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, 2015, 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 "lu_object.h"
100 #include <linux/atomic.h>
101 #include "linux/lustre_compat25.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 vvp_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
165 * "Data attributes" of cl_object. Data attributes can be updated
166 * independently for a sub-object, and top-object's attributes are calculated
167 * from sub-objects' ones.
170 /** Object size, in bytes */
173 * Known minimal size, in bytes.
175 * This is only valid when at least one DLM lock is held.
178 /** Modification time. Measured in seconds since epoch. */
180 /** Access time. Measured in seconds since epoch. */
182 /** Change time. Measured in seconds since epoch. */
185 * Blocks allocated to this cl_object on the server file system.
187 * \todo XXX An interface for block size is needed.
191 * User identifier for quota purposes.
195 * Group identifier for quota purposes.
201 * Fields in cl_attr that are being set.
215 * Sub-class of lu_object with methods common for objects on the client
218 * cl_object: represents a regular file system object, both a file and a
219 * stripe. cl_object is based on lu_object: it is identified by a fid,
220 * layered, cached, hashed, and lrued. Important distinction with the server
221 * side, where md_object and dt_object are used, is that cl_object "fans out"
222 * at the lov/sns level: depending on the file layout, single file is
223 * represented as a set of "sub-objects" (stripes). At the implementation
224 * level, struct lov_object contains an array of cl_objects. Each sub-object
225 * is a full-fledged cl_object, having its fid, living in the lru and hash
228 * This leads to the next important difference with the server side: on the
229 * client, it's quite usual to have objects with the different sequence of
230 * layers. For example, typical top-object is composed of the following
236 * whereas its sub-objects are composed of
241 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
242 * track of the object-subobject relationship.
244 * Sub-objects are not cached independently: when top-object is about to
245 * be discarded from the memory, all its sub-objects are torn-down and
248 * \see vvp_object, lov_object, lovsub_object, osc_object
252 struct lu_object co_lu;
253 /** per-object-layer operations */
254 const struct cl_object_operations *co_ops;
255 /** offset of page slice in cl_page buffer */
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;
283 * Layout lock handle.
285 struct ldlm_lock *coc_lock;
287 * Operation to handle layout, OBJECT_CONF_XYZ.
293 /** configure layout, set up a new stripe, must be called while
294 * holding layout lock.
297 /** 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 set up. */
306 * Operations implemented for each cl object layer.
308 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
310 struct cl_object_operations {
312 * Initialize page slice for this layer. Called top-to-bottom through
313 * every object layer when a new cl_page is instantiated. Layer
314 * keeping private per-page data, or requiring its own page operations
315 * vector should allocate these data here, and attach then to the page
316 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
319 * \retval NULL success.
321 * \retval ERR_PTR(errno) failure code.
323 * \retval valid-pointer pointer to already existing referenced page
324 * to be used instead of newly created.
326 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
327 struct cl_page *page, pgoff_t index);
329 * Initialize lock slice for this layer. Called top-to-bottom through
330 * every object layer when a new cl_lock is instantiated. Layer
331 * keeping private per-lock data, or requiring its own lock operations
332 * vector should allocate these data here, and attach then to the lock
333 * by calling cl_lock_slice_add(). Mandatory.
335 int (*coo_lock_init)(const struct lu_env *env,
336 struct cl_object *obj, struct cl_lock *lock,
337 const struct cl_io *io);
339 * Initialize io state for a given layer.
341 * called top-to-bottom once per io existence to initialize io
342 * state. If layer wants to keep some state for this type of io, it
343 * has to embed struct cl_io_slice in lu_env::le_ses, and register
344 * slice with cl_io_slice_add(). It is guaranteed that all threads
345 * participating in this io share the same session.
347 int (*coo_io_init)(const struct lu_env *env,
348 struct cl_object *obj, struct cl_io *io);
350 * Fill portion of \a attr that this layer controls. This method is
351 * called top-to-bottom through all object layers.
353 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
355 * \return 0: to continue
356 * \return +ve: to stop iterating through layers (but 0 is returned
357 * from enclosing cl_object_attr_get())
358 * \return -ve: to signal error
360 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
361 struct cl_attr *attr);
365 * \a valid is a bitmask composed from enum #cl_attr_valid, and
366 * indicating what attributes are to be set.
368 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
370 * \return the same convention as for
371 * cl_object_operations::coo_attr_get() is used.
373 int (*coo_attr_set)(const struct lu_env *env, struct cl_object *obj,
374 const struct cl_attr *attr, unsigned valid);
376 * Update object configuration. Called top-to-bottom to modify object
379 * XXX error conditions and handling.
381 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
382 const struct cl_object_conf *conf);
384 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
385 * object. Layers are supposed to fill parts of \a lvb that will be
386 * shipped to the glimpse originator as a glimpse result.
388 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
389 * \see osc_object_glimpse()
391 int (*coo_glimpse)(const struct lu_env *env,
392 const struct cl_object *obj, struct ost_lvb *lvb);
394 * Object prune method. Called when the layout is going to change on
395 * this object, therefore each layer has to clean up their cache,
396 * mainly pages and locks.
398 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
402 * Extended header for client object.
404 struct cl_object_header {
405 /** Standard lu_object_header. cl_object::co_lu::lo_header points
408 struct lu_object_header coh_lu;
411 * Parent object. It is assumed that an object has a well-defined
412 * parent, but not a well-defined child (there may be multiple
413 * sub-objects, for the same top-object). cl_object_header::coh_parent
414 * field allows certain code to be written generically, without
415 * limiting possible cl_object layouts unduly.
417 struct cl_object_header *coh_parent;
419 * Protects consistency between cl_attr of parent object and
420 * attributes of sub-objects, that the former is calculated ("merged")
423 * \todo XXX this can be read/write lock if needed.
425 spinlock_t coh_attr_guard;
427 * Size of cl_page + page slices
429 unsigned short coh_page_bufsize;
431 * Number of objects above this one: 0 for a top-object, 1 for its
434 unsigned char coh_nesting;
438 * Helper macro: iterate over all layers of the object \a obj, assigning every
439 * layer top-to-bottom to \a slice.
441 #define cl_object_for_each(slice, obj) \
442 list_for_each_entry((slice), \
443 &(obj)->co_lu.lo_header->loh_layers, \
446 * Helper macro: iterate over all layers of the object \a obj, assigning every
447 * layer bottom-to-top to \a slice.
449 #define cl_object_for_each_reverse(slice, obj) \
450 list_for_each_entry_reverse((slice), \
451 &(obj)->co_lu.lo_header->loh_layers, \
455 #define CL_PAGE_EOF ((pgoff_t)~0ull)
457 /** \addtogroup cl_page cl_page
462 * Layered client page.
464 * cl_page: represents a portion of a file, cached in the memory. All pages
465 * of the given file are of the same size, and are kept in the radix tree
466 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
467 * of the top-level file object are first class cl_objects, they have their
468 * own radix trees of pages and hence page is implemented as a sequence of
469 * struct cl_pages's, linked into double-linked list through
470 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
471 * corresponding radix tree at the corresponding logical offset.
473 * cl_page is associated with VM page of the hosting environment (struct
474 * page in Linux kernel, for example), struct page. It is assumed, that this
475 * association is implemented by one of cl_page layers (top layer in the
476 * current design) that
478 * - intercepts per-VM-page call-backs made by the environment (e.g.,
481 * - translates state (page flag bits) and locking between lustre and
484 * The association between cl_page and struct page is immutable and
485 * established when cl_page is created.
487 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
488 * this io an exclusive access to this page w.r.t. other io attempts and
489 * various events changing page state (such as transfer completion, or
490 * eviction of the page from the memory). Note, that in general cl_io
491 * cannot be identified with a particular thread, and page ownership is not
492 * exactly equal to the current thread holding a lock on the page. Layer
493 * implementing association between cl_page and struct page has to implement
494 * ownership on top of available synchronization mechanisms.
496 * While lustre client maintains the notion of an page ownership by io,
497 * hosting MM/VM usually has its own page concurrency control
498 * mechanisms. For example, in Linux, page access is synchronized by the
499 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
500 * takes care to acquire and release such locks as necessary around the
501 * calls to the file system methods (->readpage(), ->prepare_write(),
502 * ->commit_write(), etc.). This leads to the situation when there are two
503 * different ways to own a page in the client:
505 * - client code explicitly and voluntary owns the page (cl_page_own());
507 * - VM locks a page and then calls the client, that has "to assume"
508 * the ownership from the VM (cl_page_assume()).
510 * Dual methods to release ownership are cl_page_disown() and
511 * cl_page_unassume().
513 * cl_page is reference counted (cl_page::cp_ref). When reference counter
514 * drops to 0, the page is returned to the cache, unless it is in
515 * cl_page_state::CPS_FREEING state, in which case it is immediately
518 * The general logic guaranteeing the absence of "existential races" for
519 * pages is the following:
521 * - there are fixed known ways for a thread to obtain a new reference
524 * - by doing a lookup in the cl_object radix tree, protected by the
527 * - by starting from VM-locked struct page and following some
528 * hosting environment method (e.g., following ->private pointer in
529 * the case of Linux kernel), see cl_vmpage_page();
531 * - when the page enters cl_page_state::CPS_FREEING state, all these
532 * ways are severed with the proper synchronization
533 * (cl_page_delete());
535 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
538 * - no new references to the page in cl_page_state::CPS_FREEING state
539 * are allowed (checked in cl_page_get()).
541 * Together this guarantees that when last reference to a
542 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
543 * page, as neither references to it can be acquired at that point, nor
546 * cl_page is a state machine. States are enumerated in enum
547 * cl_page_state. Possible state transitions are enumerated in
548 * cl_page_state_set(). State transition process (i.e., actual changing of
549 * cl_page::cp_state field) is protected by the lock on the underlying VM
552 * Linux Kernel implementation.
554 * Binding between cl_page and struct page (which is a typedef for
555 * struct page) is implemented in the vvp layer. cl_page is attached to the
556 * ->private pointer of the struct page, together with the setting of
557 * PG_private bit in page->flags, and acquiring additional reference on the
558 * struct page (much like struct buffer_head, or any similar file system
559 * private data structures).
561 * PG_locked lock is used to implement both ownership and transfer
562 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
563 * states. No additional references are acquired for the duration of the
566 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
567 * write-out is "protected" by the special PG_writeback bit.
571 * States of cl_page. cl_page.c assumes particular order here.
573 * The page state machine is rather crude, as it doesn't recognize finer page
574 * states like "dirty" or "up to date". This is because such states are not
575 * always well defined for the whole stack (see, for example, the
576 * implementation of the read-ahead, that hides page up-to-dateness to track
577 * cache hits accurately). Such sub-states are maintained by the layers that
578 * are interested in them.
582 * Page is in the cache, un-owned. Page leaves cached state in the
585 * - [cl_page_state::CPS_OWNED] io comes across the page and
588 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
589 * req-formation engine decides that it wants to include this page
590 * into an cl_req being constructed, and yanks it from the cache;
592 * - [cl_page_state::CPS_FREEING] VM callback is executed to
593 * evict the page form the memory;
595 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
599 * Page is exclusively owned by some cl_io. Page may end up in this
600 * state as a result of
602 * - io creating new page and immediately owning it;
604 * - [cl_page_state::CPS_CACHED] io finding existing cached page
607 * - [cl_page_state::CPS_OWNED] io finding existing owned page
608 * and waiting for owner to release the page;
610 * Page leaves owned state in the following cases:
612 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
613 * the cache, doing nothing;
615 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
618 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
619 * transfer for this page;
621 * - [cl_page_state::CPS_FREEING] io decides to destroy this
622 * page (e.g., as part of truncate or extent lock cancellation).
624 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
628 * Page is being written out, as a part of a transfer. This state is
629 * entered when req-formation logic decided that it wants this page to
630 * be sent through the wire _now_. Specifically, it means that once
631 * this state is achieved, transfer completion handler (with either
632 * success or failure indication) is guaranteed to be executed against
633 * this page independently of any locks and any scheduling decisions
634 * made by the hosting environment (that effectively means that the
635 * page is never put into cl_page_state::CPS_PAGEOUT state "in
636 * advance". This property is mentioned, because it is important when
637 * reasoning about possible dead-locks in the system). The page can
638 * enter this state as a result of
640 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
641 * write-out of this page, or
643 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
644 * that it has enough dirty pages cached to issue a "good"
647 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
648 * is completed---it is moved into cl_page_state::CPS_CACHED state.
650 * Underlying VM page is locked for the duration of transfer.
652 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
656 * Page is being read in, as a part of a transfer. This is quite
657 * similar to the cl_page_state::CPS_PAGEOUT state, except that
658 * read-in is always "immediate"---there is no such thing a sudden
659 * construction of read cl_req from cached, presumably not up to date,
662 * Underlying VM page is locked for the duration of transfer.
664 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
668 * Page is being destroyed. This state is entered when client decides
669 * that page has to be deleted from its host object, as, e.g., a part
672 * Once this state is reached, there is no way to escape it.
674 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
681 /** Host page, the page is from the host inode which the cl_page
686 /** Transient page, the transient cl_page is used to bind a cl_page
687 * to vmpage which is not belonging to the same object of cl_page.
688 * it is used in DirectIO and lockless IO.
694 * Flags maintained for every cl_page.
698 * Set when pagein completes. Used for debugging (read completes at
699 * most once for a page).
701 CPF_READ_COMPLETED = 1 << 0
705 * Fields are protected by the lock on struct page, except for atomics and
708 * \invariant Data type invariants are in cl_page_invariant(). Basically:
709 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
710 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
711 * cl_page::cp_owner (when set).
714 /** Reference counter. */
716 /** An object this page is a part of. Immutable after creation. */
717 struct cl_object *cp_obj;
718 /** List of slices. Immutable after creation. */
719 struct list_head cp_layers;
721 struct page *cp_vmpage;
723 * Page state. This field is const to avoid accidental update, it is
724 * modified only internally within cl_page.c. Protected by a VM lock.
726 const enum cl_page_state cp_state;
727 /** Linkage of pages within group. Protected by cl_page::cp_mutex. */
728 struct list_head cp_batch;
729 /** Mutex serializing membership of a page in a batch. */
730 struct mutex cp_mutex;
731 /** Linkage of pages within cl_req. */
732 struct list_head cp_flight;
733 /** Transfer error. */
737 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
740 enum cl_page_type cp_type;
743 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
744 * by sub-io. Protected by a VM lock.
746 struct cl_io *cp_owner;
748 * Debug information, the task is owning the page.
750 struct task_struct *cp_task;
752 * Owning IO request in cl_page_state::CPS_PAGEOUT and
753 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
754 * the top-level pages. Protected by a VM lock.
756 struct cl_req *cp_req;
757 /** List of references to this page, for debugging. */
758 struct lu_ref cp_reference;
759 /** Link to an object, for debugging. */
760 struct lu_ref_link cp_obj_ref;
761 /** Link to a queue, for debugging. */
762 struct lu_ref_link cp_queue_ref;
763 /** Per-page flags from enum cl_page_flags. Protected by a VM lock. */
765 /** Assigned if doing a sync_io */
766 struct cl_sync_io *cp_sync_io;
770 * Per-layer part of cl_page.
772 * \see vvp_page, lov_page, osc_page
774 struct cl_page_slice {
775 struct cl_page *cpl_page;
778 * Object slice corresponding to this page slice. Immutable after
781 struct cl_object *cpl_obj;
782 const struct cl_page_operations *cpl_ops;
783 /** Linkage into cl_page::cp_layers. Immutable after creation. */
784 struct list_head cpl_linkage;
788 * Lock mode. For the client extent locks.
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<->struct page 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 * Called when \a io acquires this page into the exclusive
826 * ownership. When this method returns, it is guaranteed that the is
827 * not owned by other io, and no transfer is going on against
831 * \see vvp_page_own(), lov_page_own()
833 int (*cpo_own)(const struct lu_env *env,
834 const struct cl_page_slice *slice,
835 struct cl_io *io, int nonblock);
836 /** Called when ownership it yielded. Optional.
838 * \see cl_page_disown()
839 * \see vvp_page_disown()
841 void (*cpo_disown)(const struct lu_env *env,
842 const struct cl_page_slice *slice, struct cl_io *io);
844 * Called for a page that is already "owned" by \a io from VM point of
847 * \see cl_page_assume()
848 * \see vvp_page_assume(), lov_page_assume()
850 void (*cpo_assume)(const struct lu_env *env,
851 const struct cl_page_slice *slice, struct cl_io *io);
852 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
853 * bottom-to-top when IO releases a page without actually unlocking
856 * \see cl_page_unassume()
857 * \see vvp_page_unassume()
859 void (*cpo_unassume)(const struct lu_env *env,
860 const struct cl_page_slice *slice,
863 * Announces whether the page contains valid data or not by \a uptodate.
865 * \see cl_page_export()
866 * \see vvp_page_export()
868 void (*cpo_export)(const struct lu_env *env,
869 const struct cl_page_slice *slice, int uptodate);
871 * Checks whether underlying VM page is locked (in the suitable
872 * sense). Used for assertions.
874 * \retval -EBUSY: page is protected by a lock of a given mode;
875 * \retval -ENODATA: page is not protected by a lock;
876 * \retval 0: this layer cannot decide. (Should never happen.)
878 int (*cpo_is_vmlocked)(const struct lu_env *env,
879 const struct cl_page_slice *slice);
885 * Called when page is truncated from the object. Optional.
887 * \see cl_page_discard()
888 * \see vvp_page_discard(), osc_page_discard()
890 void (*cpo_discard)(const struct lu_env *env,
891 const struct cl_page_slice *slice,
894 * Called when page is removed from the cache, and is about to being
895 * destroyed. Optional.
897 * \see cl_page_delete()
898 * \see vvp_page_delete(), osc_page_delete()
900 void (*cpo_delete)(const struct lu_env *env,
901 const struct cl_page_slice *slice);
902 /** Destructor. Frees resources and slice itself. */
903 void (*cpo_fini)(const struct lu_env *env,
904 struct cl_page_slice *slice);
907 * Checks whether the page is protected by a cl_lock. This is a
908 * per-layer method, because certain layers have ways to check for the
909 * lock much more efficiently than through the generic locks scan, or
910 * implement locking mechanisms separate from cl_lock, e.g.,
911 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
912 * being canceled, or scheduled for cancellation as soon as the last
913 * user goes away, too.
915 * \retval -EBUSY: page is protected by a lock of a given mode;
916 * \retval -ENODATA: page is not protected by a lock;
917 * \retval 0: this layer cannot decide.
919 * \see cl_page_is_under_lock()
921 int (*cpo_is_under_lock)(const struct lu_env *env,
922 const struct cl_page_slice *slice,
923 struct cl_io *io, pgoff_t *max);
926 * Optional debugging helper. Prints given page slice.
928 * \see cl_page_print()
930 int (*cpo_print)(const struct lu_env *env,
931 const struct cl_page_slice *slice,
932 void *cookie, lu_printer_t p);
936 * Transfer methods. See comment on cl_req for a description of
937 * transfer formation and life-cycle.
942 * Request type dependent vector of operations.
944 * Transfer operations depend on transfer mode (cl_req_type). To avoid
945 * passing transfer mode to each and every of these methods, and to
946 * avoid branching on request type inside of the methods, separate
947 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
948 * provided. That is, method invocation usually looks like
950 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
954 * Called when a page is submitted for a transfer as a part of
957 * \return 0 : page is eligible for submission;
958 * \return -EALREADY : skip this page;
959 * \return -ve : error.
961 * \see cl_page_prep()
963 int (*cpo_prep)(const struct lu_env *env,
964 const struct cl_page_slice *slice,
967 * Completion handler. This is guaranteed to be eventually
968 * fired after cl_page_operations::cpo_prep() or
969 * cl_page_operations::cpo_make_ready() call.
971 * This method can be called in a non-blocking context. It is
972 * guaranteed however, that the page involved and its object
973 * are pinned in memory (and, hence, calling cl_page_put() is
976 * \see cl_page_completion()
978 void (*cpo_completion)(const struct lu_env *env,
979 const struct cl_page_slice *slice,
982 * Called when cached page is about to be added to the
983 * cl_req as a part of req formation.
985 * \return 0 : proceed with this page;
986 * \return -EAGAIN : skip this page;
987 * \return -ve : error.
989 * \see cl_page_make_ready()
991 int (*cpo_make_ready)(const struct lu_env *env,
992 const struct cl_page_slice *slice);
995 * Tell transfer engine that only [to, from] part of a page should be
998 * This is used for immediate transfers.
1000 * \todo XXX this is not very good interface. It would be much better
1001 * if all transfer parameters were supplied as arguments to
1002 * cl_io_operations::cio_submit() call, but it is not clear how to do
1003 * this for page queues.
1005 * \see cl_page_clip()
1007 void (*cpo_clip)(const struct lu_env *env,
1008 const struct cl_page_slice *slice,
1011 * \pre the page was queued for transferring.
1012 * \post page is removed from client's pending list, or -EBUSY
1013 * is returned if it has already been in transferring.
1015 * This is one of seldom page operation which is:
1016 * 0. called from top level;
1017 * 1. don't have vmpage locked;
1018 * 2. every layer should synchronize execution of its ->cpo_cancel()
1019 * with completion handlers. Osc uses client obd lock for this
1020 * purpose. Based on there is no vvp_page_cancel and
1021 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1023 * \see osc_page_cancel().
1025 int (*cpo_cancel)(const struct lu_env *env,
1026 const struct cl_page_slice *slice);
1028 * Write out a page by kernel. This is only called by ll_writepage
1031 * \see cl_page_flush()
1033 int (*cpo_flush)(const struct lu_env *env,
1034 const struct cl_page_slice *slice,
1040 * Helper macro, dumping detailed information about \a page into a log.
1042 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1044 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1045 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1046 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1047 CDEBUG(mask, format, ## __VA_ARGS__); \
1052 * Helper macro, dumping shorter information about \a page into a log.
1054 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1056 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1057 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1058 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1059 CDEBUG(mask, format, ## __VA_ARGS__); \
1063 static inline int __page_in_use(const struct cl_page *page, int refc)
1065 if (page->cp_type == CPT_CACHEABLE)
1067 LASSERT(atomic_read(&page->cp_ref) > 0);
1068 return (atomic_read(&page->cp_ref) > refc);
1071 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1072 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1074 static inline struct page *cl_page_vmpage(struct cl_page *page)
1076 LASSERT(page->cp_vmpage);
1077 return page->cp_vmpage;
1082 /** \addtogroup cl_lock cl_lock
1087 * Extent locking on the client.
1091 * The locking model of the new client code is built around
1095 * data-type representing an extent lock on a regular file. cl_lock is a
1096 * layered object (much like cl_object and cl_page), it consists of a header
1097 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1098 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1100 * Typical cl_lock consists of the two layers:
1102 * - vvp_lock (vvp specific data), and
1103 * - lov_lock (lov specific data).
1105 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1106 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1108 * - lovsub_lock, and
1111 * Each sub-lock is associated with a cl_object (representing stripe
1112 * sub-object or the file to which top-level cl_lock is associated to), and is
1113 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1114 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1115 * is different from cl_page, that doesn't fan out (there is usually exactly
1116 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1117 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1121 * cl_lock is a cacheless data container for the requirements of locks to
1122 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1125 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1126 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1128 * INTERFACE AND USAGE
1130 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1131 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1132 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1133 * consists of multiple sub cl_locks, each sub locks will be enqueued
1134 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1135 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1138 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1139 * method will be called for each layer to release the resource held by this
1140 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1141 * clo_enqueue time, is released.
1143 * LDLM lock can only be canceled if there is no cl_lock using it.
1145 * Overall process of the locking during IO operation is as following:
1147 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1148 * is called on each layer. Responsibility of this method is to add locks,
1149 * needed by a given layer into cl_io.ci_lockset.
1151 * - once locks for all layers were collected, they are sorted to avoid
1152 * dead-locks (cl_io_locks_sort()), and enqueued.
1154 * - when all locks are acquired, IO is performed;
1156 * - locks are released after IO is complete.
1158 * Striping introduces major additional complexity into locking. The
1159 * fundamental problem is that it is generally unsafe to actively use (hold)
1160 * two locks on the different OST servers at the same time, as this introduces
1161 * inter-server dependency and can lead to cascading evictions.
1163 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1164 * that no multi-stripe locks are taken (note that this design abandons POSIX
1165 * read/write semantics). Such pieces ideally can be executed concurrently. At
1166 * the same time, certain types of IO cannot be sub-divived, without
1167 * sacrificing correctness. This includes:
1169 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1172 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1174 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1175 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1176 * has to be held together with the usual lock on [offset, offset + count].
1178 * Interaction with DLM
1180 * In the expected setup, cl_lock is ultimately backed up by a collection of
1181 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1182 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1183 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1184 * description of interaction with DLM.
1190 struct cl_lock_descr {
1191 /** Object this lock is granted for. */
1192 struct cl_object *cld_obj;
1193 /** Index of the first page protected by this lock. */
1195 /** Index of the last page (inclusive) protected by this lock. */
1197 /** Group ID, for group lock */
1200 enum cl_lock_mode cld_mode;
1202 * flags to enqueue lock. A combination of bit-flags from
1203 * enum cl_enq_flags.
1205 __u32 cld_enq_flags;
1208 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1209 #define PDESCR(descr) \
1210 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1211 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1213 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1216 * Layered client lock.
1219 /** List of slices. Immutable after creation. */
1220 struct list_head cll_layers;
1221 /** lock attribute, extent, cl_object, etc. */
1222 struct cl_lock_descr cll_descr;
1226 * Per-layer part of cl_lock
1228 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1230 struct cl_lock_slice {
1231 struct cl_lock *cls_lock;
1232 /** Object slice corresponding to this lock slice. Immutable after
1235 struct cl_object *cls_obj;
1236 const struct cl_lock_operations *cls_ops;
1237 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1238 struct list_head cls_linkage;
1243 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1245 struct cl_lock_operations {
1248 * Attempts to enqueue the lock. Called top-to-bottom.
1250 * \retval 0 this layer has enqueued the lock successfully
1251 * \retval >0 this layer has enqueued the lock, but need to wait on
1252 * @anchor for resources
1253 * \retval -ve failure
1255 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1256 * \see osc_lock_enqueue()
1258 int (*clo_enqueue)(const struct lu_env *env,
1259 const struct cl_lock_slice *slice,
1260 struct cl_io *io, struct cl_sync_io *anchor);
1262 * Cancel a lock, release its DLM lock ref, while does not cancel the
1265 void (*clo_cancel)(const struct lu_env *env,
1266 const struct cl_lock_slice *slice);
1269 * Destructor. Frees resources and the slice.
1271 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1272 * \see osc_lock_fini()
1274 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1276 * Optional debugging helper. Prints given lock slice.
1278 int (*clo_print)(const struct lu_env *env,
1279 void *cookie, lu_printer_t p,
1280 const struct cl_lock_slice *slice);
1283 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1285 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1287 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1288 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1289 CDEBUG(mask, format, ## __VA_ARGS__); \
1293 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1297 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1303 /** \addtogroup cl_page_list cl_page_list
1304 * Page list used to perform collective operations on a group of pages.
1306 * Pages are added to the list one by one. cl_page_list acquires a reference
1307 * for every page in it. Page list is used to perform collective operations on
1310 * - submit pages for an immediate transfer,
1312 * - own pages on behalf of certain io (waiting for each page in turn),
1316 * When list is finalized, it releases references on all pages it still has.
1318 * \todo XXX concurrency control.
1322 struct cl_page_list {
1324 struct list_head pl_pages;
1325 struct task_struct *pl_owner;
1329 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1330 * contains an incoming page list and an outgoing page list.
1333 struct cl_page_list c2_qin;
1334 struct cl_page_list c2_qout;
1337 /** @} cl_page_list */
1339 /** \addtogroup cl_io cl_io
1345 * cl_io represents a high level I/O activity like
1346 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1349 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1350 * important distinction. We want to minimize number of calls to the allocator
1351 * in the fast path, e.g., in the case of read(2) when everything is cached:
1352 * client already owns the lock over region being read, and data are cached
1353 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1354 * per-layer io state is stored in the session, associated with the io, see
1355 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1356 * by using free-lists, see cl_env_get().
1358 * There is a small predefined number of possible io types, enumerated in enum
1361 * cl_io is a state machine, that can be advanced concurrently by the multiple
1362 * threads. It is up to these threads to control the concurrency and,
1363 * specifically, to detect when io is done, and its state can be safely
1366 * For read/write io overall execution plan is as following:
1368 * (0) initialize io state through all layers;
1370 * (1) loop: prepare chunk of work to do
1372 * (2) call all layers to collect locks they need to process current chunk
1374 * (3) sort all locks to avoid dead-locks, and acquire them
1376 * (4) process the chunk: call per-page methods
1377 * (cl_io_operations::cio_read_page() for read,
1378 * cl_io_operations::cio_prepare_write(),
1379 * cl_io_operations::cio_commit_write() for write)
1385 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1386 * address allocation efficiency issues mentioned above), and returns with the
1387 * special error condition from per-page method when current sub-io has to
1388 * block. This causes io loop to be repeated, and lov switches to the next
1389 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1394 /** read system call */
1396 /** write system call */
1398 /** truncate, utime system calls */
1401 * page fault handling
1405 * fsync system call handling
1406 * To write out a range of file
1410 * Miscellaneous io. This is used for occasional io activity that
1411 * doesn't fit into other types. Currently this is used for:
1413 * - cancellation of an extent lock. This io exists as a context
1414 * to write dirty pages from under the lock being canceled back
1417 * - VM induced page write-out. An io context for writing page out
1418 * for memory cleansing;
1420 * - glimpse. An io context to acquire glimpse lock.
1422 * - grouplock. An io context to acquire group lock.
1424 * CIT_MISC io is used simply as a context in which locks and pages
1425 * are manipulated. Such io has no internal "process", that is,
1426 * cl_io_loop() is never called for it.
1433 * States of cl_io state machine
1436 /** Not initialized. */
1440 /** IO iteration started. */
1444 /** Actual IO is in progress. */
1446 /** IO for the current iteration finished. */
1448 /** Locks released. */
1450 /** Iteration completed. */
1452 /** cl_io finalized. */
1457 * IO state private for a layer.
1459 * This is usually embedded into layer session data, rather than allocated
1462 * \see vvp_io, lov_io, osc_io
1464 struct cl_io_slice {
1465 struct cl_io *cis_io;
1466 /** corresponding object slice. Immutable after creation. */
1467 struct cl_object *cis_obj;
1468 /** io operations. Immutable after creation. */
1469 const struct cl_io_operations *cis_iop;
1471 * linkage into a list of all slices for a given cl_io, hanging off
1472 * cl_io::ci_layers. Immutable after creation.
1474 struct list_head cis_linkage;
1477 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1480 * Per-layer io operations.
1481 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1483 struct cl_io_operations {
1485 * Vector of io state transition methods for every io type.
1487 * \see cl_page_operations::io
1491 * Prepare io iteration at a given layer.
1493 * Called top-to-bottom at the beginning of each iteration of
1494 * "io loop" (if it makes sense for this type of io). Here
1495 * layer selects what work it will do during this iteration.
1497 * \see cl_io_operations::cio_iter_fini()
1499 int (*cio_iter_init)(const struct lu_env *env,
1500 const struct cl_io_slice *slice);
1502 * Finalize io iteration.
1504 * Called bottom-to-top at the end of each iteration of "io
1505 * loop". Here layers can decide whether IO has to be
1508 * \see cl_io_operations::cio_iter_init()
1510 void (*cio_iter_fini)(const struct lu_env *env,
1511 const struct cl_io_slice *slice);
1513 * Collect locks for the current iteration of io.
1515 * Called top-to-bottom to collect all locks necessary for
1516 * this iteration. This methods shouldn't actually enqueue
1517 * anything, instead it should post a lock through
1518 * cl_io_lock_add(). Once all locks are collected, they are
1519 * sorted and enqueued in the proper order.
1521 int (*cio_lock)(const struct lu_env *env,
1522 const struct cl_io_slice *slice);
1524 * Finalize unlocking.
1526 * Called bottom-to-top to finish layer specific unlocking
1527 * functionality, after generic code released all locks
1528 * acquired by cl_io_operations::cio_lock().
1530 void (*cio_unlock)(const struct lu_env *env,
1531 const struct cl_io_slice *slice);
1533 * Start io iteration.
1535 * Once all locks are acquired, called top-to-bottom to
1536 * commence actual IO. In the current implementation,
1537 * top-level vvp_io_{read,write}_start() does all the work
1538 * synchronously by calling generic_file_*(), so other layers
1539 * are called when everything is done.
1541 int (*cio_start)(const struct lu_env *env,
1542 const struct cl_io_slice *slice);
1544 * Called top-to-bottom at the end of io loop. Here layer
1545 * might wait for an unfinished asynchronous io.
1547 void (*cio_end)(const struct lu_env *env,
1548 const struct cl_io_slice *slice);
1550 * Called bottom-to-top to notify layers that read/write IO
1551 * iteration finished, with \a nob bytes transferred.
1553 void (*cio_advance)(const struct lu_env *env,
1554 const struct cl_io_slice *slice,
1557 * Called once per io, bottom-to-top to release io resources.
1559 void (*cio_fini)(const struct lu_env *env,
1560 const struct cl_io_slice *slice);
1564 * Submit pages from \a queue->c2_qin for IO, and move
1565 * successfully submitted pages into \a queue->c2_qout. Return
1566 * non-zero if failed to submit even the single page. If
1567 * submission failed after some pages were moved into \a
1568 * queue->c2_qout, completion callback with non-zero ioret is
1571 int (*cio_submit)(const struct lu_env *env,
1572 const struct cl_io_slice *slice,
1573 enum cl_req_type crt,
1574 struct cl_2queue *queue);
1576 * Queue async page for write.
1577 * The difference between cio_submit and cio_queue is that
1578 * cio_submit is for urgent request.
1580 int (*cio_commit_async)(const struct lu_env *env,
1581 const struct cl_io_slice *slice,
1582 struct cl_page_list *queue, int from, int to,
1585 * Read missing page.
1587 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
1588 * method, when it hits not-up-to-date page in the range. Optional.
1590 * \pre io->ci_type == CIT_READ
1592 int (*cio_read_page)(const struct lu_env *env,
1593 const struct cl_io_slice *slice,
1594 const struct cl_page_slice *page);
1596 * Optional debugging helper. Print given io slice.
1598 int (*cio_print)(const struct lu_env *env, void *cookie,
1599 lu_printer_t p, const struct cl_io_slice *slice);
1603 * Flags to lock enqueue procedure.
1608 * instruct server to not block, if conflicting lock is found. Instead
1609 * -EWOULDBLOCK is returned immediately.
1611 CEF_NONBLOCK = 0x00000001,
1613 * take lock asynchronously (out of order), as it cannot
1614 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1616 CEF_ASYNC = 0x00000002,
1618 * tell the server to instruct (though a flag in the blocking ast) an
1619 * owner of the conflicting lock, that it can drop dirty pages
1620 * protected by this lock, without sending them to the server.
1622 CEF_DISCARD_DATA = 0x00000004,
1624 * tell the sub layers that it must be a `real' lock. This is used for
1625 * mmapped-buffer locks and glimpse locks that must be never converted
1626 * into lockless mode.
1628 * \see vvp_mmap_locks(), cl_glimpse_lock().
1630 CEF_MUST = 0x00000008,
1632 * tell the sub layers that never request a `real' lock. This flag is
1633 * not used currently.
1635 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1636 * conversion policy: ci_lockreq describes generic information of lock
1637 * requirement for this IO, especially for locks which belong to the
1638 * object doing IO; however, lock itself may have precise requirements
1639 * that are described by the enqueue flags.
1641 CEF_NEVER = 0x00000010,
1643 * for async glimpse lock.
1645 CEF_AGL = 0x00000020,
1647 * enqueue a lock to test DLM lock existence.
1649 CEF_PEEK = 0x00000040,
1651 * mask of enq_flags.
1653 CEF_MASK = 0x0000007f,
1657 * Link between lock and io. Intermediate structure is needed, because the
1658 * same lock can be part of multiple io's simultaneously.
1660 struct cl_io_lock_link {
1661 /** linkage into one of cl_lockset lists. */
1662 struct list_head cill_linkage;
1663 struct cl_lock cill_lock;
1664 /** optional destructor */
1665 void (*cill_fini)(const struct lu_env *env,
1666 struct cl_io_lock_link *link);
1668 #define cill_descr cill_lock.cll_descr
1671 * Lock-set represents a collection of locks, that io needs at a
1672 * time. Generally speaking, client tries to avoid holding multiple locks when
1675 * - holding extent locks over multiple ost's introduces the danger of
1676 * "cascading timeouts";
1678 * - holding multiple locks over the same ost is still dead-lock prone,
1679 * see comment in osc_lock_enqueue(),
1681 * but there are certain situations where this is unavoidable:
1683 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1685 * - truncate has to take [new-size, EOF] lock for correctness;
1687 * - SNS has to take locks across full stripe for correctness;
1689 * - in the case when user level buffer, supplied to {read,write}(file0),
1690 * is a part of a memory mapped lustre file, client has to take a dlm
1691 * locks on file0, and all files that back up the buffer (or a part of
1692 * the buffer, that is being processed in the current chunk, in any
1693 * case, there are situations where at least 2 locks are necessary).
1695 * In such cases we at least try to take locks in the same consistent
1696 * order. To this end, all locks are first collected, then sorted, and then
1700 /** locks to be acquired. */
1701 struct list_head cls_todo;
1702 /** locks acquired. */
1703 struct list_head cls_done;
1707 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1708 * but 'req' is always to be thought as 'request' :-)
1710 enum cl_io_lock_dmd {
1711 /** Always lock data (e.g., O_APPEND). */
1713 /** Layers are free to decide between local and global locking. */
1715 /** Never lock: there is no cache (e.g., lockless IO). */
1719 enum cl_fsync_mode {
1720 /** start writeback, do not wait for them to finish */
1722 /** start writeback and wait for them to finish */
1724 /** discard all of dirty pages in a specific file range */
1725 CL_FSYNC_DISCARD = 2,
1726 /** start writeback and make sure they have reached storage before
1727 * return. OST_SYNC RPC must be issued and finished
1732 struct cl_io_rw_common {
1741 * cl_io is shared by all threads participating in this IO (in current
1742 * implementation only one thread advances IO, but parallel IO design and
1743 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1744 * is up to these threads to serialize their activities, including updates to
1745 * mutable cl_io fields.
1748 /** type of this IO. Immutable after creation. */
1749 enum cl_io_type ci_type;
1750 /** current state of cl_io state machine. */
1751 enum cl_io_state ci_state;
1752 /** main object this io is against. Immutable after creation. */
1753 struct cl_object *ci_obj;
1755 * Upper layer io, of which this io is a part of. Immutable after
1758 struct cl_io *ci_parent;
1759 /** List of slices. Immutable after creation. */
1760 struct list_head ci_layers;
1761 /** list of locks (to be) acquired by this io. */
1762 struct cl_lockset ci_lockset;
1763 /** lock requirements, this is just a help info for sublayers. */
1764 enum cl_io_lock_dmd ci_lockreq;
1767 struct cl_io_rw_common rd;
1770 struct cl_io_rw_common wr;
1774 struct cl_io_rw_common ci_rw;
1775 struct cl_setattr_io {
1776 struct ost_lvb sa_attr;
1777 unsigned int sa_valid;
1779 struct cl_fault_io {
1780 /** page index within file. */
1782 /** bytes valid byte on a faulted page. */
1784 /** writable page? for nopage() only */
1786 /** page of an executable? */
1788 /** page_mkwrite() */
1790 /** resulting page */
1791 struct cl_page *ft_page;
1793 struct cl_fsync_io {
1796 /** file system level fid */
1797 struct lu_fid *fi_fid;
1798 enum cl_fsync_mode fi_mode;
1799 /* how many pages were written/discarded */
1800 unsigned int fi_nr_written;
1803 struct cl_2queue ci_queue;
1806 unsigned int ci_continue:1,
1808 * This io has held grouplock, to inform sublayers that
1809 * don't do lockless i/o.
1813 * The whole IO need to be restarted because layout has been changed
1817 * to not refresh layout - the IO issuer knows that the layout won't
1818 * change(page operations, layout change causes all page to be
1819 * discarded), or it doesn't matter if it changes(sync).
1823 * Check if layout changed after the IO finishes. Mainly for HSM
1824 * requirement. If IO occurs to openning files, it doesn't need to
1825 * verify layout because HSM won't release openning files.
1826 * Right now, only two operations need to verify layout: glimpse
1831 * file is released, restore has to to be triggered by vvp layer
1833 ci_restore_needed:1,
1839 * Number of pages owned by this IO. For invariant checking.
1841 unsigned ci_owned_nr;
1846 /** \addtogroup cl_req cl_req
1852 * There are two possible modes of transfer initiation on the client:
1854 * - immediate transfer: this is started when a high level io wants a page
1855 * or a collection of pages to be transferred right away. Examples:
1856 * read-ahead, synchronous read in the case of non-page aligned write,
1857 * page write-out as a part of extent lock cancellation, page write-out
1858 * as a part of memory cleansing. Immediate transfer can be both
1859 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1861 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1862 * when io wants to transfer a page to the server some time later, when
1863 * it can be done efficiently. Example: pages dirtied by the write(2)
1866 * In any case, transfer takes place in the form of a cl_req, which is a
1867 * representation for a network RPC.
1869 * Pages queued for an opportunistic transfer are cached until it is decided
1870 * that efficient RPC can be composed of them. This decision is made by "a
1871 * req-formation engine", currently implemented as a part of osc
1872 * layer. Req-formation depends on many factors: the size of the resulting
1873 * RPC, whether or not multi-object RPCs are supported by the server,
1874 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1876 * For the immediate transfer io submits a cl_page_list, that req-formation
1877 * engine slices into cl_req's, possibly adding cached pages to some of
1878 * the resulting req's.
1880 * Whenever a page from cl_page_list is added to a newly constructed req, its
1881 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1882 * page state is atomically changed from cl_page_state::CPS_OWNED to
1883 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1884 * is zeroed, and cl_page::cp_req is set to the
1885 * req. cl_page_operations::cpo_prep() method at the particular layer might
1886 * return -EALREADY to indicate that it does not need to submit this page
1887 * at all. This is possible, for example, if page, submitted for read,
1888 * became up-to-date in the meantime; and for write, the page don't have
1889 * dirty bit marked. \see cl_io_submit_rw()
1891 * Whenever a cached page is added to a newly constructed req, its
1892 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1893 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1894 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1895 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1896 * might return -EAGAIN to indicate that this page is not eligible for the
1897 * transfer right now.
1901 * Plan is to divide transfers into "priority bands" (indicated when
1902 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1903 * and allow glueing of cached pages to immediate transfers only within single
1904 * band. This would make high priority transfers (like lock cancellation or
1905 * memory pressure induced write-out) really high priority.
1910 * Per-transfer attributes.
1912 struct cl_req_attr {
1913 /** Generic attributes for the server consumption. */
1914 struct obdo *cra_oa;
1916 char cra_jobid[JOBSTATS_JOBID_SIZE];
1920 * Transfer request operations definable at every layer.
1922 * Concurrency: transfer formation engine synchronizes calls to all transfer
1925 struct cl_req_operations {
1927 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1928 * complete (all pages are added).
1930 * \see osc_req_prep()
1932 int (*cro_prep)(const struct lu_env *env,
1933 const struct cl_req_slice *slice);
1935 * Called top-to-bottom to fill in \a oa fields. This is called twice
1936 * with different flags, see bug 10150 and osc_build_req().
1938 * \param obj an object from cl_req which attributes are to be set in
1941 * \param oa struct obdo where attributes are placed
1943 * \param flags \a oa fields to be filled.
1945 void (*cro_attr_set)(const struct lu_env *env,
1946 const struct cl_req_slice *slice,
1947 const struct cl_object *obj,
1948 struct cl_req_attr *attr, u64 flags);
1950 * Called top-to-bottom from cl_req_completion() to notify layers that
1951 * transfer completed. Has to free all state allocated by
1952 * cl_device_operations::cdo_req_init().
1954 void (*cro_completion)(const struct lu_env *env,
1955 const struct cl_req_slice *slice, int ioret);
1959 * A per-object state that (potentially multi-object) transfer request keeps.
1962 /** object itself */
1963 struct cl_object *ro_obj;
1964 /** reference to cl_req_obj::ro_obj. For debugging. */
1965 struct lu_ref_link ro_obj_ref;
1966 /* something else? Number of pages for a given object? */
1972 * Transfer requests are not reference counted, because IO sub-system owns
1973 * them exclusively and knows when to free them.
1977 * cl_req is created by cl_req_alloc() that calls
1978 * cl_device_operations::cdo_req_init() device methods to allocate per-req
1979 * state in every layer.
1981 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
1982 * contains pages for.
1984 * Once all pages were collected, cl_page_operations::cpo_prep() method is
1985 * called top-to-bottom. At that point layers can modify req, let it pass, or
1986 * deny it completely. This is to support things like SNS that have transfer
1987 * ordering requirements invisible to the individual req-formation engine.
1989 * On transfer completion (or transfer timeout, or failure to initiate the
1990 * transfer of an allocated req), cl_req_operations::cro_completion() method
1991 * is called, after execution of cl_page_operations::cpo_completion() of all
1995 enum cl_req_type crq_type;
1996 /** A list of pages being transferred */
1997 struct list_head crq_pages;
1998 /** Number of pages in cl_req::crq_pages */
1999 unsigned crq_nrpages;
2000 /** An array of objects which pages are in ->crq_pages */
2001 struct cl_req_obj *crq_o;
2002 /** Number of elements in cl_req::crq_objs[] */
2003 unsigned crq_nrobjs;
2004 struct list_head crq_layers;
2008 * Per-layer state for request.
2010 struct cl_req_slice {
2011 struct cl_req *crs_req;
2012 struct cl_device *crs_dev;
2013 struct list_head crs_linkage;
2014 const struct cl_req_operations *crs_ops;
2019 enum cache_stats_item {
2020 /** how many cache lookups were performed */
2022 /** how many times cache lookup resulted in a hit */
2024 /** how many entities are in the cache right now */
2026 /** how many entities in the cache are actively used (and cannot be
2027 * evicted) right now
2030 /** how many entities were created at all */
2035 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2038 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2040 struct cache_stats {
2041 const char *cs_name;
2042 atomic_t cs_stats[CS_NR];
2045 /** These are not exported so far */
2046 void cache_stats_init(struct cache_stats *cs, const char *name);
2049 * Client-side site. This represents particular client stack. "Global"
2050 * variables should (directly or indirectly) be added here to allow multiple
2051 * clients to co-exist in the single address space.
2054 struct lu_site cs_lu;
2056 * Statistical counters. Atomics do not scale, something better like
2057 * per-cpu counters is needed.
2059 * These are exported as /sys/kernel/debug/lustre/llite/.../site
2061 * When interpreting keep in mind that both sub-locks (and sub-pages)
2062 * and top-locks (and top-pages) are accounted here.
2064 struct cache_stats cs_pages;
2065 atomic_t cs_pages_state[CPS_NR];
2068 int cl_site_init(struct cl_site *s, struct cl_device *top);
2069 void cl_site_fini(struct cl_site *s);
2070 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2073 * Output client site statistical counters into a buffer. Suitable for
2074 * ll_rd_*()-style functions.
2076 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2081 * Type conversion and accessory functions.
2085 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2087 return container_of(site, struct cl_site, cs_lu);
2090 static inline int lu_device_is_cl(const struct lu_device *d)
2092 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2095 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2097 LASSERT(!d || IS_ERR(d) || lu_device_is_cl(d));
2098 return container_of0(d, struct cl_device, cd_lu_dev);
2101 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2103 return &d->cd_lu_dev;
2106 static inline struct cl_object *lu2cl(const struct lu_object *o)
2108 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2109 return container_of0(o, struct cl_object, co_lu);
2112 static inline const struct cl_object_conf *
2113 lu2cl_conf(const struct lu_object_conf *conf)
2115 return container_of0(conf, struct cl_object_conf, coc_lu);
2118 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2120 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2123 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2125 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2126 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2129 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2131 return container_of0(h, struct cl_object_header, coh_lu);
2134 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2136 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2140 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2142 return luh2coh(obj->co_lu.lo_header);
2145 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2147 return lu_device_init(&d->cd_lu_dev, t);
2150 static inline void cl_device_fini(struct cl_device *d)
2152 lu_device_fini(&d->cd_lu_dev);
2155 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2156 struct cl_object *obj, pgoff_t index,
2157 const struct cl_page_operations *ops);
2158 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2159 struct cl_object *obj,
2160 const struct cl_lock_operations *ops);
2161 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2162 struct cl_object *obj, const struct cl_io_operations *ops);
2163 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2164 struct cl_device *dev,
2165 const struct cl_req_operations *ops);
2168 /** \defgroup cl_object cl_object
2171 struct cl_object *cl_object_top(struct cl_object *o);
2172 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2173 const struct lu_fid *fid,
2174 const struct cl_object_conf *c);
2176 int cl_object_header_init(struct cl_object_header *h);
2177 void cl_object_put(const struct lu_env *env, struct cl_object *o);
2178 void cl_object_get(struct cl_object *o);
2179 void cl_object_attr_lock(struct cl_object *o);
2180 void cl_object_attr_unlock(struct cl_object *o);
2181 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2182 struct cl_attr *attr);
2183 int cl_object_attr_set(const struct lu_env *env, struct cl_object *obj,
2184 const struct cl_attr *attr, unsigned valid);
2185 int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj,
2186 struct ost_lvb *lvb);
2187 int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
2188 const struct cl_object_conf *conf);
2189 int cl_object_prune(const struct lu_env *env, struct cl_object *obj);
2190 void cl_object_kill(const struct lu_env *env, struct cl_object *obj);
2193 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2195 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2197 return cl_object_header(o0) == cl_object_header(o1);
2200 static inline void cl_object_page_init(struct cl_object *clob, int size)
2202 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2203 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2206 static inline void *cl_object_page_slice(struct cl_object *clob,
2207 struct cl_page *page)
2209 return (void *)((char *)page + clob->co_slice_off);
2213 * Return refcount of cl_object.
2215 static inline int cl_object_refc(struct cl_object *clob)
2217 struct lu_object_header *header = clob->co_lu.lo_header;
2219 return atomic_read(&header->loh_ref);
2224 /** \defgroup cl_page cl_page
2234 /* callback of cl_page_gang_lookup() */
2235 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *obj,
2236 pgoff_t idx, struct page *vmpage,
2237 enum cl_page_type type);
2238 struct cl_page *cl_page_alloc(const struct lu_env *env,
2239 struct cl_object *o, pgoff_t ind,
2240 struct page *vmpage,
2241 enum cl_page_type type);
2242 void cl_page_get(struct cl_page *page);
2243 void cl_page_put(const struct lu_env *env, struct cl_page *page);
2244 void cl_page_print(const struct lu_env *env, void *cookie, lu_printer_t printer,
2245 const struct cl_page *pg);
2246 void cl_page_header_print(const struct lu_env *env, void *cookie,
2247 lu_printer_t printer, const struct cl_page *pg);
2248 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj);
2250 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2251 const struct lu_device_type *dtype);
2256 * Functions dealing with the ownership of page by io.
2260 int cl_page_own(const struct lu_env *env,
2261 struct cl_io *io, struct cl_page *page);
2262 int cl_page_own_try(const struct lu_env *env,
2263 struct cl_io *io, struct cl_page *page);
2264 void cl_page_assume(const struct lu_env *env,
2265 struct cl_io *io, struct cl_page *page);
2266 void cl_page_unassume(const struct lu_env *env,
2267 struct cl_io *io, struct cl_page *pg);
2268 void cl_page_disown(const struct lu_env *env,
2269 struct cl_io *io, struct cl_page *page);
2270 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io);
2277 * Functions dealing with the preparation of a page for a transfer, and
2278 * tracking transfer state.
2281 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
2282 struct cl_page *pg, enum cl_req_type crt);
2283 void cl_page_completion(const struct lu_env *env,
2284 struct cl_page *pg, enum cl_req_type crt, int ioret);
2285 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
2286 enum cl_req_type crt);
2287 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
2288 struct cl_page *pg, enum cl_req_type crt);
2289 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
2291 int cl_page_cancel(const struct lu_env *env, struct cl_page *page);
2292 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
2293 struct cl_page *pg);
2298 * \name helper routines
2299 * Functions to discard, delete and export a cl_page.
2302 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2303 struct cl_page *pg);
2304 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2305 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg);
2306 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate);
2307 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2308 struct cl_page *page, pgoff_t *max_index);
2309 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2310 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2311 int cl_page_size(const struct cl_object *obj);
2312 int cl_pages_prune(const struct lu_env *env, struct cl_object *obj);
2314 void cl_lock_print(const struct lu_env *env, void *cookie,
2315 lu_printer_t printer, const struct cl_lock *lock);
2316 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2317 lu_printer_t printer,
2318 const struct cl_lock_descr *descr);
2322 * Data structure managing a client's cached pages. A count of
2323 * "unstable" pages is maintained, and an LRU of clean pages is
2324 * maintained. "unstable" pages are pages pinned by the ptlrpc
2325 * layer for recovery purposes.
2327 struct cl_client_cache {
2333 * # of threads are doing shrinking
2335 unsigned int ccc_lru_shrinkers;
2337 * # of LRU entries available
2339 atomic_t ccc_lru_left;
2341 * List of entities(OSCs) for this LRU cache
2343 struct list_head ccc_lru;
2345 * Max # of LRU entries
2347 unsigned long ccc_lru_max;
2349 * Lock to protect ccc_lru list
2351 spinlock_t ccc_lru_lock;
2353 * # of unstable pages for this mount point
2355 atomic_t ccc_unstable_nr;
2357 * Waitq for awaiting unstable pages to reach zero.
2358 * Used at umounting time and signaled on BRW commit
2360 wait_queue_head_t ccc_unstable_waitq;
2366 /** \defgroup cl_lock cl_lock
2370 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2371 struct cl_lock *lock);
2372 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2373 const struct cl_io *io);
2374 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2375 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2376 const struct lu_device_type *dtype);
2377 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2378 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2379 struct cl_lock *lock, struct cl_sync_io *anchor);
2380 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2384 /** \defgroup cl_io cl_io
2388 int cl_io_init(const struct lu_env *env, struct cl_io *io,
2389 enum cl_io_type iot, struct cl_object *obj);
2390 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
2391 enum cl_io_type iot, struct cl_object *obj);
2392 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
2393 enum cl_io_type iot, loff_t pos, size_t count);
2394 int cl_io_loop(const struct lu_env *env, struct cl_io *io);
2396 void cl_io_fini(const struct lu_env *env, struct cl_io *io);
2397 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io);
2398 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io);
2399 int cl_io_lock(const struct lu_env *env, struct cl_io *io);
2400 void cl_io_unlock(const struct lu_env *env, struct cl_io *io);
2401 int cl_io_start(const struct lu_env *env, struct cl_io *io);
2402 void cl_io_end(const struct lu_env *env, struct cl_io *io);
2403 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
2404 struct cl_io_lock_link *link);
2405 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2406 struct cl_lock_descr *descr);
2407 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
2408 struct cl_page *page);
2409 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
2410 enum cl_req_type iot, struct cl_2queue *queue);
2411 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
2412 enum cl_req_type iot, struct cl_2queue *queue,
2414 int cl_io_commit_async(const struct lu_env *env, struct cl_io *io,
2415 struct cl_page_list *queue, int from, int to,
2417 int cl_io_is_going(const struct lu_env *env);
2420 * True, iff \a io is an O_APPEND write(2).
2422 static inline int cl_io_is_append(const struct cl_io *io)
2424 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2427 static inline int cl_io_is_sync_write(const struct cl_io *io)
2429 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2432 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2434 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2438 * True, iff \a io is a truncate(2).
2440 static inline int cl_io_is_trunc(const struct cl_io *io)
2442 return io->ci_type == CIT_SETATTR &&
2443 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2446 struct cl_io *cl_io_top(struct cl_io *io);
2448 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2450 typeof(foo_io) __foo_io = (foo_io); \
2452 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2453 memset(&__foo_io->base + 1, 0, \
2454 sizeof(*__foo_io) - sizeof(__foo_io->base)); \
2459 /** \defgroup cl_page_list cl_page_list
2464 * Last page in the page list.
2466 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2468 LASSERT(plist->pl_nr > 0);
2469 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2472 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2474 LASSERT(plist->pl_nr > 0);
2475 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2479 * Iterate over pages in a page list.
2481 #define cl_page_list_for_each(page, list) \
2482 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2485 * Iterate over pages in a page list, taking possible removals into account.
2487 #define cl_page_list_for_each_safe(page, temp, list) \
2488 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2490 void cl_page_list_init(struct cl_page_list *plist);
2491 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page);
2492 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2493 struct cl_page *page);
2494 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2495 struct cl_page *page);
2496 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head);
2497 void cl_page_list_del(const struct lu_env *env, struct cl_page_list *plist,
2498 struct cl_page *page);
2499 void cl_page_list_disown(const struct lu_env *env,
2500 struct cl_io *io, struct cl_page_list *plist);
2501 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2503 void cl_2queue_init(struct cl_2queue *queue);
2504 void cl_2queue_disown(const struct lu_env *env,
2505 struct cl_io *io, struct cl_2queue *queue);
2506 void cl_2queue_discard(const struct lu_env *env,
2507 struct cl_io *io, struct cl_2queue *queue);
2508 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2509 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2511 /** @} cl_page_list */
2513 /** \defgroup cl_req cl_req
2516 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2517 enum cl_req_type crt, int nr_objects);
2519 void cl_req_page_add(const struct lu_env *env, struct cl_req *req,
2520 struct cl_page *page);
2521 void cl_req_page_done(const struct lu_env *env, struct cl_page *page);
2522 int cl_req_prep(const struct lu_env *env, struct cl_req *req);
2523 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2524 struct cl_req_attr *attr, u64 flags);
2525 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2527 /** \defgroup cl_sync_io cl_sync_io
2532 * Anchor for synchronous transfer. This is allocated on a stack by thread
2533 * doing synchronous transfer, and a pointer to this structure is set up in
2534 * every page submitted for transfer. Transfer completion routine updates
2535 * anchor and wakes up waiting thread when transfer is complete.
2538 /** number of pages yet to be transferred. */
2539 atomic_t csi_sync_nr;
2542 /** barrier of destroy this structure */
2543 atomic_t csi_barrier;
2544 /** completion to be signaled when transfer is complete. */
2545 wait_queue_head_t csi_waitq;
2546 /** callback to invoke when this IO is finished */
2547 void (*csi_end_io)(const struct lu_env *,
2548 struct cl_sync_io *);
2551 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2552 void (*end)(const struct lu_env *, struct cl_sync_io *));
2553 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2555 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2557 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2559 /** @} cl_sync_io */
2563 /** \defgroup cl_env cl_env
2565 * lu_env handling for a client.
2567 * lu_env is an environment within which lustre code executes. Its major part
2568 * is lu_context---a fast memory allocation mechanism that is used to conserve
2569 * precious kernel stack space. Originally lu_env was designed for a server,
2572 * - there is a (mostly) fixed number of threads, and
2574 * - call chains have no non-lustre portions inserted between lustre code.
2576 * On a client both these assumption fails, because every user thread can
2577 * potentially execute lustre code as part of a system call, and lustre calls
2578 * into VFS or MM that call back into lustre.
2580 * To deal with that, cl_env wrapper functions implement the following
2583 * - allocation and destruction of environment is amortized by caching no
2584 * longer used environments instead of destroying them;
2586 * - there is a notion of "current" environment, attached to the kernel
2587 * data structure representing current thread Top-level lustre code
2588 * allocates an environment and makes it current, then calls into
2589 * non-lustre code, that in turn calls lustre back. Low-level lustre
2590 * code thus called can fetch environment created by the top-level code
2591 * and reuse it, avoiding additional environment allocation.
2592 * Right now, three interfaces can attach the cl_env to running thread:
2595 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2597 * \see lu_env, lu_context, lu_context_key
2601 struct cl_env_nest {
2606 struct lu_env *cl_env_get(int *refcheck);
2607 struct lu_env *cl_env_alloc(int *refcheck, __u32 tags);
2608 struct lu_env *cl_env_nested_get(struct cl_env_nest *nest);
2609 void cl_env_put(struct lu_env *env, int *refcheck);
2610 void cl_env_nested_put(struct cl_env_nest *nest, struct lu_env *env);
2611 void *cl_env_reenter(void);
2612 void cl_env_reexit(void *cookie);
2613 void cl_env_implant(struct lu_env *env, int *refcheck);
2614 void cl_env_unplant(struct lu_env *env, int *refcheck);
2615 unsigned int cl_env_cache_purge(unsigned int nr);
2616 struct lu_env *cl_env_percpu_get(void);
2617 void cl_env_percpu_put(struct lu_env *env);
2624 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2626 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2627 struct lu_device_type *ldt,
2628 struct lu_device *next);
2631 int cl_global_init(void);
2632 void cl_global_fini(void);
2634 #endif /* _LINUX_CL_OBJECT_H */