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 lu_buf coc_layout;
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. */
323 * Operations implemented for each cl object layer.
325 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
327 struct cl_object_operations {
329 * Initialize page slice for this layer. Called top-to-bottom through
330 * every object layer when a new cl_page is instantiated. Layer
331 * keeping private per-page data, or requiring its own page operations
332 * vector should allocate these data here, and attach then to the page
333 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
336 * \retval NULL success.
338 * \retval ERR_PTR(errno) failure code.
340 * \retval valid-pointer pointer to already existing referenced page
341 * to be used instead of newly created.
343 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
344 struct cl_page *page, pgoff_t index);
346 * Initialize lock slice for this layer. Called top-to-bottom through
347 * every object layer when a new cl_lock is instantiated. Layer
348 * keeping private per-lock data, or requiring its own lock operations
349 * vector should allocate these data here, and attach then to the lock
350 * by calling cl_lock_slice_add(). Mandatory.
352 int (*coo_lock_init)(const struct lu_env *env,
353 struct cl_object *obj, struct cl_lock *lock,
354 const struct cl_io *io);
356 * Initialize io state for a given layer.
358 * called top-to-bottom once per io existence to initialize io
359 * state. If layer wants to keep some state for this type of io, it
360 * has to embed struct cl_io_slice in lu_env::le_ses, and register
361 * slice with cl_io_slice_add(). It is guaranteed that all threads
362 * participating in this io share the same session.
364 int (*coo_io_init)(const struct lu_env *env,
365 struct cl_object *obj, struct cl_io *io);
367 * Fill portion of \a attr that this layer controls. This method is
368 * called top-to-bottom through all object layers.
370 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
372 * \return 0: to continue
373 * \return +ve: to stop iterating through layers (but 0 is returned
374 * from enclosing cl_object_attr_get())
375 * \return -ve: to signal error
377 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
378 struct cl_attr *attr);
382 * \a valid is a bitmask composed from enum #cl_attr_valid, and
383 * indicating what attributes are to be set.
385 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
387 * \return the same convention as for
388 * cl_object_operations::coo_attr_get() is used.
390 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
391 const struct cl_attr *attr, unsigned valid);
393 * Update object configuration. Called top-to-bottom to modify object
396 * XXX error conditions and handling.
398 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
399 const struct cl_object_conf *conf);
401 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
402 * object. Layers are supposed to fill parts of \a lvb that will be
403 * shipped to the glimpse originator as a glimpse result.
405 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
406 * \see osc_object_glimpse()
408 int (*coo_glimpse)(const struct lu_env *env,
409 const struct cl_object *obj, struct ost_lvb *lvb);
411 * Object prune method. Called when the layout is going to change on
412 * this object, therefore each layer has to clean up their cache,
413 * mainly pages and locks.
415 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
417 * Object getstripe method.
419 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
420 struct lov_user_md __user *lum);
422 * Find whether there is any callback data (ldlm lock) attached upon
425 int (*coo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
426 ldlm_iterator_t iter, void *data);
428 * Get FIEMAP mapping from the object.
430 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
431 struct ll_fiemap_info_key *fmkey,
432 struct fiemap *fiemap, size_t *buflen);
434 * Get layout and generation of the object.
436 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
437 struct cl_layout *layout);
439 * Get maximum size of the object.
441 loff_t (*coo_maxbytes)(struct cl_object *obj);
445 * Extended header for client object.
447 struct cl_object_header {
448 /** Standard lu_object_header. cl_object::co_lu::lo_header points
450 struct lu_object_header coh_lu;
453 * Parent object. It is assumed that an object has a well-defined
454 * parent, but not a well-defined child (there may be multiple
455 * sub-objects, for the same top-object). cl_object_header::coh_parent
456 * field allows certain code to be written generically, without
457 * limiting possible cl_object layouts unduly.
459 struct cl_object_header *coh_parent;
461 * Protects consistency between cl_attr of parent object and
462 * attributes of sub-objects, that the former is calculated ("merged")
465 * \todo XXX this can be read/write lock if needed.
467 spinlock_t coh_attr_guard;
469 * Size of cl_page + page slices
471 unsigned short coh_page_bufsize;
473 * Number of objects above this one: 0 for a top-object, 1 for its
476 unsigned char coh_nesting;
480 * Helper macro: iterate over all layers of the object \a obj, assigning every
481 * layer top-to-bottom to \a slice.
483 #define cl_object_for_each(slice, obj) \
484 list_for_each_entry((slice), \
485 &(obj)->co_lu.lo_header->loh_layers,\
489 * Helper macro: iterate over all layers of the object \a obj, assigning every
490 * layer bottom-to-top to \a slice.
492 #define cl_object_for_each_reverse(slice, obj) \
493 list_for_each_entry_reverse((slice), \
494 &(obj)->co_lu.lo_header->loh_layers,\
499 #define CL_PAGE_EOF ((pgoff_t)~0ull)
501 /** \addtogroup cl_page cl_page
505 * Layered client page.
507 * cl_page: represents a portion of a file, cached in the memory. All pages
508 * of the given file are of the same size, and are kept in the radix tree
509 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
510 * of the top-level file object are first class cl_objects, they have their
511 * own radix trees of pages and hence page is implemented as a sequence of
512 * struct cl_pages's, linked into double-linked list through
513 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
514 * corresponding radix tree at the corresponding logical offset.
516 * cl_page is associated with VM page of the hosting environment (struct
517 * page in Linux kernel, for example), struct page. It is assumed, that this
518 * association is implemented by one of cl_page layers (top layer in the
519 * current design) that
521 * - intercepts per-VM-page call-backs made by the environment (e.g.,
524 * - translates state (page flag bits) and locking between lustre and
527 * The association between cl_page and struct page is immutable and
528 * established when cl_page is created.
530 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
531 * this io an exclusive access to this page w.r.t. other io attempts and
532 * various events changing page state (such as transfer completion, or
533 * eviction of the page from the memory). Note, that in general cl_io
534 * cannot be identified with a particular thread, and page ownership is not
535 * exactly equal to the current thread holding a lock on the page. Layer
536 * implementing association between cl_page and struct page has to implement
537 * ownership on top of available synchronization mechanisms.
539 * While lustre client maintains the notion of an page ownership by io,
540 * hosting MM/VM usually has its own page concurrency control
541 * mechanisms. For example, in Linux, page access is synchronized by the
542 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
543 * takes care to acquire and release such locks as necessary around the
544 * calls to the file system methods (->readpage(), ->prepare_write(),
545 * ->commit_write(), etc.). This leads to the situation when there are two
546 * different ways to own a page in the client:
548 * - client code explicitly and voluntary owns the page (cl_page_own());
550 * - VM locks a page and then calls the client, that has "to assume"
551 * the ownership from the VM (cl_page_assume()).
553 * Dual methods to release ownership are cl_page_disown() and
554 * cl_page_unassume().
556 * cl_page is reference counted (cl_page::cp_ref). When reference counter
557 * drops to 0, the page is returned to the cache, unless it is in
558 * cl_page_state::CPS_FREEING state, in which case it is immediately
561 * The general logic guaranteeing the absence of "existential races" for
562 * pages is the following:
564 * - there are fixed known ways for a thread to obtain a new reference
567 * - by doing a lookup in the cl_object radix tree, protected by the
570 * - by starting from VM-locked struct page and following some
571 * hosting environment method (e.g., following ->private pointer in
572 * the case of Linux kernel), see cl_vmpage_page();
574 * - when the page enters cl_page_state::CPS_FREEING state, all these
575 * ways are severed with the proper synchronization
576 * (cl_page_delete());
578 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
581 * - no new references to the page in cl_page_state::CPS_FREEING state
582 * are allowed (checked in cl_page_get()).
584 * Together this guarantees that when last reference to a
585 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
586 * page, as neither references to it can be acquired at that point, nor
589 * cl_page is a state machine. States are enumerated in enum
590 * cl_page_state. Possible state transitions are enumerated in
591 * cl_page_state_set(). State transition process (i.e., actual changing of
592 * cl_page::cp_state field) is protected by the lock on the underlying VM
595 * Linux Kernel implementation.
597 * Binding between cl_page and struct page (which is a typedef for
598 * struct page) is implemented in the vvp layer. cl_page is attached to the
599 * ->private pointer of the struct page, together with the setting of
600 * PG_private bit in page->flags, and acquiring additional reference on the
601 * struct page (much like struct buffer_head, or any similar file system
602 * private data structures).
604 * PG_locked lock is used to implement both ownership and transfer
605 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
606 * states. No additional references are acquired for the duration of the
609 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
610 * write-out is "protected" by the special PG_writeback bit.
614 * States of cl_page. cl_page.c assumes particular order here.
616 * The page state machine is rather crude, as it doesn't recognize finer page
617 * states like "dirty" or "up to date". This is because such states are not
618 * always well defined for the whole stack (see, for example, the
619 * implementation of the read-ahead, that hides page up-to-dateness to track
620 * cache hits accurately). Such sub-states are maintained by the layers that
621 * are interested in them.
625 * Page is in the cache, un-owned. Page leaves cached state in the
628 * - [cl_page_state::CPS_OWNED] io comes across the page and
631 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
632 * req-formation engine decides that it wants to include this page
633 * into an cl_req being constructed, and yanks it from the cache;
635 * - [cl_page_state::CPS_FREEING] VM callback is executed to
636 * evict the page form the memory;
638 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
642 * Page is exclusively owned by some cl_io. Page may end up in this
643 * state as a result of
645 * - io creating new page and immediately owning it;
647 * - [cl_page_state::CPS_CACHED] io finding existing cached page
650 * - [cl_page_state::CPS_OWNED] io finding existing owned page
651 * and waiting for owner to release the page;
653 * Page leaves owned state in the following cases:
655 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
656 * the cache, doing nothing;
658 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
661 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
662 * transfer for this page;
664 * - [cl_page_state::CPS_FREEING] io decides to destroy this
665 * page (e.g., as part of truncate or extent lock cancellation).
667 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
671 * Page is being written out, as a part of a transfer. This state is
672 * entered when req-formation logic decided that it wants this page to
673 * be sent through the wire _now_. Specifically, it means that once
674 * this state is achieved, transfer completion handler (with either
675 * success or failure indication) is guaranteed to be executed against
676 * this page independently of any locks and any scheduling decisions
677 * made by the hosting environment (that effectively means that the
678 * page is never put into cl_page_state::CPS_PAGEOUT state "in
679 * advance". This property is mentioned, because it is important when
680 * reasoning about possible dead-locks in the system). The page can
681 * enter this state as a result of
683 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
684 * write-out of this page, or
686 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
687 * that it has enough dirty pages cached to issue a "good"
690 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
691 * is completed---it is moved into cl_page_state::CPS_CACHED state.
693 * Underlying VM page is locked for the duration of transfer.
695 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
699 * Page is being read in, as a part of a transfer. This is quite
700 * similar to the cl_page_state::CPS_PAGEOUT state, except that
701 * read-in is always "immediate"---there is no such thing a sudden
702 * construction of read cl_req from cached, presumably not up to date,
705 * Underlying VM page is locked for the duration of transfer.
707 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
711 * Page is being destroyed. This state is entered when client decides
712 * that page has to be deleted from its host object, as, e.g., a part
715 * Once this state is reached, there is no way to escape it.
717 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
724 /** Host page, the page is from the host inode which the cl_page
728 /** Transient page, the transient cl_page is used to bind a cl_page
729 * to vmpage which is not belonging to the same object of cl_page.
730 * it is used in DirectIO, lockless IO and liblustre. */
735 * Fields are protected by the lock on struct page, except for atomics and
738 * \invariant Data type invariants are in cl_page_invariant(). Basically:
739 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
740 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
741 * cl_page::cp_owner (when set).
744 /** Reference counter. */
746 /** Transfer error. */
748 /** An object this page is a part of. Immutable after creation. */
749 struct cl_object *cp_obj;
751 struct page *cp_vmpage;
752 /** Linkage of pages within group. Pages must be owned */
753 struct list_head cp_batch;
754 /** List of slices. Immutable after creation. */
755 struct list_head cp_layers;
756 /** Linkage of pages within cl_req. */
757 struct list_head cp_flight;
759 * Page state. This field is const to avoid accidental update, it is
760 * modified only internally within cl_page.c. Protected by a VM lock.
762 const enum cl_page_state cp_state;
764 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
767 enum cl_page_type cp_type;
770 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
771 * by sub-io. Protected by a VM lock.
773 struct cl_io *cp_owner;
775 * Owning IO request in cl_page_state::CPS_PAGEOUT and
776 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
777 * the top-level pages. Protected by a VM lock.
779 struct cl_req *cp_req;
780 /** List of references to this page, for debugging. */
781 struct lu_ref cp_reference;
782 /** Link to an object, for debugging. */
783 struct lu_ref_link cp_obj_ref;
784 /** Link to a queue, for debugging. */
785 struct lu_ref_link cp_queue_ref;
786 /** Assigned if doing a sync_io */
787 struct cl_sync_io *cp_sync_io;
791 * Per-layer part of cl_page.
793 * \see vvp_page, lov_page, osc_page
795 struct cl_page_slice {
796 struct cl_page *cpl_page;
799 * Object slice corresponding to this page slice. Immutable after
802 struct cl_object *cpl_obj;
803 const struct cl_page_operations *cpl_ops;
804 /** Linkage into cl_page::cp_layers. Immutable after creation. */
805 struct list_head cpl_linkage;
809 * Lock mode. For the client extent locks.
821 * Requested transfer type.
831 * Per-layer page operations.
833 * Methods taking an \a io argument are for the activity happening in the
834 * context of given \a io. Page is assumed to be owned by that io, except for
835 * the obvious cases (like cl_page_operations::cpo_own()).
837 * \see vvp_page_ops, lov_page_ops, osc_page_ops
839 struct cl_page_operations {
841 * cl_page<->struct page methods. Only one layer in the stack has to
842 * implement these. Current code assumes that this functionality is
843 * provided by the topmost layer, see cl_page_disown0() as an example.
847 * Called when \a io acquires this page into the exclusive
848 * ownership. When this method returns, it is guaranteed that the is
849 * not owned by other io, and no transfer is going on against
853 * \see vvp_page_own(), lov_page_own()
855 int (*cpo_own)(const struct lu_env *env,
856 const struct cl_page_slice *slice,
857 struct cl_io *io, int nonblock);
858 /** Called when ownership it yielded. Optional.
860 * \see cl_page_disown()
861 * \see vvp_page_disown()
863 void (*cpo_disown)(const struct lu_env *env,
864 const struct cl_page_slice *slice, struct cl_io *io);
866 * Called for a page that is already "owned" by \a io from VM point of
869 * \see cl_page_assume()
870 * \see vvp_page_assume(), lov_page_assume()
872 void (*cpo_assume)(const struct lu_env *env,
873 const struct cl_page_slice *slice, struct cl_io *io);
874 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
875 * bottom-to-top when IO releases a page without actually unlocking
878 * \see cl_page_unassume()
879 * \see vvp_page_unassume()
881 void (*cpo_unassume)(const struct lu_env *env,
882 const struct cl_page_slice *slice,
885 * Announces whether the page contains valid data or not by \a uptodate.
887 * \see cl_page_export()
888 * \see vvp_page_export()
890 void (*cpo_export)(const struct lu_env *env,
891 const struct cl_page_slice *slice, int uptodate);
893 * Checks whether underlying VM page is locked (in the suitable
894 * sense). Used for assertions.
896 * \retval -EBUSY: page is protected by a lock of a given mode;
897 * \retval -ENODATA: page is not protected by a lock;
898 * \retval 0: this layer cannot decide. (Should never happen.)
900 int (*cpo_is_vmlocked)(const struct lu_env *env,
901 const struct cl_page_slice *slice);
907 * Called when page is truncated from the object. Optional.
909 * \see cl_page_discard()
910 * \see vvp_page_discard(), osc_page_discard()
912 void (*cpo_discard)(const struct lu_env *env,
913 const struct cl_page_slice *slice,
916 * Called when page is removed from the cache, and is about to being
917 * destroyed. Optional.
919 * \see cl_page_delete()
920 * \see vvp_page_delete(), osc_page_delete()
922 void (*cpo_delete)(const struct lu_env *env,
923 const struct cl_page_slice *slice);
924 /** Destructor. Frees resources and slice itself. */
925 void (*cpo_fini)(const struct lu_env *env,
926 struct cl_page_slice *slice);
928 * Optional debugging helper. Prints given page slice.
930 * \see cl_page_print()
932 int (*cpo_print)(const struct lu_env *env,
933 const struct cl_page_slice *slice,
934 void *cookie, lu_printer_t p);
938 * Transfer methods. See comment on cl_req for a description of
939 * transfer formation and life-cycle.
944 * Request type dependent vector of operations.
946 * Transfer operations depend on transfer mode (cl_req_type). To avoid
947 * passing transfer mode to each and every of these methods, and to
948 * avoid branching on request type inside of the methods, separate
949 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
950 * provided. That is, method invocation usually looks like
952 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
956 * Called when a page is submitted for a transfer as a part of
959 * \return 0 : page is eligible for submission;
960 * \return -EALREADY : skip this page;
961 * \return -ve : error.
963 * \see cl_page_prep()
965 int (*cpo_prep)(const struct lu_env *env,
966 const struct cl_page_slice *slice,
969 * Completion handler. This is guaranteed to be eventually
970 * fired after cl_page_operations::cpo_prep() or
971 * cl_page_operations::cpo_make_ready() call.
973 * This method can be called in a non-blocking context. It is
974 * guaranteed however, that the page involved and its object
975 * are pinned in memory (and, hence, calling cl_page_put() is
978 * \see cl_page_completion()
980 void (*cpo_completion)(const struct lu_env *env,
981 const struct cl_page_slice *slice,
984 * Called when cached page is about to be added to the
985 * cl_req as a part of req formation.
987 * \return 0 : proceed with this page;
988 * \return -EAGAIN : skip this page;
989 * \return -ve : error.
991 * \see cl_page_make_ready()
993 int (*cpo_make_ready)(const struct lu_env *env,
994 const struct cl_page_slice *slice);
997 * Tell transfer engine that only [to, from] part of a page should be
1000 * This is used for immediate transfers.
1002 * \todo XXX this is not very good interface. It would be much better
1003 * if all transfer parameters were supplied as arguments to
1004 * cl_io_operations::cio_submit() call, but it is not clear how to do
1005 * this for page queues.
1007 * \see cl_page_clip()
1009 void (*cpo_clip)(const struct lu_env *env,
1010 const struct cl_page_slice *slice,
1013 * \pre the page was queued for transferring.
1014 * \post page is removed from client's pending list, or -EBUSY
1015 * is returned if it has already been in transferring.
1017 * This is one of seldom page operation which is:
1018 * 0. called from top level;
1019 * 1. don't have vmpage locked;
1020 * 2. every layer should synchronize execution of its ->cpo_cancel()
1021 * with completion handlers. Osc uses client obd lock for this
1022 * purpose. Based on there is no vvp_page_cancel and
1023 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1025 * \see osc_page_cancel().
1027 int (*cpo_cancel)(const struct lu_env *env,
1028 const struct cl_page_slice *slice);
1030 * Write out a page by kernel. This is only called by ll_writepage
1033 * \see cl_page_flush()
1035 int (*cpo_flush)(const struct lu_env *env,
1036 const struct cl_page_slice *slice,
1042 * Helper macro, dumping detailed information about \a page into a log.
1044 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1046 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1047 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1048 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1049 CDEBUG(mask, format , ## __VA_ARGS__); \
1054 * Helper macro, dumping shorter information about \a page into a log.
1056 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1058 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1059 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1060 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1061 CDEBUG(mask, format , ## __VA_ARGS__); \
1065 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1067 LASSERT(page->cp_vmpage != NULL);
1068 return page->cp_vmpage;
1072 * Check if a cl_page is in use.
1074 * Client cache holds a refcount, this refcount will be dropped when
1075 * the page is taken out of cache, see vvp_page_delete().
1077 static inline bool __page_in_use(const struct cl_page *page, int refc)
1079 return (atomic_read(&page->cp_ref) > refc + 1);
1083 * Caller itself holds a refcount of cl_page.
1085 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1087 * Caller doesn't hold a refcount.
1089 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1093 /** \addtogroup cl_lock cl_lock
1097 * Extent locking on the client.
1101 * The locking model of the new client code is built around
1105 * data-type representing an extent lock on a regular file. cl_lock is a
1106 * layered object (much like cl_object and cl_page), it consists of a header
1107 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1108 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1110 * Typical cl_lock consists of the two layers:
1112 * - vvp_lock (vvp specific data), and
1113 * - lov_lock (lov specific data).
1115 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1116 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1118 * - lovsub_lock, and
1121 * Each sub-lock is associated with a cl_object (representing stripe
1122 * sub-object or the file to which top-level cl_lock is associated to), and is
1123 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1124 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1125 * is different from cl_page, that doesn't fan out (there is usually exactly
1126 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1127 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1131 * cl_lock is a cacheless data container for the requirements of locks to
1132 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1135 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1136 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1138 * INTERFACE AND USAGE
1140 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1141 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1142 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1143 * consists of multiple sub cl_locks, each sub locks will be enqueued
1144 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1145 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1148 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1149 * method will be called for each layer to release the resource held by this
1150 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1151 * clo_enqueue time, is released.
1153 * LDLM lock can only be canceled if there is no cl_lock using it.
1155 * Overall process of the locking during IO operation is as following:
1157 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1158 * is called on each layer. Responsibility of this method is to add locks,
1159 * needed by a given layer into cl_io.ci_lockset.
1161 * - once locks for all layers were collected, they are sorted to avoid
1162 * dead-locks (cl_io_locks_sort()), and enqueued.
1164 * - when all locks are acquired, IO is performed;
1166 * - locks are released after IO is complete.
1168 * Striping introduces major additional complexity into locking. The
1169 * fundamental problem is that it is generally unsafe to actively use (hold)
1170 * two locks on the different OST servers at the same time, as this introduces
1171 * inter-server dependency and can lead to cascading evictions.
1173 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1174 * that no multi-stripe locks are taken (note that this design abandons POSIX
1175 * read/write semantics). Such pieces ideally can be executed concurrently. At
1176 * the same time, certain types of IO cannot be sub-divived, without
1177 * sacrificing correctness. This includes:
1179 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1182 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1184 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1185 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1186 * has to be held together with the usual lock on [offset, offset + count].
1188 * Interaction with DLM
1190 * In the expected setup, cl_lock is ultimately backed up by a collection of
1191 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1192 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1193 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1194 * description of interaction with DLM.
1200 struct cl_lock_descr {
1201 /** Object this lock is granted for. */
1202 struct cl_object *cld_obj;
1203 /** Index of the first page protected by this lock. */
1205 /** Index of the last page (inclusive) protected by this lock. */
1207 /** Group ID, for group lock */
1210 enum cl_lock_mode cld_mode;
1212 * flags to enqueue lock. A combination of bit-flags from
1213 * enum cl_enq_flags.
1215 __u32 cld_enq_flags;
1218 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1219 #define PDESCR(descr) \
1220 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1221 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1223 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1226 * Layered client lock.
1229 /** List of slices. Immutable after creation. */
1230 struct list_head cll_layers;
1231 /** lock attribute, extent, cl_object, etc. */
1232 struct cl_lock_descr cll_descr;
1236 * Per-layer part of cl_lock
1238 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1240 struct cl_lock_slice {
1241 struct cl_lock *cls_lock;
1242 /** Object slice corresponding to this lock slice. Immutable after
1244 struct cl_object *cls_obj;
1245 const struct cl_lock_operations *cls_ops;
1246 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1247 struct list_head cls_linkage;
1252 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1254 struct cl_lock_operations {
1257 * Attempts to enqueue the lock. Called top-to-bottom.
1259 * \retval 0 this layer has enqueued the lock successfully
1260 * \retval >0 this layer has enqueued the lock, but need to wait on
1261 * @anchor for resources
1262 * \retval -ve failure
1264 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1265 * \see osc_lock_enqueue()
1267 int (*clo_enqueue)(const struct lu_env *env,
1268 const struct cl_lock_slice *slice,
1269 struct cl_io *io, struct cl_sync_io *anchor);
1271 * Cancel a lock, release its DLM lock ref, while does not cancel the
1274 void (*clo_cancel)(const struct lu_env *env,
1275 const struct cl_lock_slice *slice);
1278 * Destructor. Frees resources and the slice.
1280 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1281 * \see osc_lock_fini()
1283 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1285 * Optional debugging helper. Prints given lock slice.
1287 int (*clo_print)(const struct lu_env *env,
1288 void *cookie, lu_printer_t p,
1289 const struct cl_lock_slice *slice);
1292 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1294 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1295 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1296 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1297 CDEBUG(mask, format , ## __VA_ARGS__); \
1301 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1305 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1311 /** \addtogroup cl_page_list cl_page_list
1312 * Page list used to perform collective operations on a group of pages.
1314 * Pages are added to the list one by one. cl_page_list acquires a reference
1315 * for every page in it. Page list is used to perform collective operations on
1318 * - submit pages for an immediate transfer,
1320 * - own pages on behalf of certain io (waiting for each page in turn),
1324 * When list is finalized, it releases references on all pages it still has.
1326 * \todo XXX concurrency control.
1330 struct cl_page_list {
1332 struct list_head pl_pages;
1333 struct task_struct *pl_owner;
1337 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1338 * contains an incoming page list and an outgoing page list.
1341 struct cl_page_list c2_qin;
1342 struct cl_page_list c2_qout;
1345 /** @} cl_page_list */
1347 /** \addtogroup cl_io cl_io
1352 * cl_io represents a high level I/O activity like
1353 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1356 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1357 * important distinction. We want to minimize number of calls to the allocator
1358 * in the fast path, e.g., in the case of read(2) when everything is cached:
1359 * client already owns the lock over region being read, and data are cached
1360 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1361 * per-layer io state is stored in the session, associated with the io, see
1362 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1363 * by using free-lists, see cl_env_get().
1365 * There is a small predefined number of possible io types, enumerated in enum
1368 * cl_io is a state machine, that can be advanced concurrently by the multiple
1369 * threads. It is up to these threads to control the concurrency and,
1370 * specifically, to detect when io is done, and its state can be safely
1373 * For read/write io overall execution plan is as following:
1375 * (0) initialize io state through all layers;
1377 * (1) loop: prepare chunk of work to do
1379 * (2) call all layers to collect locks they need to process current chunk
1381 * (3) sort all locks to avoid dead-locks, and acquire them
1383 * (4) process the chunk: call per-page methods
1384 * cl_io_operations::cio_prepare_write(),
1385 * cl_io_operations::cio_commit_write() for write)
1391 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1392 * address allocation efficiency issues mentioned above), and returns with the
1393 * special error condition from per-page method when current sub-io has to
1394 * block. This causes io loop to be repeated, and lov switches to the next
1395 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1400 /** read system call */
1402 /** write system call */
1404 /** truncate, utime system calls */
1406 /** get data version */
1409 * page fault handling
1413 * fsync system call handling
1414 * To write out a range of file
1418 * Miscellaneous io. This is used for occasional io activity that
1419 * doesn't fit into other types. Currently this is used for:
1421 * - cancellation of an extent lock. This io exists as a context
1422 * to write dirty pages from under the lock being canceled back
1425 * - VM induced page write-out. An io context for writing page out
1426 * for memory cleansing;
1428 * - glimpse. An io context to acquire glimpse lock.
1430 * - grouplock. An io context to acquire group lock.
1432 * CIT_MISC io is used simply as a context in which locks and pages
1433 * are manipulated. Such io has no internal "process", that is,
1434 * cl_io_loop() is never called for it.
1441 * States of cl_io state machine
1444 /** Not initialized. */
1448 /** IO iteration started. */
1452 /** Actual IO is in progress. */
1454 /** IO for the current iteration finished. */
1456 /** Locks released. */
1458 /** Iteration completed. */
1460 /** cl_io finalized. */
1465 * IO state private for a layer.
1467 * This is usually embedded into layer session data, rather than allocated
1470 * \see vvp_io, lov_io, osc_io
1472 struct cl_io_slice {
1473 struct cl_io *cis_io;
1474 /** corresponding object slice. Immutable after creation. */
1475 struct cl_object *cis_obj;
1476 /** io operations. Immutable after creation. */
1477 const struct cl_io_operations *cis_iop;
1479 * linkage into a list of all slices for a given cl_io, hanging off
1480 * cl_io::ci_layers. Immutable after creation.
1482 struct list_head cis_linkage;
1485 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1488 struct cl_read_ahead {
1489 /* Maximum page index the readahead window will end.
1490 * This is determined DLM lock coverage, RPC and stripe boundary.
1491 * cra_end is included. */
1493 /* Release routine. If readahead holds resources underneath, this
1494 * function should be called to release it. */
1495 void (*cra_release)(const struct lu_env *env, void *cbdata);
1496 /* Callback data for cra_release routine */
1500 static inline void cl_read_ahead_release(const struct lu_env *env,
1501 struct cl_read_ahead *ra)
1503 if (ra->cra_release != NULL)
1504 ra->cra_release(env, ra->cra_cbdata);
1505 memset(ra, 0, sizeof(*ra));
1510 * Per-layer io operations.
1511 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1513 struct cl_io_operations {
1515 * Vector of io state transition methods for every io type.
1517 * \see cl_page_operations::io
1521 * Prepare io iteration at a given layer.
1523 * Called top-to-bottom at the beginning of each iteration of
1524 * "io loop" (if it makes sense for this type of io). Here
1525 * layer selects what work it will do during this iteration.
1527 * \see cl_io_operations::cio_iter_fini()
1529 int (*cio_iter_init) (const struct lu_env *env,
1530 const struct cl_io_slice *slice);
1532 * Finalize io iteration.
1534 * Called bottom-to-top at the end of each iteration of "io
1535 * loop". Here layers can decide whether IO has to be
1538 * \see cl_io_operations::cio_iter_init()
1540 void (*cio_iter_fini) (const struct lu_env *env,
1541 const struct cl_io_slice *slice);
1543 * Collect locks for the current iteration of io.
1545 * Called top-to-bottom to collect all locks necessary for
1546 * this iteration. This methods shouldn't actually enqueue
1547 * anything, instead it should post a lock through
1548 * cl_io_lock_add(). Once all locks are collected, they are
1549 * sorted and enqueued in the proper order.
1551 int (*cio_lock) (const struct lu_env *env,
1552 const struct cl_io_slice *slice);
1554 * Finalize unlocking.
1556 * Called bottom-to-top to finish layer specific unlocking
1557 * functionality, after generic code released all locks
1558 * acquired by cl_io_operations::cio_lock().
1560 void (*cio_unlock)(const struct lu_env *env,
1561 const struct cl_io_slice *slice);
1563 * Start io iteration.
1565 * Once all locks are acquired, called top-to-bottom to
1566 * commence actual IO. In the current implementation,
1567 * top-level vvp_io_{read,write}_start() does all the work
1568 * synchronously by calling generic_file_*(), so other layers
1569 * are called when everything is done.
1571 int (*cio_start)(const struct lu_env *env,
1572 const struct cl_io_slice *slice);
1574 * Called top-to-bottom at the end of io loop. Here layer
1575 * might wait for an unfinished asynchronous io.
1577 void (*cio_end) (const struct lu_env *env,
1578 const struct cl_io_slice *slice);
1580 * Called bottom-to-top to notify layers that read/write IO
1581 * iteration finished, with \a nob bytes transferred.
1583 void (*cio_advance)(const struct lu_env *env,
1584 const struct cl_io_slice *slice,
1587 * Called once per io, bottom-to-top to release io resources.
1589 void (*cio_fini) (const struct lu_env *env,
1590 const struct cl_io_slice *slice);
1594 * Submit pages from \a queue->c2_qin for IO, and move
1595 * successfully submitted pages into \a queue->c2_qout. Return
1596 * non-zero if failed to submit even the single page. If
1597 * submission failed after some pages were moved into \a
1598 * queue->c2_qout, completion callback with non-zero ioret is
1601 int (*cio_submit)(const struct lu_env *env,
1602 const struct cl_io_slice *slice,
1603 enum cl_req_type crt,
1604 struct cl_2queue *queue);
1606 * Queue async page for write.
1607 * The difference between cio_submit and cio_queue is that
1608 * cio_submit is for urgent request.
1610 int (*cio_commit_async)(const struct lu_env *env,
1611 const struct cl_io_slice *slice,
1612 struct cl_page_list *queue, int from, int to,
1615 * Decide maximum read ahead extent
1617 * \pre io->ci_type == CIT_READ
1619 int (*cio_read_ahead)(const struct lu_env *env,
1620 const struct cl_io_slice *slice,
1621 pgoff_t start, struct cl_read_ahead *ra);
1623 * Optional debugging helper. Print given io slice.
1625 int (*cio_print)(const struct lu_env *env, void *cookie,
1626 lu_printer_t p, const struct cl_io_slice *slice);
1630 * Flags to lock enqueue procedure.
1635 * instruct server to not block, if conflicting lock is found. Instead
1636 * -EWOULDBLOCK is returned immediately.
1638 CEF_NONBLOCK = 0x00000001,
1640 * take lock asynchronously (out of order), as it cannot
1641 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1643 CEF_ASYNC = 0x00000002,
1645 * tell the server to instruct (though a flag in the blocking ast) an
1646 * owner of the conflicting lock, that it can drop dirty pages
1647 * protected by this lock, without sending them to the server.
1649 CEF_DISCARD_DATA = 0x00000004,
1651 * tell the sub layers that it must be a `real' lock. This is used for
1652 * mmapped-buffer locks and glimpse locks that must be never converted
1653 * into lockless mode.
1655 * \see vvp_mmap_locks(), cl_glimpse_lock().
1657 CEF_MUST = 0x00000008,
1659 * tell the sub layers that never request a `real' lock. This flag is
1660 * not used currently.
1662 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1663 * conversion policy: ci_lockreq describes generic information of lock
1664 * requirement for this IO, especially for locks which belong to the
1665 * object doing IO; however, lock itself may have precise requirements
1666 * that are described by the enqueue flags.
1668 CEF_NEVER = 0x00000010,
1670 * for async glimpse lock.
1672 CEF_AGL = 0x00000020,
1674 * enqueue a lock to test DLM lock existence.
1676 CEF_PEEK = 0x00000040,
1678 * mask of enq_flags.
1680 CEF_MASK = 0x0000007f,
1684 * Link between lock and io. Intermediate structure is needed, because the
1685 * same lock can be part of multiple io's simultaneously.
1687 struct cl_io_lock_link {
1688 /** linkage into one of cl_lockset lists. */
1689 struct list_head cill_linkage;
1690 struct cl_lock cill_lock;
1691 /** optional destructor */
1692 void (*cill_fini)(const struct lu_env *env,
1693 struct cl_io_lock_link *link);
1695 #define cill_descr cill_lock.cll_descr
1698 * Lock-set represents a collection of locks, that io needs at a
1699 * time. Generally speaking, client tries to avoid holding multiple locks when
1702 * - holding extent locks over multiple ost's introduces the danger of
1703 * "cascading timeouts";
1705 * - holding multiple locks over the same ost is still dead-lock prone,
1706 * see comment in osc_lock_enqueue(),
1708 * but there are certain situations where this is unavoidable:
1710 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1712 * - truncate has to take [new-size, EOF] lock for correctness;
1714 * - SNS has to take locks across full stripe for correctness;
1716 * - in the case when user level buffer, supplied to {read,write}(file0),
1717 * is a part of a memory mapped lustre file, client has to take a dlm
1718 * locks on file0, and all files that back up the buffer (or a part of
1719 * the buffer, that is being processed in the current chunk, in any
1720 * case, there are situations where at least 2 locks are necessary).
1722 * In such cases we at least try to take locks in the same consistent
1723 * order. To this end, all locks are first collected, then sorted, and then
1727 /** locks to be acquired. */
1728 struct list_head cls_todo;
1729 /** locks acquired. */
1730 struct list_head cls_done;
1734 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1735 * but 'req' is always to be thought as 'request' :-)
1737 enum cl_io_lock_dmd {
1738 /** Always lock data (e.g., O_APPEND). */
1740 /** Layers are free to decide between local and global locking. */
1742 /** Never lock: there is no cache (e.g., liblustre). */
1746 enum cl_fsync_mode {
1747 /** start writeback, do not wait for them to finish */
1749 /** start writeback and wait for them to finish */
1751 /** discard all of dirty pages in a specific file range */
1752 CL_FSYNC_DISCARD = 2,
1753 /** start writeback and make sure they have reached storage before
1754 * return. OST_SYNC RPC must be issued and finished */
1758 struct cl_io_rw_common {
1768 * cl_io is shared by all threads participating in this IO (in current
1769 * implementation only one thread advances IO, but parallel IO design and
1770 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1771 * is up to these threads to serialize their activities, including updates to
1772 * mutable cl_io fields.
1775 /** type of this IO. Immutable after creation. */
1776 enum cl_io_type ci_type;
1777 /** current state of cl_io state machine. */
1778 enum cl_io_state ci_state;
1779 /** main object this io is against. Immutable after creation. */
1780 struct cl_object *ci_obj;
1782 * Upper layer io, of which this io is a part of. Immutable after
1785 struct cl_io *ci_parent;
1786 /** List of slices. Immutable after creation. */
1787 struct list_head ci_layers;
1788 /** list of locks (to be) acquired by this io. */
1789 struct cl_lockset ci_lockset;
1790 /** lock requirements, this is just a help info for sublayers. */
1791 enum cl_io_lock_dmd ci_lockreq;
1794 struct cl_io_rw_common rd;
1797 struct cl_io_rw_common wr;
1801 struct cl_io_rw_common ci_rw;
1802 struct cl_setattr_io {
1803 struct ost_lvb sa_attr;
1804 unsigned int sa_attr_flags;
1805 unsigned int sa_valid;
1806 int sa_stripe_index;
1807 const struct lu_fid *sa_parent_fid;
1809 struct cl_data_version_io {
1810 u64 dv_data_version;
1813 struct cl_fault_io {
1814 /** page index within file. */
1816 /** bytes valid byte on a faulted page. */
1818 /** writable page? for nopage() only */
1820 /** page of an executable? */
1822 /** page_mkwrite() */
1824 /** resulting page */
1825 struct cl_page *ft_page;
1827 struct cl_fsync_io {
1830 /** file system level fid */
1831 struct lu_fid *fi_fid;
1832 enum cl_fsync_mode fi_mode;
1833 /* how many pages were written/discarded */
1834 unsigned int fi_nr_written;
1837 struct cl_2queue ci_queue;
1840 unsigned int ci_continue:1,
1842 * This io has held grouplock, to inform sublayers that
1843 * don't do lockless i/o.
1847 * The whole IO need to be restarted because layout has been changed
1851 * to not refresh layout - the IO issuer knows that the layout won't
1852 * change(page operations, layout change causes all page to be
1853 * discarded), or it doesn't matter if it changes(sync).
1857 * Check if layout changed after the IO finishes. Mainly for HSM
1858 * requirement. If IO occurs to openning files, it doesn't need to
1859 * verify layout because HSM won't release openning files.
1860 * Right now, only two opertaions need to verify layout: glimpse
1865 * file is released, restore has to to be triggered by vvp layer
1867 ci_restore_needed:1,
1873 * Number of pages owned by this IO. For invariant checking.
1875 unsigned ci_owned_nr;
1880 /** \addtogroup cl_req cl_req
1885 * There are two possible modes of transfer initiation on the client:
1887 * - immediate transfer: this is started when a high level io wants a page
1888 * or a collection of pages to be transferred right away. Examples:
1889 * read-ahead, synchronous read in the case of non-page aligned write,
1890 * page write-out as a part of extent lock cancellation, page write-out
1891 * as a part of memory cleansing. Immediate transfer can be both
1892 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1894 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1895 * when io wants to transfer a page to the server some time later, when
1896 * it can be done efficiently. Example: pages dirtied by the write(2)
1899 * In any case, transfer takes place in the form of a cl_req, which is a
1900 * representation for a network RPC.
1902 * Pages queued for an opportunistic transfer are cached until it is decided
1903 * that efficient RPC can be composed of them. This decision is made by "a
1904 * req-formation engine", currently implemented as a part of osc
1905 * layer. Req-formation depends on many factors: the size of the resulting
1906 * RPC, whether or not multi-object RPCs are supported by the server,
1907 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1909 * For the immediate transfer io submits a cl_page_list, that req-formation
1910 * engine slices into cl_req's, possibly adding cached pages to some of
1911 * the resulting req's.
1913 * Whenever a page from cl_page_list is added to a newly constructed req, its
1914 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1915 * page state is atomically changed from cl_page_state::CPS_OWNED to
1916 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1917 * is zeroed, and cl_page::cp_req is set to the
1918 * req. cl_page_operations::cpo_prep() method at the particular layer might
1919 * return -EALREADY to indicate that it does not need to submit this page
1920 * at all. This is possible, for example, if page, submitted for read,
1921 * became up-to-date in the meantime; and for write, the page don't have
1922 * dirty bit marked. \see cl_io_submit_rw()
1924 * Whenever a cached page is added to a newly constructed req, its
1925 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1926 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1927 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1928 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1929 * might return -EAGAIN to indicate that this page is not eligible for the
1930 * transfer right now.
1934 * Plan is to divide transfers into "priority bands" (indicated when
1935 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1936 * and allow glueing of cached pages to immediate transfers only within single
1937 * band. This would make high priority transfers (like lock cancellation or
1938 * memory pressure induced write-out) really high priority.
1943 * Per-transfer attributes.
1945 struct cl_req_attr {
1946 /** Generic attributes for the server consumption. */
1947 struct obdo *cra_oa;
1949 char cra_jobid[LUSTRE_JOBID_SIZE];
1953 * Transfer request operations definable at every layer.
1955 * Concurrency: transfer formation engine synchronizes calls to all transfer
1958 struct cl_req_operations {
1960 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1961 * complete (all pages are added).
1963 * \see osc_req_prep()
1965 int (*cro_prep)(const struct lu_env *env,
1966 const struct cl_req_slice *slice);
1968 * Called top-to-bottom to fill in \a oa fields. This is called twice
1969 * with different flags, see bug 10150 and osc_build_req().
1971 * \param obj an object from cl_req which attributes are to be set in
1974 * \param oa struct obdo where attributes are placed
1976 * \param flags \a oa fields to be filled.
1978 void (*cro_attr_set)(const struct lu_env *env,
1979 const struct cl_req_slice *slice,
1980 const struct cl_object *obj,
1981 struct cl_req_attr *attr, u64 flags);
1983 * Called top-to-bottom from cl_req_completion() to notify layers that
1984 * transfer completed. Has to free all state allocated by
1985 * cl_device_operations::cdo_req_init().
1987 void (*cro_completion)(const struct lu_env *env,
1988 const struct cl_req_slice *slice, int ioret);
1992 * A per-object state that (potentially multi-object) transfer request keeps.
1995 /** object itself */
1996 struct cl_object *ro_obj;
1997 /** reference to cl_req_obj::ro_obj. For debugging. */
1998 struct lu_ref_link ro_obj_ref;
1999 /* something else? Number of pages for a given object? */
2005 * Transfer requests are not reference counted, because IO sub-system owns
2006 * them exclusively and knows when to free them.
2010 * cl_req is created by cl_req_alloc() that calls
2011 * cl_device_operations::cdo_req_init() device methods to allocate per-req
2012 * state in every layer.
2014 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
2015 * contains pages for.
2017 * Once all pages were collected, cl_page_operations::cpo_prep() method is
2018 * called top-to-bottom. At that point layers can modify req, let it pass, or
2019 * deny it completely. This is to support things like SNS that have transfer
2020 * ordering requirements invisible to the individual req-formation engine.
2022 * On transfer completion (or transfer timeout, or failure to initiate the
2023 * transfer of an allocated req), cl_req_operations::cro_completion() method
2024 * is called, after execution of cl_page_operations::cpo_completion() of all
2028 enum cl_req_type crq_type;
2029 /** A list of pages being transferred */
2030 struct list_head crq_pages;
2031 /** Number of pages in cl_req::crq_pages */
2032 unsigned crq_nrpages;
2033 /** An array of objects which pages are in ->crq_pages */
2034 struct cl_req_obj *crq_o;
2035 /** Number of elements in cl_req::crq_objs[] */
2036 unsigned crq_nrobjs;
2037 struct list_head crq_layers;
2041 * Per-layer state for request.
2043 struct cl_req_slice {
2044 struct cl_req *crs_req;
2045 struct cl_device *crs_dev;
2046 struct list_head crs_linkage;
2047 const struct cl_req_operations *crs_ops;
2052 enum cache_stats_item {
2053 /** how many cache lookups were performed */
2055 /** how many times cache lookup resulted in a hit */
2057 /** how many entities are in the cache right now */
2059 /** how many entities in the cache are actively used (and cannot be
2060 * evicted) right now */
2062 /** how many entities were created at all */
2067 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2070 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2072 struct cache_stats {
2073 const char *cs_name;
2074 atomic_t cs_stats[CS_NR];
2077 /** These are not exported so far */
2078 void cache_stats_init (struct cache_stats *cs, const char *name);
2081 * Client-side site. This represents particular client stack. "Global"
2082 * variables should (directly or indirectly) be added here to allow multiple
2083 * clients to co-exist in the single address space.
2086 struct lu_site cs_lu;
2088 * Statistical counters. Atomics do not scale, something better like
2089 * per-cpu counters is needed.
2091 * These are exported as /proc/fs/lustre/llite/.../site
2093 * When interpreting keep in mind that both sub-locks (and sub-pages)
2094 * and top-locks (and top-pages) are accounted here.
2096 struct cache_stats cs_pages;
2097 atomic_t cs_pages_state[CPS_NR];
2100 int cl_site_init(struct cl_site *s, struct cl_device *top);
2101 void cl_site_fini(struct cl_site *s);
2102 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2105 * Output client site statistical counters into a buffer. Suitable for
2106 * ll_rd_*()-style functions.
2108 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2113 * Type conversion and accessory functions.
2117 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2119 return container_of(site, struct cl_site, cs_lu);
2122 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2124 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2125 return container_of0(d, struct cl_device, cd_lu_dev);
2128 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2130 return &d->cd_lu_dev;
2133 static inline struct cl_object *lu2cl(const struct lu_object *o)
2135 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2136 return container_of0(o, struct cl_object, co_lu);
2139 static inline const struct cl_object_conf *
2140 lu2cl_conf(const struct lu_object_conf *conf)
2142 return container_of0(conf, struct cl_object_conf, coc_lu);
2145 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2147 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2150 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2152 return container_of0(h, struct cl_object_header, coh_lu);
2155 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2157 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2161 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2163 return luh2coh(obj->co_lu.lo_header);
2166 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2168 return lu_device_init(&d->cd_lu_dev, t);
2171 static inline void cl_device_fini(struct cl_device *d)
2173 lu_device_fini(&d->cd_lu_dev);
2176 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2177 struct cl_object *obj, pgoff_t index,
2178 const struct cl_page_operations *ops);
2179 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2180 struct cl_object *obj,
2181 const struct cl_lock_operations *ops);
2182 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2183 struct cl_object *obj, const struct cl_io_operations *ops);
2184 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2185 struct cl_device *dev,
2186 const struct cl_req_operations *ops);
2189 /** \defgroup cl_object cl_object
2191 struct cl_object *cl_object_top (struct cl_object *o);
2192 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2193 const struct lu_fid *fid,
2194 const struct cl_object_conf *c);
2196 int cl_object_header_init(struct cl_object_header *h);
2197 void cl_object_header_fini(struct cl_object_header *h);
2198 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2199 void cl_object_get (struct cl_object *o);
2200 void cl_object_attr_lock (struct cl_object *o);
2201 void cl_object_attr_unlock(struct cl_object *o);
2202 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2203 struct cl_attr *attr);
2204 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2205 const struct cl_attr *attr, unsigned valid);
2206 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2207 struct ost_lvb *lvb);
2208 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2209 const struct cl_object_conf *conf);
2210 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2211 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2212 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2213 struct lov_user_md __user *lum);
2214 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2215 ldlm_iterator_t iter, void *data);
2216 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2217 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2219 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2220 struct cl_layout *cl);
2221 loff_t cl_object_maxbytes(struct cl_object *obj);
2224 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2226 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2228 return cl_object_header(o0) == cl_object_header(o1);
2231 static inline void cl_object_page_init(struct cl_object *clob, int size)
2233 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2234 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2235 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2238 static inline void *cl_object_page_slice(struct cl_object *clob,
2239 struct cl_page *page)
2241 return (void *)((char *)page + clob->co_slice_off);
2245 * Return refcount of cl_object.
2247 static inline int cl_object_refc(struct cl_object *clob)
2249 struct lu_object_header *header = clob->co_lu.lo_header;
2250 return atomic_read(&header->loh_ref);
2255 /** \defgroup cl_page cl_page
2263 /* callback of cl_page_gang_lookup() */
2265 struct cl_page *cl_page_find (const struct lu_env *env,
2266 struct cl_object *obj,
2267 pgoff_t idx, struct page *vmpage,
2268 enum cl_page_type type);
2269 struct cl_page *cl_page_alloc (const struct lu_env *env,
2270 struct cl_object *o, pgoff_t ind,
2271 struct page *vmpage,
2272 enum cl_page_type type);
2273 void cl_page_get (struct cl_page *page);
2274 void cl_page_put (const struct lu_env *env,
2275 struct cl_page *page);
2276 void cl_page_print (const struct lu_env *env, void *cookie,
2277 lu_printer_t printer,
2278 const struct cl_page *pg);
2279 void cl_page_header_print(const struct lu_env *env, void *cookie,
2280 lu_printer_t printer,
2281 const struct cl_page *pg);
2282 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2283 struct cl_page *cl_page_top (struct cl_page *page);
2285 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2286 const struct lu_device_type *dtype);
2291 * Functions dealing with the ownership of page by io.
2295 int cl_page_own (const struct lu_env *env,
2296 struct cl_io *io, struct cl_page *page);
2297 int cl_page_own_try (const struct lu_env *env,
2298 struct cl_io *io, struct cl_page *page);
2299 void cl_page_assume (const struct lu_env *env,
2300 struct cl_io *io, struct cl_page *page);
2301 void cl_page_unassume (const struct lu_env *env,
2302 struct cl_io *io, struct cl_page *pg);
2303 void cl_page_disown (const struct lu_env *env,
2304 struct cl_io *io, struct cl_page *page);
2305 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2312 * Functions dealing with the preparation of a page for a transfer, and
2313 * tracking transfer state.
2316 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2317 struct cl_page *pg, enum cl_req_type crt);
2318 void cl_page_completion (const struct lu_env *env,
2319 struct cl_page *pg, enum cl_req_type crt, int ioret);
2320 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2321 enum cl_req_type crt);
2322 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2323 struct cl_page *pg, enum cl_req_type crt);
2324 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2326 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2327 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2328 struct cl_page *pg);
2334 * \name helper routines
2335 * Functions to discard, delete and export a cl_page.
2338 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2339 struct cl_page *pg);
2340 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2341 int cl_page_is_vmlocked(const struct lu_env *env,
2342 const struct cl_page *pg);
2343 void cl_page_export(const struct lu_env *env,
2344 struct cl_page *pg, int uptodate);
2345 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2346 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2347 size_t cl_page_size(const struct cl_object *obj);
2349 void cl_lock_print(const struct lu_env *env, void *cookie,
2350 lu_printer_t printer, const struct cl_lock *lock);
2351 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2352 lu_printer_t printer,
2353 const struct cl_lock_descr *descr);
2357 * Data structure managing a client's cached pages. A count of
2358 * "unstable" pages is maintained, and an LRU of clean pages is
2359 * maintained. "unstable" pages are pages pinned by the ptlrpc
2360 * layer for recovery purposes.
2362 struct cl_client_cache {
2364 * # of client cache refcount
2365 * # of users (OSCs) + 2 (held by llite and lov)
2369 * # of threads are doing shrinking
2371 unsigned int ccc_lru_shrinkers;
2373 * # of LRU entries available
2375 atomic_long_t ccc_lru_left;
2377 * List of entities(OSCs) for this LRU cache
2379 struct list_head ccc_lru;
2381 * Max # of LRU entries
2383 unsigned long ccc_lru_max;
2385 * Lock to protect ccc_lru list
2387 spinlock_t ccc_lru_lock;
2389 * Set if unstable check is enabled
2391 unsigned int ccc_unstable_check:1;
2393 * # of unstable pages for this mount point
2395 atomic_long_t ccc_unstable_nr;
2397 * Waitq for awaiting unstable pages to reach zero.
2398 * Used at umounting time and signaled on BRW commit
2400 wait_queue_head_t ccc_unstable_waitq;
2403 * cl_cache functions
2405 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2406 void cl_cache_incref(struct cl_client_cache *cache);
2407 void cl_cache_decref(struct cl_client_cache *cache);
2411 /** \defgroup cl_lock cl_lock
2413 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2414 struct cl_lock *lock);
2415 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2416 const struct cl_io *io);
2417 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2418 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2419 const struct lu_device_type *dtype);
2420 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2422 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2423 struct cl_lock *lock, struct cl_sync_io *anchor);
2424 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2428 /** \defgroup cl_io cl_io
2431 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2432 enum cl_io_type iot, struct cl_object *obj);
2433 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2434 enum cl_io_type iot, struct cl_object *obj);
2435 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2436 enum cl_io_type iot, loff_t pos, size_t count);
2437 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2439 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2440 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2441 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2442 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2443 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2444 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2445 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2446 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2447 struct cl_io_lock_link *link);
2448 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2449 struct cl_lock_descr *descr);
2450 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2451 enum cl_req_type iot, struct cl_2queue *queue);
2452 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2453 enum cl_req_type iot, struct cl_2queue *queue,
2455 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2456 struct cl_page_list *queue, int from, int to,
2458 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2459 pgoff_t start, struct cl_read_ahead *ra);
2460 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2462 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2463 struct cl_page_list *queue);
2464 int cl_io_is_going (const struct lu_env *env);
2467 * True, iff \a io is an O_APPEND write(2).
2469 static inline int cl_io_is_append(const struct cl_io *io)
2471 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2474 static inline int cl_io_is_sync_write(const struct cl_io *io)
2476 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2479 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2481 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2485 * True, iff \a io is a truncate(2).
2487 static inline int cl_io_is_trunc(const struct cl_io *io)
2489 return io->ci_type == CIT_SETATTR &&
2490 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2493 struct cl_io *cl_io_top(struct cl_io *io);
2495 void cl_io_print(const struct lu_env *env, void *cookie,
2496 lu_printer_t printer, const struct cl_io *io);
2498 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2500 typeof(foo_io) __foo_io = (foo_io); \
2502 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2503 memset(&__foo_io->base + 1, 0, \
2504 (sizeof *__foo_io) - sizeof __foo_io->base); \
2509 /** \defgroup cl_page_list cl_page_list
2513 * Last page in the page list.
2515 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2517 LASSERT(plist->pl_nr > 0);
2518 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2521 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2523 LASSERT(plist->pl_nr > 0);
2524 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2528 * Iterate over pages in a page list.
2530 #define cl_page_list_for_each(page, list) \
2531 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2534 * Iterate over pages in a page list, taking possible removals into account.
2536 #define cl_page_list_for_each_safe(page, temp, list) \
2537 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2539 void cl_page_list_init (struct cl_page_list *plist);
2540 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2541 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2542 struct cl_page *page);
2543 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2544 struct cl_page *page);
2545 void cl_page_list_splice (struct cl_page_list *list,
2546 struct cl_page_list *head);
2547 void cl_page_list_del (const struct lu_env *env,
2548 struct cl_page_list *plist, struct cl_page *page);
2549 void cl_page_list_disown (const struct lu_env *env,
2550 struct cl_io *io, struct cl_page_list *plist);
2551 int cl_page_list_own (const struct lu_env *env,
2552 struct cl_io *io, struct cl_page_list *plist);
2553 void cl_page_list_assume (const struct lu_env *env,
2554 struct cl_io *io, struct cl_page_list *plist);
2555 void cl_page_list_discard(const struct lu_env *env,
2556 struct cl_io *io, struct cl_page_list *plist);
2557 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2559 void cl_2queue_init (struct cl_2queue *queue);
2560 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2561 void cl_2queue_disown (const struct lu_env *env,
2562 struct cl_io *io, struct cl_2queue *queue);
2563 void cl_2queue_assume (const struct lu_env *env,
2564 struct cl_io *io, struct cl_2queue *queue);
2565 void cl_2queue_discard (const struct lu_env *env,
2566 struct cl_io *io, struct cl_2queue *queue);
2567 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2568 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2570 /** @} cl_page_list */
2572 /** \defgroup cl_req cl_req
2574 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2575 enum cl_req_type crt, int nr_objects);
2577 void cl_req_page_add (const struct lu_env *env, struct cl_req *req,
2578 struct cl_page *page);
2579 void cl_req_page_done (const struct lu_env *env, struct cl_page *page);
2580 int cl_req_prep (const struct lu_env *env, struct cl_req *req);
2581 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2582 struct cl_req_attr *attr, u64 flags);
2583 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2585 /** \defgroup cl_sync_io cl_sync_io
2589 * Anchor for synchronous transfer. This is allocated on a stack by thread
2590 * doing synchronous transfer, and a pointer to this structure is set up in
2591 * every page submitted for transfer. Transfer completion routine updates
2592 * anchor and wakes up waiting thread when transfer is complete.
2595 /** number of pages yet to be transferred. */
2596 atomic_t csi_sync_nr;
2599 /** barrier of destroy this structure */
2600 atomic_t csi_barrier;
2601 /** completion to be signaled when transfer is complete. */
2602 wait_queue_head_t csi_waitq;
2603 /** callback to invoke when this IO is finished */
2604 void (*csi_end_io)(const struct lu_env *,
2605 struct cl_sync_io *);
2608 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2609 void (*end)(const struct lu_env *, struct cl_sync_io *));
2610 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2612 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2614 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2616 /** @} cl_sync_io */
2620 /** \defgroup cl_env cl_env
2622 * lu_env handling for a client.
2624 * lu_env is an environment within which lustre code executes. Its major part
2625 * is lu_context---a fast memory allocation mechanism that is used to conserve
2626 * precious kernel stack space. Originally lu_env was designed for a server,
2629 * - there is a (mostly) fixed number of threads, and
2631 * - call chains have no non-lustre portions inserted between lustre code.
2633 * On a client both these assumtpion fails, because every user thread can
2634 * potentially execute lustre code as part of a system call, and lustre calls
2635 * into VFS or MM that call back into lustre.
2637 * To deal with that, cl_env wrapper functions implement the following
2640 * - allocation and destruction of environment is amortized by caching no
2641 * longer used environments instead of destroying them;
2643 * - there is a notion of "current" environment, attached to the kernel
2644 * data structure representing current thread Top-level lustre code
2645 * allocates an environment and makes it current, then calls into
2646 * non-lustre code, that in turn calls lustre back. Low-level lustre
2647 * code thus called can fetch environment created by the top-level code
2648 * and reuse it, avoiding additional environment allocation.
2649 * Right now, three interfaces can attach the cl_env to running thread:
2652 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2654 * \see lu_env, lu_context, lu_context_key
2657 struct cl_env_nest {
2662 struct lu_env *cl_env_peek (int *refcheck);
2663 struct lu_env *cl_env_get (int *refcheck);
2664 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2665 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2666 void cl_env_put (struct lu_env *env, int *refcheck);
2667 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2668 void *cl_env_reenter (void);
2669 void cl_env_reexit (void *cookie);
2670 void cl_env_implant (struct lu_env *env, int *refcheck);
2671 void cl_env_unplant (struct lu_env *env, int *refcheck);
2672 unsigned cl_env_cache_purge(unsigned nr);
2673 struct lu_env *cl_env_percpu_get (void);
2674 void cl_env_percpu_put (struct lu_env *env);
2681 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2682 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2684 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2685 struct lu_device_type *ldt,
2686 struct lu_device *next);
2689 int cl_global_init(void);
2690 void cl_global_fini(void);
2692 #endif /* _LINUX_CL_OBJECT_H */