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, 2015, 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).
62 * - cl_req represents a collection of pages for a transfer. cl_req is
63 * constructed by req-forming engine that tries to saturate
64 * transport with large and continuous transfers.
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 "lu_object.h"
96 #include "lustre_compat.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>
106 struct cl_device_operations;
109 struct cl_object_page_operations;
110 struct cl_object_lock_operations;
113 struct cl_page_slice;
115 struct cl_lock_slice;
117 struct cl_lock_operations;
118 struct cl_page_operations;
127 * Operations for each data device in the client stack.
129 * \see vvp_cl_ops, lov_cl_ops, lovsub_cl_ops, osc_cl_ops
131 struct cl_device_operations {
133 * Initialize cl_req. This method is called top-to-bottom on all
134 * devices in the stack to get them a chance to allocate layer-private
135 * data, and to attach them to the cl_req by calling
136 * cl_req_slice_add().
138 * \see osc_req_init(), lov_req_init(), lovsub_req_init()
139 * \see vvp_req_init()
141 int (*cdo_req_init)(const struct lu_env *env, struct cl_device *dev,
146 * Device in the client stack.
148 * \see vvp_device, lov_device, lovsub_device, osc_device
152 struct lu_device cd_lu_dev;
153 /** Per-layer operation vector. */
154 const struct cl_device_operations *cd_ops;
157 /** \addtogroup cl_object cl_object
161 * "Data attributes" of cl_object. Data attributes can be updated
162 * independently for a sub-object, and top-object's attributes are calculated
163 * from sub-objects' ones.
166 /** Object size, in bytes */
169 * Known minimal size, in bytes.
171 * This is only valid when at least one DLM lock is held.
174 /** Modification time. Measured in seconds since epoch. */
176 /** Access time. Measured in seconds since epoch. */
178 /** Change time. Measured in seconds since epoch. */
181 * Blocks allocated to this cl_object on the server file system.
183 * \todo XXX An interface for block size is needed.
187 * User identifier for quota purposes.
191 * Group identifier for quota purposes.
195 /* nlink of the directory */
200 * Fields in cl_attr that are being set.
214 * Sub-class of lu_object with methods common for objects on the client
217 * cl_object: represents a regular file system object, both a file and a
218 * stripe. cl_object is based on lu_object: it is identified by a fid,
219 * layered, cached, hashed, and lrued. Important distinction with the server
220 * side, where md_object and dt_object are used, is that cl_object "fans out"
221 * at the lov/sns level: depending on the file layout, single file is
222 * represented as a set of "sub-objects" (stripes). At the implementation
223 * level, struct lov_object contains an array of cl_objects. Each sub-object
224 * is a full-fledged cl_object, having its fid, living in the lru and hash
227 * This leads to the next important difference with the server side: on the
228 * client, it's quite usual to have objects with the different sequence of
229 * layers. For example, typical top-object is composed of the following
235 * whereas its sub-objects are composed of
240 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
241 * track of the object-subobject relationship.
243 * Sub-objects are not cached independently: when top-object is about to
244 * be discarded from the memory, all its sub-objects are torn-down and
247 * \see vvp_object, lov_object, lovsub_object, osc_object
251 struct lu_object co_lu;
252 /** per-object-layer operations */
253 const struct cl_object_operations *co_ops;
254 /** offset of page slice in cl_page buffer */
259 * Description of the client object configuration. This is used for the
260 * creation of a new client object that is identified by a more state than
263 struct cl_object_conf {
265 struct lu_object_conf coc_lu;
268 * Object layout. This is consumed by lov.
270 struct lustre_md *coc_md;
272 * Description of particular stripe location in the
273 * cluster. This is consumed by osc.
275 struct lov_oinfo *coc_oinfo;
278 * VFS inode. This is consumed by vvp.
280 struct inode *coc_inode;
282 * Layout lock handle.
284 struct ldlm_lock *coc_lock;
286 * Operation to handle layout, OBJECT_CONF_XYZ.
292 /** configure layout, set up a new stripe, must be called while
293 * holding layout lock.
296 /** invalidate the current stripe configuration due to losing
299 OBJECT_CONF_INVALIDATE = 1,
300 /** wait for old layout to go away so that new layout can be set up. */
305 * Operations implemented for each cl object layer.
307 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
309 struct cl_object_operations {
311 * Initialize page slice for this layer. Called top-to-bottom through
312 * every object layer when a new cl_page is instantiated. Layer
313 * keeping private per-page data, or requiring its own page operations
314 * vector should allocate these data here, and attach then to the page
315 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
318 * \retval NULL success.
320 * \retval ERR_PTR(errno) failure code.
322 * \retval valid-pointer pointer to already existing referenced page
323 * to be used instead of newly created.
325 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
326 struct cl_page *page, pgoff_t index);
328 * Initialize lock slice for this layer. Called top-to-bottom through
329 * every object layer when a new cl_lock is instantiated. Layer
330 * keeping private per-lock data, or requiring its own lock operations
331 * vector should allocate these data here, and attach then to the lock
332 * by calling cl_lock_slice_add(). Mandatory.
334 int (*coo_lock_init)(const struct lu_env *env,
335 struct cl_object *obj, struct cl_lock *lock,
336 const struct cl_io *io);
338 * Initialize io state for a given layer.
340 * called top-to-bottom once per io existence to initialize io
341 * state. If layer wants to keep some state for this type of io, it
342 * has to embed struct cl_io_slice in lu_env::le_ses, and register
343 * slice with cl_io_slice_add(). It is guaranteed that all threads
344 * participating in this io share the same session.
346 int (*coo_io_init)(const struct lu_env *env,
347 struct cl_object *obj, struct cl_io *io);
349 * Fill portion of \a attr that this layer controls. This method is
350 * called top-to-bottom through all object layers.
352 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
354 * \return 0: to continue
355 * \return +ve: to stop iterating through layers (but 0 is returned
356 * from enclosing cl_object_attr_get())
357 * \return -ve: to signal error
359 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
360 struct cl_attr *attr);
364 * \a valid is a bitmask composed from enum #cl_attr_valid, and
365 * indicating what attributes are to be set.
367 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
369 * \return the same convention as for
370 * cl_object_operations::coo_attr_get() is used.
372 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
373 const struct cl_attr *attr, unsigned int valid);
375 * Update object configuration. Called top-to-bottom to modify object
378 * XXX error conditions and handling.
380 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
381 const struct cl_object_conf *conf);
383 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
384 * object. Layers are supposed to fill parts of \a lvb that will be
385 * shipped to the glimpse originator as a glimpse result.
387 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
388 * \see osc_object_glimpse()
390 int (*coo_glimpse)(const struct lu_env *env,
391 const struct cl_object *obj, struct ost_lvb *lvb);
393 * Object prune method. Called when the layout is going to change on
394 * this object, therefore each layer has to clean up their cache,
395 * mainly pages and locks.
397 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
399 * Object getstripe method.
401 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
402 struct lov_user_md __user *lum);
406 * Extended header for client object.
408 struct cl_object_header {
409 /** Standard lu_object_header. cl_object::co_lu::lo_header points
412 struct lu_object_header coh_lu;
415 * Parent object. It is assumed that an object has a well-defined
416 * parent, but not a well-defined child (there may be multiple
417 * sub-objects, for the same top-object). cl_object_header::coh_parent
418 * field allows certain code to be written generically, without
419 * limiting possible cl_object layouts unduly.
421 struct cl_object_header *coh_parent;
423 * Protects consistency between cl_attr of parent object and
424 * attributes of sub-objects, that the former is calculated ("merged")
427 * \todo XXX this can be read/write lock if needed.
429 spinlock_t coh_attr_guard;
431 * Size of cl_page + page slices
433 unsigned short coh_page_bufsize;
435 * Number of objects above this one: 0 for a top-object, 1 for its
438 unsigned char coh_nesting;
442 * Helper macro: iterate over all layers of the object \a obj, assigning every
443 * layer top-to-bottom to \a slice.
445 #define cl_object_for_each(slice, obj) \
446 list_for_each_entry((slice), \
447 &(obj)->co_lu.lo_header->loh_layers, \
450 * Helper macro: iterate over all layers of the object \a obj, assigning every
451 * layer bottom-to-top to \a slice.
453 #define cl_object_for_each_reverse(slice, obj) \
454 list_for_each_entry_reverse((slice), \
455 &(obj)->co_lu.lo_header->loh_layers, \
459 #define CL_PAGE_EOF ((pgoff_t)~0ull)
461 /** \addtogroup cl_page cl_page
466 * Layered client page.
468 * cl_page: represents a portion of a file, cached in the memory. All pages
469 * of the given file are of the same size, and are kept in the radix tree
470 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
471 * of the top-level file object are first class cl_objects, they have their
472 * own radix trees of pages and hence page is implemented as a sequence of
473 * struct cl_pages's, linked into double-linked list through
474 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
475 * corresponding radix tree at the corresponding logical offset.
477 * cl_page is associated with VM page of the hosting environment (struct
478 * page in Linux kernel, for example), struct page. It is assumed, that this
479 * association is implemented by one of cl_page layers (top layer in the
480 * current design) that
482 * - intercepts per-VM-page call-backs made by the environment (e.g.,
485 * - translates state (page flag bits) and locking between lustre and
488 * The association between cl_page and struct page is immutable and
489 * established when cl_page is created.
491 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
492 * this io an exclusive access to this page w.r.t. other io attempts and
493 * various events changing page state (such as transfer completion, or
494 * eviction of the page from the memory). Note, that in general cl_io
495 * cannot be identified with a particular thread, and page ownership is not
496 * exactly equal to the current thread holding a lock on the page. Layer
497 * implementing association between cl_page and struct page has to implement
498 * ownership on top of available synchronization mechanisms.
500 * While lustre client maintains the notion of an page ownership by io,
501 * hosting MM/VM usually has its own page concurrency control
502 * mechanisms. For example, in Linux, page access is synchronized by the
503 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
504 * takes care to acquire and release such locks as necessary around the
505 * calls to the file system methods (->readpage(), ->prepare_write(),
506 * ->commit_write(), etc.). This leads to the situation when there are two
507 * different ways to own a page in the client:
509 * - client code explicitly and voluntary owns the page (cl_page_own());
511 * - VM locks a page and then calls the client, that has "to assume"
512 * the ownership from the VM (cl_page_assume()).
514 * Dual methods to release ownership are cl_page_disown() and
515 * cl_page_unassume().
517 * cl_page is reference counted (cl_page::cp_ref). When reference counter
518 * drops to 0, the page is returned to the cache, unless it is in
519 * cl_page_state::CPS_FREEING state, in which case it is immediately
522 * The general logic guaranteeing the absence of "existential races" for
523 * pages is the following:
525 * - there are fixed known ways for a thread to obtain a new reference
528 * - by doing a lookup in the cl_object radix tree, protected by the
531 * - by starting from VM-locked struct page and following some
532 * hosting environment method (e.g., following ->private pointer in
533 * the case of Linux kernel), see cl_vmpage_page();
535 * - when the page enters cl_page_state::CPS_FREEING state, all these
536 * ways are severed with the proper synchronization
537 * (cl_page_delete());
539 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
542 * - no new references to the page in cl_page_state::CPS_FREEING state
543 * are allowed (checked in cl_page_get()).
545 * Together this guarantees that when last reference to a
546 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
547 * page, as neither references to it can be acquired at that point, nor
550 * cl_page is a state machine. States are enumerated in enum
551 * cl_page_state. Possible state transitions are enumerated in
552 * cl_page_state_set(). State transition process (i.e., actual changing of
553 * cl_page::cp_state field) is protected by the lock on the underlying VM
556 * Linux Kernel implementation.
558 * Binding between cl_page and struct page (which is a typedef for
559 * struct page) is implemented in the vvp layer. cl_page is attached to the
560 * ->private pointer of the struct page, together with the setting of
561 * PG_private bit in page->flags, and acquiring additional reference on the
562 * struct page (much like struct buffer_head, or any similar file system
563 * private data structures).
565 * PG_locked lock is used to implement both ownership and transfer
566 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
567 * states. No additional references are acquired for the duration of the
570 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
571 * write-out is "protected" by the special PG_writeback bit.
575 * States of cl_page. cl_page.c assumes particular order here.
577 * The page state machine is rather crude, as it doesn't recognize finer page
578 * states like "dirty" or "up to date". This is because such states are not
579 * always well defined for the whole stack (see, for example, the
580 * implementation of the read-ahead, that hides page up-to-dateness to track
581 * cache hits accurately). Such sub-states are maintained by the layers that
582 * are interested in them.
586 * Page is in the cache, un-owned. Page leaves cached state in the
589 * - [cl_page_state::CPS_OWNED] io comes across the page and
592 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
593 * req-formation engine decides that it wants to include this page
594 * into an cl_req being constructed, and yanks it from the cache;
596 * - [cl_page_state::CPS_FREEING] VM callback is executed to
597 * evict the page form the memory;
599 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
603 * Page is exclusively owned by some cl_io. Page may end up in this
604 * state as a result of
606 * - io creating new page and immediately owning it;
608 * - [cl_page_state::CPS_CACHED] io finding existing cached page
611 * - [cl_page_state::CPS_OWNED] io finding existing owned page
612 * and waiting for owner to release the page;
614 * Page leaves owned state in the following cases:
616 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
617 * the cache, doing nothing;
619 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
622 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
623 * transfer for this page;
625 * - [cl_page_state::CPS_FREEING] io decides to destroy this
626 * page (e.g., as part of truncate or extent lock cancellation).
628 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
632 * Page is being written out, as a part of a transfer. This state is
633 * entered when req-formation logic decided that it wants this page to
634 * be sent through the wire _now_. Specifically, it means that once
635 * this state is achieved, transfer completion handler (with either
636 * success or failure indication) is guaranteed to be executed against
637 * this page independently of any locks and any scheduling decisions
638 * made by the hosting environment (that effectively means that the
639 * page is never put into cl_page_state::CPS_PAGEOUT state "in
640 * advance". This property is mentioned, because it is important when
641 * reasoning about possible dead-locks in the system). The page can
642 * enter this state as a result of
644 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
645 * write-out of this page, or
647 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
648 * that it has enough dirty pages cached to issue a "good"
651 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
652 * is completed---it is moved into cl_page_state::CPS_CACHED state.
654 * Underlying VM page is locked for the duration of transfer.
656 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
660 * Page is being read in, as a part of a transfer. This is quite
661 * similar to the cl_page_state::CPS_PAGEOUT state, except that
662 * read-in is always "immediate"---there is no such thing a sudden
663 * construction of read cl_req from cached, presumably not up to date,
666 * Underlying VM page is locked for the duration of transfer.
668 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
672 * Page is being destroyed. This state is entered when client decides
673 * that page has to be deleted from its host object, as, e.g., a part
676 * Once this state is reached, there is no way to escape it.
678 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
685 /** Host page, the page is from the host inode which the cl_page
690 /** Transient page, the transient cl_page is used to bind a cl_page
691 * to vmpage which is not belonging to the same object of cl_page.
692 * it is used in DirectIO and lockless IO.
698 * Fields are protected by the lock on struct page, except for atomics and
701 * \invariant Data type invariants are in cl_page_invariant(). Basically:
702 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
703 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
704 * cl_page::cp_owner (when set).
707 /** Reference counter. */
709 /** An object this page is a part of. Immutable after creation. */
710 struct cl_object *cp_obj;
712 struct page *cp_vmpage;
713 /** Linkage of pages within group. Pages must be owned */
714 struct list_head cp_batch;
715 /** List of slices. Immutable after creation. */
716 struct list_head cp_layers;
717 /** Linkage of pages within cl_req. */
718 struct list_head cp_flight;
720 * Page state. This field is const to avoid accidental update, it is
721 * modified only internally within cl_page.c. Protected by a VM lock.
723 const enum cl_page_state cp_state;
725 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
728 enum cl_page_type cp_type;
731 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
732 * by sub-io. Protected by a VM lock.
734 struct cl_io *cp_owner;
736 * Owning IO request in cl_page_state::CPS_PAGEOUT and
737 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
738 * the top-level pages. Protected by a VM lock.
740 struct cl_req *cp_req;
741 /** List of references to this page, for debugging. */
742 struct lu_ref cp_reference;
743 /** Link to an object, for debugging. */
744 struct lu_ref_link cp_obj_ref;
745 /** Link to a queue, for debugging. */
746 struct lu_ref_link cp_queue_ref;
747 /** Assigned if doing a sync_io */
748 struct cl_sync_io *cp_sync_io;
752 * Per-layer part of cl_page.
754 * \see vvp_page, lov_page, osc_page
756 struct cl_page_slice {
757 struct cl_page *cpl_page;
760 * Object slice corresponding to this page slice. Immutable after
763 struct cl_object *cpl_obj;
764 const struct cl_page_operations *cpl_ops;
765 /** Linkage into cl_page::cp_layers. Immutable after creation. */
766 struct list_head cpl_linkage;
770 * Lock mode. For the client extent locks.
781 * Requested transfer type.
791 * Per-layer page operations.
793 * Methods taking an \a io argument are for the activity happening in the
794 * context of given \a io. Page is assumed to be owned by that io, except for
795 * the obvious cases (like cl_page_operations::cpo_own()).
797 * \see vvp_page_ops, lov_page_ops, osc_page_ops
799 struct cl_page_operations {
801 * cl_page<->struct page methods. Only one layer in the stack has to
802 * implement these. Current code assumes that this functionality is
803 * provided by the topmost layer, see cl_page_disown0() as an example.
807 * Called when \a io acquires this page into the exclusive
808 * ownership. When this method returns, it is guaranteed that the is
809 * not owned by other io, and no transfer is going on against
813 * \see vvp_page_own(), lov_page_own()
815 int (*cpo_own)(const struct lu_env *env,
816 const struct cl_page_slice *slice,
817 struct cl_io *io, int nonblock);
818 /** Called when ownership it yielded. Optional.
820 * \see cl_page_disown()
821 * \see vvp_page_disown()
823 void (*cpo_disown)(const struct lu_env *env,
824 const struct cl_page_slice *slice, struct cl_io *io);
826 * Called for a page that is already "owned" by \a io from VM point of
829 * \see cl_page_assume()
830 * \see vvp_page_assume(), lov_page_assume()
832 void (*cpo_assume)(const struct lu_env *env,
833 const struct cl_page_slice *slice, struct cl_io *io);
834 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
835 * bottom-to-top when IO releases a page without actually unlocking
838 * \see cl_page_unassume()
839 * \see vvp_page_unassume()
841 void (*cpo_unassume)(const struct lu_env *env,
842 const struct cl_page_slice *slice,
845 * Announces whether the page contains valid data or not by \a uptodate.
847 * \see cl_page_export()
848 * \see vvp_page_export()
850 void (*cpo_export)(const struct lu_env *env,
851 const struct cl_page_slice *slice, int uptodate);
853 * Checks whether underlying VM page is locked (in the suitable
854 * sense). Used for assertions.
856 * \retval -EBUSY: page is protected by a lock of a given mode;
857 * \retval -ENODATA: page is not protected by a lock;
858 * \retval 0: this layer cannot decide. (Should never happen.)
860 int (*cpo_is_vmlocked)(const struct lu_env *env,
861 const struct cl_page_slice *slice);
867 * Called when page is truncated from the object. Optional.
869 * \see cl_page_discard()
870 * \see vvp_page_discard(), osc_page_discard()
872 void (*cpo_discard)(const struct lu_env *env,
873 const struct cl_page_slice *slice,
876 * Called when page is removed from the cache, and is about to being
877 * destroyed. Optional.
879 * \see cl_page_delete()
880 * \see vvp_page_delete(), osc_page_delete()
882 void (*cpo_delete)(const struct lu_env *env,
883 const struct cl_page_slice *slice);
884 /** Destructor. Frees resources and slice itself. */
885 void (*cpo_fini)(const struct lu_env *env,
886 struct cl_page_slice *slice);
889 * Checks whether the page is protected by a cl_lock. This is a
890 * per-layer method, because certain layers have ways to check for the
891 * lock much more efficiently than through the generic locks scan, or
892 * implement locking mechanisms separate from cl_lock, e.g.,
893 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
894 * being canceled, or scheduled for cancellation as soon as the last
895 * user goes away, too.
897 * \retval -EBUSY: page is protected by a lock of a given mode;
898 * \retval -ENODATA: page is not protected by a lock;
899 * \retval 0: this layer cannot decide.
901 * \see cl_page_is_under_lock()
903 int (*cpo_is_under_lock)(const struct lu_env *env,
904 const struct cl_page_slice *slice,
905 struct cl_io *io, pgoff_t *max);
908 * Optional debugging helper. Prints given page slice.
910 * \see cl_page_print()
912 int (*cpo_print)(const struct lu_env *env,
913 const struct cl_page_slice *slice,
914 void *cookie, lu_printer_t p);
918 * Transfer methods. See comment on cl_req for a description of
919 * transfer formation and life-cycle.
924 * Request type dependent vector of operations.
926 * Transfer operations depend on transfer mode (cl_req_type). To avoid
927 * passing transfer mode to each and every of these methods, and to
928 * avoid branching on request type inside of the methods, separate
929 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
930 * provided. That is, method invocation usually looks like
932 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
936 * Called when a page is submitted for a transfer as a part of
939 * \return 0 : page is eligible for submission;
940 * \return -EALREADY : skip this page;
941 * \return -ve : error.
943 * \see cl_page_prep()
945 int (*cpo_prep)(const struct lu_env *env,
946 const struct cl_page_slice *slice,
949 * Completion handler. This is guaranteed to be eventually
950 * fired after cl_page_operations::cpo_prep() or
951 * cl_page_operations::cpo_make_ready() call.
953 * This method can be called in a non-blocking context. It is
954 * guaranteed however, that the page involved and its object
955 * are pinned in memory (and, hence, calling cl_page_put() is
958 * \see cl_page_completion()
960 void (*cpo_completion)(const struct lu_env *env,
961 const struct cl_page_slice *slice,
964 * Called when cached page is about to be added to the
965 * cl_req as a part of req formation.
967 * \return 0 : proceed with this page;
968 * \return -EAGAIN : skip this page;
969 * \return -ve : error.
971 * \see cl_page_make_ready()
973 int (*cpo_make_ready)(const struct lu_env *env,
974 const struct cl_page_slice *slice);
977 * Tell transfer engine that only [to, from] part of a page should be
980 * This is used for immediate transfers.
982 * \todo XXX this is not very good interface. It would be much better
983 * if all transfer parameters were supplied as arguments to
984 * cl_io_operations::cio_submit() call, but it is not clear how to do
985 * this for page queues.
987 * \see cl_page_clip()
989 void (*cpo_clip)(const struct lu_env *env,
990 const struct cl_page_slice *slice,
993 * \pre the page was queued for transferring.
994 * \post page is removed from client's pending list, or -EBUSY
995 * is returned if it has already been in transferring.
997 * This is one of seldom page operation which is:
998 * 0. called from top level;
999 * 1. don't have vmpage locked;
1000 * 2. every layer should synchronize execution of its ->cpo_cancel()
1001 * with completion handlers. Osc uses client obd lock for this
1002 * purpose. Based on there is no vvp_page_cancel and
1003 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
1005 * \see osc_page_cancel().
1007 int (*cpo_cancel)(const struct lu_env *env,
1008 const struct cl_page_slice *slice);
1010 * Write out a page by kernel. This is only called by ll_writepage
1013 * \see cl_page_flush()
1015 int (*cpo_flush)(const struct lu_env *env,
1016 const struct cl_page_slice *slice,
1022 * Helper macro, dumping detailed information about \a page into a log.
1024 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1026 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1027 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1028 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1029 CDEBUG(mask, format, ## __VA_ARGS__); \
1034 * Helper macro, dumping shorter information about \a page into a log.
1036 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1038 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1039 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1040 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1041 CDEBUG(mask, format, ## __VA_ARGS__); \
1045 static inline struct page *cl_page_vmpage(struct cl_page *page)
1047 LASSERT(page->cp_vmpage);
1048 return page->cp_vmpage;
1052 * Check if a cl_page is in use.
1054 * Client cache holds a refcount, this refcount will be dropped when
1055 * the page is taken out of cache, see vvp_page_delete().
1057 static inline bool __page_in_use(const struct cl_page *page, int refc)
1059 return (atomic_read(&page->cp_ref) > refc + 1);
1063 * Caller itself holds a refcount of cl_page.
1065 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1067 * Caller doesn't hold a refcount.
1069 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1073 /** \addtogroup cl_lock cl_lock
1078 * Extent locking on the client.
1082 * The locking model of the new client code is built around
1086 * data-type representing an extent lock on a regular file. cl_lock is a
1087 * layered object (much like cl_object and cl_page), it consists of a header
1088 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1089 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1091 * Typical cl_lock consists of the two layers:
1093 * - vvp_lock (vvp specific data), and
1094 * - lov_lock (lov specific data).
1096 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1097 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1099 * - lovsub_lock, and
1102 * Each sub-lock is associated with a cl_object (representing stripe
1103 * sub-object or the file to which top-level cl_lock is associated to), and is
1104 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1105 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1106 * is different from cl_page, that doesn't fan out (there is usually exactly
1107 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1108 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1112 * cl_lock is a cacheless data container for the requirements of locks to
1113 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1116 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1117 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1119 * INTERFACE AND USAGE
1121 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1122 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1123 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1124 * consists of multiple sub cl_locks, each sub locks will be enqueued
1125 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1126 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1129 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1130 * method will be called for each layer to release the resource held by this
1131 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1132 * clo_enqueue time, is released.
1134 * LDLM lock can only be canceled if there is no cl_lock using it.
1136 * Overall process of the locking during IO operation is as following:
1138 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1139 * is called on each layer. Responsibility of this method is to add locks,
1140 * needed by a given layer into cl_io.ci_lockset.
1142 * - once locks for all layers were collected, they are sorted to avoid
1143 * dead-locks (cl_io_locks_sort()), and enqueued.
1145 * - when all locks are acquired, IO is performed;
1147 * - locks are released after IO is complete.
1149 * Striping introduces major additional complexity into locking. The
1150 * fundamental problem is that it is generally unsafe to actively use (hold)
1151 * two locks on the different OST servers at the same time, as this introduces
1152 * inter-server dependency and can lead to cascading evictions.
1154 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1155 * that no multi-stripe locks are taken (note that this design abandons POSIX
1156 * read/write semantics). Such pieces ideally can be executed concurrently. At
1157 * the same time, certain types of IO cannot be sub-divived, without
1158 * sacrificing correctness. This includes:
1160 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1163 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1165 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1166 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1167 * has to be held together with the usual lock on [offset, offset + count].
1169 * Interaction with DLM
1171 * In the expected setup, cl_lock is ultimately backed up by a collection of
1172 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1173 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1174 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1175 * description of interaction with DLM.
1181 struct cl_lock_descr {
1182 /** Object this lock is granted for. */
1183 struct cl_object *cld_obj;
1184 /** Index of the first page protected by this lock. */
1186 /** Index of the last page (inclusive) protected by this lock. */
1188 /** Group ID, for group lock */
1191 enum cl_lock_mode cld_mode;
1193 * flags to enqueue lock. A combination of bit-flags from
1194 * enum cl_enq_flags.
1196 __u32 cld_enq_flags;
1199 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1200 #define PDESCR(descr) \
1201 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1202 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1204 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1207 * Layered client lock.
1210 /** List of slices. Immutable after creation. */
1211 struct list_head cll_layers;
1212 /** lock attribute, extent, cl_object, etc. */
1213 struct cl_lock_descr cll_descr;
1217 * Per-layer part of cl_lock
1219 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1221 struct cl_lock_slice {
1222 struct cl_lock *cls_lock;
1223 /** Object slice corresponding to this lock slice. Immutable after
1226 struct cl_object *cls_obj;
1227 const struct cl_lock_operations *cls_ops;
1228 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1229 struct list_head cls_linkage;
1234 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1236 struct cl_lock_operations {
1239 * Attempts to enqueue the lock. Called top-to-bottom.
1241 * \retval 0 this layer has enqueued the lock successfully
1242 * \retval >0 this layer has enqueued the lock, but need to wait on
1243 * @anchor for resources
1244 * \retval -ve failure
1246 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1247 * \see osc_lock_enqueue()
1249 int (*clo_enqueue)(const struct lu_env *env,
1250 const struct cl_lock_slice *slice,
1251 struct cl_io *io, struct cl_sync_io *anchor);
1253 * Cancel a lock, release its DLM lock ref, while does not cancel the
1256 void (*clo_cancel)(const struct lu_env *env,
1257 const struct cl_lock_slice *slice);
1260 * Destructor. Frees resources and the slice.
1262 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1263 * \see osc_lock_fini()
1265 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1267 * Optional debugging helper. Prints given lock slice.
1269 int (*clo_print)(const struct lu_env *env,
1270 void *cookie, lu_printer_t p,
1271 const struct cl_lock_slice *slice);
1274 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1276 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1278 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1279 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1280 CDEBUG(mask, format, ## __VA_ARGS__); \
1284 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1288 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1294 /** \addtogroup cl_page_list cl_page_list
1295 * Page list used to perform collective operations on a group of pages.
1297 * Pages are added to the list one by one. cl_page_list acquires a reference
1298 * for every page in it. Page list is used to perform collective operations on
1301 * - submit pages for an immediate transfer,
1303 * - own pages on behalf of certain io (waiting for each page in turn),
1307 * When list is finalized, it releases references on all pages it still has.
1309 * \todo XXX concurrency control.
1313 struct cl_page_list {
1315 struct list_head pl_pages;
1316 struct task_struct *pl_owner;
1320 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1321 * contains an incoming page list and an outgoing page list.
1324 struct cl_page_list c2_qin;
1325 struct cl_page_list c2_qout;
1328 /** @} cl_page_list */
1330 /** \addtogroup cl_io cl_io
1336 * cl_io represents a high level I/O activity like
1337 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1340 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1341 * important distinction. We want to minimize number of calls to the allocator
1342 * in the fast path, e.g., in the case of read(2) when everything is cached:
1343 * client already owns the lock over region being read, and data are cached
1344 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1345 * per-layer io state is stored in the session, associated with the io, see
1346 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1347 * by using free-lists, see cl_env_get().
1349 * There is a small predefined number of possible io types, enumerated in enum
1352 * cl_io is a state machine, that can be advanced concurrently by the multiple
1353 * threads. It is up to these threads to control the concurrency and,
1354 * specifically, to detect when io is done, and its state can be safely
1357 * For read/write io overall execution plan is as following:
1359 * (0) initialize io state through all layers;
1361 * (1) loop: prepare chunk of work to do
1363 * (2) call all layers to collect locks they need to process current chunk
1365 * (3) sort all locks to avoid dead-locks, and acquire them
1367 * (4) process the chunk: call per-page methods
1368 * (cl_io_operations::cio_read_page() for read,
1369 * cl_io_operations::cio_prepare_write(),
1370 * cl_io_operations::cio_commit_write() for write)
1376 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1377 * address allocation efficiency issues mentioned above), and returns with the
1378 * special error condition from per-page method when current sub-io has to
1379 * block. This causes io loop to be repeated, and lov switches to the next
1380 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1385 /** read system call */
1387 /** write system call */
1389 /** truncate, utime system calls */
1392 * page fault handling
1396 * fsync system call handling
1397 * To write out a range of file
1401 * Miscellaneous io. This is used for occasional io activity that
1402 * doesn't fit into other types. Currently this is used for:
1404 * - cancellation of an extent lock. This io exists as a context
1405 * to write dirty pages from under the lock being canceled back
1408 * - VM induced page write-out. An io context for writing page out
1409 * for memory cleansing;
1411 * - glimpse. An io context to acquire glimpse lock.
1413 * - grouplock. An io context to acquire group lock.
1415 * CIT_MISC io is used simply as a context in which locks and pages
1416 * are manipulated. Such io has no internal "process", that is,
1417 * cl_io_loop() is never called for it.
1424 * States of cl_io state machine
1427 /** Not initialized. */
1431 /** IO iteration started. */
1435 /** Actual IO is in progress. */
1437 /** IO for the current iteration finished. */
1439 /** Locks released. */
1441 /** Iteration completed. */
1443 /** cl_io finalized. */
1448 * IO state private for a layer.
1450 * This is usually embedded into layer session data, rather than allocated
1453 * \see vvp_io, lov_io, osc_io
1455 struct cl_io_slice {
1456 struct cl_io *cis_io;
1457 /** corresponding object slice. Immutable after creation. */
1458 struct cl_object *cis_obj;
1459 /** io operations. Immutable after creation. */
1460 const struct cl_io_operations *cis_iop;
1462 * linkage into a list of all slices for a given cl_io, hanging off
1463 * cl_io::ci_layers. Immutable after creation.
1465 struct list_head cis_linkage;
1468 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1471 * Per-layer io operations.
1472 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1474 struct cl_io_operations {
1476 * Vector of io state transition methods for every io type.
1478 * \see cl_page_operations::io
1482 * Prepare io iteration at a given layer.
1484 * Called top-to-bottom at the beginning of each iteration of
1485 * "io loop" (if it makes sense for this type of io). Here
1486 * layer selects what work it will do during this iteration.
1488 * \see cl_io_operations::cio_iter_fini()
1490 int (*cio_iter_init)(const struct lu_env *env,
1491 const struct cl_io_slice *slice);
1493 * Finalize io iteration.
1495 * Called bottom-to-top at the end of each iteration of "io
1496 * loop". Here layers can decide whether IO has to be
1499 * \see cl_io_operations::cio_iter_init()
1501 void (*cio_iter_fini)(const struct lu_env *env,
1502 const struct cl_io_slice *slice);
1504 * Collect locks for the current iteration of io.
1506 * Called top-to-bottom to collect all locks necessary for
1507 * this iteration. This methods shouldn't actually enqueue
1508 * anything, instead it should post a lock through
1509 * cl_io_lock_add(). Once all locks are collected, they are
1510 * sorted and enqueued in the proper order.
1512 int (*cio_lock)(const struct lu_env *env,
1513 const struct cl_io_slice *slice);
1515 * Finalize unlocking.
1517 * Called bottom-to-top to finish layer specific unlocking
1518 * functionality, after generic code released all locks
1519 * acquired by cl_io_operations::cio_lock().
1521 void (*cio_unlock)(const struct lu_env *env,
1522 const struct cl_io_slice *slice);
1524 * Start io iteration.
1526 * Once all locks are acquired, called top-to-bottom to
1527 * commence actual IO. In the current implementation,
1528 * top-level vvp_io_{read,write}_start() does all the work
1529 * synchronously by calling generic_file_*(), so other layers
1530 * are called when everything is done.
1532 int (*cio_start)(const struct lu_env *env,
1533 const struct cl_io_slice *slice);
1535 * Called top-to-bottom at the end of io loop. Here layer
1536 * might wait for an unfinished asynchronous io.
1538 void (*cio_end)(const struct lu_env *env,
1539 const struct cl_io_slice *slice);
1541 * Called bottom-to-top to notify layers that read/write IO
1542 * iteration finished, with \a nob bytes transferred.
1544 void (*cio_advance)(const struct lu_env *env,
1545 const struct cl_io_slice *slice,
1548 * Called once per io, bottom-to-top to release io resources.
1550 void (*cio_fini)(const struct lu_env *env,
1551 const struct cl_io_slice *slice);
1555 * Submit pages from \a queue->c2_qin for IO, and move
1556 * successfully submitted pages into \a queue->c2_qout. Return
1557 * non-zero if failed to submit even the single page. If
1558 * submission failed after some pages were moved into \a
1559 * queue->c2_qout, completion callback with non-zero ioret is
1562 int (*cio_submit)(const struct lu_env *env,
1563 const struct cl_io_slice *slice,
1564 enum cl_req_type crt,
1565 struct cl_2queue *queue);
1567 * Queue async page for write.
1568 * The difference between cio_submit and cio_queue is that
1569 * cio_submit is for urgent request.
1571 int (*cio_commit_async)(const struct lu_env *env,
1572 const struct cl_io_slice *slice,
1573 struct cl_page_list *queue, int from, int to,
1576 * Read missing page.
1578 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
1579 * method, when it hits not-up-to-date page in the range. Optional.
1581 * \pre io->ci_type == CIT_READ
1583 int (*cio_read_page)(const struct lu_env *env,
1584 const struct cl_io_slice *slice,
1585 const struct cl_page_slice *page);
1587 * Optional debugging helper. Print given io slice.
1589 int (*cio_print)(const struct lu_env *env, void *cookie,
1590 lu_printer_t p, const struct cl_io_slice *slice);
1594 * Flags to lock enqueue procedure.
1599 * instruct server to not block, if conflicting lock is found. Instead
1600 * -EWOULDBLOCK is returned immediately.
1602 CEF_NONBLOCK = 0x00000001,
1604 * take lock asynchronously (out of order), as it cannot
1605 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1607 CEF_ASYNC = 0x00000002,
1609 * tell the server to instruct (though a flag in the blocking ast) an
1610 * owner of the conflicting lock, that it can drop dirty pages
1611 * protected by this lock, without sending them to the server.
1613 CEF_DISCARD_DATA = 0x00000004,
1615 * tell the sub layers that it must be a `real' lock. This is used for
1616 * mmapped-buffer locks and glimpse locks that must be never converted
1617 * into lockless mode.
1619 * \see vvp_mmap_locks(), cl_glimpse_lock().
1621 CEF_MUST = 0x00000008,
1623 * tell the sub layers that never request a `real' lock. This flag is
1624 * not used currently.
1626 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1627 * conversion policy: ci_lockreq describes generic information of lock
1628 * requirement for this IO, especially for locks which belong to the
1629 * object doing IO; however, lock itself may have precise requirements
1630 * that are described by the enqueue flags.
1632 CEF_NEVER = 0x00000010,
1634 * for async glimpse lock.
1636 CEF_AGL = 0x00000020,
1638 * enqueue a lock to test DLM lock existence.
1640 CEF_PEEK = 0x00000040,
1642 * mask of enq_flags.
1644 CEF_MASK = 0x0000007f,
1648 * Link between lock and io. Intermediate structure is needed, because the
1649 * same lock can be part of multiple io's simultaneously.
1651 struct cl_io_lock_link {
1652 /** linkage into one of cl_lockset lists. */
1653 struct list_head cill_linkage;
1654 struct cl_lock cill_lock;
1655 /** optional destructor */
1656 void (*cill_fini)(const struct lu_env *env,
1657 struct cl_io_lock_link *link);
1659 #define cill_descr cill_lock.cll_descr
1662 * Lock-set represents a collection of locks, that io needs at a
1663 * time. Generally speaking, client tries to avoid holding multiple locks when
1666 * - holding extent locks over multiple ost's introduces the danger of
1667 * "cascading timeouts";
1669 * - holding multiple locks over the same ost is still dead-lock prone,
1670 * see comment in osc_lock_enqueue(),
1672 * but there are certain situations where this is unavoidable:
1674 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1676 * - truncate has to take [new-size, EOF] lock for correctness;
1678 * - SNS has to take locks across full stripe for correctness;
1680 * - in the case when user level buffer, supplied to {read,write}(file0),
1681 * is a part of a memory mapped lustre file, client has to take a dlm
1682 * locks on file0, and all files that back up the buffer (or a part of
1683 * the buffer, that is being processed in the current chunk, in any
1684 * case, there are situations where at least 2 locks are necessary).
1686 * In such cases we at least try to take locks in the same consistent
1687 * order. To this end, all locks are first collected, then sorted, and then
1691 /** locks to be acquired. */
1692 struct list_head cls_todo;
1693 /** locks acquired. */
1694 struct list_head cls_done;
1698 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1699 * but 'req' is always to be thought as 'request' :-)
1701 enum cl_io_lock_dmd {
1702 /** Always lock data (e.g., O_APPEND). */
1704 /** Layers are free to decide between local and global locking. */
1706 /** Never lock: there is no cache (e.g., lockless IO). */
1710 enum cl_fsync_mode {
1711 /** start writeback, do not wait for them to finish */
1713 /** start writeback and wait for them to finish */
1715 /** discard all of dirty pages in a specific file range */
1716 CL_FSYNC_DISCARD = 2,
1717 /** start writeback and make sure they have reached storage before
1718 * return. OST_SYNC RPC must be issued and finished
1723 struct cl_io_rw_common {
1732 * cl_io is shared by all threads participating in this IO (in current
1733 * implementation only one thread advances IO, but parallel IO design and
1734 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1735 * is up to these threads to serialize their activities, including updates to
1736 * mutable cl_io fields.
1739 /** type of this IO. Immutable after creation. */
1740 enum cl_io_type ci_type;
1741 /** current state of cl_io state machine. */
1742 enum cl_io_state ci_state;
1743 /** main object this io is against. Immutable after creation. */
1744 struct cl_object *ci_obj;
1746 * Upper layer io, of which this io is a part of. Immutable after
1749 struct cl_io *ci_parent;
1750 /** List of slices. Immutable after creation. */
1751 struct list_head ci_layers;
1752 /** list of locks (to be) acquired by this io. */
1753 struct cl_lockset ci_lockset;
1754 /** lock requirements, this is just a help info for sublayers. */
1755 enum cl_io_lock_dmd ci_lockreq;
1758 struct cl_io_rw_common rd;
1761 struct cl_io_rw_common wr;
1765 struct cl_io_rw_common ci_rw;
1766 struct cl_setattr_io {
1767 struct ost_lvb sa_attr;
1768 unsigned int sa_valid;
1769 int sa_stripe_index;
1770 struct lu_fid *sa_parent_fid;
1772 struct cl_fault_io {
1773 /** page index within file. */
1775 /** bytes valid byte on a faulted page. */
1777 /** writable page? for nopage() only */
1779 /** page of an executable? */
1781 /** page_mkwrite() */
1783 /** resulting page */
1784 struct cl_page *ft_page;
1786 struct cl_fsync_io {
1789 /** file system level fid */
1790 struct lu_fid *fi_fid;
1791 enum cl_fsync_mode fi_mode;
1792 /* how many pages were written/discarded */
1793 unsigned int fi_nr_written;
1796 struct cl_2queue ci_queue;
1799 unsigned int ci_continue:1,
1801 * This io has held grouplock, to inform sublayers that
1802 * don't do lockless i/o.
1806 * The whole IO need to be restarted because layout has been changed
1810 * to not refresh layout - the IO issuer knows that the layout won't
1811 * change(page operations, layout change causes all page to be
1812 * discarded), or it doesn't matter if it changes(sync).
1816 * Check if layout changed after the IO finishes. Mainly for HSM
1817 * requirement. If IO occurs to openning files, it doesn't need to
1818 * verify layout because HSM won't release openning files.
1819 * Right now, only two operations need to verify layout: glimpse
1824 * file is released, restore has to to be triggered by vvp layer
1826 ci_restore_needed:1,
1832 * Number of pages owned by this IO. For invariant checking.
1834 unsigned ci_owned_nr;
1839 /** \addtogroup cl_req cl_req
1845 * There are two possible modes of transfer initiation on the client:
1847 * - immediate transfer: this is started when a high level io wants a page
1848 * or a collection of pages to be transferred right away. Examples:
1849 * read-ahead, synchronous read in the case of non-page aligned write,
1850 * page write-out as a part of extent lock cancellation, page write-out
1851 * as a part of memory cleansing. Immediate transfer can be both
1852 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1854 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1855 * when io wants to transfer a page to the server some time later, when
1856 * it can be done efficiently. Example: pages dirtied by the write(2)
1859 * In any case, transfer takes place in the form of a cl_req, which is a
1860 * representation for a network RPC.
1862 * Pages queued for an opportunistic transfer are cached until it is decided
1863 * that efficient RPC can be composed of them. This decision is made by "a
1864 * req-formation engine", currently implemented as a part of osc
1865 * layer. Req-formation depends on many factors: the size of the resulting
1866 * RPC, whether or not multi-object RPCs are supported by the server,
1867 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1869 * For the immediate transfer io submits a cl_page_list, that req-formation
1870 * engine slices into cl_req's, possibly adding cached pages to some of
1871 * the resulting req's.
1873 * Whenever a page from cl_page_list is added to a newly constructed req, its
1874 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1875 * page state is atomically changed from cl_page_state::CPS_OWNED to
1876 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1877 * is zeroed, and cl_page::cp_req is set to the
1878 * req. cl_page_operations::cpo_prep() method at the particular layer might
1879 * return -EALREADY to indicate that it does not need to submit this page
1880 * at all. This is possible, for example, if page, submitted for read,
1881 * became up-to-date in the meantime; and for write, the page don't have
1882 * dirty bit marked. \see cl_io_submit_rw()
1884 * Whenever a cached page is added to a newly constructed req, its
1885 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1886 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1887 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1888 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1889 * might return -EAGAIN to indicate that this page is not eligible for the
1890 * transfer right now.
1894 * Plan is to divide transfers into "priority bands" (indicated when
1895 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1896 * and allow glueing of cached pages to immediate transfers only within single
1897 * band. This would make high priority transfers (like lock cancellation or
1898 * memory pressure induced write-out) really high priority.
1903 * Per-transfer attributes.
1905 struct cl_req_attr {
1906 /** Generic attributes for the server consumption. */
1907 struct obdo *cra_oa;
1909 char cra_jobid[LUSTRE_JOBID_SIZE];
1913 * Transfer request operations definable at every layer.
1915 * Concurrency: transfer formation engine synchronizes calls to all transfer
1918 struct cl_req_operations {
1920 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1921 * complete (all pages are added).
1923 * \see osc_req_prep()
1925 int (*cro_prep)(const struct lu_env *env,
1926 const struct cl_req_slice *slice);
1928 * Called top-to-bottom to fill in \a oa fields. This is called twice
1929 * with different flags, see bug 10150 and osc_build_req().
1931 * \param obj an object from cl_req which attributes are to be set in
1934 * \param oa struct obdo where attributes are placed
1936 * \param flags \a oa fields to be filled.
1938 void (*cro_attr_set)(const struct lu_env *env,
1939 const struct cl_req_slice *slice,
1940 const struct cl_object *obj,
1941 struct cl_req_attr *attr, u64 flags);
1943 * Called top-to-bottom from cl_req_completion() to notify layers that
1944 * transfer completed. Has to free all state allocated by
1945 * cl_device_operations::cdo_req_init().
1947 void (*cro_completion)(const struct lu_env *env,
1948 const struct cl_req_slice *slice, int ioret);
1952 * A per-object state that (potentially multi-object) transfer request keeps.
1955 /** object itself */
1956 struct cl_object *ro_obj;
1957 /** reference to cl_req_obj::ro_obj. For debugging. */
1958 struct lu_ref_link ro_obj_ref;
1959 /* something else? Number of pages for a given object? */
1965 * Transfer requests are not reference counted, because IO sub-system owns
1966 * them exclusively and knows when to free them.
1970 * cl_req is created by cl_req_alloc() that calls
1971 * cl_device_operations::cdo_req_init() device methods to allocate per-req
1972 * state in every layer.
1974 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
1975 * contains pages for.
1977 * Once all pages were collected, cl_page_operations::cpo_prep() method is
1978 * called top-to-bottom. At that point layers can modify req, let it pass, or
1979 * deny it completely. This is to support things like SNS that have transfer
1980 * ordering requirements invisible to the individual req-formation engine.
1982 * On transfer completion (or transfer timeout, or failure to initiate the
1983 * transfer of an allocated req), cl_req_operations::cro_completion() method
1984 * is called, after execution of cl_page_operations::cpo_completion() of all
1988 enum cl_req_type crq_type;
1989 /** A list of pages being transferred */
1990 struct list_head crq_pages;
1991 /** Number of pages in cl_req::crq_pages */
1992 unsigned crq_nrpages;
1993 /** An array of objects which pages are in ->crq_pages */
1994 struct cl_req_obj *crq_o;
1995 /** Number of elements in cl_req::crq_objs[] */
1996 unsigned crq_nrobjs;
1997 struct list_head crq_layers;
2001 * Per-layer state for request.
2003 struct cl_req_slice {
2004 struct cl_req *crs_req;
2005 struct cl_device *crs_dev;
2006 struct list_head crs_linkage;
2007 const struct cl_req_operations *crs_ops;
2012 enum cache_stats_item {
2013 /** how many cache lookups were performed */
2015 /** how many times cache lookup resulted in a hit */
2017 /** how many entities are in the cache right now */
2019 /** how many entities in the cache are actively used (and cannot be
2020 * evicted) right now
2023 /** how many entities were created at all */
2028 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2031 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2033 struct cache_stats {
2034 const char *cs_name;
2035 atomic_t cs_stats[CS_NR];
2038 /** These are not exported so far */
2039 void cache_stats_init(struct cache_stats *cs, const char *name);
2042 * Client-side site. This represents particular client stack. "Global"
2043 * variables should (directly or indirectly) be added here to allow multiple
2044 * clients to co-exist in the single address space.
2047 struct lu_site cs_lu;
2049 * Statistical counters. Atomics do not scale, something better like
2050 * per-cpu counters is needed.
2052 * These are exported as /sys/kernel/debug/lustre/llite/.../site
2054 * When interpreting keep in mind that both sub-locks (and sub-pages)
2055 * and top-locks (and top-pages) are accounted here.
2057 struct cache_stats cs_pages;
2058 atomic_t cs_pages_state[CPS_NR];
2061 int cl_site_init(struct cl_site *s, struct cl_device *top);
2062 void cl_site_fini(struct cl_site *s);
2063 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2066 * Output client site statistical counters into a buffer. Suitable for
2067 * ll_rd_*()-style functions.
2069 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2074 * Type conversion and accessory functions.
2078 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2080 return container_of(site, struct cl_site, cs_lu);
2083 static inline int lu_device_is_cl(const struct lu_device *d)
2085 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2088 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2090 LASSERT(!d || IS_ERR(d) || lu_device_is_cl(d));
2091 return container_of0(d, struct cl_device, cd_lu_dev);
2094 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2096 return &d->cd_lu_dev;
2099 static inline struct cl_object *lu2cl(const struct lu_object *o)
2101 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2102 return container_of0(o, struct cl_object, co_lu);
2105 static inline const struct cl_object_conf *
2106 lu2cl_conf(const struct lu_object_conf *conf)
2108 return container_of0(conf, struct cl_object_conf, coc_lu);
2111 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2113 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2116 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2118 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2119 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2122 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2124 return container_of0(h, struct cl_object_header, coh_lu);
2127 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2129 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2133 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2135 return luh2coh(obj->co_lu.lo_header);
2138 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2140 return lu_device_init(&d->cd_lu_dev, t);
2143 static inline void cl_device_fini(struct cl_device *d)
2145 lu_device_fini(&d->cd_lu_dev);
2148 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2149 struct cl_object *obj, pgoff_t index,
2150 const struct cl_page_operations *ops);
2151 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2152 struct cl_object *obj,
2153 const struct cl_lock_operations *ops);
2154 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2155 struct cl_object *obj, const struct cl_io_operations *ops);
2156 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2157 struct cl_device *dev,
2158 const struct cl_req_operations *ops);
2161 /** \defgroup cl_object cl_object
2164 struct cl_object *cl_object_top(struct cl_object *o);
2165 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2166 const struct lu_fid *fid,
2167 const struct cl_object_conf *c);
2169 int cl_object_header_init(struct cl_object_header *h);
2170 void cl_object_put(const struct lu_env *env, struct cl_object *o);
2171 void cl_object_get(struct cl_object *o);
2172 void cl_object_attr_lock(struct cl_object *o);
2173 void cl_object_attr_unlock(struct cl_object *o);
2174 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2175 struct cl_attr *attr);
2176 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2177 const struct cl_attr *attr, unsigned int valid);
2178 int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj,
2179 struct ost_lvb *lvb);
2180 int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
2181 const struct cl_object_conf *conf);
2182 int cl_object_prune(const struct lu_env *env, struct cl_object *obj);
2183 void cl_object_kill(const struct lu_env *env, struct cl_object *obj);
2184 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2185 struct lov_user_md __user *lum);
2188 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2190 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2192 return cl_object_header(o0) == cl_object_header(o1);
2195 static inline void cl_object_page_init(struct cl_object *clob, int size)
2197 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2198 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2199 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2202 static inline void *cl_object_page_slice(struct cl_object *clob,
2203 struct cl_page *page)
2205 return (void *)((char *)page + clob->co_slice_off);
2209 * Return refcount of cl_object.
2211 static inline int cl_object_refc(struct cl_object *clob)
2213 struct lu_object_header *header = clob->co_lu.lo_header;
2215 return atomic_read(&header->loh_ref);
2220 /** \defgroup cl_page cl_page
2230 /* callback of cl_page_gang_lookup() */
2231 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *obj,
2232 pgoff_t idx, struct page *vmpage,
2233 enum cl_page_type type);
2234 struct cl_page *cl_page_alloc(const struct lu_env *env,
2235 struct cl_object *o, pgoff_t ind,
2236 struct page *vmpage,
2237 enum cl_page_type type);
2238 void cl_page_get(struct cl_page *page);
2239 void cl_page_put(const struct lu_env *env, struct cl_page *page);
2240 void cl_page_print(const struct lu_env *env, void *cookie, lu_printer_t printer,
2241 const struct cl_page *pg);
2242 void cl_page_header_print(const struct lu_env *env, void *cookie,
2243 lu_printer_t printer, const struct cl_page *pg);
2244 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj);
2246 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2247 const struct lu_device_type *dtype);
2252 * Functions dealing with the ownership of page by io.
2256 int cl_page_own(const struct lu_env *env,
2257 struct cl_io *io, struct cl_page *page);
2258 int cl_page_own_try(const struct lu_env *env,
2259 struct cl_io *io, struct cl_page *page);
2260 void cl_page_assume(const struct lu_env *env,
2261 struct cl_io *io, struct cl_page *page);
2262 void cl_page_unassume(const struct lu_env *env,
2263 struct cl_io *io, struct cl_page *pg);
2264 void cl_page_disown(const struct lu_env *env,
2265 struct cl_io *io, struct cl_page *page);
2266 void cl_page_disown0(const struct lu_env *env,
2267 struct cl_io *io, struct cl_page *pg);
2268 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io);
2275 * Functions dealing with the preparation of a page for a transfer, and
2276 * tracking transfer state.
2279 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
2280 struct cl_page *pg, enum cl_req_type crt);
2281 void cl_page_completion(const struct lu_env *env,
2282 struct cl_page *pg, enum cl_req_type crt, int ioret);
2283 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
2284 enum cl_req_type crt);
2285 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
2286 struct cl_page *pg, enum cl_req_type crt);
2287 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
2289 int cl_page_cancel(const struct lu_env *env, struct cl_page *page);
2290 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
2291 struct cl_page *pg);
2296 * \name helper routines
2297 * Functions to discard, delete and export a cl_page.
2300 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2301 struct cl_page *pg);
2302 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2303 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg);
2304 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate);
2305 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2306 struct cl_page *page, pgoff_t *max_index);
2307 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2308 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2309 size_t cl_page_size(const struct cl_object *obj);
2310 int cl_pages_prune(const struct lu_env *env, struct cl_object *obj);
2312 void cl_lock_print(const struct lu_env *env, void *cookie,
2313 lu_printer_t printer, const struct cl_lock *lock);
2314 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2315 lu_printer_t printer,
2316 const struct cl_lock_descr *descr);
2320 * Data structure managing a client's cached pages. A count of
2321 * "unstable" pages is maintained, and an LRU of clean pages is
2322 * maintained. "unstable" pages are pages pinned by the ptlrpc
2323 * layer for recovery purposes.
2325 struct cl_client_cache {
2327 * # of client cache refcount
2328 * # of users (OSCs) + 2 (held by llite and lov)
2332 * # of threads are doing shrinking
2334 unsigned int ccc_lru_shrinkers;
2336 * # of LRU entries available
2338 atomic_long_t ccc_lru_left;
2340 * List of entities(OSCs) for this LRU cache
2342 struct list_head ccc_lru;
2344 * Max # of LRU entries
2346 unsigned long ccc_lru_max;
2348 * Lock to protect ccc_lru list
2350 spinlock_t ccc_lru_lock;
2352 * Set if unstable check is enabled
2354 unsigned int ccc_unstable_check:1;
2356 * # of unstable pages for this mount point
2358 atomic_long_t ccc_unstable_nr;
2360 * Waitq for awaiting unstable pages to reach zero.
2361 * Used at umounting time and signaled on BRW commit
2363 wait_queue_head_t ccc_unstable_waitq;
2368 * cl_cache functions
2370 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2371 void cl_cache_incref(struct cl_client_cache *cache);
2372 void cl_cache_decref(struct cl_client_cache *cache);
2376 /** \defgroup cl_lock cl_lock
2380 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2381 struct cl_lock *lock);
2382 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2383 const struct cl_io *io);
2384 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2385 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2386 const struct lu_device_type *dtype);
2387 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2388 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2389 struct cl_lock *lock, struct cl_sync_io *anchor);
2390 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2394 /** \defgroup cl_io cl_io
2398 int cl_io_init(const struct lu_env *env, struct cl_io *io,
2399 enum cl_io_type iot, struct cl_object *obj);
2400 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
2401 enum cl_io_type iot, struct cl_object *obj);
2402 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
2403 enum cl_io_type iot, loff_t pos, size_t count);
2404 int cl_io_loop(const struct lu_env *env, struct cl_io *io);
2406 void cl_io_fini(const struct lu_env *env, struct cl_io *io);
2407 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io);
2408 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io);
2409 int cl_io_lock(const struct lu_env *env, struct cl_io *io);
2410 void cl_io_unlock(const struct lu_env *env, struct cl_io *io);
2411 int cl_io_start(const struct lu_env *env, struct cl_io *io);
2412 void cl_io_end(const struct lu_env *env, struct cl_io *io);
2413 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
2414 struct cl_io_lock_link *link);
2415 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2416 struct cl_lock_descr *descr);
2417 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
2418 struct cl_page *page);
2419 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
2420 enum cl_req_type iot, struct cl_2queue *queue);
2421 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
2422 enum cl_req_type iot, struct cl_2queue *queue,
2424 int cl_io_commit_async(const struct lu_env *env, struct cl_io *io,
2425 struct cl_page_list *queue, int from, int to,
2427 int cl_io_is_going(const struct lu_env *env);
2430 * True, iff \a io is an O_APPEND write(2).
2432 static inline int cl_io_is_append(const struct cl_io *io)
2434 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2437 static inline int cl_io_is_sync_write(const struct cl_io *io)
2439 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2442 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2444 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2448 * True, iff \a io is a truncate(2).
2450 static inline int cl_io_is_trunc(const struct cl_io *io)
2452 return io->ci_type == CIT_SETATTR &&
2453 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2456 struct cl_io *cl_io_top(struct cl_io *io);
2458 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2460 typeof(foo_io) __foo_io = (foo_io); \
2462 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2463 memset(&__foo_io->base + 1, 0, \
2464 sizeof(*__foo_io) - sizeof(__foo_io->base)); \
2469 /** \defgroup cl_page_list cl_page_list
2474 * Last page in the page list.
2476 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2478 LASSERT(plist->pl_nr > 0);
2479 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2482 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2484 LASSERT(plist->pl_nr > 0);
2485 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2489 * Iterate over pages in a page list.
2491 #define cl_page_list_for_each(page, list) \
2492 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2495 * Iterate over pages in a page list, taking possible removals into account.
2497 #define cl_page_list_for_each_safe(page, temp, list) \
2498 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2500 void cl_page_list_init(struct cl_page_list *plist);
2501 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page);
2502 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2503 struct cl_page *page);
2504 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2505 struct cl_page *page);
2506 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head);
2507 void cl_page_list_del(const struct lu_env *env, struct cl_page_list *plist,
2508 struct cl_page *page);
2509 void cl_page_list_disown(const struct lu_env *env,
2510 struct cl_io *io, struct cl_page_list *plist);
2511 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2513 void cl_2queue_init(struct cl_2queue *queue);
2514 void cl_2queue_disown(const struct lu_env *env,
2515 struct cl_io *io, struct cl_2queue *queue);
2516 void cl_2queue_discard(const struct lu_env *env,
2517 struct cl_io *io, struct cl_2queue *queue);
2518 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2519 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2521 /** @} cl_page_list */
2523 /** \defgroup cl_req cl_req
2526 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2527 enum cl_req_type crt, int nr_objects);
2529 void cl_req_page_add(const struct lu_env *env, struct cl_req *req,
2530 struct cl_page *page);
2531 void cl_req_page_done(const struct lu_env *env, struct cl_page *page);
2532 int cl_req_prep(const struct lu_env *env, struct cl_req *req);
2533 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2534 struct cl_req_attr *attr, u64 flags);
2535 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2537 /** \defgroup cl_sync_io cl_sync_io
2542 * Anchor for synchronous transfer. This is allocated on a stack by thread
2543 * doing synchronous transfer, and a pointer to this structure is set up in
2544 * every page submitted for transfer. Transfer completion routine updates
2545 * anchor and wakes up waiting thread when transfer is complete.
2548 /** number of pages yet to be transferred. */
2549 atomic_t csi_sync_nr;
2552 /** barrier of destroy this structure */
2553 atomic_t csi_barrier;
2554 /** completion to be signaled when transfer is complete. */
2555 wait_queue_head_t csi_waitq;
2556 /** callback to invoke when this IO is finished */
2557 void (*csi_end_io)(const struct lu_env *,
2558 struct cl_sync_io *);
2561 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2562 void (*end)(const struct lu_env *, struct cl_sync_io *));
2563 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2565 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2567 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2569 /** @} cl_sync_io */
2573 /** \defgroup cl_env cl_env
2575 * lu_env handling for a client.
2577 * lu_env is an environment within which lustre code executes. Its major part
2578 * is lu_context---a fast memory allocation mechanism that is used to conserve
2579 * precious kernel stack space. Originally lu_env was designed for a server,
2582 * - there is a (mostly) fixed number of threads, and
2584 * - call chains have no non-lustre portions inserted between lustre code.
2586 * On a client both these assumption fails, because every user thread can
2587 * potentially execute lustre code as part of a system call, and lustre calls
2588 * into VFS or MM that call back into lustre.
2590 * To deal with that, cl_env wrapper functions implement the following
2593 * - allocation and destruction of environment is amortized by caching no
2594 * longer used environments instead of destroying them;
2596 * - there is a notion of "current" environment, attached to the kernel
2597 * data structure representing current thread Top-level lustre code
2598 * allocates an environment and makes it current, then calls into
2599 * non-lustre code, that in turn calls lustre back. Low-level lustre
2600 * code thus called can fetch environment created by the top-level code
2601 * and reuse it, avoiding additional environment allocation.
2602 * Right now, three interfaces can attach the cl_env to running thread:
2605 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2607 * \see lu_env, lu_context, lu_context_key
2611 struct cl_env_nest {
2616 struct lu_env *cl_env_get(int *refcheck);
2617 struct lu_env *cl_env_alloc(int *refcheck, __u32 tags);
2618 struct lu_env *cl_env_nested_get(struct cl_env_nest *nest);
2619 void cl_env_put(struct lu_env *env, int *refcheck);
2620 void cl_env_nested_put(struct cl_env_nest *nest, struct lu_env *env);
2621 void *cl_env_reenter(void);
2622 void cl_env_reexit(void *cookie);
2623 void cl_env_implant(struct lu_env *env, int *refcheck);
2624 void cl_env_unplant(struct lu_env *env, int *refcheck);
2625 unsigned int cl_env_cache_purge(unsigned int nr);
2626 struct lu_env *cl_env_percpu_get(void);
2627 void cl_env_percpu_put(struct lu_env *env);
2634 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2636 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2637 struct lu_device_type *ldt,
2638 struct lu_device *next);
2641 int cl_global_init(void);
2642 void cl_global_fini(void);
2644 #endif /* _LINUX_CL_OBJECT_H */