj;
1186 /** Index of the first page protected by this lock. */
1188 /** Index of the last page (inclusive) protected by this lock. */
1190 /** Group ID, for group lock */
1193 enum cl_lock_mode cld_mode;
1195 * flags to enqueue lock. A combination of bit-flags from
1196 * enum cl_enq_flags.
1198 __u32 cld_enq_flags;
1201 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1202 #define PDESCR(descr) \
1203 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1204 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1206 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1209 * Layered client lock.
1212 /** List of slices. Immutable after creation. */
1213 struct list_head cll_layers;
1214 /** lock attribute, extent, cl_object, etc. */
1215 struct cl_lock_descr cll_descr;
1219 * Per-layer part of cl_lock
1221 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1223 struct cl_lock_slice {
1224 struct cl_lock *cls_lock;
1225 /** Object slice corresponding to this lock slice. Immutable after
1227 struct cl_object *cls_obj;
1228 const struct cl_lock_operations *cls_ops;
1229 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1230 struct list_head cls_linkage;
1235 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1237 struct cl_lock_operations {
1240 * Attempts to enqueue the lock. Called top-to-bottom.
1242 * \retval 0 this layer has enqueued the lock successfully
1243 * \retval >0 this layer has enqueued the lock, but need to wait on
1244 * @anchor for resources
1245 * \retval -ve failure
1247 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1248 * \see osc_lock_enqueue()
1250 int (*clo_enqueue)(const struct lu_env *env,
1251 const struct cl_lock_slice *slice,
1252 struct cl_io *io, struct cl_sync_io *anchor);
1254 * Cancel a lock, release its DLM lock ref, while does not cancel the
1257 void (*clo_cancel)(const struct lu_env *env,
1258 const struct cl_lock_slice *slice);
1261 * Destructor. Frees resources and the slice.
1263 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1264 * \see osc_lock_fini()
1266 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1268 * Optional debugging helper. Prints given lock slice.
1270 int (*clo_print)(const struct lu_env *env,
1271 void *cookie, lu_printer_t p,
1272 const struct cl_lock_slice *slice);
1275 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1277 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1278 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1279 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1280 CDEBUG(mask, format , ## __VA_ARGS__); \
1284 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1288 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1294 /** \addtogroup cl_page_list cl_page_list
1295 * Page list used to perform collective operations on a group of pages.
1297 * Pages are added to the list one by one. cl_page_list acquires a reference
1298 * for every page in it. Page list is used to perform collective operations on
1301 * - submit pages for an immediate transfer,
1303 * - own pages on behalf of certain io (waiting for each page in turn),
1307 * When list is finalized, it releases references on all pages it still has.
1309 * \todo XXX concurrency control.
1313 struct cl_page_list {
1315 struct list_head pl_pages;
1316 struct task_struct *pl_owner;
1320 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1321 * contains an incoming page list and an outgoing page list.
1324 struct cl_page_list c2_qin;
1325 struct cl_page_list c2_qout;
1328 /** @} cl_page_list */
1330 /** \addtogroup cl_io cl_io
1335 * cl_io represents a high level I/O activity like
1336 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1339 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1340 * important distinction. We want to minimize number of calls to the allocator
1341 * in the fast path, e.g., in the case of read(2) when everything is cached:
1342 * client already owns the lock over region being read, and data are cached
1343 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1344 * per-layer io state is stored in the session, associated with the io, see
1345 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1346 * by using free-lists, see cl_env_get().
1348 * There is a small predefined number of possible io types, enumerated in enum
1351 * cl_io is a state machine, that can be advanced concurrently by the multiple
1352 * threads. It is up to these threads to control the concurrency and,
1353 * specifically, to detect when io is done, and its state can be safely
1356 * For read/write io overall execution plan is as following:
1358 * (0) initialize io state through all layers;
1360 * (1) loop: prepare chunk of work to do
1362 * (2) call all layers to collect locks they need to process current chunk
1364 * (3) sort all locks to avoid dead-locks, and acquire them
1366 * (4) process the chunk: call per-page methods
1367 * (cl_io_operations::cio_read_page() for read,
1368 * cl_io_operations::cio_prepare_write(),
1369 * cl_io_operations::cio_commit_write() for write)
1375 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1376 * address allocation efficiency issues mentioned above), and returns with the
1377 * special error condition from per-page method when current sub-io has to
1378 * block. This causes io loop to be repeated, and lov switches to the next
1379 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1384 /** read system call */
1386 /** write system call */
1388 /** truncate, utime system calls */
1391 * page fault handling
1395 * fsync system call handling
1396 * To write out a range of file
1400 * Miscellaneous io. This is used for occasional io activity that
1401 * doesn't fit into other types. Currently this is used for:
1403 * - cancellation of an extent lock. This io exists as a context
1404 * to write dirty pages from under the lock being canceled back
1407 * - VM induced page write-out. An io context for writing page out
1408 * for memory cleansing;
1410 * - glimpse. An io context to acquire glimpse lock.
1412 * - grouplock. An io context to acquire group lock.
1414 * CIT_MISC io is used simply as a context in which locks and pages
1415 * are manipulated. Such io has no internal "process", that is,
1416 * cl_io_loop() is never called for it.
1423 * States of cl_io state machine
1426 /** Not initialized. */
1430 /** IO iteration started. */
1434 /** Actual IO is in progress. */
1436 /** IO for the current iteration finished. */
1438 /** Locks released. */
1440 /** Iteration completed. */
1442 /** cl_io finalized. */
1447 * IO state private for a layer.
1449 * This is usually embedded into layer session data, rather than allocated
1452 * \see vvp_io, lov_io, osc_io
1454 struct cl_io_slice {
1455 struct cl_io *cis_io;
1456 /** corresponding object slice. Immutable after creation. */
1457 struct cl_object *cis_obj;
1458 /** io operations. Immutable after creation. */
1459 const struct cl_io_operations *cis_iop;
1461 * linkage into a list of all slices for a given cl_io, hanging off
1462 * cl_io::ci_layers. Immutable after creation.
1464 struct list_head cis_linkage;
1467 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1471 * Per-layer io operations.
1472 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1474 struct cl_io_operations {
1476 * Vector of io state transition methods for every io type.
1478 * \see cl_page_operations::io
1482 * Prepare io iteration at a given layer.
1484 * Called top-to-bottom at the beginning of each iteration of
1485 * "io loop" (if it makes sense for this type of io). Here
1486 * layer selects what work it will do during this iteration.
1488 * \see cl_io_operations::cio_iter_fini()
1490 int (*cio_iter_init) (const struct lu_env *env,
1491 const struct cl_io_slice *slice);
1493 * Finalize io iteration.
1495 * Called bottom-to-top at the end of each iteration of "io
1496 * loop". Here layers can decide whether IO has to be
1499 * \see cl_io_operations::cio_iter_init()
1501 void (*cio_iter_fini) (const struct lu_env *env,
1502 const struct cl_io_slice *slice);
1504 * Collect locks for the current iteration of io.
1506 * Called top-to-bottom to collect all locks necessary for
1507 * this iteration. This methods shouldn't actually enqueue
1508 * anything, instead it should post a lock through
1509 * cl_io_lock_add(). Once all locks are collected, they are
1510 * sorted and enqueued in the proper order.
1512 int (*cio_lock) (const struct lu_env *env,
1513 const struct cl_io_slice *slice);
1515 * Finalize unlocking.
1517 * Called bottom-to-top to finish layer specific unlocking
1518 * functionality, after generic code released all locks
1519 * acquired by cl_io_operations::cio_lock().
1521 void (*cio_unlock)(const struct lu_env *env,
1522 const struct cl_io_slice *slice);
1524 * Start io iteration.
1526 * Once all locks are acquired, called top-to-bottom to
1527 * commence actual IO. In the current implementation,
1528 * top-level vvp_io_{read,write}_start() does all the work
1529 * synchronously by calling generic_file_*(), so other layers
1530 * are called when everything is done.
1532 int (*cio_start)(const struct lu_env *env,
1533 const struct cl_io_slice *slice);
1535 * Called top-to-bottom at the end of io loop. Here layer
1536 * might wait for an unfinished asynchronous io.
1538 void (*cio_end) (const struct lu_env *env,
1539 const struct cl_io_slice *slice);
1541 * Called bottom-to-top to notify layers that read/write IO
1542 * iteration finished, with \a nob bytes transferred.
1544 void (*cio_advance)(const struct lu_env *env,
1545 const struct cl_io_slice *slice,
1548 * Called once per io, bottom-to-top to release io resources.
1550 void (*cio_fini) (const struct lu_env *env,
1551 const struct cl_io_slice *slice);
1555 * Submit pages from \a queue->c2_qin for IO, and move
1556 * successfully submitted pages into \a queue->c2_qout. Return
1557 * non-zero if failed to submit even the single page. If
1558 * submission failed after some pages were moved into \a
1559 * queue->c2_qout, completion callback with non-zero ioret is
1562 int (*cio_submit)(const struct lu_env *env,
1563 const struct cl_io_slice *slice,
1564 enum cl_req_type crt,
1565 struct cl_2queue *queue);
1567 * Queue async page for write.
1568 * The difference between cio_submit and cio_queue is that
1569 * cio_submit is for urgent request.
1571 int (*cio_commit_async)(const struct lu_env *env,
1572 const struct cl_io_slice *slice,
1573 struct cl_page_list *queue, int from, int to,
1576 * Read missing page.
1578 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
1579 * method, when it hits not-up-to-date page in the range. Optional.
1581 * \pre io->ci_type == CIT_READ
1583 int (*cio_read_page)(const struct lu_env *env,
1584 const struct cl_io_slice *slice,
1585 const struct cl_page_slice *page);
1587 * Optional debugging helper. Print given io slice.
1589 int (*cio_print)(const struct lu_env *env, void *cookie,
1590 lu_printer_t p, const struct cl_io_slice *slice);
1594 * Flags to lock enqueue procedure.
1599 * instruct server to not block, if conflicting lock is found. Instead
1600 * -EWOULDBLOCK is returned immediately.
1602 CEF_NONBLOCK = 0x00000001,
1604 * take lock asynchronously (out of order), as it cannot
1605 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1607 CEF_ASYNC = 0x00000002,
1609 * tell the server to instruct (though a flag in the blocking ast) an
1610 * owner of the conflicting lock, that it can drop dirty pages
1611 * protected by this lock, without sending them to the server.
1613 CEF_DISCARD_DATA = 0x00000004,
1615 * tell the sub layers that it must be a `real' lock. This is used for
1616 * mmapped-buffer locks and glimpse locks that must be never converted
1617 * into lockless mode.
1619 * \see vvp_mmap_locks(), cl_glimpse_lock().
1621 CEF_MUST = 0x00000008,
1623 * tell the sub layers that never request a `real' lock. This flag is
1624 * not used currently.
1626 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1627 * conversion policy: ci_lockreq describes generic information of lock
1628 * requirement for this IO, especially for locks which belong to the
1629 * object doing IO; however, lock itself may have precise requirements
1630 * that are described by the enqueue flags.
1632 CEF_NEVER = 0x00000010,
1634 * for async glimpse lock.
1636 CEF_AGL = 0x00000020,
1638 * enqueue a lock to test DLM lock existence.
1640 CEF_PEEK = 0x00000040,
1642 * mask of enq_flags.
1644 CEF_MASK = 0x0000007f,
1648 * Link between lock and io. Intermediate structure is needed, because the
1649 * same lock can be part of multiple io's simultaneously.
1651 struct cl_io_lock_link {
1652 /** linkage into one of cl_lockset lists. */
1653 struct list_head cill_linkage;
1654 struct cl_lock cill_lock;
1655 /** optional destructor */
1656 void (*cill_fini)(const struct lu_env *env,
1657 struct cl_io_lock_link *link);
1659 #define cill_descr cill_lock.cll_descr
1662 * Lock-set represents a collection of locks, that io needs at a
1663 * time. Generally speaking, client tries to avoid holding multiple locks when
1666 * - holding extent locks over multiple ost's introduces the danger of
1667 * "cascading timeouts";
1669 * - holding multiple locks over the same ost is still dead-lock prone,
1670 * see comment in osc_lock_enqueue(),
1672 * but there are certain situations where this is unavoidable:
1674 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1676 * - truncate has to take [new-size, EOF] lock for correctness;
1678 * - SNS has to take locks across full stripe for correctness;
1680 * - in the case when user level buffer, supplied to {read,write}(file0),
1681 * is a part of a memory mapped lustre file, client has to take a dlm
1682 * locks on file0, and all files that back up the buffer (or a part of
1683 * the buffer, that is being processed in the current chunk, in any
1684 * case, there are situations where at least 2 locks are necessary).
1686 * In such cases we at least try to take locks in the same consistent
1687 * order. To this end, all locks are first collected, then sorted, and then
1691 /** locks to be acquired. */
1692 struct list_head cls_todo;
1693 /** locks acquired. */
1694 struct list_head cls_done;
1698 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1699 * but 'req' is always to be thought as 'request' :-)
1701 enum cl_io_lock_dmd {
1702 /** Always lock data (e.g., O_APPEND). */
1704 /** Layers are free to decide between local and global locking. */
1706 /** Never lock: there is no cache (e.g., liblustre). */
1710 enum cl_fsync_mode {
1711 /** start writeback, do not wait for them to finish */
1713 /** start writeback and wait for them to finish */
1715 /** discard all of dirty pages in a specific file range */
1716 CL_FSYNC_DISCARD = 2,
1717 /** start writeback and make sure they have reached storage before
1718 * return. OST_SYNC RPC must be issued and finished */
1722 struct cl_io_rw_common {
1732 * cl_io is shared by all threads participating in this IO (in current
1733 * implementation only one thread advances IO, but parallel IO design and
1734 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1735 * is up to these threads to serialize their activities, including updates to
1736 * mutable cl_io fields.
1739 /** type of this IO. Immutable after creation. */
1740 enum cl_io_type ci_type;
1741 /** current state of cl_io state machine. */
1742 enum cl_io_state ci_state;
1743 /** main object this io is against. Immutable after creation. */
1744 struct cl_object *ci_obj;
1746 * Upper layer io, of which this io is a part of. Immutable after
1749 struct cl_io *ci_parent;
1750 /** List of slices. Immutable after creation. */
1751 struct list_head ci_layers;
1752 /** list of locks (to be) acquired by this io. */
1753 struct cl_lockset ci_lockset;
1754 /** lock requirements, this is just a help info for sublayers. */
1755 enum cl_io_lock_dmd ci_lockreq;
1758 struct cl_io_rw_common rd;
1761 struct cl_io_rw_common wr;
1765 struct cl_io_rw_common ci_rw;
1766 struct cl_setattr_io {
1767 struct ost_lvb sa_attr;
1768 unsigned int sa_valid;
1769 int sa_stripe_index;
1770 struct lu_fid *sa_parent_fid;
1771 struct obd_capa *sa_capa;
1773 struct cl_fault_io {
1774 /** page index within file. */
1776 /** bytes valid byte on a faulted page. */
1778 /** writable page? for nopage() only */
1780 /** page of an executable? */
1782 /** page_mkwrite() */
1784 /** resulting page */
1785 struct cl_page *ft_page;
1787 struct cl_fsync_io {
1790 struct obd_capa *fi_capa;
1791 /** file system level fid */
1792 struct lu_fid *fi_fid;
1793 enum cl_fsync_mode fi_mode;
1794 /* how many pages were written/discarded */
1795 unsigned int fi_nr_written;
1798 struct cl_2queue ci_queue;
1801 unsigned int ci_continue:1,
1803 * This io has held grouplock, to inform sublayers that
1804 * don't do lockless i/o.
1808 * The whole IO need to be restarted because layout has been changed
1812 * to not refresh layout - the IO issuer knows that the layout won't
1813 * change(page operations, layout change causes all page to be
1814 * discarded), or it doesn't matter if it changes(sync).
1818 * Check if layout changed after the IO finishes. Mainly for HSM
1819 * requirement. If IO occurs to openning files, it doesn't need to
1820 * verify layout because HSM won't release openning files.
1821 * Right now, only two opertaions need to verify layout: glimpse
1826 * file is released, restore has to to be triggered by vvp layer
1828 ci_restore_needed:1,
1834 * Number of pages owned by this IO. For invariant checking.
1836 unsigned ci_owned_nr;
1841 /** \addtogroup cl_req cl_req
1846 * There are two possible modes of transfer initiation on the client:
1848 * - immediate transfer: this is started when a high level io wants a page
1849 * or a collection of pages to be transferred right away. Examples:
1850 * read-ahead, synchronous read in the case of non-page aligned write,
1851 * page write-out as a part of extent lock cancellation, page write-out
1852 * as a part of memory cleansing. Immediate transfer can be both
1853 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1855 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1856 * when io wants to transfer a page to the server some time later, when
1857 * it can be done efficiently. Example: pages dirtied by the write(2)
1860 * In any case, transfer takes place in the form of a cl_req, which is a
1861 * representation for a network RPC.
1863 * Pages queued for an opportunistic transfer are cached until it is decided
1864 * that efficient RPC can be composed of them. This decision is made by "a
1865 * req-formation engine", currently implemented as a part of osc
1866 * layer. Req-formation depends on many factors: the size of the resulting
1867 * RPC, whether or not multi-object RPCs are supported by the server,
1868 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1870 * For the immediate transfer io submits a cl_page_list, that req-formation
1871 * engine slices into cl_req's, possibly adding cached pages to some of
1872 * the resulting req's.
1874 * Whenever a page from cl_page_list is added to a newly constructed req, its
1875 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1876 * page state is atomically changed from cl_page_state::CPS_OWNED to
1877 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1878 * is zeroed, and cl_page::cp_req is set to the
1879 * req. cl_page_operations::cpo_prep() method at the particular layer might
1880 * return -EALREADY to indicate that it does not need to submit this page
1881 * at all. This is possible, for example, if page, submitted for read,
1882 * became up-to-date in the meantime; and for write, the page don't have
1883 * dirty bit marked. \see cl_io_submit_rw()
1885 * Whenever a cached page is added to a newly constructed req, its
1886 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1887 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1888 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1889 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1890 * might return -EAGAIN to indicate that this page is not eligible for the
1891 * transfer right now.
1895 * Plan is to divide transfers into "priority bands" (indicated when
1896 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1897 * and allow glueing of cached pages to immediate transfers only within single
1898 * band. This would make high priority transfers (like lock cancellation or
1899 * memory pressure induced write-out) really high priority.
1904 * Per-transfer attributes.
1906 struct cl_req_attr {
1907 /** Generic attributes for the server consumption. */
1908 struct obdo *cra_oa;
1910 struct obd_capa *cra_capa;
1912 char cra_jobid[LUSTRE_JOBID_SIZE];
1916 * Transfer request operations definable at every layer.
1918 * Concurrency: transfer formation engine synchronizes calls to all transfer
1921 struct cl_req_operations {
1923 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1924 * complete (all pages are added).
1926 * \see osc_req_prep()
1928 int (*cro_prep)(const struct lu_env *env,
1929 const struct cl_req_slice *slice);
1931 * Called top-to-bottom to fill in \a oa fields. This is called twice
1932 * with different flags, see bug 10150 and osc_build_req().
1934 * \param obj an object from cl_req which attributes are to be set in
1937 * \param oa struct obdo where attributes are placed
1939 * \param flags \a oa fields to be filled.
1941 void (*cro_attr_set)(const struct lu_env *env,
1942 const struct cl_req_slice *slice,
1943 const struct cl_object *obj,
1944 struct cl_req_attr *attr, u64 flags);
1946 * Called top-to-bottom from cl_req_completion() to notify layers that
1947 * transfer completed. Has to free all state allocated by
1948 * cl_device_operations::cdo_req_init().
1950 void (*cro_completion)(const struct lu_env *env,
1951 const struct cl_req_slice *slice, int ioret);
1955 * A per-object state that (potentially multi-object) transfer request keeps.
1958 /** object itself */
1959 struct cl_object *ro_obj;
1960 /** reference to cl_req_obj::ro_obj. For debugging. */
1961 struct lu_ref_link ro_obj_ref;
1962 /* something else? Number of pages for a given object? */
1968 * Transfer requests are not reference counted, because IO sub-system owns
1969 * them exclusively and knows when to free them.
1973 * cl_req is created by cl_req_alloc() that calls
1974 * cl_device_operations::cdo_req_init() device methods to allocate per-req
1975 * state in every layer.
1977 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
1978 * contains pages for.
1980 * Once all pages were collected, cl_page_operations::cpo_prep() method is
1981 * called top-to-bottom. At that point layers can modify req, let it pass, or
1982 * deny it completely. This is to support things like SNS that have transfer
1983 * ordering requirements invisible to the individual req-formation engine.
1985 * On transfer completion (or transfer timeout, or failure to initiate the
1986 * transfer of an allocated req), cl_req_operations::cro_completion() method
1987 * is called, after execution of cl_page_operations::cpo_completion() of all
1991 enum cl_req_type crq_type;
1992 /** A list of pages being transfered */
1993 struct list_head crq_pages;
1994 /** Number of pages in cl_req::crq_pages */
1995 unsigned crq_nrpages;
1996 /** An array of objects which pages are in ->crq_pages */
1997 struct cl_req_obj *crq_o;
1998 /** Number of elements in cl_req::crq_objs[] */
1999 unsigned crq_nrobjs;
2000 struct list_head crq_layers;
2004 * Per-layer state for request.
2006 struct cl_req_slice {
2007 struct cl_req *crs_req;
2008 struct cl_device *crs_dev;
2009 struct list_head crs_linkage;
2010 const struct cl_req_operations *crs_ops;
2015 enum cache_stats_item {
2016 /** how many cache lookups were performed */
2018 /** how many times cache lookup resulted in a hit */
2020 /** how many entities are in the cache right now */
2022 /** how many entities in the cache are actively used (and cannot be
2023 * evicted) right now */
2025 /** how many entities were created at all */
2030 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2033 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2035 struct cache_stats {
2036 const char *cs_name;
2037 atomic_t cs_stats[CS_NR];
2040 /** These are not exported so far */
2041 void cache_stats_init (struct cache_stats *cs, const char *name);
2044 * Client-side site. This represents particular client stack. "Global"
2045 * variables should (directly or indirectly) be added here to allow multiple
2046 * clients to co-exist in the single address space.
2049 struct lu_site cs_lu;
2051 * Statistical counters. Atomics do not scale, something better like
2052 * per-cpu counters is needed.
2054 * These are exported as /proc/fs/lustre/llite/.../site
2056 * When interpreting keep in mind that both sub-locks (and sub-pages)
2057 * and top-locks (and top-pages) are accounted here.
2059 struct cache_stats cs_pages;
2060 atomic_t cs_pages_state[CPS_NR];
2063 int cl_site_init(struct cl_site *s, struct cl_device *top);
2064 void cl_site_fini(struct cl_site *s);
2065 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2068 * Output client site statistical counters into a buffer. Suitable for
2069 * ll_rd_*()-style functions.
2071 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2076 * Type conversion and accessory functions.
2080 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2082 return container_of(site, struct cl_site, cs_lu);
2085 static inline int lu_device_is_cl(const struct lu_device *d)
2087 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2090 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2092 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2093 return container_of0(d, struct cl_device, cd_lu_dev);
2096 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2098 return &d->cd_lu_dev;
2101 static inline struct cl_object *lu2cl(const struct lu_object *o)
2103 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2104 return container_of0(o, struct cl_object, co_lu);
2107 static inline const struct cl_object_conf *
2108 lu2cl_conf(const struct lu_object_conf *conf)
2110 return container_of0(conf, struct cl_object_conf, coc_lu);
2113 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2115 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2118 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2120 return container_of0(h, struct cl_object_header, coh_lu);
2123 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2125 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2129 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2131 return luh2coh(obj->co_lu.lo_header);
2134 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2136 return lu_device_init(&d->cd_lu_dev, t);
2139 static inline void cl_device_fini(struct cl_device *d)
2141 lu_device_fini(&d->cd_lu_dev);
2144 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2145 struct cl_object *obj, pgoff_t index,
2146 const struct cl_page_operations *ops);
2147 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2148 struct cl_object *obj,
2149 const struct cl_lock_operations *ops);
2150 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2151 struct cl_object *obj, const struct cl_io_operations *ops);
2152 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2153 struct cl_device *dev,
2154 const struct cl_req_operations *ops);
2157 /** \defgroup cl_object cl_object
2159 struct cl_object *cl_object_top (struct cl_object *o);
2160 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2161 const struct lu_fid *fid,
2162 const struct cl_object_conf *c);
2164 int cl_object_header_init(struct cl_object_header *h);
2165 void cl_object_header_fini(struct cl_object_header *h);
2166 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2167 void cl_object_get (struct cl_object *o);
2168 void cl_object_attr_lock (struct cl_object *o);
2169 void cl_object_attr_unlock(struct cl_object *o);
2170 int cl_object_attr_get (const struct lu_env *env, struct cl_object *obj,
2171 struct cl_attr *attr);
2172 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2173 const struct cl_attr *attr, unsigned valid);
2174 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2175 struct ost_lvb *lvb);
2176 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2177 const struct cl_object_conf *conf);
2178 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2179 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2180 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2181 struct lov_user_md __user *lum);
2184 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2186 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2188 return cl_object_header(o0) == cl_object_header(o1);
2191 static inline void cl_object_page_init(struct cl_object *clob, int size)
2193 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2194 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2195 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2198 static inline void *cl_object_page_slice(struct cl_object *clob,
2199 struct cl_page *page)
2201 return (void *)((char *)page + clob->co_slice_off);
2205 * Return refcount of cl_object.
2207 static inline int cl_object_refc(struct cl_object *clob)
2209 struct lu_object_header *header = clob->co_lu.lo_header;
2210 return atomic_read(&header->loh_ref);
2215 /** \defgroup cl_page cl_page
2223 /* callback of cl_page_gang_lookup() */
2225 struct cl_page *cl_page_find (const struct lu_env *env,
2226 struct cl_object *obj,
2227 pgoff_t idx, struct page *vmpage,
2228 enum cl_page_type type);
2229 struct cl_page *cl_page_alloc (const struct lu_env *env,
2230 struct cl_object *o, pgoff_t ind,
2231 struct page *vmpage,
2232 enum cl_page_type type);
2233 void cl_page_get (struct cl_page *page);
2234 void cl_page_put (const struct lu_env *env,
2235 struct cl_page *page);
2236 void cl_page_print (const struct lu_env *env, void *cookie,
2237 lu_printer_t printer,
2238 const struct cl_page *pg);
2239 void cl_page_header_print(const struct lu_env *env, void *cookie,
2240 lu_printer_t printer,
2241 const struct cl_page *pg);
2242 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2243 struct cl_page *cl_page_top (struct cl_page *page);
2245 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2246 const struct lu_device_type *dtype);
2251 * Functions dealing with the ownership of page by io.
2255 int cl_page_own (const struct lu_env *env,
2256 struct cl_io *io, struct cl_page *page);
2257 int cl_page_own_try (const struct lu_env *env,
2258 struct cl_io *io, struct cl_page *page);
2259 void cl_page_assume (const struct lu_env *env,
2260 struct cl_io *io, struct cl_page *page);
2261 void cl_page_unassume (const struct lu_env *env,
2262 struct cl_io *io, struct cl_page *pg);
2263 void cl_page_disown (const struct lu_env *env,
2264 struct cl_io *io, struct cl_page *page);
2265 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2272 * Functions dealing with the preparation of a page for a transfer, and
2273 * tracking transfer state.
2276 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2277 struct cl_page *pg, enum cl_req_type crt);
2278 void cl_page_completion (const struct lu_env *env,
2279 struct cl_page *pg, enum cl_req_type crt, int ioret);
2280 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2281 enum cl_req_type crt);
2282 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2283 struct cl_page *pg, enum cl_req_type crt);
2284 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2286 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2287 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2288 struct cl_page *pg);
2294 * \name helper routines
2295 * Functions to discard, delete and export a cl_page.
2298 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2299 struct cl_page *pg);
2300 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2301 int cl_page_is_vmlocked(const struct lu_env *env,
2302 const struct cl_page *pg);
2303 void cl_page_export(const struct lu_env *env,
2304 struct cl_page *pg, int uptodate);
2305 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2306 struct cl_page *page, pgoff_t *max_index);
2307 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2308 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2309 size_t cl_page_size(const struct cl_object *obj);
2311 void cl_lock_print(const struct lu_env *env, void *cookie,
2312 lu_printer_t printer, const struct cl_lock *lock);
2313 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2314 lu_printer_t printer,
2315 const struct cl_lock_descr *descr);
2319 * Data structure managing a client's cached pages. A count of
2320 * "unstable" pages is maintained, and an LRU of clean pages is
2321 * maintained. "unstable" pages are pages pinned by the ptlrpc
2322 * layer for recovery purposes.
2324 struct cl_client_cache {
2330 * # of threads are doing shrinking
2332 unsigned int ccc_lru_shrinkers;
2334 * # of LRU entries available
2336 atomic_long_t ccc_lru_left;
2338 * List of entities(OSCs) for this LRU cache
2340 struct list_head ccc_lru;
2342 * Max # of LRU entries
2344 unsigned long ccc_lru_max;
2346 * Lock to protect ccc_lru list
2348 spinlock_t ccc_lru_lock;
2350 * Set if unstable check is enabled
2352 unsigned int ccc_unstable_check:1;
2354 * # of unstable pages for this mount point
2356 atomic_long_t ccc_unstable_nr;
2358 * Waitq for awaiting unstable pages to reach zero.
2359 * Used at umounting time and signaled on BRW commit
2361 wait_queue_head_t ccc_unstable_waitq;
2366 /** \defgroup cl_lock cl_lock
2368 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2369 struct cl_lock *lock);
2370 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2371 const struct cl_io *io);
2372 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2373 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2374 const struct lu_device_type *dtype);
2375 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2377 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2378 struct cl_lock *lock, struct cl_sync_io *anchor);
2379 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2383 /** \defgroup cl_io cl_io
2386 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2387 enum cl_io_type iot, struct cl_object *obj);
2388 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2389 enum cl_io_type iot, struct cl_object *obj);
2390 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2391 enum cl_io_type iot, loff_t pos, size_t count);
2392 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2394 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2395 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2396 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2397 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2398 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2399 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2400 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2401 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2402 struct cl_io_lock_link *link);
2403 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2404 struct cl_lock_descr *descr);
2405 int cl_io_read_page (const struct lu_env *env, struct cl_io *io,
2406 struct cl_page *page);
2407 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2408 enum cl_req_type iot, struct cl_2queue *queue);
2409 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2410 enum cl_req_type iot, struct cl_2queue *queue,
2412 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2413 struct cl_page_list *queue, int from, int to,
2415 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2417 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2418 struct cl_page_list *queue);
2419 int cl_io_is_going (const struct lu_env *env);
2422 * True, iff \a io is an O_APPEND write(2).
2424 static inline int cl_io_is_append(const struct cl_io *io)
2426 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2429 static inline int cl_io_is_sync_write(const struct cl_io *io)
2431 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2434 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2436 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2440 * True, iff \a io is a truncate(2).
2442 static inline int cl_io_is_trunc(const struct cl_io *io)
2444 return io->ci_type == CIT_SETATTR &&
2445 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2448 struct cl_io *cl_io_top(struct cl_io *io);
2450 void cl_io_print(const struct lu_env *env, void *cookie,
2451 lu_printer_t printer, const struct cl_io *io);
2453 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2455 typeof(foo_io) __foo_io = (foo_io); \
2457 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2458 memset(&__foo_io->base + 1, 0, \
2459 (sizeof *__foo_io) - sizeof __foo_io->base); \
2464 /** \defgroup cl_page_list cl_page_list
2468 * Last page in the page list.
2470 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2472 LASSERT(plist->pl_nr > 0);
2473 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2476 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2478 LASSERT(plist->pl_nr > 0);
2479 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2483 * Iterate over pages in a page list.
2485 #define cl_page_list_for_each(page, list) \
2486 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2489 * Iterate over pages in a page list, taking possible removals into account.
2491 #define cl_page_list_for_each_safe(page, temp, list) \
2492 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2494 void cl_page_list_init (struct cl_page_list *plist);
2495 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2496 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2497 struct cl_page *page);
2498 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2499 struct cl_page *page);
2500 void cl_page_list_splice (struct cl_page_list *list,
2501 struct cl_page_list *head);
2502 void cl_page_list_del (const struct lu_env *env,
2503 struct cl_page_list *plist, struct cl_page *page);
2504 void cl_page_list_disown (const struct lu_env *env,
2505 struct cl_io *io, struct cl_page_list *plist);
2506 int cl_page_list_own (const struct lu_env *env,
2507 struct cl_io *io, struct cl_page_list *plist);
2508 void cl_page_list_assume (const struct lu_env *env,
2509 struct cl_io *io, struct cl_page_list *plist);
2510 void cl_page_list_discard(const struct lu_env *env,
2511 struct cl_io *io, struct cl_page_list *plist);
2512 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2514 void cl_2queue_init (struct cl_2queue *queue);
2515 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2516 void cl_2queue_disown (const struct lu_env *env,
2517 struct cl_io *io, struct cl_2queue *queue);
2518 void cl_2queue_assume (const struct lu_env *env,
2519 struct cl_io *io, struct cl_2queue *queue);
2520 void cl_2queue_discard (const struct lu_env *env,
2521 struct cl_io *io, struct cl_2queue *queue);
2522 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2523 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2525 /** @} cl_page_list */
2527 /** \defgroup cl_req cl_req
2529 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2530 enum cl_req_type crt, int nr_objects);
2532 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2533 struct cl_page *page);
2534 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2535 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2536 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2537 struct cl_req_attr *attr, u64 flags);
2538 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2540 /** \defgroup cl_sync_io cl_sync_io
2544 * Anchor for synchronous transfer. This is allocated on a stack by thread
2545 * doing synchronous transfer, and a pointer to this structure is set up in
2546 * every page submitted for transfer. Transfer completion routine updates
2547 * anchor and wakes up waiting thread when transfer is complete.
2550 /** number of pages yet to be transferred. */
2551 atomic_t csi_sync_nr;
2554 /** barrier of destroy this structure */
2555 atomic_t csi_barrier;
2556 /** completion to be signaled when transfer is complete. */
2557 wait_queue_head_t csi_waitq;
2558 /** callback to invoke when this IO is finished */
2559 void (*csi_end_io)(const struct lu_env *,
2560 struct cl_sync_io *);
2563 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2564 void (*end)(const struct lu_env *, struct cl_sync_io *));
2565 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2567 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2569 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2571 /** @} cl_sync_io */
2575 /** \defgroup cl_env cl_env
2577 * lu_env handling for a client.
2579 * lu_env is an environment within which lustre code executes. Its major part
2580 * is lu_context---a fast memory allocation mechanism that is used to conserve
2581 * precious kernel stack space. Originally lu_env was designed for a server,
2584 * - there is a (mostly) fixed number of threads, and
2586 * - call chains have no non-lustre portions inserted between lustre code.
2588 * On a client both these assumtpion fails, because every user thread can
2589 * potentially execute lustre code as part of a system call, and lustre calls
2590 * into VFS or MM that call back into lustre.
2592 * To deal with that, cl_env wrapper functions implement the following
2595 * - allocation and destruction of environment is amortized by caching no
2596 * longer used environments instead of destroying them;
2598 * - there is a notion of "current" environment, attached to the kernel
2599 * data structure representing current thread Top-level lustre code
2600 * allocates an environment and makes it current, then calls into
2601 * non-lustre code, that in turn calls lustre back. Low-level lustre
2602 * code thus called can fetch environment created by the top-level code
2603 * and reuse it, avoiding additional environment allocation.
2604 * Right now, three interfaces can attach the cl_env to running thread:
2607 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2609 * \see lu_env, lu_context, lu_context_key
2612 struct cl_env_nest {
2617 struct lu_env *cl_env_peek (int *refcheck);
2618 struct lu_env *cl_env_get (int *refcheck);
2619 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2620 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2621 void cl_env_put (struct lu_env *env, int *refcheck);
2622 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2623 void *cl_env_reenter (void);
2624 void cl_env_reexit (void *cookie);
2625 void cl_env_implant (struct lu_env *env, int *refcheck);
2626 void cl_env_unplant (struct lu_env *env, int *refcheck);
2627 unsigned cl_env_cache_purge(unsigned nr);
2628 struct lu_env *cl_env_percpu_get (void);
2629 void cl_env_percpu_put (struct lu_env *env);
2636 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2637 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2639 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2640 struct lu_device_type *ldt,
2641 struct lu_device *next);
2644 int cl_global_init(void);
2645 void cl_global_fini(void);
2647 #endif /* _LINUX_CL_OBJECT_H */