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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 #ifndef _LUSTRE_CL_OBJECT_H
33 #define _LUSTRE_CL_OBJECT_H
35 /** \defgroup clio clio
37 * Client objects implement io operations and cache pages.
39 * Examples: lov and osc are implementations of cl interface.
41 * Big Theory Statement.
45 * Client implementation is based on the following data-types:
51 * - cl_lock represents an extent lock on an object.
53 * - cl_io represents high-level i/o activity such as whole read/write
54 * system call, or write-out of pages from under the lock being
55 * canceled. cl_io has sub-ios that can be stopped and resumed
56 * independently, thus achieving high degree of transfer
57 * parallelism. Single cl_io can be advanced forward by
58 * the multiple threads (although in the most usual case of
59 * read/write system call it is associated with the single user
60 * thread, that issued the system call).
64 * - to avoid confusion high-level I/O operation like read or write system
65 * call is referred to as "an io", whereas low-level I/O operation, like
66 * RPC, is referred to as "a transfer"
68 * - "generic code" means generic (not file system specific) code in the
69 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
70 * is not layer specific.
76 * - cl_object_header::coh_page_guard
79 * See the top comment in cl_object.c for the description of overall locking and
80 * reference-counting design.
82 * See comments below for the description of i/o, page, and dlm-locking
89 * super-class definitions.
91 #include <libcfs/libcfs.h>
92 #include <lu_object.h>
93 #include <linux/atomic.h>
94 #include <linux/mutex.h>
95 #include <linux/radix-tree.h>
96 #include <linux/spinlock.h>
97 #include <linux/wait.h>
98 #include <lustre_dlm.h>
108 struct cl_page_slice;
110 struct cl_lock_slice;
112 struct cl_lock_operations;
113 struct cl_page_operations;
121 * Device in the client stack.
123 * \see vvp_device, lov_device, lovsub_device, osc_device
127 struct lu_device cd_lu_dev;
130 /** \addtogroup cl_object cl_object
133 * "Data attributes" of cl_object. Data attributes can be updated
134 * independently for a sub-object, and top-object's attributes are calculated
135 * from sub-objects' ones.
138 /** Object size, in bytes */
141 * Known minimal size, in bytes.
143 * This is only valid when at least one DLM lock is held.
146 /** Modification time. Measured in seconds since epoch. */
148 /** Access time. Measured in seconds since epoch. */
150 /** Change time. Measured in seconds since epoch. */
153 * Blocks allocated to this cl_object on the server file system.
155 * \todo XXX An interface for block size is needed.
159 * User identifier for quota purposes.
163 * Group identifier for quota purposes.
167 /* nlink of the directory */
170 /* Project identifier for quota purpose. */
175 * Fields in cl_attr that are being set.
190 * Sub-class of lu_object with methods common for objects on the client
193 * cl_object: represents a regular file system object, both a file and a
194 * stripe. cl_object is based on lu_object: it is identified by a fid,
195 * layered, cached, hashed, and lrued. Important distinction with the server
196 * side, where md_object and dt_object are used, is that cl_object "fans out"
197 * at the lov/sns level: depending on the file layout, single file is
198 * represented as a set of "sub-objects" (stripes). At the implementation
199 * level, struct lov_object contains an array of cl_objects. Each sub-object
200 * is a full-fledged cl_object, having its fid, living in the lru and hash
203 * This leads to the next important difference with the server side: on the
204 * client, it's quite usual to have objects with the different sequence of
205 * layers. For example, typical top-object is composed of the following
211 * whereas its sub-objects are composed of
216 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
217 * track of the object-subobject relationship.
219 * Sub-objects are not cached independently: when top-object is about to
220 * be discarded from the memory, all its sub-objects are torn-down and
223 * \see vvp_object, lov_object, lovsub_object, osc_object
227 struct lu_object co_lu;
228 /** per-object-layer operations */
229 const struct cl_object_operations *co_ops;
230 /** offset of page slice in cl_page buffer */
235 * Description of the client object configuration. This is used for the
236 * creation of a new client object that is identified by a more state than
239 struct cl_object_conf {
241 struct lu_object_conf coc_lu;
244 * Object layout. This is consumed by lov.
246 struct lu_buf coc_layout;
248 * Description of particular stripe location in the
249 * cluster. This is consumed by osc.
251 struct lov_oinfo *coc_oinfo;
254 * VFS inode. This is consumed by vvp.
256 struct inode *coc_inode;
258 * Layout lock handle.
260 struct ldlm_lock *coc_lock;
262 * Operation to handle layout, OBJECT_CONF_XYZ.
268 /** configure layout, set up a new stripe, must be called while
269 * holding layout lock. */
271 /** invalidate the current stripe configuration due to losing
273 OBJECT_CONF_INVALIDATE = 1,
274 /** wait for old layout to go away so that new layout can be
280 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
281 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
285 /** the buffer to return the layout in lov_mds_md format. */
286 struct lu_buf cl_buf;
287 /** size of layout in lov_mds_md format. */
289 /** Layout generation. */
291 /** whether layout is a composite one */
292 bool cl_is_composite;
296 * Operations implemented for each cl object layer.
298 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
300 struct cl_object_operations {
302 * Initialize page slice for this layer. Called top-to-bottom through
303 * every object layer when a new cl_page is instantiated. Layer
304 * keeping private per-page data, or requiring its own page operations
305 * vector should allocate these data here, and attach then to the page
306 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
309 * \retval NULL success.
311 * \retval ERR_PTR(errno) failure code.
313 * \retval valid-pointer pointer to already existing referenced page
314 * to be used instead of newly created.
316 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
317 struct cl_page *page, pgoff_t index);
319 * Initialize lock slice for this layer. Called top-to-bottom through
320 * every object layer when a new cl_lock is instantiated. Layer
321 * keeping private per-lock data, or requiring its own lock operations
322 * vector should allocate these data here, and attach then to the lock
323 * by calling cl_lock_slice_add(). Mandatory.
325 int (*coo_lock_init)(const struct lu_env *env,
326 struct cl_object *obj, struct cl_lock *lock,
327 const struct cl_io *io);
329 * Initialize io state for a given layer.
331 * called top-to-bottom once per io existence to initialize io
332 * state. If layer wants to keep some state for this type of io, it
333 * has to embed struct cl_io_slice in lu_env::le_ses, and register
334 * slice with cl_io_slice_add(). It is guaranteed that all threads
335 * participating in this io share the same session.
337 int (*coo_io_init)(const struct lu_env *env,
338 struct cl_object *obj, struct cl_io *io);
340 * Fill portion of \a attr that this layer controls. This method is
341 * called top-to-bottom through all object layers.
343 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
345 * \return 0: to continue
346 * \return +ve: to stop iterating through layers (but 0 is returned
347 * from enclosing cl_object_attr_get())
348 * \return -ve: to signal error
350 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
351 struct cl_attr *attr);
355 * \a valid is a bitmask composed from enum #cl_attr_valid, and
356 * indicating what attributes are to be set.
358 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
360 * \return the same convention as for
361 * cl_object_operations::coo_attr_get() is used.
363 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
364 const struct cl_attr *attr, unsigned valid);
366 * Update object configuration. Called top-to-bottom to modify object
369 * XXX error conditions and handling.
371 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
372 const struct cl_object_conf *conf);
374 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
375 * object. Layers are supposed to fill parts of \a lvb that will be
376 * shipped to the glimpse originator as a glimpse result.
378 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
379 * \see osc_object_glimpse()
381 int (*coo_glimpse)(const struct lu_env *env,
382 const struct cl_object *obj, struct ost_lvb *lvb);
384 * Object prune method. Called when the layout is going to change on
385 * this object, therefore each layer has to clean up their cache,
386 * mainly pages and locks.
388 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
390 * Object getstripe method.
392 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
393 struct lov_user_md __user *lum);
395 * Get FIEMAP mapping from the object.
397 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
398 struct ll_fiemap_info_key *fmkey,
399 struct fiemap *fiemap, size_t *buflen);
401 * Get layout and generation of the object.
403 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
404 struct cl_layout *layout);
406 * Get maximum size of the object.
408 loff_t (*coo_maxbytes)(struct cl_object *obj);
410 * Set request attributes.
412 void (*coo_req_attr_set)(const struct lu_env *env,
413 struct cl_object *obj,
414 struct cl_req_attr *attr);
418 * Extended header for client object.
420 struct cl_object_header {
421 /** Standard lu_object_header. cl_object::co_lu::lo_header points
423 struct lu_object_header coh_lu;
426 * Parent object. It is assumed that an object has a well-defined
427 * parent, but not a well-defined child (there may be multiple
428 * sub-objects, for the same top-object). cl_object_header::coh_parent
429 * field allows certain code to be written generically, without
430 * limiting possible cl_object layouts unduly.
432 struct cl_object_header *coh_parent;
434 * Protects consistency between cl_attr of parent object and
435 * attributes of sub-objects, that the former is calculated ("merged")
438 * \todo XXX this can be read/write lock if needed.
440 spinlock_t coh_attr_guard;
442 * Size of cl_page + page slices
444 unsigned short coh_page_bufsize;
446 * Number of objects above this one: 0 for a top-object, 1 for its
449 unsigned char coh_nesting;
453 * Helper macro: iterate over all layers of the object \a obj, assigning every
454 * layer top-to-bottom to \a slice.
456 #define cl_object_for_each(slice, obj) \
457 list_for_each_entry((slice), \
458 &(obj)->co_lu.lo_header->loh_layers,\
462 * Helper macro: iterate over all layers of the object \a obj, assigning every
463 * layer bottom-to-top to \a slice.
465 #define cl_object_for_each_reverse(slice, obj) \
466 list_for_each_entry_reverse((slice), \
467 &(obj)->co_lu.lo_header->loh_layers,\
472 #define CL_PAGE_EOF ((pgoff_t)~0ull)
474 /** \addtogroup cl_page cl_page
478 * Layered client page.
480 * cl_page: represents a portion of a file, cached in the memory. All pages
481 * of the given file are of the same size, and are kept in the radix tree
482 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
483 * of the top-level file object are first class cl_objects, they have their
484 * own radix trees of pages and hence page is implemented as a sequence of
485 * struct cl_pages's, linked into double-linked list through
486 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
487 * corresponding radix tree at the corresponding logical offset.
489 * cl_page is associated with VM page of the hosting environment (struct
490 * page in Linux kernel, for example), struct page. It is assumed, that this
491 * association is implemented by one of cl_page layers (top layer in the
492 * current design) that
494 * - intercepts per-VM-page call-backs made by the environment (e.g.,
497 * - translates state (page flag bits) and locking between lustre and
500 * The association between cl_page and struct page is immutable and
501 * established when cl_page is created.
503 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
504 * this io an exclusive access to this page w.r.t. other io attempts and
505 * various events changing page state (such as transfer completion, or
506 * eviction of the page from the memory). Note, that in general cl_io
507 * cannot be identified with a particular thread, and page ownership is not
508 * exactly equal to the current thread holding a lock on the page. Layer
509 * implementing association between cl_page and struct page has to implement
510 * ownership on top of available synchronization mechanisms.
512 * While lustre client maintains the notion of an page ownership by io,
513 * hosting MM/VM usually has its own page concurrency control
514 * mechanisms. For example, in Linux, page access is synchronized by the
515 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
516 * takes care to acquire and release such locks as necessary around the
517 * calls to the file system methods (->readpage(), ->prepare_write(),
518 * ->commit_write(), etc.). This leads to the situation when there are two
519 * different ways to own a page in the client:
521 * - client code explicitly and voluntary owns the page (cl_page_own());
523 * - VM locks a page and then calls the client, that has "to assume"
524 * the ownership from the VM (cl_page_assume()).
526 * Dual methods to release ownership are cl_page_disown() and
527 * cl_page_unassume().
529 * cl_page is reference counted (cl_page::cp_ref). When reference counter
530 * drops to 0, the page is returned to the cache, unless it is in
531 * cl_page_state::CPS_FREEING state, in which case it is immediately
534 * The general logic guaranteeing the absence of "existential races" for
535 * pages is the following:
537 * - there are fixed known ways for a thread to obtain a new reference
540 * - by doing a lookup in the cl_object radix tree, protected by the
543 * - by starting from VM-locked struct page and following some
544 * hosting environment method (e.g., following ->private pointer in
545 * the case of Linux kernel), see cl_vmpage_page();
547 * - when the page enters cl_page_state::CPS_FREEING state, all these
548 * ways are severed with the proper synchronization
549 * (cl_page_delete());
551 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
554 * - no new references to the page in cl_page_state::CPS_FREEING state
555 * are allowed (checked in cl_page_get()).
557 * Together this guarantees that when last reference to a
558 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
559 * page, as neither references to it can be acquired at that point, nor
562 * cl_page is a state machine. States are enumerated in enum
563 * cl_page_state. Possible state transitions are enumerated in
564 * cl_page_state_set(). State transition process (i.e., actual changing of
565 * cl_page::cp_state field) is protected by the lock on the underlying VM
568 * Linux Kernel implementation.
570 * Binding between cl_page and struct page (which is a typedef for
571 * struct page) is implemented in the vvp layer. cl_page is attached to the
572 * ->private pointer of the struct page, together with the setting of
573 * PG_private bit in page->flags, and acquiring additional reference on the
574 * struct page (much like struct buffer_head, or any similar file system
575 * private data structures).
577 * PG_locked lock is used to implement both ownership and transfer
578 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
579 * states. No additional references are acquired for the duration of the
582 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
583 * write-out is "protected" by the special PG_writeback bit.
587 * States of cl_page. cl_page.c assumes particular order here.
589 * The page state machine is rather crude, as it doesn't recognize finer page
590 * states like "dirty" or "up to date". This is because such states are not
591 * always well defined for the whole stack (see, for example, the
592 * implementation of the read-ahead, that hides page up-to-dateness to track
593 * cache hits accurately). Such sub-states are maintained by the layers that
594 * are interested in them.
598 * Page is in the cache, un-owned. Page leaves cached state in the
601 * - [cl_page_state::CPS_OWNED] io comes across the page and
604 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
605 * req-formation engine decides that it wants to include this page
606 * into an RPC being constructed, and yanks it from the cache;
608 * - [cl_page_state::CPS_FREEING] VM callback is executed to
609 * evict the page form the memory;
611 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
615 * Page is exclusively owned by some cl_io. Page may end up in this
616 * state as a result of
618 * - io creating new page and immediately owning it;
620 * - [cl_page_state::CPS_CACHED] io finding existing cached page
623 * - [cl_page_state::CPS_OWNED] io finding existing owned page
624 * and waiting for owner to release the page;
626 * Page leaves owned state in the following cases:
628 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
629 * the cache, doing nothing;
631 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
634 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
635 * transfer for this page;
637 * - [cl_page_state::CPS_FREEING] io decides to destroy this
638 * page (e.g., as part of truncate or extent lock cancellation).
640 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
644 * Page is being written out, as a part of a transfer. This state is
645 * entered when req-formation logic decided that it wants this page to
646 * be sent through the wire _now_. Specifically, it means that once
647 * this state is achieved, transfer completion handler (with either
648 * success or failure indication) is guaranteed to be executed against
649 * this page independently of any locks and any scheduling decisions
650 * made by the hosting environment (that effectively means that the
651 * page is never put into cl_page_state::CPS_PAGEOUT state "in
652 * advance". This property is mentioned, because it is important when
653 * reasoning about possible dead-locks in the system). The page can
654 * enter this state as a result of
656 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
657 * write-out of this page, or
659 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
660 * that it has enough dirty pages cached to issue a "good"
663 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
664 * is completed---it is moved into cl_page_state::CPS_CACHED state.
666 * Underlying VM page is locked for the duration of transfer.
668 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
672 * Page is being read in, as a part of a transfer. This is quite
673 * similar to the cl_page_state::CPS_PAGEOUT state, except that
674 * read-in is always "immediate"---there is no such thing a sudden
675 * construction of read request from cached, presumably not up to date,
678 * Underlying VM page is locked for the duration of transfer.
680 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
684 * Page is being destroyed. This state is entered when client decides
685 * that page has to be deleted from its host object, as, e.g., a part
688 * Once this state is reached, there is no way to escape it.
690 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
697 /** Host page, the page is from the host inode which the cl_page
701 /** Transient page, the transient cl_page is used to bind a cl_page
702 * to vmpage which is not belonging to the same object of cl_page.
703 * it is used in DirectIO, lockless IO and liblustre. */
708 * Fields are protected by the lock on struct page, except for atomics and
711 * \invariant Data type invariants are in cl_page_invariant(). Basically:
712 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
713 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
714 * cl_page::cp_owner (when set).
717 /** Reference counter. */
719 /** An object this page is a part of. Immutable after creation. */
720 struct cl_object *cp_obj;
722 struct page *cp_vmpage;
723 /** Linkage of pages within group. Pages must be owned */
724 struct list_head cp_batch;
725 /** List of slices. Immutable after creation. */
726 struct list_head cp_layers;
728 * Page state. This field is const to avoid accidental update, it is
729 * modified only internally within cl_page.c. Protected by a VM lock.
731 const enum cl_page_state cp_state;
733 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
736 enum cl_page_type cp_type;
739 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
740 * by sub-io. Protected by a VM lock.
742 struct cl_io *cp_owner;
743 /** List of references to this page, for debugging. */
744 struct lu_ref cp_reference;
745 /** Link to an object, for debugging. */
746 struct lu_ref_link cp_obj_ref;
747 /** Link to a queue, for debugging. */
748 struct lu_ref_link cp_queue_ref;
749 /** Assigned if doing a sync_io */
750 struct cl_sync_io *cp_sync_io;
754 * Per-layer part of cl_page.
756 * \see vvp_page, lov_page, osc_page
758 struct cl_page_slice {
759 struct cl_page *cpl_page;
762 * Object slice corresponding to this page slice. Immutable after
765 struct cl_object *cpl_obj;
766 const struct cl_page_operations *cpl_ops;
767 /** Linkage into cl_page::cp_layers. Immutable after creation. */
768 struct list_head cpl_linkage;
772 * Lock mode. For the client extent locks.
784 * Requested transfer type.
793 * Per-layer page operations.
795 * Methods taking an \a io argument are for the activity happening in the
796 * context of given \a io. Page is assumed to be owned by that io, except for
797 * the obvious cases (like cl_page_operations::cpo_own()).
799 * \see vvp_page_ops, lov_page_ops, osc_page_ops
801 struct cl_page_operations {
803 * cl_page<->struct page methods. Only one layer in the stack has to
804 * implement these. Current code assumes that this functionality is
805 * provided by the topmost layer, see cl_page_disown0() as an example.
809 * Called when \a io acquires this page into the exclusive
810 * ownership. When this method returns, it is guaranteed that the is
811 * not owned by other io, and no transfer is going on against
815 * \see vvp_page_own(), lov_page_own()
817 int (*cpo_own)(const struct lu_env *env,
818 const struct cl_page_slice *slice,
819 struct cl_io *io, int nonblock);
820 /** Called when ownership it yielded. Optional.
822 * \see cl_page_disown()
823 * \see vvp_page_disown()
825 void (*cpo_disown)(const struct lu_env *env,
826 const struct cl_page_slice *slice, struct cl_io *io);
828 * Called for a page that is already "owned" by \a io from VM point of
831 * \see cl_page_assume()
832 * \see vvp_page_assume(), lov_page_assume()
834 void (*cpo_assume)(const struct lu_env *env,
835 const struct cl_page_slice *slice, struct cl_io *io);
836 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
837 * bottom-to-top when IO releases a page without actually unlocking
840 * \see cl_page_unassume()
841 * \see vvp_page_unassume()
843 void (*cpo_unassume)(const struct lu_env *env,
844 const struct cl_page_slice *slice,
847 * Announces whether the page contains valid data or not by \a uptodate.
849 * \see cl_page_export()
850 * \see vvp_page_export()
852 void (*cpo_export)(const struct lu_env *env,
853 const struct cl_page_slice *slice, int uptodate);
855 * Checks whether underlying VM page is locked (in the suitable
856 * sense). Used for assertions.
858 * \retval -EBUSY: page is protected by a lock of a given mode;
859 * \retval -ENODATA: page is not protected by a lock;
860 * \retval 0: this layer cannot decide. (Should never happen.)
862 int (*cpo_is_vmlocked)(const struct lu_env *env,
863 const struct cl_page_slice *slice);
869 * Called when page is truncated from the object. Optional.
871 * \see cl_page_discard()
872 * \see vvp_page_discard(), osc_page_discard()
874 void (*cpo_discard)(const struct lu_env *env,
875 const struct cl_page_slice *slice,
878 * Called when page is removed from the cache, and is about to being
879 * destroyed. Optional.
881 * \see cl_page_delete()
882 * \see vvp_page_delete(), osc_page_delete()
884 void (*cpo_delete)(const struct lu_env *env,
885 const struct cl_page_slice *slice);
886 /** Destructor. Frees resources and slice itself. */
887 void (*cpo_fini)(const struct lu_env *env,
888 struct cl_page_slice *slice);
890 * Optional debugging helper. Prints given page slice.
892 * \see cl_page_print()
894 int (*cpo_print)(const struct lu_env *env,
895 const struct cl_page_slice *slice,
896 void *cookie, lu_printer_t p);
905 * Request type dependent vector of operations.
907 * Transfer operations depend on transfer mode (cl_req_type). To avoid
908 * passing transfer mode to each and every of these methods, and to
909 * avoid branching on request type inside of the methods, separate
910 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
911 * provided. That is, method invocation usually looks like
913 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
917 * Called when a page is submitted for a transfer as a part of
920 * \return 0 : page is eligible for submission;
921 * \return -EALREADY : skip this page;
922 * \return -ve : error.
924 * \see cl_page_prep()
926 int (*cpo_prep)(const struct lu_env *env,
927 const struct cl_page_slice *slice,
930 * Completion handler. This is guaranteed to be eventually
931 * fired after cl_page_operations::cpo_prep() or
932 * cl_page_operations::cpo_make_ready() call.
934 * This method can be called in a non-blocking context. It is
935 * guaranteed however, that the page involved and its object
936 * are pinned in memory (and, hence, calling cl_page_put() is
939 * \see cl_page_completion()
941 void (*cpo_completion)(const struct lu_env *env,
942 const struct cl_page_slice *slice,
945 * Called when cached page is about to be added to the
946 * ptlrpc request as a part of req formation.
948 * \return 0 : proceed with this page;
949 * \return -EAGAIN : skip this page;
950 * \return -ve : error.
952 * \see cl_page_make_ready()
954 int (*cpo_make_ready)(const struct lu_env *env,
955 const struct cl_page_slice *slice);
958 * Tell transfer engine that only [to, from] part of a page should be
961 * This is used for immediate transfers.
963 * \todo XXX this is not very good interface. It would be much better
964 * if all transfer parameters were supplied as arguments to
965 * cl_io_operations::cio_submit() call, but it is not clear how to do
966 * this for page queues.
968 * \see cl_page_clip()
970 void (*cpo_clip)(const struct lu_env *env,
971 const struct cl_page_slice *slice,
974 * \pre the page was queued for transferring.
975 * \post page is removed from client's pending list, or -EBUSY
976 * is returned if it has already been in transferring.
978 * This is one of seldom page operation which is:
979 * 0. called from top level;
980 * 1. don't have vmpage locked;
981 * 2. every layer should synchronize execution of its ->cpo_cancel()
982 * with completion handlers. Osc uses client obd lock for this
983 * purpose. Based on there is no vvp_page_cancel and
984 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
986 * \see osc_page_cancel().
988 int (*cpo_cancel)(const struct lu_env *env,
989 const struct cl_page_slice *slice);
991 * Write out a page by kernel. This is only called by ll_writepage
994 * \see cl_page_flush()
996 int (*cpo_flush)(const struct lu_env *env,
997 const struct cl_page_slice *slice,
1003 * Helper macro, dumping detailed information about \a page into a log.
1005 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1007 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1008 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1009 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1010 CDEBUG(mask, format , ## __VA_ARGS__); \
1015 * Helper macro, dumping shorter information about \a page into a log.
1017 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1019 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1020 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1021 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1022 CDEBUG(mask, format , ## __VA_ARGS__); \
1026 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1028 LASSERT(page->cp_vmpage != NULL);
1029 return page->cp_vmpage;
1033 * Check if a cl_page is in use.
1035 * Client cache holds a refcount, this refcount will be dropped when
1036 * the page is taken out of cache, see vvp_page_delete().
1038 static inline bool __page_in_use(const struct cl_page *page, int refc)
1040 return (atomic_read(&page->cp_ref) > refc + 1);
1044 * Caller itself holds a refcount of cl_page.
1046 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1048 * Caller doesn't hold a refcount.
1050 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1054 /** \addtogroup cl_lock cl_lock
1058 * Extent locking on the client.
1062 * The locking model of the new client code is built around
1066 * data-type representing an extent lock on a regular file. cl_lock is a
1067 * layered object (much like cl_object and cl_page), it consists of a header
1068 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1069 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1071 * Typical cl_lock consists of the two layers:
1073 * - vvp_lock (vvp specific data), and
1074 * - lov_lock (lov specific data).
1076 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1077 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1079 * - lovsub_lock, and
1082 * Each sub-lock is associated with a cl_object (representing stripe
1083 * sub-object or the file to which top-level cl_lock is associated to), and is
1084 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1085 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1086 * is different from cl_page, that doesn't fan out (there is usually exactly
1087 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1088 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1092 * cl_lock is a cacheless data container for the requirements of locks to
1093 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1096 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1097 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1099 * INTERFACE AND USAGE
1101 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1102 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1103 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1104 * consists of multiple sub cl_locks, each sub locks will be enqueued
1105 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1106 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1109 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1110 * method will be called for each layer to release the resource held by this
1111 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1112 * clo_enqueue time, is released.
1114 * LDLM lock can only be canceled if there is no cl_lock using it.
1116 * Overall process of the locking during IO operation is as following:
1118 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1119 * is called on each layer. Responsibility of this method is to add locks,
1120 * needed by a given layer into cl_io.ci_lockset.
1122 * - once locks for all layers were collected, they are sorted to avoid
1123 * dead-locks (cl_io_locks_sort()), and enqueued.
1125 * - when all locks are acquired, IO is performed;
1127 * - locks are released after IO is complete.
1129 * Striping introduces major additional complexity into locking. The
1130 * fundamental problem is that it is generally unsafe to actively use (hold)
1131 * two locks on the different OST servers at the same time, as this introduces
1132 * inter-server dependency and can lead to cascading evictions.
1134 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1135 * that no multi-stripe locks are taken (note that this design abandons POSIX
1136 * read/write semantics). Such pieces ideally can be executed concurrently. At
1137 * the same time, certain types of IO cannot be sub-divived, without
1138 * sacrificing correctness. This includes:
1140 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1143 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1145 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1146 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1147 * has to be held together with the usual lock on [offset, offset + count].
1149 * Interaction with DLM
1151 * In the expected setup, cl_lock is ultimately backed up by a collection of
1152 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1153 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1154 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1155 * description of interaction with DLM.
1161 struct cl_lock_descr {
1162 /** Object this lock is granted for. */
1163 struct cl_object *cld_obj;
1164 /** Index of the first page protected by this lock. */
1166 /** Index of the last page (inclusive) protected by this lock. */
1168 /** Group ID, for group lock */
1171 enum cl_lock_mode cld_mode;
1173 * flags to enqueue lock. A combination of bit-flags from
1174 * enum cl_enq_flags.
1176 __u32 cld_enq_flags;
1179 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1180 #define PDESCR(descr) \
1181 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1182 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1184 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1187 * Layered client lock.
1190 /** List of slices. Immutable after creation. */
1191 struct list_head cll_layers;
1192 /** lock attribute, extent, cl_object, etc. */
1193 struct cl_lock_descr cll_descr;
1197 * Per-layer part of cl_lock
1199 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1201 struct cl_lock_slice {
1202 struct cl_lock *cls_lock;
1203 /** Object slice corresponding to this lock slice. Immutable after
1205 struct cl_object *cls_obj;
1206 const struct cl_lock_operations *cls_ops;
1207 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1208 struct list_head cls_linkage;
1213 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1215 struct cl_lock_operations {
1218 * Attempts to enqueue the lock. Called top-to-bottom.
1220 * \retval 0 this layer has enqueued the lock successfully
1221 * \retval >0 this layer has enqueued the lock, but need to wait on
1222 * @anchor for resources
1223 * \retval -ve failure
1225 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1226 * \see osc_lock_enqueue()
1228 int (*clo_enqueue)(const struct lu_env *env,
1229 const struct cl_lock_slice *slice,
1230 struct cl_io *io, struct cl_sync_io *anchor);
1232 * Cancel a lock, release its DLM lock ref, while does not cancel the
1235 void (*clo_cancel)(const struct lu_env *env,
1236 const struct cl_lock_slice *slice);
1239 * Destructor. Frees resources and the slice.
1241 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1242 * \see osc_lock_fini()
1244 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1246 * Optional debugging helper. Prints given lock slice.
1248 int (*clo_print)(const struct lu_env *env,
1249 void *cookie, lu_printer_t p,
1250 const struct cl_lock_slice *slice);
1253 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1255 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1256 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1257 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1258 CDEBUG(mask, format , ## __VA_ARGS__); \
1262 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1266 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1272 /** \addtogroup cl_page_list cl_page_list
1273 * Page list used to perform collective operations on a group of pages.
1275 * Pages are added to the list one by one. cl_page_list acquires a reference
1276 * for every page in it. Page list is used to perform collective operations on
1279 * - submit pages for an immediate transfer,
1281 * - own pages on behalf of certain io (waiting for each page in turn),
1285 * When list is finalized, it releases references on all pages it still has.
1287 * \todo XXX concurrency control.
1291 struct cl_page_list {
1293 struct list_head pl_pages;
1294 struct task_struct *pl_owner;
1298 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1299 * contains an incoming page list and an outgoing page list.
1302 struct cl_page_list c2_qin;
1303 struct cl_page_list c2_qout;
1306 /** @} cl_page_list */
1308 /** \addtogroup cl_io cl_io
1313 * cl_io represents a high level I/O activity like
1314 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1317 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1318 * important distinction. We want to minimize number of calls to the allocator
1319 * in the fast path, e.g., in the case of read(2) when everything is cached:
1320 * client already owns the lock over region being read, and data are cached
1321 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1322 * per-layer io state is stored in the session, associated with the io, see
1323 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1324 * by using free-lists, see cl_env_get().
1326 * There is a small predefined number of possible io types, enumerated in enum
1329 * cl_io is a state machine, that can be advanced concurrently by the multiple
1330 * threads. It is up to these threads to control the concurrency and,
1331 * specifically, to detect when io is done, and its state can be safely
1334 * For read/write io overall execution plan is as following:
1336 * (0) initialize io state through all layers;
1338 * (1) loop: prepare chunk of work to do
1340 * (2) call all layers to collect locks they need to process current chunk
1342 * (3) sort all locks to avoid dead-locks, and acquire them
1344 * (4) process the chunk: call per-page methods
1345 * cl_io_operations::cio_prepare_write(),
1346 * cl_io_operations::cio_commit_write() for write)
1352 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1353 * address allocation efficiency issues mentioned above), and returns with the
1354 * special error condition from per-page method when current sub-io has to
1355 * block. This causes io loop to be repeated, and lov switches to the next
1356 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1361 /** read system call */
1363 /** write system call */
1365 /** truncate, utime system calls */
1367 /** get data version */
1370 * page fault handling
1374 * fsync system call handling
1375 * To write out a range of file
1379 * Miscellaneous io. This is used for occasional io activity that
1380 * doesn't fit into other types. Currently this is used for:
1382 * - cancellation of an extent lock. This io exists as a context
1383 * to write dirty pages from under the lock being canceled back
1386 * - VM induced page write-out. An io context for writing page out
1387 * for memory cleansing;
1389 * - glimpse. An io context to acquire glimpse lock.
1391 * - grouplock. An io context to acquire group lock.
1393 * CIT_MISC io is used simply as a context in which locks and pages
1394 * are manipulated. Such io has no internal "process", that is,
1395 * cl_io_loop() is never called for it.
1400 * To give advice about access of a file
1407 * States of cl_io state machine
1410 /** Not initialized. */
1414 /** IO iteration started. */
1418 /** Actual IO is in progress. */
1420 /** IO for the current iteration finished. */
1422 /** Locks released. */
1424 /** Iteration completed. */
1426 /** cl_io finalized. */
1431 * IO state private for a layer.
1433 * This is usually embedded into layer session data, rather than allocated
1436 * \see vvp_io, lov_io, osc_io
1438 struct cl_io_slice {
1439 struct cl_io *cis_io;
1440 /** corresponding object slice. Immutable after creation. */
1441 struct cl_object *cis_obj;
1442 /** io operations. Immutable after creation. */
1443 const struct cl_io_operations *cis_iop;
1445 * linkage into a list of all slices for a given cl_io, hanging off
1446 * cl_io::ci_layers. Immutable after creation.
1448 struct list_head cis_linkage;
1451 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1454 struct cl_read_ahead {
1455 /* Maximum page index the readahead window will end.
1456 * This is determined DLM lock coverage, RPC and stripe boundary.
1457 * cra_end is included. */
1459 /* optimal RPC size for this read, by pages */
1460 unsigned long cra_rpc_size;
1461 /* Release callback. If readahead holds resources underneath, this
1462 * function should be called to release it. */
1463 void (*cra_release)(const struct lu_env *env, void *cbdata);
1464 /* Callback data for cra_release routine */
1468 static inline void cl_read_ahead_release(const struct lu_env *env,
1469 struct cl_read_ahead *ra)
1471 if (ra->cra_release != NULL)
1472 ra->cra_release(env, ra->cra_cbdata);
1473 memset(ra, 0, sizeof(*ra));
1478 * Per-layer io operations.
1479 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1481 struct cl_io_operations {
1483 * Vector of io state transition methods for every io type.
1485 * \see cl_page_operations::io
1489 * Prepare io iteration at a given layer.
1491 * Called top-to-bottom at the beginning of each iteration of
1492 * "io loop" (if it makes sense for this type of io). Here
1493 * layer selects what work it will do during this iteration.
1495 * \see cl_io_operations::cio_iter_fini()
1497 int (*cio_iter_init) (const struct lu_env *env,
1498 const struct cl_io_slice *slice);
1500 * Finalize io iteration.
1502 * Called bottom-to-top at the end of each iteration of "io
1503 * loop". Here layers can decide whether IO has to be
1506 * \see cl_io_operations::cio_iter_init()
1508 void (*cio_iter_fini) (const struct lu_env *env,
1509 const struct cl_io_slice *slice);
1511 * Collect locks for the current iteration of io.
1513 * Called top-to-bottom to collect all locks necessary for
1514 * this iteration. This methods shouldn't actually enqueue
1515 * anything, instead it should post a lock through
1516 * cl_io_lock_add(). Once all locks are collected, they are
1517 * sorted and enqueued in the proper order.
1519 int (*cio_lock) (const struct lu_env *env,
1520 const struct cl_io_slice *slice);
1522 * Finalize unlocking.
1524 * Called bottom-to-top to finish layer specific unlocking
1525 * functionality, after generic code released all locks
1526 * acquired by cl_io_operations::cio_lock().
1528 void (*cio_unlock)(const struct lu_env *env,
1529 const struct cl_io_slice *slice);
1531 * Start io iteration.
1533 * Once all locks are acquired, called top-to-bottom to
1534 * commence actual IO. In the current implementation,
1535 * top-level vvp_io_{read,write}_start() does all the work
1536 * synchronously by calling generic_file_*(), so other layers
1537 * are called when everything is done.
1539 int (*cio_start)(const struct lu_env *env,
1540 const struct cl_io_slice *slice);
1542 * Called top-to-bottom at the end of io loop. Here layer
1543 * might wait for an unfinished asynchronous io.
1545 void (*cio_end) (const struct lu_env *env,
1546 const struct cl_io_slice *slice);
1548 * Called bottom-to-top to notify layers that read/write IO
1549 * iteration finished, with \a nob bytes transferred.
1551 void (*cio_advance)(const struct lu_env *env,
1552 const struct cl_io_slice *slice,
1555 * Called once per io, bottom-to-top to release io resources.
1557 void (*cio_fini) (const struct lu_env *env,
1558 const struct cl_io_slice *slice);
1562 * Submit pages from \a queue->c2_qin for IO, and move
1563 * successfully submitted pages into \a queue->c2_qout. Return
1564 * non-zero if failed to submit even the single page. If
1565 * submission failed after some pages were moved into \a
1566 * queue->c2_qout, completion callback with non-zero ioret is
1569 int (*cio_submit)(const struct lu_env *env,
1570 const struct cl_io_slice *slice,
1571 enum cl_req_type crt,
1572 struct cl_2queue *queue);
1574 * Queue async page for write.
1575 * The difference between cio_submit and cio_queue is that
1576 * cio_submit is for urgent request.
1578 int (*cio_commit_async)(const struct lu_env *env,
1579 const struct cl_io_slice *slice,
1580 struct cl_page_list *queue, int from, int to,
1583 * Decide maximum read ahead extent
1585 * \pre io->ci_type == CIT_READ
1587 int (*cio_read_ahead)(const struct lu_env *env,
1588 const struct cl_io_slice *slice,
1589 pgoff_t start, struct cl_read_ahead *ra);
1591 * Optional debugging helper. Print given io slice.
1593 int (*cio_print)(const struct lu_env *env, void *cookie,
1594 lu_printer_t p, const struct cl_io_slice *slice);
1598 * Flags to lock enqueue procedure.
1603 * instruct server to not block, if conflicting lock is found. Instead
1604 * -EWOULDBLOCK is returned immediately.
1606 CEF_NONBLOCK = 0x00000001,
1608 * take lock asynchronously (out of order), as it cannot
1609 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1611 CEF_ASYNC = 0x00000002,
1613 * tell the server to instruct (though a flag in the blocking ast) an
1614 * owner of the conflicting lock, that it can drop dirty pages
1615 * protected by this lock, without sending them to the server.
1617 CEF_DISCARD_DATA = 0x00000004,
1619 * tell the sub layers that it must be a `real' lock. This is used for
1620 * mmapped-buffer locks and glimpse locks that must be never converted
1621 * into lockless mode.
1623 * \see vvp_mmap_locks(), cl_glimpse_lock().
1625 CEF_MUST = 0x00000008,
1627 * tell the sub layers that never request a `real' lock. This flag is
1628 * not used currently.
1630 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1631 * conversion policy: ci_lockreq describes generic information of lock
1632 * requirement for this IO, especially for locks which belong to the
1633 * object doing IO; however, lock itself may have precise requirements
1634 * that are described by the enqueue flags.
1636 CEF_NEVER = 0x00000010,
1638 * for async glimpse lock.
1640 CEF_AGL = 0x00000020,
1642 * enqueue a lock to test DLM lock existence.
1644 CEF_PEEK = 0x00000040,
1646 * Lock match only. Used by group lock in I/O as group lock
1647 * is known to exist.
1649 CEF_LOCK_MATCH = 0x00000080,
1651 * mask of enq_flags.
1653 CEF_MASK = 0x000000ff,
1657 * Link between lock and io. Intermediate structure is needed, because the
1658 * same lock can be part of multiple io's simultaneously.
1660 struct cl_io_lock_link {
1661 /** linkage into one of cl_lockset lists. */
1662 struct list_head cill_linkage;
1663 struct cl_lock cill_lock;
1664 /** optional destructor */
1665 void (*cill_fini)(const struct lu_env *env,
1666 struct cl_io_lock_link *link);
1668 #define cill_descr cill_lock.cll_descr
1671 * Lock-set represents a collection of locks, that io needs at a
1672 * time. Generally speaking, client tries to avoid holding multiple locks when
1675 * - holding extent locks over multiple ost's introduces the danger of
1676 * "cascading timeouts";
1678 * - holding multiple locks over the same ost is still dead-lock prone,
1679 * see comment in osc_lock_enqueue(),
1681 * but there are certain situations where this is unavoidable:
1683 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1685 * - truncate has to take [new-size, EOF] lock for correctness;
1687 * - SNS has to take locks across full stripe for correctness;
1689 * - in the case when user level buffer, supplied to {read,write}(file0),
1690 * is a part of a memory mapped lustre file, client has to take a dlm
1691 * locks on file0, and all files that back up the buffer (or a part of
1692 * the buffer, that is being processed in the current chunk, in any
1693 * case, there are situations where at least 2 locks are necessary).
1695 * In such cases we at least try to take locks in the same consistent
1696 * order. To this end, all locks are first collected, then sorted, and then
1700 /** locks to be acquired. */
1701 struct list_head cls_todo;
1702 /** locks acquired. */
1703 struct list_head cls_done;
1707 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1708 * but 'req' is always to be thought as 'request' :-)
1710 enum cl_io_lock_dmd {
1711 /** Always lock data (e.g., O_APPEND). */
1713 /** Layers are free to decide between local and global locking. */
1715 /** Never lock: there is no cache (e.g., liblustre). */
1719 enum cl_fsync_mode {
1720 /** start writeback, do not wait for them to finish */
1722 /** start writeback and wait for them to finish */
1724 /** discard all of dirty pages in a specific file range */
1725 CL_FSYNC_DISCARD = 2,
1726 /** start writeback and make sure they have reached storage before
1727 * return. OST_SYNC RPC must be issued and finished */
1731 struct cl_io_rw_common {
1740 * cl_io is shared by all threads participating in this IO (in current
1741 * implementation only one thread advances IO, but parallel IO design and
1742 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1743 * is up to these threads to serialize their activities, including updates to
1744 * mutable cl_io fields.
1747 /** type of this IO. Immutable after creation. */
1748 enum cl_io_type ci_type;
1749 /** current state of cl_io state machine. */
1750 enum cl_io_state ci_state;
1751 /** main object this io is against. Immutable after creation. */
1752 struct cl_object *ci_obj;
1754 * Upper layer io, of which this io is a part of. Immutable after
1757 struct cl_io *ci_parent;
1758 /** List of slices. Immutable after creation. */
1759 struct list_head ci_layers;
1760 /** list of locks (to be) acquired by this io. */
1761 struct cl_lockset ci_lockset;
1762 /** lock requirements, this is just a help info for sublayers. */
1763 enum cl_io_lock_dmd ci_lockreq;
1766 struct cl_io_rw_common rd;
1769 struct cl_io_rw_common wr;
1773 struct cl_io_rw_common ci_rw;
1774 struct cl_setattr_io {
1775 struct ost_lvb sa_attr;
1776 unsigned int sa_attr_flags;
1777 unsigned int sa_valid;
1778 int sa_stripe_index;
1779 struct ost_layout sa_layout;
1780 const struct lu_fid *sa_parent_fid;
1782 struct cl_data_version_io {
1783 u64 dv_data_version;
1786 struct cl_fault_io {
1787 /** page index within file. */
1789 /** bytes valid byte on a faulted page. */
1791 /** writable page? for nopage() only */
1793 /** page of an executable? */
1795 /** page_mkwrite() */
1797 /** resulting page */
1798 struct cl_page *ft_page;
1800 struct cl_fsync_io {
1803 /** file system level fid */
1804 struct lu_fid *fi_fid;
1805 enum cl_fsync_mode fi_mode;
1806 /* how many pages were written/discarded */
1807 unsigned int fi_nr_written;
1809 struct cl_ladvise_io {
1812 /** file system level fid */
1813 struct lu_fid *li_fid;
1814 enum lu_ladvise_type li_advice;
1818 struct cl_2queue ci_queue;
1821 unsigned int ci_continue:1,
1823 * This io has held grouplock, to inform sublayers that
1824 * don't do lockless i/o.
1828 * The whole IO need to be restarted because layout has been changed
1832 * to not refresh layout - the IO issuer knows that the layout won't
1833 * change(page operations, layout change causes all page to be
1834 * discarded), or it doesn't matter if it changes(sync).
1838 * Need MDS intervention to complete a write. This usually means the
1839 * corresponding component is not initialized for the writing extent.
1841 ci_need_write_intent:1,
1843 * Check if layout changed after the IO finishes. Mainly for HSM
1844 * requirement. If IO occurs to openning files, it doesn't need to
1845 * verify layout because HSM won't release openning files.
1846 * Right now, only two opertaions need to verify layout: glimpse
1851 * file is released, restore has to to be triggered by vvp layer
1853 ci_restore_needed:1,
1859 * Number of pages owned by this IO. For invariant checking.
1861 unsigned ci_owned_nr;
1867 * Per-transfer attributes.
1869 struct cl_req_attr {
1870 enum cl_req_type cra_type;
1872 struct cl_page *cra_page;
1873 /** Generic attributes for the server consumption. */
1874 struct obdo *cra_oa;
1876 char cra_jobid[LUSTRE_JOBID_SIZE];
1879 enum cache_stats_item {
1880 /** how many cache lookups were performed */
1882 /** how many times cache lookup resulted in a hit */
1884 /** how many entities are in the cache right now */
1886 /** how many entities in the cache are actively used (and cannot be
1887 * evicted) right now */
1889 /** how many entities were created at all */
1894 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
1897 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
1899 struct cache_stats {
1900 const char *cs_name;
1901 atomic_t cs_stats[CS_NR];
1904 /** These are not exported so far */
1905 void cache_stats_init (struct cache_stats *cs, const char *name);
1908 * Client-side site. This represents particular client stack. "Global"
1909 * variables should (directly or indirectly) be added here to allow multiple
1910 * clients to co-exist in the single address space.
1913 struct lu_site cs_lu;
1915 * Statistical counters. Atomics do not scale, something better like
1916 * per-cpu counters is needed.
1918 * These are exported as /proc/fs/lustre/llite/.../site
1920 * When interpreting keep in mind that both sub-locks (and sub-pages)
1921 * and top-locks (and top-pages) are accounted here.
1923 struct cache_stats cs_pages;
1924 atomic_t cs_pages_state[CPS_NR];
1927 int cl_site_init(struct cl_site *s, struct cl_device *top);
1928 void cl_site_fini(struct cl_site *s);
1929 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
1932 * Output client site statistical counters into a buffer. Suitable for
1933 * ll_rd_*()-style functions.
1935 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
1940 * Type conversion and accessory functions.
1944 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
1946 return container_of(site, struct cl_site, cs_lu);
1949 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
1951 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
1952 return container_of0(d, struct cl_device, cd_lu_dev);
1955 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
1957 return &d->cd_lu_dev;
1960 static inline struct cl_object *lu2cl(const struct lu_object *o)
1962 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
1963 return container_of0(o, struct cl_object, co_lu);
1966 static inline const struct cl_object_conf *
1967 lu2cl_conf(const struct lu_object_conf *conf)
1969 return container_of0(conf, struct cl_object_conf, coc_lu);
1972 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
1974 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
1977 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
1979 return container_of0(h, struct cl_object_header, coh_lu);
1982 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
1984 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
1988 struct cl_object_header *cl_object_header(const struct cl_object *obj)
1990 return luh2coh(obj->co_lu.lo_header);
1993 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
1995 return lu_device_init(&d->cd_lu_dev, t);
1998 static inline void cl_device_fini(struct cl_device *d)
2000 lu_device_fini(&d->cd_lu_dev);
2003 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2004 struct cl_object *obj, pgoff_t index,
2005 const struct cl_page_operations *ops);
2006 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2007 struct cl_object *obj,
2008 const struct cl_lock_operations *ops);
2009 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2010 struct cl_object *obj, const struct cl_io_operations *ops);
2013 /** \defgroup cl_object cl_object
2015 struct cl_object *cl_object_top (struct cl_object *o);
2016 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2017 const struct lu_fid *fid,
2018 const struct cl_object_conf *c);
2020 int cl_object_header_init(struct cl_object_header *h);
2021 void cl_object_header_fini(struct cl_object_header *h);
2022 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2023 void cl_object_get (struct cl_object *o);
2024 void cl_object_attr_lock (struct cl_object *o);
2025 void cl_object_attr_unlock(struct cl_object *o);
2026 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2027 struct cl_attr *attr);
2028 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2029 const struct cl_attr *attr, unsigned valid);
2030 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2031 struct ost_lvb *lvb);
2032 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2033 const struct cl_object_conf *conf);
2034 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2035 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2036 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2037 struct lov_user_md __user *lum);
2038 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2039 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2041 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2042 struct cl_layout *cl);
2043 loff_t cl_object_maxbytes(struct cl_object *obj);
2046 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2048 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2050 return cl_object_header(o0) == cl_object_header(o1);
2053 static inline void cl_object_page_init(struct cl_object *clob, int size)
2055 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2056 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2057 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2060 static inline void *cl_object_page_slice(struct cl_object *clob,
2061 struct cl_page *page)
2063 return (void *)((char *)page + clob->co_slice_off);
2067 * Return refcount of cl_object.
2069 static inline int cl_object_refc(struct cl_object *clob)
2071 struct lu_object_header *header = clob->co_lu.lo_header;
2072 return atomic_read(&header->loh_ref);
2077 /** \defgroup cl_page cl_page
2085 /* callback of cl_page_gang_lookup() */
2087 struct cl_page *cl_page_find (const struct lu_env *env,
2088 struct cl_object *obj,
2089 pgoff_t idx, struct page *vmpage,
2090 enum cl_page_type type);
2091 struct cl_page *cl_page_alloc (const struct lu_env *env,
2092 struct cl_object *o, pgoff_t ind,
2093 struct page *vmpage,
2094 enum cl_page_type type);
2095 void cl_page_get (struct cl_page *page);
2096 void cl_page_put (const struct lu_env *env,
2097 struct cl_page *page);
2098 void cl_page_print (const struct lu_env *env, void *cookie,
2099 lu_printer_t printer,
2100 const struct cl_page *pg);
2101 void cl_page_header_print(const struct lu_env *env, void *cookie,
2102 lu_printer_t printer,
2103 const struct cl_page *pg);
2104 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2105 struct cl_page *cl_page_top (struct cl_page *page);
2107 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2108 const struct lu_device_type *dtype);
2113 * Functions dealing with the ownership of page by io.
2117 int cl_page_own (const struct lu_env *env,
2118 struct cl_io *io, struct cl_page *page);
2119 int cl_page_own_try (const struct lu_env *env,
2120 struct cl_io *io, struct cl_page *page);
2121 void cl_page_assume (const struct lu_env *env,
2122 struct cl_io *io, struct cl_page *page);
2123 void cl_page_unassume (const struct lu_env *env,
2124 struct cl_io *io, struct cl_page *pg);
2125 void cl_page_disown (const struct lu_env *env,
2126 struct cl_io *io, struct cl_page *page);
2127 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2134 * Functions dealing with the preparation of a page for a transfer, and
2135 * tracking transfer state.
2138 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2139 struct cl_page *pg, enum cl_req_type crt);
2140 void cl_page_completion (const struct lu_env *env,
2141 struct cl_page *pg, enum cl_req_type crt, int ioret);
2142 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2143 enum cl_req_type crt);
2144 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2145 struct cl_page *pg, enum cl_req_type crt);
2146 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2148 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2149 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2150 struct cl_page *pg);
2156 * \name helper routines
2157 * Functions to discard, delete and export a cl_page.
2160 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2161 struct cl_page *pg);
2162 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2163 int cl_page_is_vmlocked(const struct lu_env *env,
2164 const struct cl_page *pg);
2165 void cl_page_export(const struct lu_env *env,
2166 struct cl_page *pg, int uptodate);
2167 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2168 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2169 size_t cl_page_size(const struct cl_object *obj);
2171 void cl_lock_print(const struct lu_env *env, void *cookie,
2172 lu_printer_t printer, const struct cl_lock *lock);
2173 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2174 lu_printer_t printer,
2175 const struct cl_lock_descr *descr);
2179 * Data structure managing a client's cached pages. A count of
2180 * "unstable" pages is maintained, and an LRU of clean pages is
2181 * maintained. "unstable" pages are pages pinned by the ptlrpc
2182 * layer for recovery purposes.
2184 struct cl_client_cache {
2186 * # of client cache refcount
2187 * # of users (OSCs) + 2 (held by llite and lov)
2191 * # of threads are doing shrinking
2193 unsigned int ccc_lru_shrinkers;
2195 * # of LRU entries available
2197 atomic_long_t ccc_lru_left;
2199 * List of entities(OSCs) for this LRU cache
2201 struct list_head ccc_lru;
2203 * Max # of LRU entries
2205 unsigned long ccc_lru_max;
2207 * Lock to protect ccc_lru list
2209 spinlock_t ccc_lru_lock;
2211 * Set if unstable check is enabled
2213 unsigned int ccc_unstable_check:1;
2215 * # of unstable pages for this mount point
2217 atomic_long_t ccc_unstable_nr;
2219 * Waitq for awaiting unstable pages to reach zero.
2220 * Used at umounting time and signaled on BRW commit
2222 wait_queue_head_t ccc_unstable_waitq;
2225 * cl_cache functions
2227 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2228 void cl_cache_incref(struct cl_client_cache *cache);
2229 void cl_cache_decref(struct cl_client_cache *cache);
2233 /** \defgroup cl_lock cl_lock
2235 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2236 struct cl_lock *lock);
2237 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2238 const struct cl_io *io);
2239 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2240 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2241 const struct lu_device_type *dtype);
2242 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2244 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2245 struct cl_lock *lock, struct cl_sync_io *anchor);
2246 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2250 /** \defgroup cl_io cl_io
2253 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2254 enum cl_io_type iot, struct cl_object *obj);
2255 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2256 enum cl_io_type iot, struct cl_object *obj);
2257 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2258 enum cl_io_type iot, loff_t pos, size_t count);
2259 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2261 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2262 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2263 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2264 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2265 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2266 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2267 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2268 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2269 struct cl_io_lock_link *link);
2270 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2271 struct cl_lock_descr *descr);
2272 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2273 enum cl_req_type iot, struct cl_2queue *queue);
2274 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2275 enum cl_req_type iot, struct cl_2queue *queue,
2277 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2278 struct cl_page_list *queue, int from, int to,
2280 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2281 pgoff_t start, struct cl_read_ahead *ra);
2282 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2284 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2285 struct cl_page_list *queue);
2288 * True, iff \a io is an O_APPEND write(2).
2290 static inline int cl_io_is_append(const struct cl_io *io)
2292 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2295 static inline int cl_io_is_sync_write(const struct cl_io *io)
2297 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2300 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2302 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2306 * True, iff \a io is a truncate(2).
2308 static inline int cl_io_is_trunc(const struct cl_io *io)
2310 return io->ci_type == CIT_SETATTR &&
2311 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2314 struct cl_io *cl_io_top(struct cl_io *io);
2316 void cl_io_print(const struct lu_env *env, void *cookie,
2317 lu_printer_t printer, const struct cl_io *io);
2319 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2321 typeof(foo_io) __foo_io = (foo_io); \
2323 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2324 memset(&__foo_io->base + 1, 0, \
2325 (sizeof *__foo_io) - sizeof __foo_io->base); \
2330 /** \defgroup cl_page_list cl_page_list
2334 * Last page in the page list.
2336 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2338 LASSERT(plist->pl_nr > 0);
2339 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2342 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2344 LASSERT(plist->pl_nr > 0);
2345 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2349 * Iterate over pages in a page list.
2351 #define cl_page_list_for_each(page, list) \
2352 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2355 * Iterate over pages in a page list, taking possible removals into account.
2357 #define cl_page_list_for_each_safe(page, temp, list) \
2358 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2360 void cl_page_list_init (struct cl_page_list *plist);
2361 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2362 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2363 struct cl_page *page);
2364 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2365 struct cl_page *page);
2366 void cl_page_list_splice (struct cl_page_list *list,
2367 struct cl_page_list *head);
2368 void cl_page_list_del (const struct lu_env *env,
2369 struct cl_page_list *plist, struct cl_page *page);
2370 void cl_page_list_disown (const struct lu_env *env,
2371 struct cl_io *io, struct cl_page_list *plist);
2372 void cl_page_list_assume (const struct lu_env *env,
2373 struct cl_io *io, struct cl_page_list *plist);
2374 void cl_page_list_discard(const struct lu_env *env,
2375 struct cl_io *io, struct cl_page_list *plist);
2376 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2378 void cl_2queue_init (struct cl_2queue *queue);
2379 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2380 void cl_2queue_disown (const struct lu_env *env,
2381 struct cl_io *io, struct cl_2queue *queue);
2382 void cl_2queue_assume (const struct lu_env *env,
2383 struct cl_io *io, struct cl_2queue *queue);
2384 void cl_2queue_discard (const struct lu_env *env,
2385 struct cl_io *io, struct cl_2queue *queue);
2386 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2387 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2389 /** @} cl_page_list */
2391 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2392 struct cl_req_attr *attr);
2394 /** \defgroup cl_sync_io cl_sync_io
2398 * Anchor for synchronous transfer. This is allocated on a stack by thread
2399 * doing synchronous transfer, and a pointer to this structure is set up in
2400 * every page submitted for transfer. Transfer completion routine updates
2401 * anchor and wakes up waiting thread when transfer is complete.
2404 /** number of pages yet to be transferred. */
2405 atomic_t csi_sync_nr;
2408 /** barrier of destroy this structure */
2409 atomic_t csi_barrier;
2410 /** completion to be signaled when transfer is complete. */
2411 wait_queue_head_t csi_waitq;
2412 /** callback to invoke when this IO is finished */
2413 void (*csi_end_io)(const struct lu_env *,
2414 struct cl_sync_io *);
2417 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2418 void (*end)(const struct lu_env *, struct cl_sync_io *));
2419 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2421 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2423 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2425 /** @} cl_sync_io */
2427 /** \defgroup cl_env cl_env
2429 * lu_env handling for a client.
2431 * lu_env is an environment within which lustre code executes. Its major part
2432 * is lu_context---a fast memory allocation mechanism that is used to conserve
2433 * precious kernel stack space. Originally lu_env was designed for a server,
2436 * - there is a (mostly) fixed number of threads, and
2438 * - call chains have no non-lustre portions inserted between lustre code.
2440 * On a client both these assumtpion fails, because every user thread can
2441 * potentially execute lustre code as part of a system call, and lustre calls
2442 * into VFS or MM that call back into lustre.
2444 * To deal with that, cl_env wrapper functions implement the following
2447 * - allocation and destruction of environment is amortized by caching no
2448 * longer used environments instead of destroying them;
2450 * \see lu_env, lu_context, lu_context_key
2453 struct lu_env *cl_env_get(__u16 *refcheck);
2454 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2455 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2456 unsigned cl_env_cache_purge(unsigned nr);
2457 struct lu_env *cl_env_percpu_get(void);
2458 void cl_env_percpu_put(struct lu_env *env);
2465 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2466 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2468 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2469 struct lu_device_type *ldt,
2470 struct lu_device *next);
2473 int cl_global_init(void);
2474 void cl_global_fini(void);
2476 #endif /* _LINUX_CL_OBJECT_H */