4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 #ifndef _LUSTRE_CL_OBJECT_H
37 #define _LUSTRE_CL_OBJECT_H
39 /** \defgroup clio clio
41 * Client objects implement io operations and cache pages.
43 * Examples: lov and osc are implementations of cl interface.
45 * Big Theory Statement.
49 * Client implementation is based on the following data-types:
55 * - cl_lock represents an extent lock on an object.
57 * - cl_io represents high-level i/o activity such as whole read/write
58 * system call, or write-out of pages from under the lock being
59 * canceled. cl_io has sub-ios that can be stopped and resumed
60 * independently, thus achieving high degree of transfer
61 * parallelism. Single cl_io can be advanced forward by
62 * the multiple threads (although in the most usual case of
63 * read/write system call it is associated with the single user
64 * thread, that issued the system call).
66 * - cl_req represents a collection of pages for a transfer. cl_req is
67 * constructed by req-forming engine that tries to saturate
68 * transport with large and continuous transfers.
72 * - to avoid confusion high-level I/O operation like read or write system
73 * call is referred to as "an io", whereas low-level I/O operation, like
74 * RPC, is referred to as "a transfer"
76 * - "generic code" means generic (not file system specific) code in the
77 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
78 * is not layer specific.
84 * - cl_object_header::coh_page_guard
87 * See the top comment in cl_object.c for the description of overall locking and
88 * reference-counting design.
90 * See comments below for the description of i/o, page, and dlm-locking
97 * super-class definitions.
99 #include <libcfs/libcfs.h>
100 #include <lu_object.h>
101 #include <linux/atomic.h>
102 #include <linux/mutex.h>
103 #include <linux/radix-tree.h>
104 #include <linux/spinlock.h>
105 #include <linux/wait.h>
106 #include <lustre_dlm.h>
112 struct cl_device_operations;
115 struct cl_object_page_operations;
116 struct cl_object_lock_operations;
119 struct cl_page_slice;
121 struct cl_lock_slice;
123 struct cl_lock_operations;
124 struct cl_page_operations;
133 * Operations for each data device in the client stack.
135 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
137 struct cl_device_operations {
139 * Initialize cl_req. This method is called top-to-bottom on all
140 * devices in the stack to get them a chance to allocate layer-private
141 * data, and to attach them to the cl_req by calling
142 * cl_req_slice_add().
144 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
145 * \see vvp_req_init()
147 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
152 * Device in the client stack.
154 * \see vvp_device, lov_device, lovsub_device, osc_device
158 struct lu_device cd_lu_dev;
159 /** Per-layer operation vector. */
160 const struct cl_device_operations *cd_ops;
163 /** \addtogroup cl_object cl_object
166 * "Data attributes" of cl_object. Data attributes can be updated
167 * independently for a sub-object, and top-object's attributes are calculated
168 * from sub-objects' ones.
171 /** Object size, in bytes */
174 * Known minimal size, in bytes.
176 * This is only valid when at least one DLM lock is held.
179 /** Modification time. Measured in seconds since epoch. */
181 /** Access time. Measured in seconds since epoch. */
183 /** Change time. Measured in seconds since epoch. */
186 * Blocks allocated to this cl_object on the server file system.
188 * \todo XXX An interface for block size is needed.
192 * User identifier for quota purposes.
196 * Group identifier for quota purposes.
200 /* nlink of the directory */
205 * Fields in cl_attr that are being set.
219 * Sub-class of lu_object with methods common for objects on the client
222 * cl_object: represents a regular file system object, both a file and a
223 * stripe. cl_object is based on lu_object: it is identified by a fid,
224 * layered, cached, hashed, and lrued. Important distinction with the server
225 * side, where md_object and dt_object are used, is that cl_object "fans out"
226 * at the lov/sns level: depending on the file layout, single file is
227 * represented as a set of "sub-objects" (stripes). At the implementation
228 * level, struct lov_object contains an array of cl_objects. Each sub-object
229 * is a full-fledged cl_object, having its fid, living in the lru and hash
232 * This leads to the next important difference with the server side: on the
233 * client, it's quite usual to have objects with the different sequence of
234 * layers. For example, typical top-object is composed of the following
240 * whereas its sub-objects are composed of
245 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
246 * track of the object-subobject relationship.
248 * Sub-objects are not cached independently: when top-object is about to
249 * be discarded from the memory, all its sub-objects are torn-down and
252 * \see vvp_object, lov_object, lovsub_object, osc_object
256 struct lu_object co_lu;
257 /** per-object-layer operations */
258 const struct cl_object_operations *co_ops;
259 /** offset of page slice in cl_page buffer */
264 * Description of the client object configuration. This is used for the
265 * creation of a new client object that is identified by a more state than
268 struct cl_object_conf {
270 struct lu_object_conf coc_lu;
273 * Object layout. This is consumed by lov.
275 struct lustre_md *coc_md;
277 * Description of particular stripe location in the
278 * cluster. This is consumed by osc.
280 struct lov_oinfo *coc_oinfo;
283 * VFS inode. This is consumed by vvp.
285 struct inode *coc_inode;
287 * Layout lock handle.
289 struct ldlm_lock *coc_lock;
291 * Operation to handle layout, OBJECT_CONF_XYZ.
297 /** configure layout, set up a new stripe, must be called while
298 * holding layout lock. */
300 /** invalidate the current stripe configuration due to losing
302 OBJECT_CONF_INVALIDATE = 1,
303 /** wait for old layout to go away so that new layout can be
309 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
310 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
314 /** the buffer to return the layout in lov_mds_md format. */
315 struct lu_buf cl_buf;
316 /** size of layout in lov_mds_md format. */
318 /** Layout generation. */
320 /** True if this is a released file.
321 * Temporarily added for released file truncate in ll_setattr_raw().
322 * It will be removed later. -Jinshan */
327 * Operations implemented for each cl object layer.
329 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
331 struct cl_object_operations {
333 * Initialize page slice for this layer. Called top-to-bottom through
334 * every object layer when a new cl_page is instantiated. Layer
335 * keeping private per-page data, or requiring its own page operations
336 * vector should allocate these data here, and attach then to the page
337 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
340 * \retval NULL success.
342 * \retval ERR_PTR(errno) failure code.
344 * \retval valid-pointer pointer to already existing referenced page
345 * to be used instead of newly created.
347 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
348 struct cl_page *page, pgoff_t index);
350 * Initialize lock slice for this layer. Called top-to-bottom through
351 * every object layer when a new cl_lock is instantiated. Layer
352 * keeping private per-lock data, or requiring its own lock operations
353 * vector should allocate these data here, and attach then to the lock
354 * by calling cl_lock_slice_add(). Mandatory.
356 int (*coo_lock_init)(const struct lu_env *env,
357 struct cl_object *obj, struct cl_lock *lock,
358 const struct cl_io *io);
360 * Initialize io state for a given layer.
362 * called top-to-bottom once per io existence to initialize io
363 * state. If layer wants to keep some state for this type of io, it
364 * has to embed struct cl_io_slice in lu_env::le_ses, and register
365 * slice with cl_io_slice_add(). It is guaranteed that all threads
366 * participating in this io share the same session.
368 int (*coo_io_init)(const struct lu_env *env,
369 struct cl_object *obj, struct cl_io *io);
371 * Fill portion of \a attr that this layer controls. This method is
372 * called top-to-bottom through all object layers.
374 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
376 * \return 0: to continue
377 * \return +ve: to stop iterating through layers (but 0 is returned
378 * from enclosing cl_object_attr_get())
379 * \return -ve: to signal error
381 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
382 struct cl_attr *attr);
386 * \a valid is a bitmask composed from enum #cl_attr_valid, and
387 * indicating what attributes are to be set.
389 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
391 * \return the same convention as for
392 * cl_object_operations::coo_attr_get() is used.
394 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
395 const struct cl_attr *attr, unsigned valid);
397 * Update object configuration. Called top-to-bottom to modify object
400 * XXX error conditions and handling.
402 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
403 const struct cl_object_conf *conf);
405 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
406 * object. Layers are supposed to fill parts of \a lvb that will be
407 * shipped to the glimpse originator as a glimpse result.
409 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
410 * \see osc_object_glimpse()
412 int (*coo_glimpse)(const struct lu_env *env,
413 const struct cl_object *obj, struct ost_lvb *lvb);
415 * Object prune method. Called when the layout is going to change on
416 * this object, therefore each layer has to clean up their cache,
417 * mainly pages and locks.
419 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
421 * Object getstripe method.
423 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
424 struct lov_user_md __user *lum);
426 * Find whether there is any callback data (ldlm lock) attached upon
429 int (*coo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
430 ldlm_iterator_t iter, void *data);
432 * Get FIEMAP mapping from the object.
434 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
435 struct ll_fiemap_info_key *fmkey,
436 struct fiemap *fiemap, size_t *buflen);
438 * Get layout and generation of the object.
440 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
441 struct cl_layout *layout);
443 * Get maximum size of the object.
445 loff_t (*coo_maxbytes)(struct cl_object *obj);
449 * Extended header for client object.
451 struct cl_object_header {
452 /** Standard lu_object_header. cl_object::co_lu::lo_header points
454 struct lu_object_header coh_lu;
457 * Parent object. It is assumed that an object has a well-defined
458 * parent, but not a well-defined child (there may be multiple
459 * sub-objects, for the same top-object). cl_object_header::coh_parent
460 * field allows certain code to be written generically, without
461 * limiting possible cl_object layouts unduly.
463 struct cl_object_header *coh_parent;
465 * Protects consistency between cl_attr of parent object and
466 * attributes of sub-objects, that the former is calculated ("merged")
469 * \todo XXX this can be read/write lock if needed.
471 spinlock_t coh_attr_guard;
473 * Size of cl_page + page slices
475 unsigned short coh_page_bufsize;
477 * Number of objects above this one: 0 for a top-object, 1 for its
480 unsigned char coh_nesting;
484 * Helper macro: iterate over all layers of the object \a obj, assigning every
485 * layer top-to-bottom to \a slice.
487 #define cl_object_for_each(slice, obj) \
488 list_for_each_entry((slice), \
489 &(obj)->co_lu.lo_header->loh_layers,\
493 * Helper macro: iterate over all layers of the object \a obj, assigning every
494 * layer bottom-to-top to \a slice.
496 #define cl_object_for_each_reverse(slice, obj) \
497 list_for_each_entry_reverse((slice), \
498 &(obj)->co_lu.lo_header->loh_layers,\
503 #define CL_PAGE_EOF ((pgoff_t)~0ull)
505 /** \addtogroup cl_page cl_page
509 * Layered client page.
511 * cl_page: represents a portion of a file, cached in the memory. All pages
512 * of the given file are of the same size, and are kept in the radix tree
513 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
514 * of the top-level file object are first class cl_objects, they have their
515 * own radix trees of pages and hence page is implemented as a sequence of
516 * struct cl_pages's, linked into double-linked list through
517 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
518 * corresponding radix tree at the corresponding logical offset.
520 * cl_page is associated with VM page of the hosting environment (struct
521 * page in Linux kernel, for example), struct page. It is assumed, that this
522 * association is implemented by one of cl_page layers (top layer in the
523 * current design) that
525 * - intercepts per-VM-page call-backs made by the environment (e.g.,
528 * - translates state (page flag bits) and locking between lustre and
531 * The association between cl_page and struct page is immutable and
532 * established when cl_page is created.
534 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
535 * this io an exclusive access to this page w.r.t. other io attempts and
536 * various events changing page state (such as transfer completion, or
537 * eviction of the page from the memory). Note, that in general cl_io
538 * cannot be identified with a particular thread, and page ownership is not
539 * exactly equal to the current thread holding a lock on the page. Layer
540 * implementing association between cl_page and struct page has to implement
541 * ownership on top of available synchronization mechanisms.
543 * While lustre client maintains the notion of an page ownership by io,
544 * hosting MM/VM usually has its own page concurrency control
545 * mechanisms. For example, in Linux, page access is synchronized by the
546 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
547 * takes care to acquire and release such locks as necessary around the
548 * calls to the file system methods (->readpage(), ->prepare_write(),
549 * ->commit_write(), etc.). This leads to the situation when there are two
550 * different ways to own a page in the client:
552 * - client code explicitly and voluntary owns the page (cl_page_own());
554 * - VM locks a page and then calls the client, that has "to assume"
555 * the ownership from the VM (cl_page_assume()).
557 * Dual methods to release ownership are cl_page_disown() and
558 * cl_page_unassume().
560 * cl_page is reference counted (cl_page::cp_ref). When reference counter
561 * drops to 0, the page is returned to the cache, unless it is in
562 * cl_page_state::CPS_FREEING state, in which case it is immediately
565 * The general logic guaranteeing the absence of "existential races" for
566 * pages is the following:
568 * - there are fixed known ways for a thread to obtain a new reference
571 * - by doing a lookup in the cl_object radix tree, protected by the
574 * - by starting from VM-locked struct page and following some
575 * hosting environment method (e.g., following ->private pointer in
576 * the case of Linux kernel), see cl_vmpage_page();
578 * - when the page enters cl_page_state::CPS_FREEING state, all these
579 * ways are severed with the proper synchronization
580 * (cl_page_delete());
582 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
585 * - no new references to the page in cl_page_state::CPS_FREEING state
586 * are allowed (checked in cl_page_get()).
588 * Together this guarantees that when last reference to a
589 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
590 * page, as neither references to it can be acquired at that point, nor
593 * cl_page is a state machine. States are enumerated in enum
594 * cl_page_state. Possible state transitions are enumerated in
595 * cl_page_state_set(). State transition process (i.e., actual changing of
596 * cl_page::cp_state field) is protected by the lock on the underlying VM
599 * Linux Kernel implementation.
601 * Binding between cl_page and struct page (which is a typedef for
602 * struct page) is implemented in the vvp layer. cl_page is attached to the
603 * ->private pointer of the struct page, together with the setting of
604 * PG_private bit in page->flags, and acquiring additional reference on the
605 * struct page (much like struct buffer_head, or any similar file system
606 * private data structures).
608 * PG_locked lock is used to implement both ownership and transfer
609 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
610 * states. No additional references are acquired for the duration of the
613 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
614 * write-out is "protected" by the special PG_writeback bit.
618 * States of cl_page. cl_page.c assumes particular order here.
620 * The page state machine is rather crude, as it doesn't recognize finer page
621 * states like "dirty" or "up to date". This is because such states are not
622 * always well defined for the whole stack (see, for example, the
623 * implementation of the read-ahead, that hides page up-to-dateness to track
624 * cache hits accurately). Such sub-states are maintained by the layers that
625 * are interested in them.
629 * Page is in the cache, un-owned. Page leaves cached state in the
632 * - [cl_page_state::CPS_OWNED] io comes across the page and
635 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
636 * req-formation engine decides that it wants to include this page
637 * into an cl_req being constructed, and yanks it from the cache;
639 * - [cl_page_state::CPS_FREEING] VM callback is executed to
640 * evict the page form the memory;
642 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
646 * Page is exclusively owned by some cl_io. Page may end up in this
647 * state as a result of
649 * - io creating new page and immediately owning it;
651 * - [cl_page_state::CPS_CACHED] io finding existing cached page
654 * - [cl_page_state::CPS_OWNED] io finding existing owned page
655 * and waiting for owner to release the page;
657 * Page leaves owned state in the following cases:
659 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
660 * the cache, doing nothing;
662 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
665 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
666 * transfer for this page;
668 * - [cl_page_state::CPS_FREEING] io decides to destroy this
669 * page (e.g., as part of truncate or extent lock cancellation).
671 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
675 * Page is being written out, as a part of a transfer. This state is
676 * entered when req-formation logic decided that it wants this page to
677 * be sent through the wire _now_. Specifically, it means that once
678 * this state is achieved, transfer completion handler (with either
679 * success or failure indication) is guaranteed to be executed against
680 * this page independently of any locks and any scheduling decisions
681 * made by the hosting environment (that effectively means that the
682 * page is never put into cl_page_state::CPS_PAGEOUT state "in
683 * advance". This property is mentioned, because it is important when
684 * reasoning about possible dead-locks in the system). The page can
685 * enter this state as a result of
687 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
688 * write-out of this page, or
690 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
691 * that it has enough dirty pages cached to issue a "good"
694 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
695 * is completed---it is moved into cl_page_state::CPS_CACHED state.
697 * Underlying VM page is locked for the duration of transfer.
699 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
703 * Page is being read in, as a part of a transfer. This is quite
704 * similar to the cl_page_state::CPS_PAGEOUT state, except that
705 * read-in is always "immediate"---there is no such thing a sudden
706 * construction of read cl_req from cached, presumably not up to date,
709 * Underlying VM page is locked for the duration of transfer.
711 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
715 * Page is being destroyed. This state is entered when client decides
716 * that page has to be deleted from its host object, as, e.g., a part
719 * Once this state is reached, there is no way to escape it.
721 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
728 /** Host page, the page is from the host inode which the cl_page
732 /** Transient page, the transient cl_page is used to bind a cl_page
733 * to vmpage which is not belonging to the same object of cl_page.
734 * it is used in DirectIO, lockless IO and liblustre. */
739 * Fields are protected by the lock on struct page, except for atomics and
742 * \invariant Data type invariants are in cl_page_invariant(). Basically:
743 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
744 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
745 * cl_page::cp_owner (when set).
748 /** Reference counter. */
750 /** Transfer error. */
752 /** An object this page is a part of. Immutable after creation. */
753 struct cl_object *cp_obj;
755 struct page *cp_vmpage;
756 /** Linkage of pages within group. Pages must be owned */
757 struct list_head cp_batch;
758 /** List of slices. Immutable after creation. */
759 struct list_head cp_layers;
760 /** Linkage of pages within cl_req. */
761 struct list_head cp_flight;
763 * Page state. This field is const to avoid accidental update, it is
764 * modified only internally within cl_page.c. Protected by a VM lock.
766 const enum cl_page_state cp_state;
768 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
771 enum cl_page_type cp_type;
774 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
775 * by sub-io. Protected by a VM lock.
777 struct cl_io *cp_owner;
779 * Owning IO request in cl_page_state::CPS_PAGEOUT and
780 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
781 * the top-level pages. Protected by a VM lock.
783 struct cl_req *cp_req;
784 /** List of references to this page, for debugging. */
785 struct lu_ref cp_reference;
786 /** Link to an object, for debugging. */
787 struct lu_ref_link cp_obj_ref;
788 /** Link to a queue, for debugging. */
789 struct lu_ref_link cp_queue_ref;
790 /** Assigned if doing a sync_io */
791 struct cl_sync_io *cp_sync_io;
795 * Per-layer part of cl_page.
797 * \see vvp_page, lov_page, osc_page
799 struct cl_page_slice {
800 struct cl_page *cpl_page;
803 * Object slice corresponding to this page slice. Immutable after
806 struct cl_object *cpl_obj;
807 const struct cl_page_operations *cpl_ops;
808 /** Linkage into cl_page::cp_layers. Immutable after creation. */
809 struct list_head cpl_linkage;
813 * Lock mode. For the client extent locks.
825 * Requested transfer type.
835 * Per-layer page operations.
837 * Methods taking an \a io argument are for the activity happening in the
838 * context of given \a io. Page is assumed to be owned by that io, except for
839 * the obvious cases (like cl_page_operations::cpo_own()).
841 * \see vvp_page_ops, lov_page_ops, osc_page_ops
843 struct cl_page_operations {
845 * cl_page<->struct page methods. Only one layer in the stack has to
846 * implement these. Current code assumes that this functionality is
847 * provided by the topmost layer, see cl_page_disown0() as an example.
851 * Called when \a io acquires this page into the exclusive
852 * ownership. When this method returns, it is guaranteed that the is
853 * not owned by other io, and no transfer is going on against
857 * \see vvp_page_own(), lov_page_own()
859 int (*cpo_own)(const struct lu_env *env,
860 const struct cl_page_slice *slice,
861 struct cl_io *io, int nonblock);
862 /** Called when ownership it yielded. Optional.
864 * \see cl_page_disown()
865 * \see vvp_page_disown()
867 void (*cpo_disown)(const struct lu_env *env,
868 const struct cl_page_slice *slice, struct cl_io *io);
870 * Called for a page that is already "owned" by \a io from VM point of
873 * \see cl_page_assume()
874 * \see vvp_page_assume(), lov_page_assume()
876 void (*cpo_assume)(const struct lu_env *env,
877 const struct cl_page_slice *slice, struct cl_io *io);
878 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
879 * bottom-to-top when IO releases a page without actually unlocking
882 * \see cl_page_unassume()
883 * \see vvp_page_unassume()
885 void (*cpo_unassume)(const struct lu_env *env,
886 const struct cl_page_slice *slice,
889 * Announces whether the page contains valid data or not by \a uptodate.
891 * \see cl_page_export()
892 * \see vvp_page_export()
894 void (*cpo_export)(const struct lu_env *env,
895 const struct cl_page_slice *slice, int uptodate);
897 * Checks whether underlying VM page is locked (in the suitable
898 * sense). Used for assertions.
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. (Should never happen.)
904 int (*cpo_is_vmlocked)(const struct lu_env *env,
905 const struct cl_page_slice *slice);
911 * Called when page is truncated from the object. Optional.
913 * \see cl_page_discard()
914 * \see vvp_page_discard(), osc_page_discard()
916 void (*cpo_discard)(const struct lu_env *env,
917 const struct cl_page_slice *slice,
920 * Called when page is removed from the cache, and is about to being
921 * destroyed. Optional.
923 * \see cl_page_delete()
924 * \see vvp_page_delete(), osc_page_delete()
926 void (*cpo_delete)(const struct lu_env *env,
927 const struct cl_page_slice *slice);
928 /** Destructor. Frees resources and slice itself. */
929 void (*cpo_fini)(const struct lu_env *env,
930 struct cl_page_slice *slice);
932 * Optional debugging helper. Prints given page slice.
934 * \see cl_page_print()
936 int (*cpo_print)(const struct lu_env *env,
937 const struct cl_page_slice *slice,
938 void *cookie, lu_printer_t p);
942 * Transfer methods. See comment on cl_req for a description of
943 * transfer formation and life-cycle.
948 * Request type dependent vector of operations.
950 * Transfer operations depend on transfer mode (cl_req_type). To avoid
951 * passing transfer mode to each and every of these methods, and to
952 * avoid branching on request type inside of the methods, separate
953 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
954 * provided. That is, method invocation usually looks like
956 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
960 * Called when a page is submitted for a transfer as a part of
963 * \return 0 : page is eligible for submission;
964 * \return -EALREADY : skip this page;
965 * \return -ve : error.
967 * \see cl_page_prep()
969 int (*cpo_prep)(const struct lu_env *env,
970 const struct cl_page_slice *slice,
973 * Completion handler. This is guaranteed to be eventually
974 * fired after cl_page_operations::cpo_prep() or
975 * cl_page_operations::cpo_make_ready() call.
977 * This method can be called in a non-blocking context. It is
978 * guaranteed however, that the page involved and its object
979 * are pinned in memory (and, hence, calling cl_page_put() is
982 * \see cl_page_completion()
984 void (*cpo_completion)(const struct lu_env *env,
985 const struct cl_page_slice *slice,
988 * Called when cached page is about to be added to the
989 * cl_req as a part of req formation.
991 * \return 0 : proceed with this page;
992 * \return -EAGAIN : skip this page;
993 * \return -ve : error.
995 * \see cl_page_make_ready()
997 int (*cpo_make_ready)(const struct lu_env *env,
998 const struct cl_page_slice *slice);
1001 * Tell transfer engine that only [to, from] part of a page should be
1004 * This is used for immediate transfers.
1006 * \todo XXX this is not very good interface. It would be much better
1007 * if all transfer parameters were supplied as arguments to
1008 * cl_io_operations::cio_submit() call, but it is not clear how to do
1009 * this for page queues.
1011 * \see cl_page_clip()
1013 void (*cpo_clip)(const struct lu_env *env,
1014 const struct cl_page_slice *slice,
1017 * \pre the page was queued for transferring.
1018 * \post page is removed from client's pending list, or -EBUSY
1019 * is returned if it has already been in transferring.
1021 * This is one of seldom page operation which is:
1022 * 0. called from top level;
1023 * 1. don't have vmpage locked;
1024 * 2. every layer should synchronize execution of its ->cpo_cancel()
1025 * with completion handlers. Osc uses client obd lock for this
1026 * purpose. Based on there is no vvp_page_cancel and
1027 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1029 * \see osc_page_cancel().
1031 int (*cpo_cancel)(const struct lu_env *env,
1032 const struct cl_page_slice *slice);
1034 * Write out a page by kernel. This is only called by ll_writepage
1037 * \see cl_page_flush()
1039 int (*cpo_flush)(const struct lu_env *env,
1040 const struct cl_page_slice *slice,
1046 * Helper macro, dumping detailed information about \a page into a log.
1048 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1050 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1051 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1052 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1053 CDEBUG(mask, format , ## __VA_ARGS__); \
1058 * Helper macro, dumping shorter information about \a page into a log.
1060 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1062 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1063 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1064 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1065 CDEBUG(mask, format , ## __VA_ARGS__); \
1069 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1071 LASSERT(page->cp_vmpage != NULL);
1072 return page->cp_vmpage;
1076 * Check if a cl_page is in use.
1078 * Client cache holds a refcount, this refcount will be dropped when
1079 * the page is taken out of cache, see vvp_page_delete().
1081 static inline bool __page_in_use(const struct cl_page *page, int refc)
1083 return (atomic_read(&page->cp_ref) > refc + 1);
1087 * Caller itself holds a refcount of cl_page.
1089 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1091 * Caller doesn't hold a refcount.
1093 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1097 /** \addtogroup cl_lock cl_lock
1101 * Extent locking on the client.
1105 * The locking model of the new client code is built around
1109 * data-type representing an extent lock on a regular file. cl_lock is a
1110 * layered object (much like cl_object and cl_page), it consists of a header
1111 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1112 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1114 * Typical cl_lock consists of the two layers:
1116 * - vvp_lock (vvp specific data), and
1117 * - lov_lock (lov specific data).
1119 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1120 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1122 * - lovsub_lock, and
1125 * Each sub-lock is associated with a cl_object (representing stripe
1126 * sub-object or the file to which top-level cl_lock is associated to), and is
1127 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1128 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1129 * is different from cl_page, that doesn't fan out (there is usually exactly
1130 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1131 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1135 * cl_lock is a cacheless data container for the requirements of locks to
1136 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1139 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1140 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1142 * INTERFACE AND USAGE
1144 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1145 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1146 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1147 * consists of multiple sub cl_locks, each sub locks will be enqueued
1148 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1149 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1152 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1153 * method will be called for each layer to release the resource held by this
1154 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1155 * clo_enqueue time, is released.
1157 * LDLM lock can only be canceled if there is no cl_lock using it.
1159 * Overall process of the locking during IO operation is as following:
1161 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1162 * is called on each layer. Responsibility of this method is to add locks,
1163 * needed by a given layer into cl_io.ci_lockset.
1165 * - once locks for all layers were collected, they are sorted to avoid
1166 * dead-locks (cl_io_locks_sort()), and enqueued.
1168 * - when all locks are acquired, IO is performed;
1170 * - locks are released after IO is complete.
1172 * Striping introduces major additional complexity into locking. The
1173 * fundamental problem is that it is generally unsafe to actively use (hold)
1174 * two locks on the different OST servers at the same time, as this introduces
1175 * inter-server dependency and can lead to cascading evictions.
1177 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1178 * that no multi-stripe locks are taken (note that this design abandons POSIX
1179 * read/write semantics). Such pieces ideally can be executed concurrently. At
1180 * the same time, certain types of IO cannot be sub-divived, without
1181 * sacrificing correctness. This includes:
1183 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1186 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1188 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1189 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1190 * has to be held together with the usual lock on [offset, offset + count].
1192 * Interaction with DLM
1194 * In the expected setup, cl_lock is ultimately backed up by a collection of
1195 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1196 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1197 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1198 * description of interaction with DLM.
1204 struct cl_lock_descr {
1205 /** Object this lock is granted for. */
1206 struct cl_object *cld_obj;
1207 /** Index of the first page protected by this lock. */
1209 /** Index of the last page (inclusive) protected by this lock. */
1211 /** Group ID, for group lock */
1214 enum cl_lock_mode cld_mode;
1216 * flags to enqueue lock. A combination of bit-flags from
1217 * enum cl_enq_flags.
1219 __u32 cld_enq_flags;
1222 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1223 #define PDESCR(descr) \
1224 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1225 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1227 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1230 * Layered client lock.
1233 /** List of slices. Immutable after creation. */
1234 struct list_head cll_layers;
1235 /** lock attribute, extent, cl_object, etc. */
1236 struct cl_lock_descr cll_descr;
1240 * Per-layer part of cl_lock
1242 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1244 struct cl_lock_slice {
1245 struct cl_lock *cls_lock;
1246 /** Object slice corresponding to this lock slice. Immutable after
1248 struct cl_object *cls_obj;
1249 const struct cl_lock_operations *cls_ops;
1250 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1251 struct list_head cls_linkage;
1256 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1258 struct cl_lock_operations {
1261 * Attempts to enqueue the lock. Called top-to-bottom.
1263 * \retval 0 this layer has enqueued the lock successfully
1264 * \retval >0 this layer has enqueued the lock, but need to wait on
1265 * @anchor for resources
1266 * \retval -ve failure
1268 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1269 * \see osc_lock_enqueue()
1271 int (*clo_enqueue)(const struct lu_env *env,
1272 const struct cl_lock_slice *slice,
1273 struct cl_io *io, struct cl_sync_io *anchor);
1275 * Cancel a lock, release its DLM lock ref, while does not cancel the
1278 void (*clo_cancel)(const struct lu_env *env,
1279 const struct cl_lock_slice *slice);
1282 * Destructor. Frees resources and the slice.
1284 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1285 * \see osc_lock_fini()
1287 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1289 * Optional debugging helper. Prints given lock slice.
1291 int (*clo_print)(const struct lu_env *env,
1292 void *cookie, lu_printer_t p,
1293 const struct cl_lock_slice *slice);
1296 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1298 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1299 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1300 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1301 CDEBUG(mask, format , ## __VA_ARGS__); \
1305 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1309 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1315 /** \addtogroup cl_page_list cl_page_list
1316 * Page list used to perform collective operations on a group of pages.
1318 * Pages are added to the list one by one. cl_page_list acquires a reference
1319 * for every page in it. Page list is used to perform collective operations on
1322 * - submit pages for an immediate transfer,
1324 * - own pages on behalf of certain io (waiting for each page in turn),
1328 * When list is finalized, it releases references on all pages it still has.
1330 * \todo XXX concurrency control.
1334 struct cl_page_list {
1336 struct list_head pl_pages;
1337 struct task_struct *pl_owner;
1341 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1342 * contains an incoming page list and an outgoing page list.
1345 struct cl_page_list c2_qin;
1346 struct cl_page_list c2_qout;
1349 /** @} cl_page_list */
1351 /** \addtogroup cl_io cl_io
1356 * cl_io represents a high level I/O activity like
1357 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1360 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1361 * important distinction. We want to minimize number of calls to the allocator
1362 * in the fast path, e.g., in the case of read(2) when everything is cached:
1363 * client already owns the lock over region being read, and data are cached
1364 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1365 * per-layer io state is stored in the session, associated with the io, see
1366 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1367 * by using free-lists, see cl_env_get().
1369 * There is a small predefined number of possible io types, enumerated in enum
1372 * cl_io is a state machine, that can be advanced concurrently by the multiple
1373 * threads. It is up to these threads to control the concurrency and,
1374 * specifically, to detect when io is done, and its state can be safely
1377 * For read/write io overall execution plan is as following:
1379 * (0) initialize io state through all layers;
1381 * (1) loop: prepare chunk of work to do
1383 * (2) call all layers to collect locks they need to process current chunk
1385 * (3) sort all locks to avoid dead-locks, and acquire them
1387 * (4) process the chunk: call per-page methods
1388 * cl_io_operations::cio_prepare_write(),
1389 * cl_io_operations::cio_commit_write() for write)
1395 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1396 * address allocation efficiency issues mentioned above), and returns with the
1397 * special error condition from per-page method when current sub-io has to
1398 * block. This causes io loop to be repeated, and lov switches to the next
1399 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1404 /** read system call */
1406 /** write system call */
1408 /** truncate, utime system calls */
1410 /** get data version */
1413 * page fault handling
1417 * fsync system call handling
1418 * To write out a range of file
1422 * Miscellaneous io. This is used for occasional io activity that
1423 * doesn't fit into other types. Currently this is used for:
1425 * - cancellation of an extent lock. This io exists as a context
1426 * to write dirty pages from under the lock being canceled back
1429 * - VM induced page write-out. An io context for writing page out
1430 * for memory cleansing;
1432 * - glimpse. An io context to acquire glimpse lock.
1434 * - grouplock. An io context to acquire group lock.
1436 * CIT_MISC io is used simply as a context in which locks and pages
1437 * are manipulated. Such io has no internal "process", that is,
1438 * cl_io_loop() is never called for it.
1445 * States of cl_io state machine
1448 /** Not initialized. */
1452 /** IO iteration started. */
1456 /** Actual IO is in progress. */
1458 /** IO for the current iteration finished. */
1460 /** Locks released. */
1462 /** Iteration completed. */
1464 /** cl_io finalized. */
1469 * IO state private for a layer.
1471 * This is usually embedded into layer session data, rather than allocated
1474 * \see vvp_io, lov_io, osc_io
1476 struct cl_io_slice {
1477 struct cl_io *cis_io;
1478 /** corresponding object slice. Immutable after creation. */
1479 struct cl_object *cis_obj;
1480 /** io operations. Immutable after creation. */
1481 const struct cl_io_operations *cis_iop;
1483 * linkage into a list of all slices for a given cl_io, hanging off
1484 * cl_io::ci_layers. Immutable after creation.
1486 struct list_head cis_linkage;
1489 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1492 struct cl_read_ahead {
1493 /* Maximum page index the readahead window will end.
1494 * This is determined DLM lock coverage, RPC and stripe boundary.
1495 * cra_end is included. */
1497 /* Release routine. If readahead holds resources underneath, this
1498 * function should be called to release it. */
1499 void (*cra_release)(const struct lu_env *env, void *cbdata);
1500 /* Callback data for cra_release routine */
1504 static inline void cl_read_ahead_release(const struct lu_env *env,
1505 struct cl_read_ahead *ra)
1507 if (ra->cra_release != NULL)
1508 ra->cra_release(env, ra->cra_cbdata);
1509 memset(ra, 0, sizeof(*ra));
1514 * Per-layer io operations.
1515 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1517 struct cl_io_operations {
1519 * Vector of io state transition methods for every io type.
1521 * \see cl_page_operations::io
1525 * Prepare io iteration at a given layer.
1527 * Called top-to-bottom at the beginning of each iteration of
1528 * "io loop" (if it makes sense for this type of io). Here
1529 * layer selects what work it will do during this iteration.
1531 * \see cl_io_operations::cio_iter_fini()
1533 int (*cio_iter_init) (const struct lu_env *env,
1534 const struct cl_io_slice *slice);
1536 * Finalize io iteration.
1538 * Called bottom-to-top at the end of each iteration of "io
1539 * loop". Here layers can decide whether IO has to be
1542 * \see cl_io_operations::cio_iter_init()
1544 void (*cio_iter_fini) (const struct lu_env *env,
1545 const struct cl_io_slice *slice);
1547 * Collect locks for the current iteration of io.
1549 * Called top-to-bottom to collect all locks necessary for
1550 * this iteration. This methods shouldn't actually enqueue
1551 * anything, instead it should post a lock through
1552 * cl_io_lock_add(). Once all locks are collected, they are
1553 * sorted and enqueued in the proper order.
1555 int (*cio_lock) (const struct lu_env *env,
1556 const struct cl_io_slice *slice);
1558 * Finalize unlocking.
1560 * Called bottom-to-top to finish layer specific unlocking
1561 * functionality, after generic code released all locks
1562 * acquired by cl_io_operations::cio_lock().
1564 void (*cio_unlock)(const struct lu_env *env,
1565 const struct cl_io_slice *slice);
1567 * Start io iteration.
1569 * Once all locks are acquired, called top-to-bottom to
1570 * commence actual IO. In the current implementation,
1571 * top-level vvp_io_{read,write}_start() does all the work
1572 * synchronously by calling generic_file_*(), so other layers
1573 * are called when everything is done.
1575 int (*cio_start)(const struct lu_env *env,
1576 const struct cl_io_slice *slice);
1578 * Called top-to-bottom at the end of io loop. Here layer
1579 * might wait for an unfinished asynchronous io.
1581 void (*cio_end) (const struct lu_env *env,
1582 const struct cl_io_slice *slice);
1584 * Called bottom-to-top to notify layers that read/write IO
1585 * iteration finished, with \a nob bytes transferred.
1587 void (*cio_advance)(const struct lu_env *env,
1588 const struct cl_io_slice *slice,
1591 * Called once per io, bottom-to-top to release io resources.
1593 void (*cio_fini) (const struct lu_env *env,
1594 const struct cl_io_slice *slice);
1598 * Submit pages from \a queue->c2_qin for IO, and move
1599 * successfully submitted pages into \a queue->c2_qout. Return
1600 * non-zero if failed to submit even the single page. If
1601 * submission failed after some pages were moved into \a
1602 * queue->c2_qout, completion callback with non-zero ioret is
1605 int (*cio_submit)(const struct lu_env *env,
1606 const struct cl_io_slice *slice,
1607 enum cl_req_type crt,
1608 struct cl_2queue *queue);
1610 * Queue async page for write.
1611 * The difference between cio_submit and cio_queue is that
1612 * cio_submit is for urgent request.
1614 int (*cio_commit_async)(const struct lu_env *env,
1615 const struct cl_io_slice *slice,
1616 struct cl_page_list *queue, int from, int to,
1619 * Decide maximum read ahead extent
1621 * \pre io->ci_type == CIT_READ
1623 int (*cio_read_ahead)(const struct lu_env *env,
1624 const struct cl_io_slice *slice,
1625 pgoff_t start, struct cl_read_ahead *ra);
1627 * Optional debugging helper. Print given io slice.
1629 int (*cio_print)(const struct lu_env *env, void *cookie,
1630 lu_printer_t p, const struct cl_io_slice *slice);
1634 * Flags to lock enqueue procedure.
1639 * instruct server to not block, if conflicting lock is found. Instead
1640 * -EWOULDBLOCK is returned immediately.
1642 CEF_NONBLOCK = 0x00000001,
1644 * take lock asynchronously (out of order), as it cannot
1645 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1647 CEF_ASYNC = 0x00000002,
1649 * tell the server to instruct (though a flag in the blocking ast) an
1650 * owner of the conflicting lock, that it can drop dirty pages
1651 * protected by this lock, without sending them to the server.
1653 CEF_DISCARD_DATA = 0x00000004,
1655 * tell the sub layers that it must be a `real' lock. This is used for
1656 * mmapped-buffer locks and glimpse locks that must be never converted
1657 * into lockless mode.
1659 * \see vvp_mmap_locks(), cl_glimpse_lock().
1661 CEF_MUST = 0x00000008,
1663 * tell the sub layers that never request a `real' lock. This flag is
1664 * not used currently.
1666 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1667 * conversion policy: ci_lockreq describes generic information of lock
1668 * requirement for this IO, especially for locks which belong to the
1669 * object doing IO; however, lock itself may have precise requirements
1670 * that are described by the enqueue flags.
1672 CEF_NEVER = 0x00000010,
1674 * for async glimpse lock.
1676 CEF_AGL = 0x00000020,
1678 * enqueue a lock to test DLM lock existence.
1680 CEF_PEEK = 0x00000040,
1682 * mask of enq_flags.
1684 CEF_MASK = 0x0000007f,
1688 * Link between lock and io. Intermediate structure is needed, because the
1689 * same lock can be part of multiple io's simultaneously.
1691 struct cl_io_lock_link {
1692 /** linkage into one of cl_lockset lists. */
1693 struct list_head cill_linkage;
1694 struct cl_lock cill_lock;
1695 /** optional destructor */
1696 void (*cill_fini)(const struct lu_env *env,
1697 struct cl_io_lock_link *link);
1699 #define cill_descr cill_lock.cll_descr
1702 * Lock-set represents a collection of locks, that io needs at a
1703 * time. Generally speaking, client tries to avoid holding multiple locks when
1706 * - holding extent locks over multiple ost's introduces the danger of
1707 * "cascading timeouts";
1709 * - holding multiple locks over the same ost is still dead-lock prone,
1710 * see comment in osc_lock_enqueue(),
1712 * but there are certain situations where this is unavoidable:
1714 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1716 * - truncate has to take [new-size, EOF] lock for correctness;
1718 * - SNS has to take locks across full stripe for correctness;
1720 * - in the case when user level buffer, supplied to {read,write}(file0),
1721 * is a part of a memory mapped lustre file, client has to take a dlm
1722 * locks on file0, and all files that back up the buffer (or a part of
1723 * the buffer, that is being processed in the current chunk, in any
1724 * case, there are situations where at least 2 locks are necessary).
1726 * In such cases we at least try to take locks in the same consistent
1727 * order. To this end, all locks are first collected, then sorted, and then
1731 /** locks to be acquired. */
1732 struct list_head cls_todo;
1733 /** locks acquired. */
1734 struct list_head cls_done;
1738 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1739 * but 'req' is always to be thought as 'request' :-)
1741 enum cl_io_lock_dmd {
1742 /** Always lock data (e.g., O_APPEND). */
1744 /** Layers are free to decide between local and global locking. */
1746 /** Never lock: there is no cache (e.g., liblustre). */
1750 enum cl_fsync_mode {
1751 /** start writeback, do not wait for them to finish */
1753 /** start writeback and wait for them to finish */
1755 /** discard all of dirty pages in a specific file range */
1756 CL_FSYNC_DISCARD = 2,
1757 /** start writeback and make sure they have reached storage before
1758 * return. OST_SYNC RPC must be issued and finished */
1762 struct cl_io_rw_common {
1772 * cl_io is shared by all threads participating in this IO (in current
1773 * implementation only one thread advances IO, but parallel IO design and
1774 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1775 * is up to these threads to serialize their activities, including updates to
1776 * mutable cl_io fields.
1779 /** type of this IO. Immutable after creation. */
1780 enum cl_io_type ci_type;
1781 /** current state of cl_io state machine. */
1782 enum cl_io_state ci_state;
1783 /** main object this io is against. Immutable after creation. */
1784 struct cl_object *ci_obj;
1786 * Upper layer io, of which this io is a part of. Immutable after
1789 struct cl_io *ci_parent;
1790 /** List of slices. Immutable after creation. */
1791 struct list_head ci_layers;
1792 /** list of locks (to be) acquired by this io. */
1793 struct cl_lockset ci_lockset;
1794 /** lock requirements, this is just a help info for sublayers. */
1795 enum cl_io_lock_dmd ci_lockreq;
1798 struct cl_io_rw_common rd;
1801 struct cl_io_rw_common wr;
1805 struct cl_io_rw_common ci_rw;
1806 struct cl_setattr_io {
1807 struct ost_lvb sa_attr;
1808 unsigned int sa_attr_flags;
1809 unsigned int sa_valid;
1810 int sa_stripe_index;
1811 const struct lu_fid *sa_parent_fid;
1812 struct obd_capa *sa_capa;
1814 struct cl_data_version_io {
1815 u64 dv_data_version;
1818 struct cl_fault_io {
1819 /** page index within file. */
1821 /** bytes valid byte on a faulted page. */
1823 /** writable page? for nopage() only */
1825 /** page of an executable? */
1827 /** page_mkwrite() */
1829 /** resulting page */
1830 struct cl_page *ft_page;
1832 struct cl_fsync_io {
1835 struct obd_capa *fi_capa;
1836 /** file system level fid */
1837 struct lu_fid *fi_fid;
1838 enum cl_fsync_mode fi_mode;
1839 /* how many pages were written/discarded */
1840 unsigned int fi_nr_written;
1843 struct cl_2queue ci_queue;
1846 unsigned int ci_continue:1,
1848 * This io has held grouplock, to inform sublayers that
1849 * don't do lockless i/o.
1853 * The whole IO need to be restarted because layout has been changed
1857 * to not refresh layout - the IO issuer knows that the layout won't
1858 * change(page operations, layout change causes all page to be
1859 * discarded), or it doesn't matter if it changes(sync).
1863 * Check if layout changed after the IO finishes. Mainly for HSM
1864 * requirement. If IO occurs to openning files, it doesn't need to
1865 * verify layout because HSM won't release openning files.
1866 * Right now, only two opertaions need to verify layout: glimpse
1871 * file is released, restore has to to be triggered by vvp layer
1873 ci_restore_needed:1,
1879 * Number of pages owned by this IO. For invariant checking.
1881 unsigned ci_owned_nr;
1886 /** \addtogroup cl_req cl_req
1891 * There are two possible modes of transfer initiation on the client:
1893 * - immediate transfer: this is started when a high level io wants a page
1894 * or a collection of pages to be transferred right away. Examples:
1895 * read-ahead, synchronous read in the case of non-page aligned write,
1896 * page write-out as a part of extent lock cancellation, page write-out
1897 * as a part of memory cleansing. Immediate transfer can be both
1898 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1900 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1901 * when io wants to transfer a page to the server some time later, when
1902 * it can be done efficiently. Example: pages dirtied by the write(2)
1905 * In any case, transfer takes place in the form of a cl_req, which is a
1906 * representation for a network RPC.
1908 * Pages queued for an opportunistic transfer are cached until it is decided
1909 * that efficient RPC can be composed of them. This decision is made by "a
1910 * req-formation engine", currently implemented as a part of osc
1911 * layer. Req-formation depends on many factors: the size of the resulting
1912 * RPC, whether or not multi-object RPCs are supported by the server,
1913 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1915 * For the immediate transfer io submits a cl_page_list, that req-formation
1916 * engine slices into cl_req's, possibly adding cached pages to some of
1917 * the resulting req's.
1919 * Whenever a page from cl_page_list is added to a newly constructed req, its
1920 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1921 * page state is atomically changed from cl_page_state::CPS_OWNED to
1922 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1923 * is zeroed, and cl_page::cp_req is set to the
1924 * req. cl_page_operations::cpo_prep() method at the particular layer might
1925 * return -EALREADY to indicate that it does not need to submit this page
1926 * at all. This is possible, for example, if page, submitted for read,
1927 * became up-to-date in the meantime; and for write, the page don't have
1928 * dirty bit marked. \see cl_io_submit_rw()
1930 * Whenever a cached page is added to a newly constructed req, its
1931 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1932 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1933 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1934 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1935 * might return -EAGAIN to indicate that this page is not eligible for the
1936 * transfer right now.
1940 * Plan is to divide transfers into "priority bands" (indicated when
1941 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1942 * and allow glueing of cached pages to immediate transfers only within single
1943 * band. This would make high priority transfers (like lock cancellation or
1944 * memory pressure induced write-out) really high priority.
1949 * Per-transfer attributes.
1951 struct cl_req_attr {
1952 /** Generic attributes for the server consumption. */
1953 struct obdo *cra_oa;
1955 struct obd_capa *cra_capa;
1957 char cra_jobid[LUSTRE_JOBID_SIZE];
1961 * Transfer request operations definable at every layer.
1963 * Concurrency: transfer formation engine synchronizes calls to all transfer
1966 struct cl_req_operations {
1968 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1969 * complete (all pages are added).
1971 * \see osc_req_prep()
1973 int (*cro_prep)(const struct lu_env *env,
1974 const struct cl_req_slice *slice);
1976 * Called top-to-bottom to fill in \a oa fields. This is called twice
1977 * with different flags, see bug 10150 and osc_build_req().
1979 * \param obj an object from cl_req which attributes are to be set in
1982 * \param oa struct obdo where attributes are placed
1984 * \param flags \a oa fields to be filled.
1986 void (*cro_attr_set)(const struct lu_env *env,
1987 const struct cl_req_slice *slice,
1988 const struct cl_object *obj,
1989 struct cl_req_attr *attr, u64 flags);
1991 * Called top-to-bottom from cl_req_completion() to notify layers that
1992 * transfer completed. Has to free all state allocated by
1993 * cl_device_operations::cdo_req_init().
1995 void (*cro_completion)(const struct lu_env *env,
1996 const struct cl_req_slice *slice, int ioret);
2000 * A per-object state that (potentially multi-object) transfer request keeps.
2003 /** object itself */
2004 struct cl_object *ro_obj;
2005 /** reference to cl_req_obj::ro_obj. For debugging. */
2006 struct lu_ref_link ro_obj_ref;
2007 /* something else? Number of pages for a given object? */
2013 * Transfer requests are not reference counted, because IO sub-system owns
2014 * them exclusively and knows when to free them.
2018 * cl_req is created by cl_req_alloc() that calls
2019 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2020 * state in every layer.
2022 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2023 * contains pages for.
2025 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2026 * called top-to-bottom. At that point layers can modify req, let it pass, or
2027 * deny it completely. This is to support things like SNS that have transfer
2028 * ordering requirements invisible to the individual req-formation engine.
2030 * On transfer completion (or transfer timeout, or failure to initiate the
2031 * transfer of an allocated req), cl_req_operations::cro_completion() method
2032 * is called, after execution of cl_page_operations::cpo_completion() of all
2036 enum cl_req_type crq_type;
2037 /** A list of pages being transferred */
2038 struct list_head crq_pages;
2039 /** Number of pages in cl_req::crq_pages */
2040 unsigned crq_nrpages;
2041 /** An array of objects which pages are in ->crq_pages */
2042 struct cl_req_obj *crq_o;
2043 /** Number of elements in cl_req::crq_objs[] */
2044 unsigned crq_nrobjs;
2045 struct list_head crq_layers;
2049 * Per-layer state for request.
2051 struct cl_req_slice {
2052 struct cl_req *crs_req;
2053 struct cl_device *crs_dev;
2054 struct list_head crs_linkage;
2055 const struct cl_req_operations *crs_ops;
2060 enum cache_stats_item {
2061 /** how many cache lookups were performed */
2063 /** how many times cache lookup resulted in a hit */
2065 /** how many entities are in the cache right now */
2067 /** how many entities in the cache are actively used (and cannot be
2068 * evicted) right now */
2070 /** how many entities were created at all */
2075 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2078 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2080 struct cache_stats {
2081 const char *cs_name;
2082 atomic_t cs_stats[CS_NR];
2085 /** These are not exported so far */
2086 void cache_stats_init (struct cache_stats *cs, const char *name);
2089 * Client-side site. This represents particular client stack. "Global"
2090 * variables should (directly or indirectly) be added here to allow multiple
2091 * clients to co-exist in the single address space.
2094 struct lu_site cs_lu;
2096 * Statistical counters. Atomics do not scale, something better like
2097 * per-cpu counters is needed.
2099 * These are exported as /proc/fs/lustre/llite/.../site
2101 * When interpreting keep in mind that both sub-locks (and sub-pages)
2102 * and top-locks (and top-pages) are accounted here.
2104 struct cache_stats cs_pages;
2105 atomic_t cs_pages_state[CPS_NR];
2108 int cl_site_init(struct cl_site *s, struct cl_device *top);
2109 void cl_site_fini(struct cl_site *s);
2110 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2113 * Output client site statistical counters into a buffer. Suitable for
2114 * ll_rd_*()-style functions.
2116 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2121 * Type conversion and accessory functions.
2125 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2127 return container_of(site, struct cl_site, cs_lu);
2130 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2132 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2133 return container_of0(d, struct cl_device, cd_lu_dev);
2136 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2138 return &d->cd_lu_dev;
2141 static inline struct cl_object *lu2cl(const struct lu_object *o)
2143 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2144 return container_of0(o, struct cl_object, co_lu);
2147 static inline const struct cl_object_conf *
2148 lu2cl_conf(const struct lu_object_conf *conf)
2150 return container_of0(conf, struct cl_object_conf, coc_lu);
2153 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2155 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2158 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2160 return container_of0(h, struct cl_object_header, coh_lu);
2163 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2165 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2169 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2171 return luh2coh(obj->co_lu.lo_header);
2174 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2176 return lu_device_init(&d->cd_lu_dev, t);
2179 static inline void cl_device_fini(struct cl_device *d)
2181 lu_device_fini(&d->cd_lu_dev);
2184 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2185 struct cl_object *obj, pgoff_t index,
2186 const struct cl_page_operations *ops);
2187 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2188 struct cl_object *obj,
2189 const struct cl_lock_operations *ops);
2190 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2191 struct cl_object *obj, const struct cl_io_operations *ops);
2192 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2193 struct cl_device *dev,
2194 const struct cl_req_operations *ops);
2197 /** \defgroup cl_object cl_object
2199 struct cl_object *cl_object_top (struct cl_object *o);
2200 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2201 const struct lu_fid *fid,
2202 const struct cl_object_conf *c);
2204 int cl_object_header_init(struct cl_object_header *h);
2205 void cl_object_header_fini(struct cl_object_header *h);
2206 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2207 void cl_object_get (struct cl_object *o);
2208 void cl_object_attr_lock (struct cl_object *o);
2209 void cl_object_attr_unlock(struct cl_object *o);
2210 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2211 struct cl_attr *attr);
2212 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2213 const struct cl_attr *attr, unsigned valid);
2214 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2215 struct ost_lvb *lvb);
2216 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2217 const struct cl_object_conf *conf);
2218 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2219 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2220 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2221 struct lov_user_md __user *lum);
2222 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2223 ldlm_iterator_t iter, void *data);
2224 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2225 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2227 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2228 struct cl_layout *cl);
2229 loff_t cl_object_maxbytes(struct cl_object *obj);
2232 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2234 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2236 return cl_object_header(o0) == cl_object_header(o1);
2239 static inline void cl_object_page_init(struct cl_object *clob, int size)
2241 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2242 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2243 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2246 static inline void *cl_object_page_slice(struct cl_object *clob,
2247 struct cl_page *page)
2249 return (void *)((char *)page + clob->co_slice_off);
2253 * Return refcount of cl_object.
2255 static inline int cl_object_refc(struct cl_object *clob)
2257 struct lu_object_header *header = clob->co_lu.lo_header;
2258 return atomic_read(&header->loh_ref);
2263 /** \defgroup cl_page cl_page
2271 /* callback of cl_page_gang_lookup() */
2273 struct cl_page *cl_page_find (const struct lu_env *env,
2274 struct cl_object *obj,
2275 pgoff_t idx, struct page *vmpage,
2276 enum cl_page_type type);
2277 struct cl_page *cl_page_alloc (const struct lu_env *env,
2278 struct cl_object *o, pgoff_t ind,
2279 struct page *vmpage,
2280 enum cl_page_type type);
2281 void cl_page_get (struct cl_page *page);
2282 void cl_page_put (const struct lu_env *env,
2283 struct cl_page *page);
2284 void cl_page_print (const struct lu_env *env, void *cookie,
2285 lu_printer_t printer,
2286 const struct cl_page *pg);
2287 void cl_page_header_print(const struct lu_env *env, void *cookie,
2288 lu_printer_t printer,
2289 const struct cl_page *pg);
2290 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2291 struct cl_page *cl_page_top (struct cl_page *page);
2293 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2294 const struct lu_device_type *dtype);
2299 * Functions dealing with the ownership of page by io.
2303 int cl_page_own (const struct lu_env *env,
2304 struct cl_io *io, struct cl_page *page);
2305 int cl_page_own_try (const struct lu_env *env,
2306 struct cl_io *io, struct cl_page *page);
2307 void cl_page_assume (const struct lu_env *env,
2308 struct cl_io *io, struct cl_page *page);
2309 void cl_page_unassume (const struct lu_env *env,
2310 struct cl_io *io, struct cl_page *pg);
2311 void cl_page_disown (const struct lu_env *env,
2312 struct cl_io *io, struct cl_page *page);
2313 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2320 * Functions dealing with the preparation of a page for a transfer, and
2321 * tracking transfer state.
2324 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2325 struct cl_page *pg, enum cl_req_type crt);
2326 void cl_page_completion (const struct lu_env *env,
2327 struct cl_page *pg, enum cl_req_type crt, int ioret);
2328 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2329 enum cl_req_type crt);
2330 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2331 struct cl_page *pg, enum cl_req_type crt);
2332 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2334 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2335 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2336 struct cl_page *pg);
2342 * \name helper routines
2343 * Functions to discard, delete and export a cl_page.
2346 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2347 struct cl_page *pg);
2348 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2349 int cl_page_is_vmlocked(const struct lu_env *env,
2350 const struct cl_page *pg);
2351 void cl_page_export(const struct lu_env *env,
2352 struct cl_page *pg, int uptodate);
2353 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2354 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2355 size_t cl_page_size(const struct cl_object *obj);
2357 void cl_lock_print(const struct lu_env *env, void *cookie,
2358 lu_printer_t printer, const struct cl_lock *lock);
2359 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2360 lu_printer_t printer,
2361 const struct cl_lock_descr *descr);
2365 * Data structure managing a client's cached pages. A count of
2366 * "unstable" pages is maintained, and an LRU of clean pages is
2367 * maintained. "unstable" pages are pages pinned by the ptlrpc
2368 * layer for recovery purposes.
2370 struct cl_client_cache {
2372 * # of client cache refcount
2373 * # of users (OSCs) + 2 (held by llite and lov)
2377 * # of threads are doing shrinking
2379 unsigned int ccc_lru_shrinkers;
2381 * # of LRU entries available
2383 atomic_long_t ccc_lru_left;
2385 * List of entities(OSCs) for this LRU cache
2387 struct list_head ccc_lru;
2389 * Max # of LRU entries
2391 unsigned long ccc_lru_max;
2393 * Lock to protect ccc_lru list
2395 spinlock_t ccc_lru_lock;
2397 * Set if unstable check is enabled
2399 unsigned int ccc_unstable_check:1;
2401 * # of unstable pages for this mount point
2403 atomic_long_t ccc_unstable_nr;
2405 * Waitq for awaiting unstable pages to reach zero.
2406 * Used at umounting time and signaled on BRW commit
2408 wait_queue_head_t ccc_unstable_waitq;
2411 * cl_cache functions
2413 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2414 void cl_cache_incref(struct cl_client_cache *cache);
2415 void cl_cache_decref(struct cl_client_cache *cache);
2419 /** \defgroup cl_lock cl_lock
2421 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2422 struct cl_lock *lock);
2423 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2424 const struct cl_io *io);
2425 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2426 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2427 const struct lu_device_type *dtype);
2428 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2430 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2431 struct cl_lock *lock, struct cl_sync_io *anchor);
2432 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2436 /** \defgroup cl_io cl_io
2439 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2440 enum cl_io_type iot, struct cl_object *obj);
2441 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2442 enum cl_io_type iot, struct cl_object *obj);
2443 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2444 enum cl_io_type iot, loff_t pos, size_t count);
2445 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2447 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2448 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2449 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2450 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2451 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2452 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2453 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2454 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2455 struct cl_io_lock_link *link);
2456 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2457 struct cl_lock_descr *descr);
2458 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2459 enum cl_req_type iot, struct cl_2queue *queue);
2460 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2461 enum cl_req_type iot, struct cl_2queue *queue,
2463 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2464 struct cl_page_list *queue, int from, int to,
2466 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2467 pgoff_t start, struct cl_read_ahead *ra);
2468 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2470 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2471 struct cl_page_list *queue);
2472 int cl_io_is_going (const struct lu_env *env);
2475 * True, iff \a io is an O_APPEND write(2).
2477 static inline int cl_io_is_append(const struct cl_io *io)
2479 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2482 static inline int cl_io_is_sync_write(const struct cl_io *io)
2484 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2487 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2489 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2493 * True, iff \a io is a truncate(2).
2495 static inline int cl_io_is_trunc(const struct cl_io *io)
2497 return io->ci_type == CIT_SETATTR &&
2498 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2501 struct cl_io *cl_io_top(struct cl_io *io);
2503 void cl_io_print(const struct lu_env *env, void *cookie,
2504 lu_printer_t printer, const struct cl_io *io);
2506 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2508 typeof(foo_io) __foo_io = (foo_io); \
2510 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2511 memset(&__foo_io->base + 1, 0, \
2512 (sizeof *__foo_io) - sizeof __foo_io->base); \
2517 /** \defgroup cl_page_list cl_page_list
2521 * Last page in the page list.
2523 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2525 LASSERT(plist->pl_nr > 0);
2526 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2529 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2531 LASSERT(plist->pl_nr > 0);
2532 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2536 * Iterate over pages in a page list.
2538 #define cl_page_list_for_each(page, list) \
2539 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2542 * Iterate over pages in a page list, taking possible removals into account.
2544 #define cl_page_list_for_each_safe(page, temp, list) \
2545 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2547 void cl_page_list_init (struct cl_page_list *plist);
2548 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2549 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2550 struct cl_page *page);
2551 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2552 struct cl_page *page);
2553 void cl_page_list_splice (struct cl_page_list *list,
2554 struct cl_page_list *head);
2555 void cl_page_list_del (const struct lu_env *env,
2556 struct cl_page_list *plist, struct cl_page *page);
2557 void cl_page_list_disown (const struct lu_env *env,
2558 struct cl_io *io, struct cl_page_list *plist);
2559 int cl_page_list_own (const struct lu_env *env,
2560 struct cl_io *io, struct cl_page_list *plist);
2561 void cl_page_list_assume (const struct lu_env *env,
2562 struct cl_io *io, struct cl_page_list *plist);
2563 void cl_page_list_discard(const struct lu_env *env,
2564 struct cl_io *io, struct cl_page_list *plist);
2565 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2567 void cl_2queue_init (struct cl_2queue *queue);
2568 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2569 void cl_2queue_disown (const struct lu_env *env,
2570 struct cl_io *io, struct cl_2queue *queue);
2571 void cl_2queue_assume (const struct lu_env *env,
2572 struct cl_io *io, struct cl_2queue *queue);
2573 void cl_2queue_discard (const struct lu_env *env,
2574 struct cl_io *io, struct cl_2queue *queue);
2575 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2576 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2578 /** @} cl_page_list */
2580 /** \defgroup cl_req cl_req
2582 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2583 enum cl_req_type crt, int nr_objects);
2585 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2586 struct cl_page *page);
2587 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2588 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2589 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2590 struct cl_req_attr *attr, u64 flags);
2591 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2593 /** \defgroup cl_sync_io cl_sync_io
2597 * Anchor for synchronous transfer. This is allocated on a stack by thread
2598 * doing synchronous transfer, and a pointer to this structure is set up in
2599 * every page submitted for transfer. Transfer completion routine updates
2600 * anchor and wakes up waiting thread when transfer is complete.
2603 /** number of pages yet to be transferred. */
2604 atomic_t csi_sync_nr;
2607 /** barrier of destroy this structure */
2608 atomic_t csi_barrier;
2609 /** completion to be signaled when transfer is complete. */
2610 wait_queue_head_t csi_waitq;
2611 /** callback to invoke when this IO is finished */
2612 void (*csi_end_io)(const struct lu_env *,
2613 struct cl_sync_io *);
2616 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2617 void (*end)(const struct lu_env *, struct cl_sync_io *));
2618 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2620 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2622 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2624 /** @} cl_sync_io */
2628 /** \defgroup cl_env cl_env
2630 * lu_env handling for a client.
2632 * lu_env is an environment within which lustre code executes. Its major part
2633 * is lu_context---a fast memory allocation mechanism that is used to conserve
2634 * precious kernel stack space. Originally lu_env was designed for a server,
2637 * - there is a (mostly) fixed number of threads, and
2639 * - call chains have no non-lustre portions inserted between lustre code.
2641 * On a client both these assumtpion fails, because every user thread can
2642 * potentially execute lustre code as part of a system call, and lustre calls
2643 * into VFS or MM that call back into lustre.
2645 * To deal with that, cl_env wrapper functions implement the following
2648 * - allocation and destruction of environment is amortized by caching no
2649 * longer used environments instead of destroying them;
2651 * - there is a notion of "current" environment, attached to the kernel
2652 * data structure representing current thread Top-level lustre code
2653 * allocates an environment and makes it current, then calls into
2654 * non-lustre code, that in turn calls lustre back. Low-level lustre
2655 * code thus called can fetch environment created by the top-level code
2656 * and reuse it, avoiding additional environment allocation.
2657 * Right now, three interfaces can attach the cl_env to running thread:
2660 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2662 * \see lu_env, lu_context, lu_context_key
2665 struct cl_env_nest {
2670 struct lu_env *cl_env_peek (int *refcheck);
2671 struct lu_env *cl_env_get (int *refcheck);
2672 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2673 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2674 void cl_env_put (struct lu_env *env, int *refcheck);
2675 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2676 void *cl_env_reenter (void);
2677 void cl_env_reexit (void *cookie);
2678 void cl_env_implant (struct lu_env *env, int *refcheck);
2679 void cl_env_unplant (struct lu_env *env, int *refcheck);
2680 unsigned cl_env_cache_purge(unsigned nr);
2681 struct lu_env *cl_env_percpu_get (void);
2682 void cl_env_percpu_put (struct lu_env *env);
2689 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2690 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2692 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2693 struct lu_device_type *ldt,
2694 struct lu_device *next);
2697 int cl_global_init(void);
2698 void cl_global_fini(void);
2700 #endif /* _LINUX_CL_OBJECT_H */