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).
68 * - to avoid confusion high-level I/O operation like read or write system
69 * call is referred to as "an io", whereas low-level I/O operation, like
70 * RPC, is referred to as "a transfer"
72 * - "generic code" means generic (not file system specific) code in the
73 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
74 * is not layer specific.
80 * - cl_object_header::coh_page_guard
83 * See the top comment in cl_object.c for the description of overall locking and
84 * reference-counting design.
86 * See comments below for the description of i/o, page, and dlm-locking
93 * super-class definitions.
95 #include <libcfs/libcfs.h>
96 #include <lu_object.h>
97 #include <linux/atomic.h>
98 #include <linux/mutex.h>
99 #include <linux/radix-tree.h>
100 #include <linux/spinlock.h>
101 #include <linux/wait.h>
102 #include <lustre_dlm.h>
112 struct cl_page_slice;
114 struct cl_lock_slice;
116 struct cl_lock_operations;
117 struct cl_page_operations;
125 * Device in the client stack.
127 * \see vvp_device, lov_device, lovsub_device, osc_device
131 struct lu_device cd_lu_dev;
134 /** \addtogroup cl_object cl_object
137 * "Data attributes" of cl_object. Data attributes can be updated
138 * independently for a sub-object, and top-object's attributes are calculated
139 * from sub-objects' ones.
142 /** Object size, in bytes */
145 * Known minimal size, in bytes.
147 * This is only valid when at least one DLM lock is held.
150 /** Modification time. Measured in seconds since epoch. */
152 /** Access time. Measured in seconds since epoch. */
154 /** Change time. Measured in seconds since epoch. */
157 * Blocks allocated to this cl_object on the server file system.
159 * \todo XXX An interface for block size is needed.
163 * User identifier for quota purposes.
167 * Group identifier for quota purposes.
171 /* nlink of the directory */
176 * 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. */
294 * Operations implemented for each cl object layer.
296 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
298 struct cl_object_operations {
300 * Initialize page slice for this layer. Called top-to-bottom through
301 * every object layer when a new cl_page is instantiated. Layer
302 * keeping private per-page data, or requiring its own page operations
303 * vector should allocate these data here, and attach then to the page
304 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
307 * \retval NULL success.
309 * \retval ERR_PTR(errno) failure code.
311 * \retval valid-pointer pointer to already existing referenced page
312 * to be used instead of newly created.
314 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
315 struct cl_page *page, pgoff_t index);
317 * Initialize lock slice for this layer. Called top-to-bottom through
318 * every object layer when a new cl_lock is instantiated. Layer
319 * keeping private per-lock data, or requiring its own lock operations
320 * vector should allocate these data here, and attach then to the lock
321 * by calling cl_lock_slice_add(). Mandatory.
323 int (*coo_lock_init)(const struct lu_env *env,
324 struct cl_object *obj, struct cl_lock *lock,
325 const struct cl_io *io);
327 * Initialize io state for a given layer.
329 * called top-to-bottom once per io existence to initialize io
330 * state. If layer wants to keep some state for this type of io, it
331 * has to embed struct cl_io_slice in lu_env::le_ses, and register
332 * slice with cl_io_slice_add(). It is guaranteed that all threads
333 * participating in this io share the same session.
335 int (*coo_io_init)(const struct lu_env *env,
336 struct cl_object *obj, struct cl_io *io);
338 * Fill portion of \a attr that this layer controls. This method is
339 * called top-to-bottom through all object layers.
341 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
343 * \return 0: to continue
344 * \return +ve: to stop iterating through layers (but 0 is returned
345 * from enclosing cl_object_attr_get())
346 * \return -ve: to signal error
348 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
349 struct cl_attr *attr);
353 * \a valid is a bitmask composed from enum #cl_attr_valid, and
354 * indicating what attributes are to be set.
356 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
358 * \return the same convention as for
359 * cl_object_operations::coo_attr_get() is used.
361 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
362 const struct cl_attr *attr, unsigned valid);
364 * Update object configuration. Called top-to-bottom to modify object
367 * XXX error conditions and handling.
369 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
370 const struct cl_object_conf *conf);
372 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
373 * object. Layers are supposed to fill parts of \a lvb that will be
374 * shipped to the glimpse originator as a glimpse result.
376 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
377 * \see osc_object_glimpse()
379 int (*coo_glimpse)(const struct lu_env *env,
380 const struct cl_object *obj, struct ost_lvb *lvb);
382 * Object prune method. Called when the layout is going to change on
383 * this object, therefore each layer has to clean up their cache,
384 * mainly pages and locks.
386 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
388 * Object getstripe method.
390 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
391 struct lov_user_md __user *lum);
393 * Find whether there is any callback data (ldlm lock) attached upon
396 int (*coo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
397 ldlm_iterator_t iter, void *data);
399 * Get FIEMAP mapping from the object.
401 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
402 struct ll_fiemap_info_key *fmkey,
403 struct fiemap *fiemap, size_t *buflen);
405 * Get layout and generation of the object.
407 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
408 struct cl_layout *layout);
410 * Get maximum size of the object.
412 loff_t (*coo_maxbytes)(struct cl_object *obj);
414 * Set request attributes.
416 void (*coo_req_attr_set)(const struct lu_env *env,
417 struct cl_object *obj,
418 struct cl_req_attr *attr);
422 * Extended header for client object.
424 struct cl_object_header {
425 /** Standard lu_object_header. cl_object::co_lu::lo_header points
427 struct lu_object_header coh_lu;
430 * Parent object. It is assumed that an object has a well-defined
431 * parent, but not a well-defined child (there may be multiple
432 * sub-objects, for the same top-object). cl_object_header::coh_parent
433 * field allows certain code to be written generically, without
434 * limiting possible cl_object layouts unduly.
436 struct cl_object_header *coh_parent;
438 * Protects consistency between cl_attr of parent object and
439 * attributes of sub-objects, that the former is calculated ("merged")
442 * \todo XXX this can be read/write lock if needed.
444 spinlock_t coh_attr_guard;
446 * Size of cl_page + page slices
448 unsigned short coh_page_bufsize;
450 * Number of objects above this one: 0 for a top-object, 1 for its
453 unsigned char coh_nesting;
457 * Helper macro: iterate over all layers of the object \a obj, assigning every
458 * layer top-to-bottom to \a slice.
460 #define cl_object_for_each(slice, obj) \
461 list_for_each_entry((slice), \
462 &(obj)->co_lu.lo_header->loh_layers,\
466 * Helper macro: iterate over all layers of the object \a obj, assigning every
467 * layer bottom-to-top to \a slice.
469 #define cl_object_for_each_reverse(slice, obj) \
470 list_for_each_entry_reverse((slice), \
471 &(obj)->co_lu.lo_header->loh_layers,\
476 #define CL_PAGE_EOF ((pgoff_t)~0ull)
478 /** \addtogroup cl_page cl_page
482 * Layered client page.
484 * cl_page: represents a portion of a file, cached in the memory. All pages
485 * of the given file are of the same size, and are kept in the radix tree
486 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
487 * of the top-level file object are first class cl_objects, they have their
488 * own radix trees of pages and hence page is implemented as a sequence of
489 * struct cl_pages's, linked into double-linked list through
490 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
491 * corresponding radix tree at the corresponding logical offset.
493 * cl_page is associated with VM page of the hosting environment (struct
494 * page in Linux kernel, for example), struct page. It is assumed, that this
495 * association is implemented by one of cl_page layers (top layer in the
496 * current design) that
498 * - intercepts per-VM-page call-backs made by the environment (e.g.,
501 * - translates state (page flag bits) and locking between lustre and
504 * The association between cl_page and struct page is immutable and
505 * established when cl_page is created.
507 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
508 * this io an exclusive access to this page w.r.t. other io attempts and
509 * various events changing page state (such as transfer completion, or
510 * eviction of the page from the memory). Note, that in general cl_io
511 * cannot be identified with a particular thread, and page ownership is not
512 * exactly equal to the current thread holding a lock on the page. Layer
513 * implementing association between cl_page and struct page has to implement
514 * ownership on top of available synchronization mechanisms.
516 * While lustre client maintains the notion of an page ownership by io,
517 * hosting MM/VM usually has its own page concurrency control
518 * mechanisms. For example, in Linux, page access is synchronized by the
519 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
520 * takes care to acquire and release such locks as necessary around the
521 * calls to the file system methods (->readpage(), ->prepare_write(),
522 * ->commit_write(), etc.). This leads to the situation when there are two
523 * different ways to own a page in the client:
525 * - client code explicitly and voluntary owns the page (cl_page_own());
527 * - VM locks a page and then calls the client, that has "to assume"
528 * the ownership from the VM (cl_page_assume()).
530 * Dual methods to release ownership are cl_page_disown() and
531 * cl_page_unassume().
533 * cl_page is reference counted (cl_page::cp_ref). When reference counter
534 * drops to 0, the page is returned to the cache, unless it is in
535 * cl_page_state::CPS_FREEING state, in which case it is immediately
538 * The general logic guaranteeing the absence of "existential races" for
539 * pages is the following:
541 * - there are fixed known ways for a thread to obtain a new reference
544 * - by doing a lookup in the cl_object radix tree, protected by the
547 * - by starting from VM-locked struct page and following some
548 * hosting environment method (e.g., following ->private pointer in
549 * the case of Linux kernel), see cl_vmpage_page();
551 * - when the page enters cl_page_state::CPS_FREEING state, all these
552 * ways are severed with the proper synchronization
553 * (cl_page_delete());
555 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
558 * - no new references to the page in cl_page_state::CPS_FREEING state
559 * are allowed (checked in cl_page_get()).
561 * Together this guarantees that when last reference to a
562 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
563 * page, as neither references to it can be acquired at that point, nor
566 * cl_page is a state machine. States are enumerated in enum
567 * cl_page_state. Possible state transitions are enumerated in
568 * cl_page_state_set(). State transition process (i.e., actual changing of
569 * cl_page::cp_state field) is protected by the lock on the underlying VM
572 * Linux Kernel implementation.
574 * Binding between cl_page and struct page (which is a typedef for
575 * struct page) is implemented in the vvp layer. cl_page is attached to the
576 * ->private pointer of the struct page, together with the setting of
577 * PG_private bit in page->flags, and acquiring additional reference on the
578 * struct page (much like struct buffer_head, or any similar file system
579 * private data structures).
581 * PG_locked lock is used to implement both ownership and transfer
582 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
583 * states. No additional references are acquired for the duration of the
586 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
587 * write-out is "protected" by the special PG_writeback bit.
591 * States of cl_page. cl_page.c assumes particular order here.
593 * The page state machine is rather crude, as it doesn't recognize finer page
594 * states like "dirty" or "up to date". This is because such states are not
595 * always well defined for the whole stack (see, for example, the
596 * implementation of the read-ahead, that hides page up-to-dateness to track
597 * cache hits accurately). Such sub-states are maintained by the layers that
598 * are interested in them.
602 * Page is in the cache, un-owned. Page leaves cached state in the
605 * - [cl_page_state::CPS_OWNED] io comes across the page and
608 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
609 * req-formation engine decides that it wants to include this page
610 * into an RPC being constructed, and yanks it from the cache;
612 * - [cl_page_state::CPS_FREEING] VM callback is executed to
613 * evict the page form the memory;
615 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
619 * Page is exclusively owned by some cl_io. Page may end up in this
620 * state as a result of
622 * - io creating new page and immediately owning it;
624 * - [cl_page_state::CPS_CACHED] io finding existing cached page
627 * - [cl_page_state::CPS_OWNED] io finding existing owned page
628 * and waiting for owner to release the page;
630 * Page leaves owned state in the following cases:
632 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
633 * the cache, doing nothing;
635 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
638 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
639 * transfer for this page;
641 * - [cl_page_state::CPS_FREEING] io decides to destroy this
642 * page (e.g., as part of truncate or extent lock cancellation).
644 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
648 * Page is being written out, as a part of a transfer. This state is
649 * entered when req-formation logic decided that it wants this page to
650 * be sent through the wire _now_. Specifically, it means that once
651 * this state is achieved, transfer completion handler (with either
652 * success or failure indication) is guaranteed to be executed against
653 * this page independently of any locks and any scheduling decisions
654 * made by the hosting environment (that effectively means that the
655 * page is never put into cl_page_state::CPS_PAGEOUT state "in
656 * advance". This property is mentioned, because it is important when
657 * reasoning about possible dead-locks in the system). The page can
658 * enter this state as a result of
660 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
661 * write-out of this page, or
663 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
664 * that it has enough dirty pages cached to issue a "good"
667 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
668 * is completed---it is moved into cl_page_state::CPS_CACHED state.
670 * Underlying VM page is locked for the duration of transfer.
672 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
676 * Page is being read in, as a part of a transfer. This is quite
677 * similar to the cl_page_state::CPS_PAGEOUT state, except that
678 * read-in is always "immediate"---there is no such thing a sudden
679 * construction of read request from cached, presumably not up to date,
682 * Underlying VM page is locked for the duration of transfer.
684 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
688 * Page is being destroyed. This state is entered when client decides
689 * that page has to be deleted from its host object, as, e.g., a part
692 * Once this state is reached, there is no way to escape it.
694 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
701 /** Host page, the page is from the host inode which the cl_page
705 /** Transient page, the transient cl_page is used to bind a cl_page
706 * to vmpage which is not belonging to the same object of cl_page.
707 * it is used in DirectIO, lockless IO and liblustre. */
712 * Fields are protected by the lock on struct page, except for atomics and
715 * \invariant Data type invariants are in cl_page_invariant(). Basically:
716 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
717 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
718 * cl_page::cp_owner (when set).
721 /** Reference counter. */
723 /** Transfer error. */
725 /** An object this page is a part of. Immutable after creation. */
726 struct cl_object *cp_obj;
728 struct page *cp_vmpage;
729 /** Linkage of pages within group. Pages must be owned */
730 struct list_head cp_batch;
731 /** List of slices. Immutable after creation. */
732 struct list_head cp_layers;
734 * Page state. This field is const to avoid accidental update, it is
735 * modified only internally within cl_page.c. Protected by a VM lock.
737 const enum cl_page_state cp_state;
739 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
742 enum cl_page_type cp_type;
745 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
746 * by sub-io. Protected by a VM lock.
748 struct cl_io *cp_owner;
749 /** List of references to this page, for debugging. */
750 struct lu_ref cp_reference;
751 /** Link to an object, for debugging. */
752 struct lu_ref_link cp_obj_ref;
753 /** Link to a queue, for debugging. */
754 struct lu_ref_link cp_queue_ref;
755 /** Assigned if doing a sync_io */
756 struct cl_sync_io *cp_sync_io;
760 * Per-layer part of cl_page.
762 * \see vvp_page, lov_page, osc_page
764 struct cl_page_slice {
765 struct cl_page *cpl_page;
768 * Object slice corresponding to this page slice. Immutable after
771 struct cl_object *cpl_obj;
772 const struct cl_page_operations *cpl_ops;
773 /** Linkage into cl_page::cp_layers. Immutable after creation. */
774 struct list_head cpl_linkage;
778 * Lock mode. For the client extent locks.
790 * Requested transfer type.
799 * Per-layer page operations.
801 * Methods taking an \a io argument are for the activity happening in the
802 * context of given \a io. Page is assumed to be owned by that io, except for
803 * the obvious cases (like cl_page_operations::cpo_own()).
805 * \see vvp_page_ops, lov_page_ops, osc_page_ops
807 struct cl_page_operations {
809 * cl_page<->struct page methods. Only one layer in the stack has to
810 * implement these. Current code assumes that this functionality is
811 * provided by the topmost layer, see cl_page_disown0() as an example.
815 * Called when \a io acquires this page into the exclusive
816 * ownership. When this method returns, it is guaranteed that the is
817 * not owned by other io, and no transfer is going on against
821 * \see vvp_page_own(), lov_page_own()
823 int (*cpo_own)(const struct lu_env *env,
824 const struct cl_page_slice *slice,
825 struct cl_io *io, int nonblock);
826 /** Called when ownership it yielded. Optional.
828 * \see cl_page_disown()
829 * \see vvp_page_disown()
831 void (*cpo_disown)(const struct lu_env *env,
832 const struct cl_page_slice *slice, struct cl_io *io);
834 * Called for a page that is already "owned" by \a io from VM point of
837 * \see cl_page_assume()
838 * \see vvp_page_assume(), lov_page_assume()
840 void (*cpo_assume)(const struct lu_env *env,
841 const struct cl_page_slice *slice, struct cl_io *io);
842 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
843 * bottom-to-top when IO releases a page without actually unlocking
846 * \see cl_page_unassume()
847 * \see vvp_page_unassume()
849 void (*cpo_unassume)(const struct lu_env *env,
850 const struct cl_page_slice *slice,
853 * Announces whether the page contains valid data or not by \a uptodate.
855 * \see cl_page_export()
856 * \see vvp_page_export()
858 void (*cpo_export)(const struct lu_env *env,
859 const struct cl_page_slice *slice, int uptodate);
861 * Checks whether underlying VM page is locked (in the suitable
862 * sense). Used for assertions.
864 * \retval -EBUSY: page is protected by a lock of a given mode;
865 * \retval -ENODATA: page is not protected by a lock;
866 * \retval 0: this layer cannot decide. (Should never happen.)
868 int (*cpo_is_vmlocked)(const struct lu_env *env,
869 const struct cl_page_slice *slice);
875 * Called when page is truncated from the object. Optional.
877 * \see cl_page_discard()
878 * \see vvp_page_discard(), osc_page_discard()
880 void (*cpo_discard)(const struct lu_env *env,
881 const struct cl_page_slice *slice,
884 * Called when page is removed from the cache, and is about to being
885 * destroyed. Optional.
887 * \see cl_page_delete()
888 * \see vvp_page_delete(), osc_page_delete()
890 void (*cpo_delete)(const struct lu_env *env,
891 const struct cl_page_slice *slice);
892 /** Destructor. Frees resources and slice itself. */
893 void (*cpo_fini)(const struct lu_env *env,
894 struct cl_page_slice *slice);
896 * Optional debugging helper. Prints given page slice.
898 * \see cl_page_print()
900 int (*cpo_print)(const struct lu_env *env,
901 const struct cl_page_slice *slice,
902 void *cookie, lu_printer_t p);
911 * Request type dependent vector of operations.
913 * Transfer operations depend on transfer mode (cl_req_type). To avoid
914 * passing transfer mode to each and every of these methods, and to
915 * avoid branching on request type inside of the methods, separate
916 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
917 * provided. That is, method invocation usually looks like
919 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
923 * Called when a page is submitted for a transfer as a part of
926 * \return 0 : page is eligible for submission;
927 * \return -EALREADY : skip this page;
928 * \return -ve : error.
930 * \see cl_page_prep()
932 int (*cpo_prep)(const struct lu_env *env,
933 const struct cl_page_slice *slice,
936 * Completion handler. This is guaranteed to be eventually
937 * fired after cl_page_operations::cpo_prep() or
938 * cl_page_operations::cpo_make_ready() call.
940 * This method can be called in a non-blocking context. It is
941 * guaranteed however, that the page involved and its object
942 * are pinned in memory (and, hence, calling cl_page_put() is
945 * \see cl_page_completion()
947 void (*cpo_completion)(const struct lu_env *env,
948 const struct cl_page_slice *slice,
951 * Called when cached page is about to be added to the
952 * ptlrpc request as a part of req formation.
954 * \return 0 : proceed with this page;
955 * \return -EAGAIN : skip this page;
956 * \return -ve : error.
958 * \see cl_page_make_ready()
960 int (*cpo_make_ready)(const struct lu_env *env,
961 const struct cl_page_slice *slice);
964 * Tell transfer engine that only [to, from] part of a page should be
967 * This is used for immediate transfers.
969 * \todo XXX this is not very good interface. It would be much better
970 * if all transfer parameters were supplied as arguments to
971 * cl_io_operations::cio_submit() call, but it is not clear how to do
972 * this for page queues.
974 * \see cl_page_clip()
976 void (*cpo_clip)(const struct lu_env *env,
977 const struct cl_page_slice *slice,
980 * \pre the page was queued for transferring.
981 * \post page is removed from client's pending list, or -EBUSY
982 * is returned if it has already been in transferring.
984 * This is one of seldom page operation which is:
985 * 0. called from top level;
986 * 1. don't have vmpage locked;
987 * 2. every layer should synchronize execution of its ->cpo_cancel()
988 * with completion handlers. Osc uses client obd lock for this
989 * purpose. Based on there is no vvp_page_cancel and
990 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
992 * \see osc_page_cancel().
994 int (*cpo_cancel)(const struct lu_env *env,
995 const struct cl_page_slice *slice);
997 * Write out a page by kernel. This is only called by ll_writepage
1000 * \see cl_page_flush()
1002 int (*cpo_flush)(const struct lu_env *env,
1003 const struct cl_page_slice *slice,
1009 * Helper macro, dumping detailed information about \a page into a log.
1011 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1013 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1014 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1015 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1016 CDEBUG(mask, format , ## __VA_ARGS__); \
1021 * Helper macro, dumping shorter information about \a page into a log.
1023 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1025 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1026 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1027 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1028 CDEBUG(mask, format , ## __VA_ARGS__); \
1032 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1034 LASSERT(page->cp_vmpage != NULL);
1035 return page->cp_vmpage;
1039 * Check if a cl_page is in use.
1041 * Client cache holds a refcount, this refcount will be dropped when
1042 * the page is taken out of cache, see vvp_page_delete().
1044 static inline bool __page_in_use(const struct cl_page *page, int refc)
1046 return (atomic_read(&page->cp_ref) > refc + 1);
1050 * Caller itself holds a refcount of cl_page.
1052 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1054 * Caller doesn't hold a refcount.
1056 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1060 /** \addtogroup cl_lock cl_lock
1064 * Extent locking on the client.
1068 * The locking model of the new client code is built around
1072 * data-type representing an extent lock on a regular file. cl_lock is a
1073 * layered object (much like cl_object and cl_page), it consists of a header
1074 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1075 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1077 * Typical cl_lock consists of the two layers:
1079 * - vvp_lock (vvp specific data), and
1080 * - lov_lock (lov specific data).
1082 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1083 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1085 * - lovsub_lock, and
1088 * Each sub-lock is associated with a cl_object (representing stripe
1089 * sub-object or the file to which top-level cl_lock is associated to), and is
1090 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1091 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1092 * is different from cl_page, that doesn't fan out (there is usually exactly
1093 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1094 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1098 * cl_lock is a cacheless data container for the requirements of locks to
1099 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1102 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1103 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1105 * INTERFACE AND USAGE
1107 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1108 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1109 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1110 * consists of multiple sub cl_locks, each sub locks will be enqueued
1111 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1112 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1115 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1116 * method will be called for each layer to release the resource held by this
1117 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1118 * clo_enqueue time, is released.
1120 * LDLM lock can only be canceled if there is no cl_lock using it.
1122 * Overall process of the locking during IO operation is as following:
1124 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1125 * is called on each layer. Responsibility of this method is to add locks,
1126 * needed by a given layer into cl_io.ci_lockset.
1128 * - once locks for all layers were collected, they are sorted to avoid
1129 * dead-locks (cl_io_locks_sort()), and enqueued.
1131 * - when all locks are acquired, IO is performed;
1133 * - locks are released after IO is complete.
1135 * Striping introduces major additional complexity into locking. The
1136 * fundamental problem is that it is generally unsafe to actively use (hold)
1137 * two locks on the different OST servers at the same time, as this introduces
1138 * inter-server dependency and can lead to cascading evictions.
1140 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1141 * that no multi-stripe locks are taken (note that this design abandons POSIX
1142 * read/write semantics). Such pieces ideally can be executed concurrently. At
1143 * the same time, certain types of IO cannot be sub-divived, without
1144 * sacrificing correctness. This includes:
1146 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1149 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1151 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1152 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1153 * has to be held together with the usual lock on [offset, offset + count].
1155 * Interaction with DLM
1157 * In the expected setup, cl_lock is ultimately backed up by a collection of
1158 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1159 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1160 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1161 * description of interaction with DLM.
1167 struct cl_lock_descr {
1168 /** Object this lock is granted for. */
1169 struct cl_object *cld_obj;
1170 /** Index of the first page protected by this lock. */
1172 /** Index of the last page (inclusive) protected by this lock. */
1174 /** Group ID, for group lock */
1177 enum cl_lock_mode cld_mode;
1179 * flags to enqueue lock. A combination of bit-flags from
1180 * enum cl_enq_flags.
1182 __u32 cld_enq_flags;
1185 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1186 #define PDESCR(descr) \
1187 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1188 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1190 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1193 * Layered client lock.
1196 /** List of slices. Immutable after creation. */
1197 struct list_head cll_layers;
1198 /** lock attribute, extent, cl_object, etc. */
1199 struct cl_lock_descr cll_descr;
1203 * Per-layer part of cl_lock
1205 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1207 struct cl_lock_slice {
1208 struct cl_lock *cls_lock;
1209 /** Object slice corresponding to this lock slice. Immutable after
1211 struct cl_object *cls_obj;
1212 const struct cl_lock_operations *cls_ops;
1213 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1214 struct list_head cls_linkage;
1219 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1221 struct cl_lock_operations {
1224 * Attempts to enqueue the lock. Called top-to-bottom.
1226 * \retval 0 this layer has enqueued the lock successfully
1227 * \retval >0 this layer has enqueued the lock, but need to wait on
1228 * @anchor for resources
1229 * \retval -ve failure
1231 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1232 * \see osc_lock_enqueue()
1234 int (*clo_enqueue)(const struct lu_env *env,
1235 const struct cl_lock_slice *slice,
1236 struct cl_io *io, struct cl_sync_io *anchor);
1238 * Cancel a lock, release its DLM lock ref, while does not cancel the
1241 void (*clo_cancel)(const struct lu_env *env,
1242 const struct cl_lock_slice *slice);
1245 * Destructor. Frees resources and the slice.
1247 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1248 * \see osc_lock_fini()
1250 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1252 * Optional debugging helper. Prints given lock slice.
1254 int (*clo_print)(const struct lu_env *env,
1255 void *cookie, lu_printer_t p,
1256 const struct cl_lock_slice *slice);
1259 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1261 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1262 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1263 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1264 CDEBUG(mask, format , ## __VA_ARGS__); \
1268 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1272 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1278 /** \addtogroup cl_page_list cl_page_list
1279 * Page list used to perform collective operations on a group of pages.
1281 * Pages are added to the list one by one. cl_page_list acquires a reference
1282 * for every page in it. Page list is used to perform collective operations on
1285 * - submit pages for an immediate transfer,
1287 * - own pages on behalf of certain io (waiting for each page in turn),
1291 * When list is finalized, it releases references on all pages it still has.
1293 * \todo XXX concurrency control.
1297 struct cl_page_list {
1299 struct list_head pl_pages;
1300 struct task_struct *pl_owner;
1304 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1305 * contains an incoming page list and an outgoing page list.
1308 struct cl_page_list c2_qin;
1309 struct cl_page_list c2_qout;
1312 /** @} cl_page_list */
1314 /** \addtogroup cl_io cl_io
1319 * cl_io represents a high level I/O activity like
1320 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1323 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1324 * important distinction. We want to minimize number of calls to the allocator
1325 * in the fast path, e.g., in the case of read(2) when everything is cached:
1326 * client already owns the lock over region being read, and data are cached
1327 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1328 * per-layer io state is stored in the session, associated with the io, see
1329 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1330 * by using free-lists, see cl_env_get().
1332 * There is a small predefined number of possible io types, enumerated in enum
1335 * cl_io is a state machine, that can be advanced concurrently by the multiple
1336 * threads. It is up to these threads to control the concurrency and,
1337 * specifically, to detect when io is done, and its state can be safely
1340 * For read/write io overall execution plan is as following:
1342 * (0) initialize io state through all layers;
1344 * (1) loop: prepare chunk of work to do
1346 * (2) call all layers to collect locks they need to process current chunk
1348 * (3) sort all locks to avoid dead-locks, and acquire them
1350 * (4) process the chunk: call per-page methods
1351 * cl_io_operations::cio_prepare_write(),
1352 * cl_io_operations::cio_commit_write() for write)
1358 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1359 * address allocation efficiency issues mentioned above), and returns with the
1360 * special error condition from per-page method when current sub-io has to
1361 * block. This causes io loop to be repeated, and lov switches to the next
1362 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1367 /** read system call */
1369 /** write system call */
1371 /** truncate, utime system calls */
1373 /** get data version */
1376 * page fault handling
1380 * fsync system call handling
1381 * To write out a range of file
1385 * Miscellaneous io. This is used for occasional io activity that
1386 * doesn't fit into other types. Currently this is used for:
1388 * - cancellation of an extent lock. This io exists as a context
1389 * to write dirty pages from under the lock being canceled back
1392 * - VM induced page write-out. An io context for writing page out
1393 * for memory cleansing;
1395 * - glimpse. An io context to acquire glimpse lock.
1397 * - grouplock. An io context to acquire group lock.
1399 * CIT_MISC io is used simply as a context in which locks and pages
1400 * are manipulated. Such io has no internal "process", that is,
1401 * cl_io_loop() is never called for it.
1408 * States of cl_io state machine
1411 /** Not initialized. */
1415 /** IO iteration started. */
1419 /** Actual IO is in progress. */
1421 /** IO for the current iteration finished. */
1423 /** Locks released. */
1425 /** Iteration completed. */
1427 /** cl_io finalized. */
1432 * IO state private for a layer.
1434 * This is usually embedded into layer session data, rather than allocated
1437 * \see vvp_io, lov_io, osc_io
1439 struct cl_io_slice {
1440 struct cl_io *cis_io;
1441 /** corresponding object slice. Immutable after creation. */
1442 struct cl_object *cis_obj;
1443 /** io operations. Immutable after creation. */
1444 const struct cl_io_operations *cis_iop;
1446 * linkage into a list of all slices for a given cl_io, hanging off
1447 * cl_io::ci_layers. Immutable after creation.
1449 struct list_head cis_linkage;
1452 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1455 struct cl_read_ahead {
1456 /* Maximum page index the readahead window will end.
1457 * This is determined DLM lock coverage, RPC and stripe boundary.
1458 * cra_end is included. */
1460 /* Release routine. If readahead holds resources underneath, this
1461 * function should be called to release it. */
1462 void (*cra_release)(const struct lu_env *env, void *cbdata);
1463 /* Callback data for cra_release routine */
1467 static inline void cl_read_ahead_release(const struct lu_env *env,
1468 struct cl_read_ahead *ra)
1470 if (ra->cra_release != NULL)
1471 ra->cra_release(env, ra->cra_cbdata);
1472 memset(ra, 0, sizeof(*ra));
1477 * Per-layer io operations.
1478 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1480 struct cl_io_operations {
1482 * Vector of io state transition methods for every io type.
1484 * \see cl_page_operations::io
1488 * Prepare io iteration at a given layer.
1490 * Called top-to-bottom at the beginning of each iteration of
1491 * "io loop" (if it makes sense for this type of io). Here
1492 * layer selects what work it will do during this iteration.
1494 * \see cl_io_operations::cio_iter_fini()
1496 int (*cio_iter_init) (const struct lu_env *env,
1497 const struct cl_io_slice *slice);
1499 * Finalize io iteration.
1501 * Called bottom-to-top at the end of each iteration of "io
1502 * loop". Here layers can decide whether IO has to be
1505 * \see cl_io_operations::cio_iter_init()
1507 void (*cio_iter_fini) (const struct lu_env *env,
1508 const struct cl_io_slice *slice);
1510 * Collect locks for the current iteration of io.
1512 * Called top-to-bottom to collect all locks necessary for
1513 * this iteration. This methods shouldn't actually enqueue
1514 * anything, instead it should post a lock through
1515 * cl_io_lock_add(). Once all locks are collected, they are
1516 * sorted and enqueued in the proper order.
1518 int (*cio_lock) (const struct lu_env *env,
1519 const struct cl_io_slice *slice);
1521 * Finalize unlocking.
1523 * Called bottom-to-top to finish layer specific unlocking
1524 * functionality, after generic code released all locks
1525 * acquired by cl_io_operations::cio_lock().
1527 void (*cio_unlock)(const struct lu_env *env,
1528 const struct cl_io_slice *slice);
1530 * Start io iteration.
1532 * Once all locks are acquired, called top-to-bottom to
1533 * commence actual IO. In the current implementation,
1534 * top-level vvp_io_{read,write}_start() does all the work
1535 * synchronously by calling generic_file_*(), so other layers
1536 * are called when everything is done.
1538 int (*cio_start)(const struct lu_env *env,
1539 const struct cl_io_slice *slice);
1541 * Called top-to-bottom at the end of io loop. Here layer
1542 * might wait for an unfinished asynchronous io.
1544 void (*cio_end) (const struct lu_env *env,
1545 const struct cl_io_slice *slice);
1547 * Called bottom-to-top to notify layers that read/write IO
1548 * iteration finished, with \a nob bytes transferred.
1550 void (*cio_advance)(const struct lu_env *env,
1551 const struct cl_io_slice *slice,
1554 * Called once per io, bottom-to-top to release io resources.
1556 void (*cio_fini) (const struct lu_env *env,
1557 const struct cl_io_slice *slice);
1561 * Submit pages from \a queue->c2_qin for IO, and move
1562 * successfully submitted pages into \a queue->c2_qout. Return
1563 * non-zero if failed to submit even the single page. If
1564 * submission failed after some pages were moved into \a
1565 * queue->c2_qout, completion callback with non-zero ioret is
1568 int (*cio_submit)(const struct lu_env *env,
1569 const struct cl_io_slice *slice,
1570 enum cl_req_type crt,
1571 struct cl_2queue *queue);
1573 * Queue async page for write.
1574 * The difference between cio_submit and cio_queue is that
1575 * cio_submit is for urgent request.
1577 int (*cio_commit_async)(const struct lu_env *env,
1578 const struct cl_io_slice *slice,
1579 struct cl_page_list *queue, int from, int to,
1582 * Decide maximum read ahead extent
1584 * \pre io->ci_type == CIT_READ
1586 int (*cio_read_ahead)(const struct lu_env *env,
1587 const struct cl_io_slice *slice,
1588 pgoff_t start, struct cl_read_ahead *ra);
1590 * Optional debugging helper. Print given io slice.
1592 int (*cio_print)(const struct lu_env *env, void *cookie,
1593 lu_printer_t p, const struct cl_io_slice *slice);
1597 * Flags to lock enqueue procedure.
1602 * instruct server to not block, if conflicting lock is found. Instead
1603 * -EWOULDBLOCK is returned immediately.
1605 CEF_NONBLOCK = 0x00000001,
1607 * take lock asynchronously (out of order), as it cannot
1608 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1610 CEF_ASYNC = 0x00000002,
1612 * tell the server to instruct (though a flag in the blocking ast) an
1613 * owner of the conflicting lock, that it can drop dirty pages
1614 * protected by this lock, without sending them to the server.
1616 CEF_DISCARD_DATA = 0x00000004,
1618 * tell the sub layers that it must be a `real' lock. This is used for
1619 * mmapped-buffer locks and glimpse locks that must be never converted
1620 * into lockless mode.
1622 * \see vvp_mmap_locks(), cl_glimpse_lock().
1624 CEF_MUST = 0x00000008,
1626 * tell the sub layers that never request a `real' lock. This flag is
1627 * not used currently.
1629 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1630 * conversion policy: ci_lockreq describes generic information of lock
1631 * requirement for this IO, especially for locks which belong to the
1632 * object doing IO; however, lock itself may have precise requirements
1633 * that are described by the enqueue flags.
1635 CEF_NEVER = 0x00000010,
1637 * for async glimpse lock.
1639 CEF_AGL = 0x00000020,
1641 * enqueue a lock to test DLM lock existence.
1643 CEF_PEEK = 0x00000040,
1645 * Lock match only. Used by group lock in I/O as group lock
1646 * is known to exist.
1648 CEF_LOCK_MATCH = 0x00000080,
1650 * mask of enq_flags.
1652 CEF_MASK = 0x000000ff,
1656 * Link between lock and io. Intermediate structure is needed, because the
1657 * same lock can be part of multiple io's simultaneously.
1659 struct cl_io_lock_link {
1660 /** linkage into one of cl_lockset lists. */
1661 struct list_head cill_linkage;
1662 struct cl_lock cill_lock;
1663 /** optional destructor */
1664 void (*cill_fini)(const struct lu_env *env,
1665 struct cl_io_lock_link *link);
1667 #define cill_descr cill_lock.cll_descr
1670 * Lock-set represents a collection of locks, that io needs at a
1671 * time. Generally speaking, client tries to avoid holding multiple locks when
1674 * - holding extent locks over multiple ost's introduces the danger of
1675 * "cascading timeouts";
1677 * - holding multiple locks over the same ost is still dead-lock prone,
1678 * see comment in osc_lock_enqueue(),
1680 * but there are certain situations where this is unavoidable:
1682 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1684 * - truncate has to take [new-size, EOF] lock for correctness;
1686 * - SNS has to take locks across full stripe for correctness;
1688 * - in the case when user level buffer, supplied to {read,write}(file0),
1689 * is a part of a memory mapped lustre file, client has to take a dlm
1690 * locks on file0, and all files that back up the buffer (or a part of
1691 * the buffer, that is being processed in the current chunk, in any
1692 * case, there are situations where at least 2 locks are necessary).
1694 * In such cases we at least try to take locks in the same consistent
1695 * order. To this end, all locks are first collected, then sorted, and then
1699 /** locks to be acquired. */
1700 struct list_head cls_todo;
1701 /** locks acquired. */
1702 struct list_head cls_done;
1706 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1707 * but 'req' is always to be thought as 'request' :-)
1709 enum cl_io_lock_dmd {
1710 /** Always lock data (e.g., O_APPEND). */
1712 /** Layers are free to decide between local and global locking. */
1714 /** Never lock: there is no cache (e.g., liblustre). */
1718 enum cl_fsync_mode {
1719 /** start writeback, do not wait for them to finish */
1721 /** start writeback and wait for them to finish */
1723 /** discard all of dirty pages in a specific file range */
1724 CL_FSYNC_DISCARD = 2,
1725 /** start writeback and make sure they have reached storage before
1726 * return. OST_SYNC RPC must be issued and finished */
1730 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 const struct lu_fid *sa_parent_fid;
1781 struct cl_data_version_io {
1782 u64 dv_data_version;
1785 struct cl_fault_io {
1786 /** page index within file. */
1788 /** bytes valid byte on a faulted page. */
1790 /** writable page? for nopage() only */
1792 /** page of an executable? */
1794 /** page_mkwrite() */
1796 /** resulting page */
1797 struct cl_page *ft_page;
1799 struct cl_fsync_io {
1802 /** file system level fid */
1803 struct lu_fid *fi_fid;
1804 enum cl_fsync_mode fi_mode;
1805 /* how many pages were written/discarded */
1806 unsigned int fi_nr_written;
1809 struct cl_2queue ci_queue;
1812 unsigned int ci_continue:1,
1814 * This io has held grouplock, to inform sublayers that
1815 * don't do lockless i/o.
1819 * The whole IO need to be restarted because layout has been changed
1823 * to not refresh layout - the IO issuer knows that the layout won't
1824 * change(page operations, layout change causes all page to be
1825 * discarded), or it doesn't matter if it changes(sync).
1829 * Check if layout changed after the IO finishes. Mainly for HSM
1830 * requirement. If IO occurs to openning files, it doesn't need to
1831 * verify layout because HSM won't release openning files.
1832 * Right now, only two opertaions need to verify layout: glimpse
1837 * file is released, restore has to to be triggered by vvp layer
1839 ci_restore_needed:1,
1845 * Number of pages owned by this IO. For invariant checking.
1847 unsigned ci_owned_nr;
1853 * Per-transfer attributes.
1855 struct cl_req_attr {
1856 enum cl_req_type cra_type;
1858 struct cl_page *cra_page;
1859 /** Generic attributes for the server consumption. */
1860 struct obdo *cra_oa;
1862 char cra_jobid[LUSTRE_JOBID_SIZE];
1865 enum cache_stats_item {
1866 /** how many cache lookups were performed */
1868 /** how many times cache lookup resulted in a hit */
1870 /** how many entities are in the cache right now */
1872 /** how many entities in the cache are actively used (and cannot be
1873 * evicted) right now */
1875 /** how many entities were created at all */
1880 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
1883 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
1885 struct cache_stats {
1886 const char *cs_name;
1887 atomic_t cs_stats[CS_NR];
1890 /** These are not exported so far */
1891 void cache_stats_init (struct cache_stats *cs, const char *name);
1894 * Client-side site. This represents particular client stack. "Global"
1895 * variables should (directly or indirectly) be added here to allow multiple
1896 * clients to co-exist in the single address space.
1899 struct lu_site cs_lu;
1901 * Statistical counters. Atomics do not scale, something better like
1902 * per-cpu counters is needed.
1904 * These are exported as /proc/fs/lustre/llite/.../site
1906 * When interpreting keep in mind that both sub-locks (and sub-pages)
1907 * and top-locks (and top-pages) are accounted here.
1909 struct cache_stats cs_pages;
1910 atomic_t cs_pages_state[CPS_NR];
1913 int cl_site_init(struct cl_site *s, struct cl_device *top);
1914 void cl_site_fini(struct cl_site *s);
1915 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
1918 * Output client site statistical counters into a buffer. Suitable for
1919 * ll_rd_*()-style functions.
1921 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
1926 * Type conversion and accessory functions.
1930 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
1932 return container_of(site, struct cl_site, cs_lu);
1935 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
1937 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
1938 return container_of0(d, struct cl_device, cd_lu_dev);
1941 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
1943 return &d->cd_lu_dev;
1946 static inline struct cl_object *lu2cl(const struct lu_object *o)
1948 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
1949 return container_of0(o, struct cl_object, co_lu);
1952 static inline const struct cl_object_conf *
1953 lu2cl_conf(const struct lu_object_conf *conf)
1955 return container_of0(conf, struct cl_object_conf, coc_lu);
1958 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
1960 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
1963 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
1965 return container_of0(h, struct cl_object_header, coh_lu);
1968 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
1970 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
1974 struct cl_object_header *cl_object_header(const struct cl_object *obj)
1976 return luh2coh(obj->co_lu.lo_header);
1979 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
1981 return lu_device_init(&d->cd_lu_dev, t);
1984 static inline void cl_device_fini(struct cl_device *d)
1986 lu_device_fini(&d->cd_lu_dev);
1989 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1990 struct cl_object *obj, pgoff_t index,
1991 const struct cl_page_operations *ops);
1992 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
1993 struct cl_object *obj,
1994 const struct cl_lock_operations *ops);
1995 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
1996 struct cl_object *obj, const struct cl_io_operations *ops);
1999 /** \defgroup cl_object cl_object
2001 struct cl_object *cl_object_top (struct cl_object *o);
2002 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2003 const struct lu_fid *fid,
2004 const struct cl_object_conf *c);
2006 int cl_object_header_init(struct cl_object_header *h);
2007 void cl_object_header_fini(struct cl_object_header *h);
2008 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2009 void cl_object_get (struct cl_object *o);
2010 void cl_object_attr_lock (struct cl_object *o);
2011 void cl_object_attr_unlock(struct cl_object *o);
2012 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2013 struct cl_attr *attr);
2014 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2015 const struct cl_attr *attr, unsigned valid);
2016 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2017 struct ost_lvb *lvb);
2018 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2019 const struct cl_object_conf *conf);
2020 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2021 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2022 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2023 struct lov_user_md __user *lum);
2024 int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
2025 ldlm_iterator_t iter, void *data);
2026 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2027 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2029 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2030 struct cl_layout *cl);
2031 loff_t cl_object_maxbytes(struct cl_object *obj);
2034 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2036 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2038 return cl_object_header(o0) == cl_object_header(o1);
2041 static inline void cl_object_page_init(struct cl_object *clob, int size)
2043 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2044 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2045 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2048 static inline void *cl_object_page_slice(struct cl_object *clob,
2049 struct cl_page *page)
2051 return (void *)((char *)page + clob->co_slice_off);
2055 * Return refcount of cl_object.
2057 static inline int cl_object_refc(struct cl_object *clob)
2059 struct lu_object_header *header = clob->co_lu.lo_header;
2060 return atomic_read(&header->loh_ref);
2065 /** \defgroup cl_page cl_page
2073 /* callback of cl_page_gang_lookup() */
2075 struct cl_page *cl_page_find (const struct lu_env *env,
2076 struct cl_object *obj,
2077 pgoff_t idx, struct page *vmpage,
2078 enum cl_page_type type);
2079 struct cl_page *cl_page_alloc (const struct lu_env *env,
2080 struct cl_object *o, pgoff_t ind,
2081 struct page *vmpage,
2082 enum cl_page_type type);
2083 void cl_page_get (struct cl_page *page);
2084 void cl_page_put (const struct lu_env *env,
2085 struct cl_page *page);
2086 void cl_page_print (const struct lu_env *env, void *cookie,
2087 lu_printer_t printer,
2088 const struct cl_page *pg);
2089 void cl_page_header_print(const struct lu_env *env, void *cookie,
2090 lu_printer_t printer,
2091 const struct cl_page *pg);
2092 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2093 struct cl_page *cl_page_top (struct cl_page *page);
2095 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2096 const struct lu_device_type *dtype);
2101 * Functions dealing with the ownership of page by io.
2105 int cl_page_own (const struct lu_env *env,
2106 struct cl_io *io, struct cl_page *page);
2107 int cl_page_own_try (const struct lu_env *env,
2108 struct cl_io *io, struct cl_page *page);
2109 void cl_page_assume (const struct lu_env *env,
2110 struct cl_io *io, struct cl_page *page);
2111 void cl_page_unassume (const struct lu_env *env,
2112 struct cl_io *io, struct cl_page *pg);
2113 void cl_page_disown (const struct lu_env *env,
2114 struct cl_io *io, struct cl_page *page);
2115 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2122 * Functions dealing with the preparation of a page for a transfer, and
2123 * tracking transfer state.
2126 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2127 struct cl_page *pg, enum cl_req_type crt);
2128 void cl_page_completion (const struct lu_env *env,
2129 struct cl_page *pg, enum cl_req_type crt, int ioret);
2130 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2131 enum cl_req_type crt);
2132 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2133 struct cl_page *pg, enum cl_req_type crt);
2134 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2136 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2137 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2138 struct cl_page *pg);
2144 * \name helper routines
2145 * Functions to discard, delete and export a cl_page.
2148 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2149 struct cl_page *pg);
2150 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2151 int cl_page_is_vmlocked(const struct lu_env *env,
2152 const struct cl_page *pg);
2153 void cl_page_export(const struct lu_env *env,
2154 struct cl_page *pg, int uptodate);
2155 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2156 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2157 size_t cl_page_size(const struct cl_object *obj);
2159 void cl_lock_print(const struct lu_env *env, void *cookie,
2160 lu_printer_t printer, const struct cl_lock *lock);
2161 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2162 lu_printer_t printer,
2163 const struct cl_lock_descr *descr);
2167 * Data structure managing a client's cached pages. A count of
2168 * "unstable" pages is maintained, and an LRU of clean pages is
2169 * maintained. "unstable" pages are pages pinned by the ptlrpc
2170 * layer for recovery purposes.
2172 struct cl_client_cache {
2174 * # of client cache refcount
2175 * # of users (OSCs) + 2 (held by llite and lov)
2179 * # of threads are doing shrinking
2181 unsigned int ccc_lru_shrinkers;
2183 * # of LRU entries available
2185 atomic_long_t ccc_lru_left;
2187 * List of entities(OSCs) for this LRU cache
2189 struct list_head ccc_lru;
2191 * Max # of LRU entries
2193 unsigned long ccc_lru_max;
2195 * Lock to protect ccc_lru list
2197 spinlock_t ccc_lru_lock;
2199 * Set if unstable check is enabled
2201 unsigned int ccc_unstable_check:1;
2203 * # of unstable pages for this mount point
2205 atomic_long_t ccc_unstable_nr;
2207 * Waitq for awaiting unstable pages to reach zero.
2208 * Used at umounting time and signaled on BRW commit
2210 wait_queue_head_t ccc_unstable_waitq;
2213 * cl_cache functions
2215 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2216 void cl_cache_incref(struct cl_client_cache *cache);
2217 void cl_cache_decref(struct cl_client_cache *cache);
2221 /** \defgroup cl_lock cl_lock
2223 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2224 struct cl_lock *lock);
2225 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2226 const struct cl_io *io);
2227 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2228 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2229 const struct lu_device_type *dtype);
2230 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2232 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2233 struct cl_lock *lock, struct cl_sync_io *anchor);
2234 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2238 /** \defgroup cl_io cl_io
2241 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2242 enum cl_io_type iot, struct cl_object *obj);
2243 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2244 enum cl_io_type iot, struct cl_object *obj);
2245 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2246 enum cl_io_type iot, loff_t pos, size_t count);
2247 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2249 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2250 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2251 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2252 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2253 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2254 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2255 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2256 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2257 struct cl_io_lock_link *link);
2258 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2259 struct cl_lock_descr *descr);
2260 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2261 enum cl_req_type iot, struct cl_2queue *queue);
2262 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2263 enum cl_req_type iot, struct cl_2queue *queue,
2265 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2266 struct cl_page_list *queue, int from, int to,
2268 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2269 pgoff_t start, struct cl_read_ahead *ra);
2270 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2272 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2273 struct cl_page_list *queue);
2274 int cl_io_is_going (const struct lu_env *env);
2277 * True, iff \a io is an O_APPEND write(2).
2279 static inline int cl_io_is_append(const struct cl_io *io)
2281 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2284 static inline int cl_io_is_sync_write(const struct cl_io *io)
2286 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2289 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2291 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2295 * True, iff \a io is a truncate(2).
2297 static inline int cl_io_is_trunc(const struct cl_io *io)
2299 return io->ci_type == CIT_SETATTR &&
2300 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2303 struct cl_io *cl_io_top(struct cl_io *io);
2305 void cl_io_print(const struct lu_env *env, void *cookie,
2306 lu_printer_t printer, const struct cl_io *io);
2308 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2310 typeof(foo_io) __foo_io = (foo_io); \
2312 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2313 memset(&__foo_io->base + 1, 0, \
2314 (sizeof *__foo_io) - sizeof __foo_io->base); \
2319 /** \defgroup cl_page_list cl_page_list
2323 * Last page in the page list.
2325 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2327 LASSERT(plist->pl_nr > 0);
2328 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2331 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2333 LASSERT(plist->pl_nr > 0);
2334 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2338 * Iterate over pages in a page list.
2340 #define cl_page_list_for_each(page, list) \
2341 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2344 * Iterate over pages in a page list, taking possible removals into account.
2346 #define cl_page_list_for_each_safe(page, temp, list) \
2347 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2349 void cl_page_list_init (struct cl_page_list *plist);
2350 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2351 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2352 struct cl_page *page);
2353 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2354 struct cl_page *page);
2355 void cl_page_list_splice (struct cl_page_list *list,
2356 struct cl_page_list *head);
2357 void cl_page_list_del (const struct lu_env *env,
2358 struct cl_page_list *plist, struct cl_page *page);
2359 void cl_page_list_disown (const struct lu_env *env,
2360 struct cl_io *io, struct cl_page_list *plist);
2361 int cl_page_list_own (const struct lu_env *env,
2362 struct cl_io *io, struct cl_page_list *plist);
2363 void cl_page_list_assume (const struct lu_env *env,
2364 struct cl_io *io, struct cl_page_list *plist);
2365 void cl_page_list_discard(const struct lu_env *env,
2366 struct cl_io *io, struct cl_page_list *plist);
2367 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2369 void cl_2queue_init (struct cl_2queue *queue);
2370 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2371 void cl_2queue_disown (const struct lu_env *env,
2372 struct cl_io *io, struct cl_2queue *queue);
2373 void cl_2queue_assume (const struct lu_env *env,
2374 struct cl_io *io, struct cl_2queue *queue);
2375 void cl_2queue_discard (const struct lu_env *env,
2376 struct cl_io *io, struct cl_2queue *queue);
2377 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2378 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2380 /** @} cl_page_list */
2382 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2383 struct cl_req_attr *attr);
2385 /** \defgroup cl_sync_io cl_sync_io
2389 * Anchor for synchronous transfer. This is allocated on a stack by thread
2390 * doing synchronous transfer, and a pointer to this structure is set up in
2391 * every page submitted for transfer. Transfer completion routine updates
2392 * anchor and wakes up waiting thread when transfer is complete.
2395 /** number of pages yet to be transferred. */
2396 atomic_t csi_sync_nr;
2399 /** barrier of destroy this structure */
2400 atomic_t csi_barrier;
2401 /** completion to be signaled when transfer is complete. */
2402 wait_queue_head_t csi_waitq;
2403 /** callback to invoke when this IO is finished */
2404 void (*csi_end_io)(const struct lu_env *,
2405 struct cl_sync_io *);
2408 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2409 void (*end)(const struct lu_env *, struct cl_sync_io *));
2410 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2412 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2414 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2416 /** @} cl_sync_io */
2418 /** \defgroup cl_env cl_env
2420 * lu_env handling for a client.
2422 * lu_env is an environment within which lustre code executes. Its major part
2423 * is lu_context---a fast memory allocation mechanism that is used to conserve
2424 * precious kernel stack space. Originally lu_env was designed for a server,
2427 * - there is a (mostly) fixed number of threads, and
2429 * - call chains have no non-lustre portions inserted between lustre code.
2431 * On a client both these assumtpion fails, because every user thread can
2432 * potentially execute lustre code as part of a system call, and lustre calls
2433 * into VFS or MM that call back into lustre.
2435 * To deal with that, cl_env wrapper functions implement the following
2438 * - allocation and destruction of environment is amortized by caching no
2439 * longer used environments instead of destroying them;
2441 * - there is a notion of "current" environment, attached to the kernel
2442 * data structure representing current thread Top-level lustre code
2443 * allocates an environment and makes it current, then calls into
2444 * non-lustre code, that in turn calls lustre back. Low-level lustre
2445 * code thus called can fetch environment created by the top-level code
2446 * and reuse it, avoiding additional environment allocation.
2447 * Right now, three interfaces can attach the cl_env to running thread:
2450 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2452 * \see lu_env, lu_context, lu_context_key
2455 struct cl_env_nest {
2460 struct lu_env *cl_env_peek (int *refcheck);
2461 struct lu_env *cl_env_get (int *refcheck);
2462 struct lu_env *cl_env_alloc (int *refcheck, __u32 tags);
2463 struct lu_env *cl_env_nested_get (struct cl_env_nest *nest);
2464 void cl_env_put (struct lu_env *env, int *refcheck);
2465 void cl_env_nested_put (struct cl_env_nest *nest, struct lu_env *env);
2466 void *cl_env_reenter (void);
2467 void cl_env_reexit (void *cookie);
2468 void cl_env_implant (struct lu_env *env, int *refcheck);
2469 void cl_env_unplant (struct lu_env *env, int *refcheck);
2470 unsigned cl_env_cache_purge(unsigned nr);
2471 struct lu_env *cl_env_percpu_get (void);
2472 void cl_env_percpu_put (struct lu_env *env);
2479 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2480 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2482 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2483 struct lu_device_type *ldt,
2484 struct lu_device *next);
2487 int cl_global_init(void);
2488 void cl_global_fini(void);
2490 #endif /* _LINUX_CL_OBJECT_H */