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, 2017, 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 <linux/aio.h>
94 #include <libcfs/libcfs.h>
95 #include <lu_object.h>
96 #include <linux/atomic.h>
97 #include <linux/mutex.h>
98 #include <linux/radix-tree.h>
99 #include <linux/spinlock.h>
100 #include <linux/wait.h>
101 #include <linux/pagevec.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 */
174 /* Project identifier for quota purpose. */
179 * Fields in cl_attr that are being set.
194 * Sub-class of lu_object with methods common for objects on the client
197 * cl_object: represents a regular file system object, both a file and a
198 * stripe. cl_object is based on lu_object: it is identified by a fid,
199 * layered, cached, hashed, and lrued. Important distinction with the server
200 * side, where md_object and dt_object are used, is that cl_object "fans out"
201 * at the lov/sns level: depending on the file layout, single file is
202 * represented as a set of "sub-objects" (stripes). At the implementation
203 * level, struct lov_object contains an array of cl_objects. Each sub-object
204 * is a full-fledged cl_object, having its fid, living in the lru and hash
207 * This leads to the next important difference with the server side: on the
208 * client, it's quite usual to have objects with the different sequence of
209 * layers. For example, typical top-object is composed of the following
215 * whereas its sub-objects are composed of
220 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
221 * track of the object-subobject relationship.
223 * Sub-objects are not cached independently: when top-object is about to
224 * be discarded from the memory, all its sub-objects are torn-down and
227 * \see vvp_object, lov_object, lovsub_object, osc_object
231 struct lu_object co_lu;
232 /** per-object-layer operations */
233 const struct cl_object_operations *co_ops;
234 /** offset of page slice in cl_page buffer */
239 * Description of the client object configuration. This is used for the
240 * creation of a new client object that is identified by a more state than
243 struct cl_object_conf {
245 struct lu_object_conf coc_lu;
248 * Object layout. This is consumed by lov.
250 struct lu_buf coc_layout;
252 * Description of particular stripe location in the
253 * cluster. This is consumed by osc.
255 struct lov_oinfo *coc_oinfo;
258 * VFS inode. This is consumed by vvp.
260 struct inode *coc_inode;
262 * Layout lock handle.
264 struct ldlm_lock *coc_lock;
266 * Operation to handle layout, OBJECT_CONF_XYZ.
272 /** configure layout, set up a new stripe, must be called while
273 * holding layout lock. */
275 /** invalidate the current stripe configuration due to losing
277 OBJECT_CONF_INVALIDATE = 1,
278 /** wait for old layout to go away so that new layout can be
284 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
285 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
289 /** the buffer to return the layout in lov_mds_md format. */
290 struct lu_buf cl_buf;
291 /** size of layout in lov_mds_md format. */
293 /** size of DoM component if exists or zero otherwise */
294 u64 cl_dom_comp_size;
295 /** Layout generation. */
297 /** whether layout is a composite one */
298 bool cl_is_composite;
302 * Operations implemented for each cl object layer.
304 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
306 struct cl_object_operations {
308 * Initialize page slice for this layer. Called top-to-bottom through
309 * every object layer when a new cl_page is instantiated. Layer
310 * keeping private per-page data, or requiring its own page operations
311 * vector should allocate these data here, and attach then to the page
312 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
315 * \retval NULL success.
317 * \retval ERR_PTR(errno) failure code.
319 * \retval valid-pointer pointer to already existing referenced page
320 * to be used instead of newly created.
322 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
323 struct cl_page *page, pgoff_t index);
325 * Initialize lock slice for this layer. Called top-to-bottom through
326 * every object layer when a new cl_lock is instantiated. Layer
327 * keeping private per-lock data, or requiring its own lock operations
328 * vector should allocate these data here, and attach then to the lock
329 * by calling cl_lock_slice_add(). Mandatory.
331 int (*coo_lock_init)(const struct lu_env *env,
332 struct cl_object *obj, struct cl_lock *lock,
333 const struct cl_io *io);
335 * Initialize io state for a given layer.
337 * called top-to-bottom once per io existence to initialize io
338 * state. If layer wants to keep some state for this type of io, it
339 * has to embed struct cl_io_slice in lu_env::le_ses, and register
340 * slice with cl_io_slice_add(). It is guaranteed that all threads
341 * participating in this io share the same session.
343 int (*coo_io_init)(const struct lu_env *env,
344 struct cl_object *obj, struct cl_io *io);
346 * Fill portion of \a attr that this layer controls. This method is
347 * called top-to-bottom through all object layers.
349 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
351 * \return 0: to continue
352 * \return +ve: to stop iterating through layers (but 0 is returned
353 * from enclosing cl_object_attr_get())
354 * \return -ve: to signal error
356 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
357 struct cl_attr *attr);
361 * \a valid is a bitmask composed from enum #cl_attr_valid, and
362 * indicating what attributes are to be set.
364 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
366 * \return the same convention as for
367 * cl_object_operations::coo_attr_get() is used.
369 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
370 const struct cl_attr *attr, unsigned valid);
372 * Update object configuration. Called top-to-bottom to modify object
375 * XXX error conditions and handling.
377 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
378 const struct cl_object_conf *conf);
380 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
381 * object. Layers are supposed to fill parts of \a lvb that will be
382 * shipped to the glimpse originator as a glimpse result.
384 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
385 * \see osc_object_glimpse()
387 int (*coo_glimpse)(const struct lu_env *env,
388 const struct cl_object *obj, struct ost_lvb *lvb);
390 * Object prune method. Called when the layout is going to change on
391 * this object, therefore each layer has to clean up their cache,
392 * mainly pages and locks.
394 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
396 * Object getstripe method.
398 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
399 struct lov_user_md __user *lum, size_t size);
401 * Get FIEMAP mapping from the object.
403 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
404 struct ll_fiemap_info_key *fmkey,
405 struct fiemap *fiemap, size_t *buflen);
407 * Get layout and generation of the object.
409 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
410 struct cl_layout *layout);
412 * Get maximum size of the object.
414 loff_t (*coo_maxbytes)(struct cl_object *obj);
416 * Set request attributes.
418 void (*coo_req_attr_set)(const struct lu_env *env,
419 struct cl_object *obj,
420 struct cl_req_attr *attr);
424 * Extended header for client object.
426 struct cl_object_header {
427 /** Standard lu_object_header. cl_object::co_lu::lo_header points
429 struct lu_object_header coh_lu;
432 * Parent object. It is assumed that an object has a well-defined
433 * parent, but not a well-defined child (there may be multiple
434 * sub-objects, for the same top-object). cl_object_header::coh_parent
435 * field allows certain code to be written generically, without
436 * limiting possible cl_object layouts unduly.
438 struct cl_object_header *coh_parent;
440 * Protects consistency between cl_attr of parent object and
441 * attributes of sub-objects, that the former is calculated ("merged")
444 * \todo XXX this can be read/write lock if needed.
446 spinlock_t coh_attr_guard;
448 * Size of cl_page + page slices
450 unsigned short coh_page_bufsize;
452 * Number of objects above this one: 0 for a top-object, 1 for its
455 unsigned char coh_nesting;
459 * Helper macro: iterate over all layers of the object \a obj, assigning every
460 * layer top-to-bottom to \a slice.
462 #define cl_object_for_each(slice, obj) \
463 list_for_each_entry((slice), \
464 &(obj)->co_lu.lo_header->loh_layers,\
468 * Helper macro: iterate over all layers of the object \a obj, assigning every
469 * layer bottom-to-top to \a slice.
471 #define cl_object_for_each_reverse(slice, obj) \
472 list_for_each_entry_reverse((slice), \
473 &(obj)->co_lu.lo_header->loh_layers,\
478 #define CL_PAGE_EOF ((pgoff_t)~0ull)
480 /** \addtogroup cl_page cl_page
484 * Layered client page.
486 * cl_page: represents a portion of a file, cached in the memory. All pages
487 * of the given file are of the same size, and are kept in the radix tree
488 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
489 * of the top-level file object are first class cl_objects, they have their
490 * own radix trees of pages and hence page is implemented as a sequence of
491 * struct cl_pages's, linked into double-linked list through
492 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
493 * corresponding radix tree at the corresponding logical offset.
495 * cl_page is associated with VM page of the hosting environment (struct
496 * page in Linux kernel, for example), struct page. It is assumed, that this
497 * association is implemented by one of cl_page layers (top layer in the
498 * current design) that
500 * - intercepts per-VM-page call-backs made by the environment (e.g.,
503 * - translates state (page flag bits) and locking between lustre and
506 * The association between cl_page and struct page is immutable and
507 * established when cl_page is created.
509 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
510 * this io an exclusive access to this page w.r.t. other io attempts and
511 * various events changing page state (such as transfer completion, or
512 * eviction of the page from the memory). Note, that in general cl_io
513 * cannot be identified with a particular thread, and page ownership is not
514 * exactly equal to the current thread holding a lock on the page. Layer
515 * implementing association between cl_page and struct page has to implement
516 * ownership on top of available synchronization mechanisms.
518 * While lustre client maintains the notion of an page ownership by io,
519 * hosting MM/VM usually has its own page concurrency control
520 * mechanisms. For example, in Linux, page access is synchronized by the
521 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
522 * takes care to acquire and release such locks as necessary around the
523 * calls to the file system methods (->readpage(), ->prepare_write(),
524 * ->commit_write(), etc.). This leads to the situation when there are two
525 * different ways to own a page in the client:
527 * - client code explicitly and voluntary owns the page (cl_page_own());
529 * - VM locks a page and then calls the client, that has "to assume"
530 * the ownership from the VM (cl_page_assume()).
532 * Dual methods to release ownership are cl_page_disown() and
533 * cl_page_unassume().
535 * cl_page is reference counted (cl_page::cp_ref). When reference counter
536 * drops to 0, the page is returned to the cache, unless it is in
537 * cl_page_state::CPS_FREEING state, in which case it is immediately
540 * The general logic guaranteeing the absence of "existential races" for
541 * pages is the following:
543 * - there are fixed known ways for a thread to obtain a new reference
546 * - by doing a lookup in the cl_object radix tree, protected by the
549 * - by starting from VM-locked struct page and following some
550 * hosting environment method (e.g., following ->private pointer in
551 * the case of Linux kernel), see cl_vmpage_page();
553 * - when the page enters cl_page_state::CPS_FREEING state, all these
554 * ways are severed with the proper synchronization
555 * (cl_page_delete());
557 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
560 * - no new references to the page in cl_page_state::CPS_FREEING state
561 * are allowed (checked in cl_page_get()).
563 * Together this guarantees that when last reference to a
564 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
565 * page, as neither references to it can be acquired at that point, nor
568 * cl_page is a state machine. States are enumerated in enum
569 * cl_page_state. Possible state transitions are enumerated in
570 * cl_page_state_set(). State transition process (i.e., actual changing of
571 * cl_page::cp_state field) is protected by the lock on the underlying VM
574 * Linux Kernel implementation.
576 * Binding between cl_page and struct page (which is a typedef for
577 * struct page) is implemented in the vvp layer. cl_page is attached to the
578 * ->private pointer of the struct page, together with the setting of
579 * PG_private bit in page->flags, and acquiring additional reference on the
580 * struct page (much like struct buffer_head, or any similar file system
581 * private data structures).
583 * PG_locked lock is used to implement both ownership and transfer
584 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
585 * states. No additional references are acquired for the duration of the
588 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
589 * write-out is "protected" by the special PG_writeback bit.
593 * States of cl_page. cl_page.c assumes particular order here.
595 * The page state machine is rather crude, as it doesn't recognize finer page
596 * states like "dirty" or "up to date". This is because such states are not
597 * always well defined for the whole stack (see, for example, the
598 * implementation of the read-ahead, that hides page up-to-dateness to track
599 * cache hits accurately). Such sub-states are maintained by the layers that
600 * are interested in them.
604 * Page is in the cache, un-owned. Page leaves cached state in the
607 * - [cl_page_state::CPS_OWNED] io comes across the page and
610 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
611 * req-formation engine decides that it wants to include this page
612 * into an RPC being constructed, and yanks it from the cache;
614 * - [cl_page_state::CPS_FREEING] VM callback is executed to
615 * evict the page form the memory;
617 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
621 * Page is exclusively owned by some cl_io. Page may end up in this
622 * state as a result of
624 * - io creating new page and immediately owning it;
626 * - [cl_page_state::CPS_CACHED] io finding existing cached page
629 * - [cl_page_state::CPS_OWNED] io finding existing owned page
630 * and waiting for owner to release the page;
632 * Page leaves owned state in the following cases:
634 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
635 * the cache, doing nothing;
637 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
640 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
641 * transfer for this page;
643 * - [cl_page_state::CPS_FREEING] io decides to destroy this
644 * page (e.g., as part of truncate or extent lock cancellation).
646 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
650 * Page is being written out, as a part of a transfer. This state is
651 * entered when req-formation logic decided that it wants this page to
652 * be sent through the wire _now_. Specifically, it means that once
653 * this state is achieved, transfer completion handler (with either
654 * success or failure indication) is guaranteed to be executed against
655 * this page independently of any locks and any scheduling decisions
656 * made by the hosting environment (that effectively means that the
657 * page is never put into cl_page_state::CPS_PAGEOUT state "in
658 * advance". This property is mentioned, because it is important when
659 * reasoning about possible dead-locks in the system). The page can
660 * enter this state as a result of
662 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
663 * write-out of this page, or
665 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
666 * that it has enough dirty pages cached to issue a "good"
669 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
670 * is completed---it is moved into cl_page_state::CPS_CACHED state.
672 * Underlying VM page is locked for the duration of transfer.
674 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
678 * Page is being read in, as a part of a transfer. This is quite
679 * similar to the cl_page_state::CPS_PAGEOUT state, except that
680 * read-in is always "immediate"---there is no such thing a sudden
681 * construction of read request from cached, presumably not up to date,
684 * Underlying VM page is locked for the duration of transfer.
686 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
690 * Page is being destroyed. This state is entered when client decides
691 * that page has to be deleted from its host object, as, e.g., a part
694 * Once this state is reached, there is no way to escape it.
696 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
703 /** Host page, the page is from the host inode which the cl_page
707 /** Transient page, the transient cl_page is used to bind a cl_page
708 * to vmpage which is not belonging to the same object of cl_page.
709 * it is used in DirectIO and lockless IO. */
714 * Fields are protected by the lock on struct page, except for atomics and
717 * \invariant Data type invariants are in cl_page_invariant(). Basically:
718 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
719 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
720 * cl_page::cp_owner (when set).
723 /** Reference counter. */
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);
872 * Update file attributes when all we have is this page. Used for tiny
873 * writes to update attributes when we don't have a full cl_io.
875 void (*cpo_page_touch)(const struct lu_env *env,
876 const struct cl_page_slice *slice, size_t to);
882 * Called when page is truncated from the object. Optional.
884 * \see cl_page_discard()
885 * \see vvp_page_discard(), osc_page_discard()
887 void (*cpo_discard)(const struct lu_env *env,
888 const struct cl_page_slice *slice,
891 * Called when page is removed from the cache, and is about to being
892 * destroyed. Optional.
894 * \see cl_page_delete()
895 * \see vvp_page_delete(), osc_page_delete()
897 void (*cpo_delete)(const struct lu_env *env,
898 const struct cl_page_slice *slice);
899 /** Destructor. Frees resources and slice itself. */
900 void (*cpo_fini)(const struct lu_env *env,
901 struct cl_page_slice *slice,
902 struct pagevec *pvec);
904 * Optional debugging helper. Prints given page slice.
906 * \see cl_page_print()
908 int (*cpo_print)(const struct lu_env *env,
909 const struct cl_page_slice *slice,
910 void *cookie, lu_printer_t p);
919 * Request type dependent vector of operations.
921 * Transfer operations depend on transfer mode (cl_req_type). To avoid
922 * passing transfer mode to each and every of these methods, and to
923 * avoid branching on request type inside of the methods, separate
924 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
925 * provided. That is, method invocation usually looks like
927 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
931 * Called when a page is submitted for a transfer as a part of
934 * \return 0 : page is eligible for submission;
935 * \return -EALREADY : skip this page;
936 * \return -ve : error.
938 * \see cl_page_prep()
940 int (*cpo_prep)(const struct lu_env *env,
941 const struct cl_page_slice *slice,
944 * Completion handler. This is guaranteed to be eventually
945 * fired after cl_page_operations::cpo_prep() or
946 * cl_page_operations::cpo_make_ready() call.
948 * This method can be called in a non-blocking context. It is
949 * guaranteed however, that the page involved and its object
950 * are pinned in memory (and, hence, calling cl_page_put() is
953 * \see cl_page_completion()
955 void (*cpo_completion)(const struct lu_env *env,
956 const struct cl_page_slice *slice,
959 * Called when cached page is about to be added to the
960 * ptlrpc request as a part of req formation.
962 * \return 0 : proceed with this page;
963 * \return -EAGAIN : skip this page;
964 * \return -ve : error.
966 * \see cl_page_make_ready()
968 int (*cpo_make_ready)(const struct lu_env *env,
969 const struct cl_page_slice *slice);
972 * Tell transfer engine that only [to, from] part of a page should be
975 * This is used for immediate transfers.
977 * \todo XXX this is not very good interface. It would be much better
978 * if all transfer parameters were supplied as arguments to
979 * cl_io_operations::cio_submit() call, but it is not clear how to do
980 * this for page queues.
982 * \see cl_page_clip()
984 void (*cpo_clip)(const struct lu_env *env,
985 const struct cl_page_slice *slice,
988 * \pre the page was queued for transferring.
989 * \post page is removed from client's pending list, or -EBUSY
990 * is returned if it has already been in transferring.
992 * This is one of seldom page operation which is:
993 * 0. called from top level;
994 * 1. don't have vmpage locked;
995 * 2. every layer should synchronize execution of its ->cpo_cancel()
996 * with completion handlers. Osc uses client obd lock for this
997 * purpose. Based on there is no vvp_page_cancel and
998 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1000 * \see osc_page_cancel().
1002 int (*cpo_cancel)(const struct lu_env *env,
1003 const struct cl_page_slice *slice);
1005 * Write out a page by kernel. This is only called by ll_writepage
1008 * \see cl_page_flush()
1010 int (*cpo_flush)(const struct lu_env *env,
1011 const struct cl_page_slice *slice,
1017 * Helper macro, dumping detailed information about \a page into a log.
1019 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1021 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1022 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1023 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1024 CDEBUG(mask, format , ## __VA_ARGS__); \
1029 * Helper macro, dumping shorter information about \a page into a log.
1031 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1033 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1034 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1035 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1036 CDEBUG(mask, format , ## __VA_ARGS__); \
1040 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1042 LASSERT(page->cp_vmpage != NULL);
1043 return page->cp_vmpage;
1047 * Check if a cl_page is in use.
1049 * Client cache holds a refcount, this refcount will be dropped when
1050 * the page is taken out of cache, see vvp_page_delete().
1052 static inline bool __page_in_use(const struct cl_page *page, int refc)
1054 return (atomic_read(&page->cp_ref) > refc + 1);
1058 * Caller itself holds a refcount of cl_page.
1060 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1062 * Caller doesn't hold a refcount.
1064 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1068 /** \addtogroup cl_lock cl_lock
1072 * Extent locking on the client.
1076 * The locking model of the new client code is built around
1080 * data-type representing an extent lock on a regular file. cl_lock is a
1081 * layered object (much like cl_object and cl_page), it consists of a header
1082 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1083 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1085 * Typical cl_lock consists of one layer:
1087 * - lov_lock (lov specific data).
1089 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1090 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1094 * Each sub-lock is associated with a cl_object (representing stripe
1095 * sub-object or the file to which top-level cl_lock is associated to), and is
1096 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1097 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1098 * is different from cl_page, that doesn't fan out (there is usually exactly
1099 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1100 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1104 * cl_lock is a cacheless data container for the requirements of locks to
1105 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1108 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1109 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1111 * INTERFACE AND USAGE
1113 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1114 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1115 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1116 * consists of multiple sub cl_locks, each sub locks will be enqueued
1117 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1118 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1121 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1122 * method will be called for each layer to release the resource held by this
1123 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1124 * clo_enqueue time, is released.
1126 * LDLM lock can only be canceled if there is no cl_lock using it.
1128 * Overall process of the locking during IO operation is as following:
1130 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1131 * is called on each layer. Responsibility of this method is to add locks,
1132 * needed by a given layer into cl_io.ci_lockset.
1134 * - once locks for all layers were collected, they are sorted to avoid
1135 * dead-locks (cl_io_locks_sort()), and enqueued.
1137 * - when all locks are acquired, IO is performed;
1139 * - locks are released after IO is complete.
1141 * Striping introduces major additional complexity into locking. The
1142 * fundamental problem is that it is generally unsafe to actively use (hold)
1143 * two locks on the different OST servers at the same time, as this introduces
1144 * inter-server dependency and can lead to cascading evictions.
1146 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1147 * that no multi-stripe locks are taken (note that this design abandons POSIX
1148 * read/write semantics). Such pieces ideally can be executed concurrently. At
1149 * the same time, certain types of IO cannot be sub-divived, without
1150 * sacrificing correctness. This includes:
1152 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1155 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1157 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1158 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1159 * has to be held together with the usual lock on [offset, offset + count].
1161 * Interaction with DLM
1163 * In the expected setup, cl_lock is ultimately backed up by a collection of
1164 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1165 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1166 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1167 * description of interaction with DLM.
1173 struct cl_lock_descr {
1174 /** Object this lock is granted for. */
1175 struct cl_object *cld_obj;
1176 /** Index of the first page protected by this lock. */
1178 /** Index of the last page (inclusive) protected by this lock. */
1180 /** Group ID, for group lock */
1183 enum cl_lock_mode cld_mode;
1185 * flags to enqueue lock. A combination of bit-flags from
1186 * enum cl_enq_flags.
1188 __u32 cld_enq_flags;
1191 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1192 #define PDESCR(descr) \
1193 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1194 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1196 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1199 * Layered client lock.
1202 /** List of slices. Immutable after creation. */
1203 struct list_head cll_layers;
1204 /** lock attribute, extent, cl_object, etc. */
1205 struct cl_lock_descr cll_descr;
1209 * Per-layer part of cl_lock
1211 * \see lov_lock, osc_lock
1213 struct cl_lock_slice {
1214 struct cl_lock *cls_lock;
1215 /** Object slice corresponding to this lock slice. Immutable after
1217 struct cl_object *cls_obj;
1218 const struct cl_lock_operations *cls_ops;
1219 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1220 struct list_head cls_linkage;
1225 * \see lov_lock_ops, osc_lock_ops
1227 struct cl_lock_operations {
1230 * Attempts to enqueue the lock. Called top-to-bottom.
1232 * \retval 0 this layer has enqueued the lock successfully
1233 * \retval >0 this layer has enqueued the lock, but need to wait on
1234 * @anchor for resources
1235 * \retval -ve failure
1237 * \see lov_lock_enqueue(), osc_lock_enqueue()
1239 int (*clo_enqueue)(const struct lu_env *env,
1240 const struct cl_lock_slice *slice,
1241 struct cl_io *io, struct cl_sync_io *anchor);
1243 * Cancel a lock, release its DLM lock ref, while does not cancel the
1246 void (*clo_cancel)(const struct lu_env *env,
1247 const struct cl_lock_slice *slice);
1250 * Destructor. Frees resources and the slice.
1252 * \see lov_lock_fini(), osc_lock_fini()
1254 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1256 * Optional debugging helper. Prints given lock slice.
1258 int (*clo_print)(const struct lu_env *env,
1259 void *cookie, lu_printer_t p,
1260 const struct cl_lock_slice *slice);
1263 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1265 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1266 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1267 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1268 CDEBUG(mask, format , ## __VA_ARGS__); \
1272 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1276 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1282 /** \addtogroup cl_page_list cl_page_list
1283 * Page list used to perform collective operations on a group of pages.
1285 * Pages are added to the list one by one. cl_page_list acquires a reference
1286 * for every page in it. Page list is used to perform collective operations on
1289 * - submit pages for an immediate transfer,
1291 * - own pages on behalf of certain io (waiting for each page in turn),
1295 * When list is finalized, it releases references on all pages it still has.
1297 * \todo XXX concurrency control.
1301 struct cl_page_list {
1303 struct list_head pl_pages;
1304 struct task_struct *pl_owner;
1308 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1309 * contains an incoming page list and an outgoing page list.
1312 struct cl_page_list c2_qin;
1313 struct cl_page_list c2_qout;
1316 /** @} cl_page_list */
1318 /** \addtogroup cl_io cl_io
1323 * cl_io represents a high level I/O activity like
1324 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1327 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1328 * important distinction. We want to minimize number of calls to the allocator
1329 * in the fast path, e.g., in the case of read(2) when everything is cached:
1330 * client already owns the lock over region being read, and data are cached
1331 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1332 * per-layer io state is stored in the session, associated with the io, see
1333 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1334 * by using free-lists, see cl_env_get().
1336 * There is a small predefined number of possible io types, enumerated in enum
1339 * cl_io is a state machine, that can be advanced concurrently by the multiple
1340 * threads. It is up to these threads to control the concurrency and,
1341 * specifically, to detect when io is done, and its state can be safely
1344 * For read/write io overall execution plan is as following:
1346 * (0) initialize io state through all layers;
1348 * (1) loop: prepare chunk of work to do
1350 * (2) call all layers to collect locks they need to process current chunk
1352 * (3) sort all locks to avoid dead-locks, and acquire them
1354 * (4) process the chunk: call per-page methods
1355 * cl_io_operations::cio_prepare_write(),
1356 * cl_io_operations::cio_commit_write() for write)
1362 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1363 * address allocation efficiency issues mentioned above), and returns with the
1364 * special error condition from per-page method when current sub-io has to
1365 * block. This causes io loop to be repeated, and lov switches to the next
1366 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1371 /** read system call */
1373 /** write system call */
1375 /** truncate, utime system calls */
1377 /** get data version */
1380 * page fault handling
1384 * fsync system call handling
1385 * To write out a range of file
1389 * glimpse. An io context to acquire glimpse lock.
1393 * Miscellaneous io. This is used for occasional io activity that
1394 * doesn't fit into other types. Currently this is used for:
1396 * - cancellation of an extent lock. This io exists as a context
1397 * to write dirty pages from under the lock being canceled back
1400 * - VM induced page write-out. An io context for writing page out
1401 * for memory cleansing;
1403 * - grouplock. An io context to acquire group lock.
1405 * CIT_MISC io is used simply as a context in which locks and pages
1406 * are manipulated. Such io has no internal "process", that is,
1407 * cl_io_loop() is never called for it.
1412 * To give advice about access of a file
1419 * States of cl_io state machine
1422 /** Not initialized. */
1426 /** IO iteration started. */
1430 /** Actual IO is in progress. */
1432 /** IO for the current iteration finished. */
1434 /** Locks released. */
1436 /** Iteration completed. */
1438 /** cl_io finalized. */
1443 * IO state private for a layer.
1445 * This is usually embedded into layer session data, rather than allocated
1448 * \see vvp_io, lov_io, osc_io
1450 struct cl_io_slice {
1451 struct cl_io *cis_io;
1452 /** corresponding object slice. Immutable after creation. */
1453 struct cl_object *cis_obj;
1454 /** io operations. Immutable after creation. */
1455 const struct cl_io_operations *cis_iop;
1457 * linkage into a list of all slices for a given cl_io, hanging off
1458 * cl_io::ci_layers. Immutable after creation.
1460 struct list_head cis_linkage;
1463 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1466 struct cl_read_ahead {
1467 /* Maximum page index the readahead window will end.
1468 * This is determined DLM lock coverage, RPC and stripe boundary.
1469 * cra_end is included. */
1471 /* optimal RPC size for this read, by pages */
1472 unsigned long cra_rpc_size;
1473 /* Release callback. If readahead holds resources underneath, this
1474 * function should be called to release it. */
1475 void (*cra_release)(const struct lu_env *env, void *cbdata);
1476 /* Callback data for cra_release routine */
1480 static inline void cl_read_ahead_release(const struct lu_env *env,
1481 struct cl_read_ahead *ra)
1483 if (ra->cra_release != NULL)
1484 ra->cra_release(env, ra->cra_cbdata);
1485 memset(ra, 0, sizeof(*ra));
1490 * Per-layer io operations.
1491 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1493 struct cl_io_operations {
1495 * Vector of io state transition methods for every io type.
1497 * \see cl_page_operations::io
1501 * Prepare io iteration at a given layer.
1503 * Called top-to-bottom at the beginning of each iteration of
1504 * "io loop" (if it makes sense for this type of io). Here
1505 * layer selects what work it will do during this iteration.
1507 * \see cl_io_operations::cio_iter_fini()
1509 int (*cio_iter_init) (const struct lu_env *env,
1510 const struct cl_io_slice *slice);
1512 * Finalize io iteration.
1514 * Called bottom-to-top at the end of each iteration of "io
1515 * loop". Here layers can decide whether IO has to be
1518 * \see cl_io_operations::cio_iter_init()
1520 void (*cio_iter_fini) (const struct lu_env *env,
1521 const struct cl_io_slice *slice);
1523 * Collect locks for the current iteration of io.
1525 * Called top-to-bottom to collect all locks necessary for
1526 * this iteration. This methods shouldn't actually enqueue
1527 * anything, instead it should post a lock through
1528 * cl_io_lock_add(). Once all locks are collected, they are
1529 * sorted and enqueued in the proper order.
1531 int (*cio_lock) (const struct lu_env *env,
1532 const struct cl_io_slice *slice);
1534 * Finalize unlocking.
1536 * Called bottom-to-top to finish layer specific unlocking
1537 * functionality, after generic code released all locks
1538 * acquired by cl_io_operations::cio_lock().
1540 void (*cio_unlock)(const struct lu_env *env,
1541 const struct cl_io_slice *slice);
1543 * Start io iteration.
1545 * Once all locks are acquired, called top-to-bottom to
1546 * commence actual IO. In the current implementation,
1547 * top-level vvp_io_{read,write}_start() does all the work
1548 * synchronously by calling generic_file_*(), so other layers
1549 * are called when everything is done.
1551 int (*cio_start)(const struct lu_env *env,
1552 const struct cl_io_slice *slice);
1554 * Called top-to-bottom at the end of io loop. Here layer
1555 * might wait for an unfinished asynchronous io.
1557 void (*cio_end) (const struct lu_env *env,
1558 const struct cl_io_slice *slice);
1560 * Called bottom-to-top to notify layers that read/write IO
1561 * iteration finished, with \a nob bytes transferred.
1563 void (*cio_advance)(const struct lu_env *env,
1564 const struct cl_io_slice *slice,
1567 * Called once per io, bottom-to-top to release io resources.
1569 void (*cio_fini) (const struct lu_env *env,
1570 const struct cl_io_slice *slice);
1574 * Submit pages from \a queue->c2_qin for IO, and move
1575 * successfully submitted pages into \a queue->c2_qout. Return
1576 * non-zero if failed to submit even the single page. If
1577 * submission failed after some pages were moved into \a
1578 * queue->c2_qout, completion callback with non-zero ioret is
1581 int (*cio_submit)(const struct lu_env *env,
1582 const struct cl_io_slice *slice,
1583 enum cl_req_type crt,
1584 struct cl_2queue *queue);
1586 * Queue async page for write.
1587 * The difference between cio_submit and cio_queue is that
1588 * cio_submit is for urgent request.
1590 int (*cio_commit_async)(const struct lu_env *env,
1591 const struct cl_io_slice *slice,
1592 struct cl_page_list *queue, int from, int to,
1595 * Decide maximum read ahead extent
1597 * \pre io->ci_type == CIT_READ
1599 int (*cio_read_ahead)(const struct lu_env *env,
1600 const struct cl_io_slice *slice,
1601 pgoff_t start, struct cl_read_ahead *ra);
1603 * Optional debugging helper. Print given io slice.
1605 int (*cio_print)(const struct lu_env *env, void *cookie,
1606 lu_printer_t p, const struct cl_io_slice *slice);
1610 * Flags to lock enqueue procedure.
1615 * instruct server to not block, if conflicting lock is found. Instead
1616 * -EWOULDBLOCK is returned immediately.
1618 CEF_NONBLOCK = 0x00000001,
1620 * Tell lower layers this is a glimpse request, translated to
1621 * LDLM_FL_HAS_INTENT at LDLM layer.
1623 * Also, because glimpse locks never block other locks, we count this
1624 * as automatically compatible with other osc locks.
1625 * (see osc_lock_compatible)
1627 CEF_GLIMPSE = 0x00000002,
1629 * tell the server to instruct (though a flag in the blocking ast) an
1630 * owner of the conflicting lock, that it can drop dirty pages
1631 * protected by this lock, without sending them to the server.
1633 CEF_DISCARD_DATA = 0x00000004,
1635 * tell the sub layers that it must be a `real' lock. This is used for
1636 * mmapped-buffer locks, glimpse locks, manually requested locks
1637 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1640 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1642 CEF_MUST = 0x00000008,
1644 * tell the sub layers that never request a `real' lock. This flag is
1645 * not used currently.
1647 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1648 * conversion policy: ci_lockreq describes generic information of lock
1649 * requirement for this IO, especially for locks which belong to the
1650 * object doing IO; however, lock itself may have precise requirements
1651 * that are described by the enqueue flags.
1653 CEF_NEVER = 0x00000010,
1655 * tell the dlm layer this is a speculative lock request
1656 * speculative lock requests are locks which are not requested as part
1657 * of an I/O operation. Instead, they are requested because we expect
1658 * to use them in the future. They are requested asynchronously at the
1661 * Currently used for asynchronous glimpse locks and manually requested
1662 * locks (LU_LADVISE_LOCKAHEAD).
1664 CEF_SPECULATIVE = 0x00000020,
1666 * enqueue a lock to test DLM lock existence.
1668 CEF_PEEK = 0x00000040,
1670 * Lock match only. Used by group lock in I/O as group lock
1671 * is known to exist.
1673 CEF_LOCK_MATCH = 0x00000080,
1675 * tell the DLM layer to lock only the requested range
1677 CEF_LOCK_NO_EXPAND = 0x00000100,
1679 * mask of enq_flags.
1681 CEF_MASK = 0x000001ff,
1685 * Link between lock and io. Intermediate structure is needed, because the
1686 * same lock can be part of multiple io's simultaneously.
1688 struct cl_io_lock_link {
1689 /** linkage into one of cl_lockset lists. */
1690 struct list_head cill_linkage;
1691 struct cl_lock cill_lock;
1692 /** optional destructor */
1693 void (*cill_fini)(const struct lu_env *env,
1694 struct cl_io_lock_link *link);
1696 #define cill_descr cill_lock.cll_descr
1699 * Lock-set represents a collection of locks, that io needs at a
1700 * time. Generally speaking, client tries to avoid holding multiple locks when
1703 * - holding extent locks over multiple ost's introduces the danger of
1704 * "cascading timeouts";
1706 * - holding multiple locks over the same ost is still dead-lock prone,
1707 * see comment in osc_lock_enqueue(),
1709 * but there are certain situations where this is unavoidable:
1711 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1713 * - truncate has to take [new-size, EOF] lock for correctness;
1715 * - SNS has to take locks across full stripe for correctness;
1717 * - in the case when user level buffer, supplied to {read,write}(file0),
1718 * is a part of a memory mapped lustre file, client has to take a dlm
1719 * locks on file0, and all files that back up the buffer (or a part of
1720 * the buffer, that is being processed in the current chunk, in any
1721 * case, there are situations where at least 2 locks are necessary).
1723 * In such cases we at least try to take locks in the same consistent
1724 * order. To this end, all locks are first collected, then sorted, and then
1728 /** locks to be acquired. */
1729 struct list_head cls_todo;
1730 /** locks acquired. */
1731 struct list_head cls_done;
1735 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1736 * but 'req' is always to be thought as 'request' :-)
1738 enum cl_io_lock_dmd {
1739 /** Always lock data (e.g., O_APPEND). */
1741 /** Layers are free to decide between local and global locking. */
1743 /** Never lock: there is no cache (e.g., liblustre). */
1747 enum cl_fsync_mode {
1748 /** start writeback, do not wait for them to finish */
1750 /** start writeback and wait for them to finish */
1752 /** discard all of dirty pages in a specific file range */
1753 CL_FSYNC_DISCARD = 2,
1754 /** start writeback and make sure they have reached storage before
1755 * return. OST_SYNC RPC must be issued and finished */
1759 struct cl_io_rw_common {
1768 * cl_io is shared by all threads participating in this IO (in current
1769 * implementation only one thread advances IO, but parallel IO design and
1770 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1771 * is up to these threads to serialize their activities, including updates to
1772 * mutable cl_io fields.
1775 /** type of this IO. Immutable after creation. */
1776 enum cl_io_type ci_type;
1777 /** current state of cl_io state machine. */
1778 enum cl_io_state ci_state;
1779 /** main object this io is against. Immutable after creation. */
1780 struct cl_object *ci_obj;
1782 * Upper layer io, of which this io is a part of. Immutable after
1785 struct cl_io *ci_parent;
1786 /** List of slices. Immutable after creation. */
1787 struct list_head ci_layers;
1788 /** list of locks (to be) acquired by this io. */
1789 struct cl_lockset ci_lockset;
1790 /** lock requirements, this is just a help info for sublayers. */
1791 enum cl_io_lock_dmd ci_lockreq;
1792 /** layout version when this IO occurs */
1793 __u32 ci_layout_version;
1796 struct cl_io_rw_common rd;
1799 struct cl_io_rw_common wr;
1803 struct cl_io_rw_common ci_rw;
1804 struct cl_setattr_io {
1805 struct ost_lvb sa_attr;
1806 unsigned int sa_attr_flags;
1807 unsigned int sa_avalid; /* ATTR_* */
1808 unsigned int sa_xvalid; /* OP_XVALID */
1809 int sa_stripe_index;
1810 struct ost_layout sa_layout;
1811 const struct lu_fid *sa_parent_fid;
1813 struct cl_data_version_io {
1814 u64 dv_data_version;
1815 u32 dv_layout_version;
1818 struct cl_fault_io {
1819 /** page index within file. */
1821 /** bytes valid byte on a faulted page. */
1823 /** writable page? for nopage() only */
1825 /** page of an executable? */
1827 /** page_mkwrite() */
1829 /** resulting page */
1830 struct cl_page *ft_page;
1832 struct cl_fsync_io {
1835 /** file system level fid */
1836 struct lu_fid *fi_fid;
1837 enum cl_fsync_mode fi_mode;
1838 /* how many pages were written/discarded */
1839 unsigned int fi_nr_written;
1841 struct cl_ladvise_io {
1844 /** file system level fid */
1845 struct lu_fid *li_fid;
1846 enum lu_ladvise_type li_advice;
1850 struct cl_2queue ci_queue;
1853 unsigned int ci_continue:1,
1855 * This io has held grouplock, to inform sublayers that
1856 * don't do lockless i/o.
1860 * The whole IO need to be restarted because layout has been changed
1864 * to not refresh layout - the IO issuer knows that the layout won't
1865 * change(page operations, layout change causes all page to be
1866 * discarded), or it doesn't matter if it changes(sync).
1870 * Need MDS intervention to complete a write.
1871 * Write intent is required for the following cases:
1872 * 1. component being written is not initialized, or
1873 * 2. the mirrored files are NOT in WRITE_PENDING state.
1875 ci_need_write_intent:1,
1877 * Check if layout changed after the IO finishes. Mainly for HSM
1878 * requirement. If IO occurs to openning files, it doesn't need to
1879 * verify layout because HSM won't release openning files.
1880 * Right now, only two opertaions need to verify layout: glimpse
1885 * file is released, restore has to to be triggered by vvp layer
1887 ci_restore_needed:1,
1892 /* Tell sublayers not to expand LDLM locks requested for this IO */
1893 ci_lock_no_expand:1,
1895 * Set if non-delay RPC should be used for this IO.
1897 * If this file has multiple mirrors, and if the OSTs of the current
1898 * mirror is inaccessible, non-delay RPC would error out quickly so
1899 * that the upper layer can try to access the next mirror.
1903 * How many times the read has retried before this one.
1904 * Set by the top level and consumed by the LOV.
1906 unsigned ci_ndelay_tried;
1908 * Designated mirror index for this I/O.
1910 unsigned ci_designated_mirror;
1912 * Number of pages owned by this IO. For invariant checking.
1914 unsigned ci_owned_nr;
1916 * Range of write intent. Valid if ci_need_write_intent is set.
1918 struct lu_extent ci_write_intent;
1924 * Per-transfer attributes.
1926 struct cl_req_attr {
1927 enum cl_req_type cra_type;
1929 struct cl_page *cra_page;
1930 /** Generic attributes for the server consumption. */
1931 struct obdo *cra_oa;
1933 char cra_jobid[LUSTRE_JOBID_SIZE];
1936 enum cache_stats_item {
1937 /** how many cache lookups were performed */
1939 /** how many times cache lookup resulted in a hit */
1941 /** how many entities are in the cache right now */
1943 /** how many entities in the cache are actively used (and cannot be
1944 * evicted) right now */
1946 /** how many entities were created at all */
1951 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
1954 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
1956 struct cache_stats {
1957 const char *cs_name;
1958 atomic_t cs_stats[CS_NR];
1961 /** These are not exported so far */
1962 void cache_stats_init (struct cache_stats *cs, const char *name);
1965 * Client-side site. This represents particular client stack. "Global"
1966 * variables should (directly or indirectly) be added here to allow multiple
1967 * clients to co-exist in the single address space.
1970 struct lu_site cs_lu;
1972 * Statistical counters. Atomics do not scale, something better like
1973 * per-cpu counters is needed.
1975 * These are exported as /proc/fs/lustre/llite/.../site
1977 * When interpreting keep in mind that both sub-locks (and sub-pages)
1978 * and top-locks (and top-pages) are accounted here.
1980 struct cache_stats cs_pages;
1981 atomic_t cs_pages_state[CPS_NR];
1984 int cl_site_init(struct cl_site *s, struct cl_device *top);
1985 void cl_site_fini(struct cl_site *s);
1986 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
1989 * Output client site statistical counters into a buffer. Suitable for
1990 * ll_rd_*()-style functions.
1992 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
1997 * Type conversion and accessory functions.
2001 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2003 return container_of(site, struct cl_site, cs_lu);
2006 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2008 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2009 return container_of0(d, struct cl_device, cd_lu_dev);
2012 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2014 return &d->cd_lu_dev;
2017 static inline struct cl_object *lu2cl(const struct lu_object *o)
2019 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2020 return container_of0(o, struct cl_object, co_lu);
2023 static inline const struct cl_object_conf *
2024 lu2cl_conf(const struct lu_object_conf *conf)
2026 return container_of0(conf, struct cl_object_conf, coc_lu);
2029 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2031 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2034 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2036 return container_of0(h, struct cl_object_header, coh_lu);
2039 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2041 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2045 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2047 return luh2coh(obj->co_lu.lo_header);
2050 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2052 return lu_device_init(&d->cd_lu_dev, t);
2055 static inline void cl_device_fini(struct cl_device *d)
2057 lu_device_fini(&d->cd_lu_dev);
2060 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2061 struct cl_object *obj, pgoff_t index,
2062 const struct cl_page_operations *ops);
2063 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2064 struct cl_object *obj,
2065 const struct cl_lock_operations *ops);
2066 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2067 struct cl_object *obj, const struct cl_io_operations *ops);
2070 /** \defgroup cl_object cl_object
2072 struct cl_object *cl_object_top (struct cl_object *o);
2073 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2074 const struct lu_fid *fid,
2075 const struct cl_object_conf *c);
2077 int cl_object_header_init(struct cl_object_header *h);
2078 void cl_object_header_fini(struct cl_object_header *h);
2079 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2080 void cl_object_get (struct cl_object *o);
2081 void cl_object_attr_lock (struct cl_object *o);
2082 void cl_object_attr_unlock(struct cl_object *o);
2083 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2084 struct cl_attr *attr);
2085 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2086 const struct cl_attr *attr, unsigned valid);
2087 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2088 struct ost_lvb *lvb);
2089 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2090 const struct cl_object_conf *conf);
2091 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2092 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2093 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2094 struct lov_user_md __user *lum, size_t size);
2095 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2096 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2098 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2099 struct cl_layout *cl);
2100 loff_t cl_object_maxbytes(struct cl_object *obj);
2103 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2105 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2107 return cl_object_header(o0) == cl_object_header(o1);
2110 static inline void cl_object_page_init(struct cl_object *clob, int size)
2112 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2113 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2114 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2117 static inline void *cl_object_page_slice(struct cl_object *clob,
2118 struct cl_page *page)
2120 return (void *)((char *)page + clob->co_slice_off);
2124 * Return refcount of cl_object.
2126 static inline int cl_object_refc(struct cl_object *clob)
2128 struct lu_object_header *header = clob->co_lu.lo_header;
2129 return atomic_read(&header->loh_ref);
2134 /** \defgroup cl_page cl_page
2142 /* callback of cl_page_gang_lookup() */
2144 struct cl_page *cl_page_find (const struct lu_env *env,
2145 struct cl_object *obj,
2146 pgoff_t idx, struct page *vmpage,
2147 enum cl_page_type type);
2148 struct cl_page *cl_page_alloc (const struct lu_env *env,
2149 struct cl_object *o, pgoff_t ind,
2150 struct page *vmpage,
2151 enum cl_page_type type);
2152 void cl_page_get (struct cl_page *page);
2153 void cl_page_put (const struct lu_env *env,
2154 struct cl_page *page);
2155 void cl_pagevec_put (const struct lu_env *env,
2156 struct cl_page *page,
2157 struct pagevec *pvec);
2158 void cl_page_print (const struct lu_env *env, void *cookie,
2159 lu_printer_t printer,
2160 const struct cl_page *pg);
2161 void cl_page_header_print(const struct lu_env *env, void *cookie,
2162 lu_printer_t printer,
2163 const struct cl_page *pg);
2164 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2165 struct cl_page *cl_page_top (struct cl_page *page);
2167 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2168 const struct lu_device_type *dtype);
2173 * Functions dealing with the ownership of page by io.
2177 int cl_page_own (const struct lu_env *env,
2178 struct cl_io *io, struct cl_page *page);
2179 int cl_page_own_try (const struct lu_env *env,
2180 struct cl_io *io, struct cl_page *page);
2181 void cl_page_assume (const struct lu_env *env,
2182 struct cl_io *io, struct cl_page *page);
2183 void cl_page_unassume (const struct lu_env *env,
2184 struct cl_io *io, struct cl_page *pg);
2185 void cl_page_disown (const struct lu_env *env,
2186 struct cl_io *io, struct cl_page *page);
2187 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2194 * Functions dealing with the preparation of a page for a transfer, and
2195 * tracking transfer state.
2198 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2199 struct cl_page *pg, enum cl_req_type crt);
2200 void cl_page_completion (const struct lu_env *env,
2201 struct cl_page *pg, enum cl_req_type crt, int ioret);
2202 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2203 enum cl_req_type crt);
2204 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2205 struct cl_page *pg, enum cl_req_type crt);
2206 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2208 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2209 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2210 struct cl_page *pg);
2216 * \name helper routines
2217 * Functions to discard, delete and export a cl_page.
2220 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2221 struct cl_page *pg);
2222 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2223 int cl_page_is_vmlocked(const struct lu_env *env,
2224 const struct cl_page *pg);
2225 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
2227 void cl_page_export(const struct lu_env *env,
2228 struct cl_page *pg, int uptodate);
2229 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2230 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2231 size_t cl_page_size(const struct cl_object *obj);
2233 void cl_lock_print(const struct lu_env *env, void *cookie,
2234 lu_printer_t printer, const struct cl_lock *lock);
2235 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2236 lu_printer_t printer,
2237 const struct cl_lock_descr *descr);
2241 * Data structure managing a client's cached pages. A count of
2242 * "unstable" pages is maintained, and an LRU of clean pages is
2243 * maintained. "unstable" pages are pages pinned by the ptlrpc
2244 * layer for recovery purposes.
2246 struct cl_client_cache {
2248 * # of client cache refcount
2249 * # of users (OSCs) + 2 (held by llite and lov)
2253 * # of threads are doing shrinking
2255 unsigned int ccc_lru_shrinkers;
2257 * # of LRU entries available
2259 atomic_long_t ccc_lru_left;
2261 * List of entities(OSCs) for this LRU cache
2263 struct list_head ccc_lru;
2265 * Max # of LRU entries
2267 unsigned long ccc_lru_max;
2269 * Lock to protect ccc_lru list
2271 spinlock_t ccc_lru_lock;
2273 * Set if unstable check is enabled
2275 unsigned int ccc_unstable_check:1;
2277 * # of unstable pages for this mount point
2279 atomic_long_t ccc_unstable_nr;
2281 * Waitq for awaiting unstable pages to reach zero.
2282 * Used at umounting time and signaled on BRW commit
2284 wait_queue_head_t ccc_unstable_waitq;
2287 * cl_cache functions
2289 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2290 void cl_cache_incref(struct cl_client_cache *cache);
2291 void cl_cache_decref(struct cl_client_cache *cache);
2295 /** \defgroup cl_lock cl_lock
2297 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2298 struct cl_lock *lock);
2299 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2300 const struct cl_io *io);
2301 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2302 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2303 const struct lu_device_type *dtype);
2304 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2306 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2307 struct cl_lock *lock, struct cl_sync_io *anchor);
2308 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2312 /** \defgroup cl_io cl_io
2315 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2316 enum cl_io_type iot, struct cl_object *obj);
2317 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2318 enum cl_io_type iot, struct cl_object *obj);
2319 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2320 enum cl_io_type iot, loff_t pos, size_t count);
2321 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2323 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2324 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2325 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2326 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2327 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2328 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2329 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2330 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2331 struct cl_io_lock_link *link);
2332 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2333 struct cl_lock_descr *descr);
2334 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2335 enum cl_req_type iot, struct cl_2queue *queue);
2336 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2337 enum cl_req_type iot, struct cl_2queue *queue,
2339 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2340 struct cl_page_list *queue, int from, int to,
2342 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2343 pgoff_t start, struct cl_read_ahead *ra);
2344 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2346 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2347 struct cl_page_list *queue);
2350 * True, iff \a io is an O_APPEND write(2).
2352 static inline int cl_io_is_append(const struct cl_io *io)
2354 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2357 static inline int cl_io_is_sync_write(const struct cl_io *io)
2359 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2362 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2364 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2368 * True, iff \a io is a truncate(2).
2370 static inline int cl_io_is_trunc(const struct cl_io *io)
2372 return io->ci_type == CIT_SETATTR &&
2373 (io->u.ci_setattr.sa_avalid & ATTR_SIZE);
2376 struct cl_io *cl_io_top(struct cl_io *io);
2378 void cl_io_print(const struct lu_env *env, void *cookie,
2379 lu_printer_t printer, const struct cl_io *io);
2381 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2383 typeof(foo_io) __foo_io = (foo_io); \
2385 memset(&__foo_io->base, 0, \
2386 sizeof(*__foo_io) - offsetof(typeof(*__foo_io), base)); \
2391 /** \defgroup cl_page_list cl_page_list
2395 * Last page in the page list.
2397 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2399 LASSERT(plist->pl_nr > 0);
2400 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2403 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2405 LASSERT(plist->pl_nr > 0);
2406 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2410 * Iterate over pages in a page list.
2412 #define cl_page_list_for_each(page, list) \
2413 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2416 * Iterate over pages in a page list, taking possible removals into account.
2418 #define cl_page_list_for_each_safe(page, temp, list) \
2419 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2421 void cl_page_list_init (struct cl_page_list *plist);
2422 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2423 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2424 struct cl_page *page);
2425 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2426 struct cl_page *page);
2427 void cl_page_list_splice (struct cl_page_list *list,
2428 struct cl_page_list *head);
2429 void cl_page_list_del (const struct lu_env *env,
2430 struct cl_page_list *plist, struct cl_page *page);
2431 void cl_page_list_disown (const struct lu_env *env,
2432 struct cl_io *io, struct cl_page_list *plist);
2433 void cl_page_list_assume (const struct lu_env *env,
2434 struct cl_io *io, struct cl_page_list *plist);
2435 void cl_page_list_discard(const struct lu_env *env,
2436 struct cl_io *io, struct cl_page_list *plist);
2437 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2439 void cl_2queue_init (struct cl_2queue *queue);
2440 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2441 void cl_2queue_disown (const struct lu_env *env,
2442 struct cl_io *io, struct cl_2queue *queue);
2443 void cl_2queue_assume (const struct lu_env *env,
2444 struct cl_io *io, struct cl_2queue *queue);
2445 void cl_2queue_discard (const struct lu_env *env,
2446 struct cl_io *io, struct cl_2queue *queue);
2447 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2448 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2450 /** @} cl_page_list */
2452 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2453 struct cl_req_attr *attr);
2455 /** \defgroup cl_sync_io cl_sync_io
2459 * Anchor for synchronous transfer. This is allocated on a stack by thread
2460 * doing synchronous transfer, and a pointer to this structure is set up in
2461 * every page submitted for transfer. Transfer completion routine updates
2462 * anchor and wakes up waiting thread when transfer is complete.
2465 /** number of pages yet to be transferred. */
2466 atomic_t csi_sync_nr;
2469 /** barrier of destroy this structure */
2470 atomic_t csi_barrier;
2471 /** completion to be signaled when transfer is complete. */
2472 wait_queue_head_t csi_waitq;
2473 /** callback to invoke when this IO is finished */
2474 void (*csi_end_io)(const struct lu_env *,
2475 struct cl_sync_io *);
2478 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2479 void (*end)(const struct lu_env *, struct cl_sync_io *));
2480 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2482 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2484 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2486 /** @} cl_sync_io */
2488 /** \defgroup cl_env cl_env
2490 * lu_env handling for a client.
2492 * lu_env is an environment within which lustre code executes. Its major part
2493 * is lu_context---a fast memory allocation mechanism that is used to conserve
2494 * precious kernel stack space. Originally lu_env was designed for a server,
2497 * - there is a (mostly) fixed number of threads, and
2499 * - call chains have no non-lustre portions inserted between lustre code.
2501 * On a client both these assumtpion fails, because every user thread can
2502 * potentially execute lustre code as part of a system call, and lustre calls
2503 * into VFS or MM that call back into lustre.
2505 * To deal with that, cl_env wrapper functions implement the following
2508 * - allocation and destruction of environment is amortized by caching no
2509 * longer used environments instead of destroying them;
2511 * \see lu_env, lu_context, lu_context_key
2514 struct lu_env *cl_env_get(__u16 *refcheck);
2515 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2516 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2517 unsigned cl_env_cache_purge(unsigned nr);
2518 struct lu_env *cl_env_percpu_get(void);
2519 void cl_env_percpu_put(struct lu_env *env);
2526 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2527 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2529 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2530 struct lu_device_type *ldt,
2531 struct lu_device *next);
2534 int cl_global_init(void);
2535 void cl_global_fini(void);
2537 #endif /* _LINUX_CL_OBJECT_H */