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 <linux/atomic.h>
97 #include "linux/lustre_compat25.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_set)(const struct lu_env *env, struct cl_object *obj,
373 const struct cl_attr *attr, unsigned 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);
401 * Extended header for client object.
403 struct cl_object_header {
404 /** Standard lu_object_header. cl_object::co_lu::lo_header points
407 struct lu_object_header coh_lu;
410 * Parent object. It is assumed that an object has a well-defined
411 * parent, but not a well-defined child (there may be multiple
412 * sub-objects, for the same top-object). cl_object_header::coh_parent
413 * field allows certain code to be written generically, without
414 * limiting possible cl_object layouts unduly.
416 struct cl_object_header *coh_parent;
418 * Protects consistency between cl_attr of parent object and
419 * attributes of sub-objects, that the former is calculated ("merged")
422 * \todo XXX this can be read/write lock if needed.
424 spinlock_t coh_attr_guard;
426 * Size of cl_page + page slices
428 unsigned short coh_page_bufsize;
430 * Number of objects above this one: 0 for a top-object, 1 for its
433 unsigned char coh_nesting;
437 * Helper macro: iterate over all layers of the object \a obj, assigning every
438 * layer top-to-bottom to \a slice.
440 #define cl_object_for_each(slice, obj) \
441 list_for_each_entry((slice), \
442 &(obj)->co_lu.lo_header->loh_layers, \
445 * Helper macro: iterate over all layers of the object \a obj, assigning every
446 * layer bottom-to-top to \a slice.
448 #define cl_object_for_each_reverse(slice, obj) \
449 list_for_each_entry_reverse((slice), \
450 &(obj)->co_lu.lo_header->loh_layers, \
454 #define CL_PAGE_EOF ((pgoff_t)~0ull)
456 /** \addtogroup cl_page cl_page
461 * Layered client page.
463 * cl_page: represents a portion of a file, cached in the memory. All pages
464 * of the given file are of the same size, and are kept in the radix tree
465 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
466 * of the top-level file object are first class cl_objects, they have their
467 * own radix trees of pages and hence page is implemented as a sequence of
468 * struct cl_pages's, linked into double-linked list through
469 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
470 * corresponding radix tree at the corresponding logical offset.
472 * cl_page is associated with VM page of the hosting environment (struct
473 * page in Linux kernel, for example), struct page. It is assumed, that this
474 * association is implemented by one of cl_page layers (top layer in the
475 * current design) that
477 * - intercepts per-VM-page call-backs made by the environment (e.g.,
480 * - translates state (page flag bits) and locking between lustre and
483 * The association between cl_page and struct page is immutable and
484 * established when cl_page is created.
486 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
487 * this io an exclusive access to this page w.r.t. other io attempts and
488 * various events changing page state (such as transfer completion, or
489 * eviction of the page from the memory). Note, that in general cl_io
490 * cannot be identified with a particular thread, and page ownership is not
491 * exactly equal to the current thread holding a lock on the page. Layer
492 * implementing association between cl_page and struct page has to implement
493 * ownership on top of available synchronization mechanisms.
495 * While lustre client maintains the notion of an page ownership by io,
496 * hosting MM/VM usually has its own page concurrency control
497 * mechanisms. For example, in Linux, page access is synchronized by the
498 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
499 * takes care to acquire and release such locks as necessary around the
500 * calls to the file system methods (->readpage(), ->prepare_write(),
501 * ->commit_write(), etc.). This leads to the situation when there are two
502 * different ways to own a page in the client:
504 * - client code explicitly and voluntary owns the page (cl_page_own());
506 * - VM locks a page and then calls the client, that has "to assume"
507 * the ownership from the VM (cl_page_assume()).
509 * Dual methods to release ownership are cl_page_disown() and
510 * cl_page_unassume().
512 * cl_page is reference counted (cl_page::cp_ref). When reference counter
513 * drops to 0, the page is returned to the cache, unless it is in
514 * cl_page_state::CPS_FREEING state, in which case it is immediately
517 * The general logic guaranteeing the absence of "existential races" for
518 * pages is the following:
520 * - there are fixed known ways for a thread to obtain a new reference
523 * - by doing a lookup in the cl_object radix tree, protected by the
526 * - by starting from VM-locked struct page and following some
527 * hosting environment method (e.g., following ->private pointer in
528 * the case of Linux kernel), see cl_vmpage_page();
530 * - when the page enters cl_page_state::CPS_FREEING state, all these
531 * ways are severed with the proper synchronization
532 * (cl_page_delete());
534 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
537 * - no new references to the page in cl_page_state::CPS_FREEING state
538 * are allowed (checked in cl_page_get()).
540 * Together this guarantees that when last reference to a
541 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
542 * page, as neither references to it can be acquired at that point, nor
545 * cl_page is a state machine. States are enumerated in enum
546 * cl_page_state. Possible state transitions are enumerated in
547 * cl_page_state_set(). State transition process (i.e., actual changing of
548 * cl_page::cp_state field) is protected by the lock on the underlying VM
551 * Linux Kernel implementation.
553 * Binding between cl_page and struct page (which is a typedef for
554 * struct page) is implemented in the vvp layer. cl_page is attached to the
555 * ->private pointer of the struct page, together with the setting of
556 * PG_private bit in page->flags, and acquiring additional reference on the
557 * struct page (much like struct buffer_head, or any similar file system
558 * private data structures).
560 * PG_locked lock is used to implement both ownership and transfer
561 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
562 * states. No additional references are acquired for the duration of the
565 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
566 * write-out is "protected" by the special PG_writeback bit.
570 * States of cl_page. cl_page.c assumes particular order here.
572 * The page state machine is rather crude, as it doesn't recognize finer page
573 * states like "dirty" or "up to date". This is because such states are not
574 * always well defined for the whole stack (see, for example, the
575 * implementation of the read-ahead, that hides page up-to-dateness to track
576 * cache hits accurately). Such sub-states are maintained by the layers that
577 * are interested in them.
581 * Page is in the cache, un-owned. Page leaves cached state in the
584 * - [cl_page_state::CPS_OWNED] io comes across the page and
587 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
588 * req-formation engine decides that it wants to include this page
589 * into an cl_req being constructed, and yanks it from the cache;
591 * - [cl_page_state::CPS_FREEING] VM callback is executed to
592 * evict the page form the memory;
594 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
598 * Page is exclusively owned by some cl_io. Page may end up in this
599 * state as a result of
601 * - io creating new page and immediately owning it;
603 * - [cl_page_state::CPS_CACHED] io finding existing cached page
606 * - [cl_page_state::CPS_OWNED] io finding existing owned page
607 * and waiting for owner to release the page;
609 * Page leaves owned state in the following cases:
611 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
612 * the cache, doing nothing;
614 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
617 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
618 * transfer for this page;
620 * - [cl_page_state::CPS_FREEING] io decides to destroy this
621 * page (e.g., as part of truncate or extent lock cancellation).
623 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
627 * Page is being written out, as a part of a transfer. This state is
628 * entered when req-formation logic decided that it wants this page to
629 * be sent through the wire _now_. Specifically, it means that once
630 * this state is achieved, transfer completion handler (with either
631 * success or failure indication) is guaranteed to be executed against
632 * this page independently of any locks and any scheduling decisions
633 * made by the hosting environment (that effectively means that the
634 * page is never put into cl_page_state::CPS_PAGEOUT state "in
635 * advance". This property is mentioned, because it is important when
636 * reasoning about possible dead-locks in the system). The page can
637 * enter this state as a result of
639 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
640 * write-out of this page, or
642 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
643 * that it has enough dirty pages cached to issue a "good"
646 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
647 * is completed---it is moved into cl_page_state::CPS_CACHED state.
649 * Underlying VM page is locked for the duration of transfer.
651 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
655 * Page is being read in, as a part of a transfer. This is quite
656 * similar to the cl_page_state::CPS_PAGEOUT state, except that
657 * read-in is always "immediate"---there is no such thing a sudden
658 * construction of read cl_req from cached, presumably not up to date,
661 * Underlying VM page is locked for the duration of transfer.
663 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
667 * Page is being destroyed. This state is entered when client decides
668 * that page has to be deleted from its host object, as, e.g., a part
671 * Once this state is reached, there is no way to escape it.
673 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
680 /** Host page, the page is from the host inode which the cl_page
685 /** Transient page, the transient cl_page is used to bind a cl_page
686 * to vmpage which is not belonging to the same object of cl_page.
687 * it is used in DirectIO and lockless IO.
693 * Fields are protected by the lock on struct page, except for atomics and
696 * \invariant Data type invariants are in cl_page_invariant(). Basically:
697 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
698 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
699 * cl_page::cp_owner (when set).
702 /** Reference counter. */
704 /** An object this page is a part of. Immutable after creation. */
705 struct cl_object *cp_obj;
707 struct page *cp_vmpage;
708 /** Linkage of pages within group. Pages must be owned */
709 struct list_head cp_batch;
710 /** List of slices. Immutable after creation. */
711 struct list_head cp_layers;
712 /** Linkage of pages within cl_req. */
713 struct list_head cp_flight;
715 * Page state. This field is const to avoid accidental update, it is
716 * modified only internally within cl_page.c. Protected by a VM lock.
718 const enum cl_page_state cp_state;
720 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
723 enum cl_page_type cp_type;
726 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
727 * by sub-io. Protected by a VM lock.
729 struct cl_io *cp_owner;
731 * Owning IO request in cl_page_state::CPS_PAGEOUT and
732 * cl_page_state::CPS_PAGEIN states. This field is maintained only in
733 * the top-level pages. Protected by a VM lock.
735 struct cl_req *cp_req;
736 /** List of references to this page, for debugging. */
737 struct lu_ref cp_reference;
738 /** Link to an object, for debugging. */
739 struct lu_ref_link cp_obj_ref;
740 /** Link to a queue, for debugging. */
741 struct lu_ref_link cp_queue_ref;
742 /** Assigned if doing a sync_io */
743 struct cl_sync_io *cp_sync_io;
747 * Per-layer part of cl_page.
749 * \see vvp_page, lov_page, osc_page
751 struct cl_page_slice {
752 struct cl_page *cpl_page;
755 * Object slice corresponding to this page slice. Immutable after
758 struct cl_object *cpl_obj;
759 const struct cl_page_operations *cpl_ops;
760 /** Linkage into cl_page::cp_layers. Immutable after creation. */
761 struct list_head cpl_linkage;
765 * Lock mode. For the client extent locks.
776 * Requested transfer type.
786 * Per-layer page operations.
788 * Methods taking an \a io argument are for the activity happening in the
789 * context of given \a io. Page is assumed to be owned by that io, except for
790 * the obvious cases (like cl_page_operations::cpo_own()).
792 * \see vvp_page_ops, lov_page_ops, osc_page_ops
794 struct cl_page_operations {
796 * cl_page<->struct page methods. Only one layer in the stack has to
797 * implement these. Current code assumes that this functionality is
798 * provided by the topmost layer, see cl_page_disown0() as an example.
802 * Called when \a io acquires this page into the exclusive
803 * ownership. When this method returns, it is guaranteed that the is
804 * not owned by other io, and no transfer is going on against
808 * \see vvp_page_own(), lov_page_own()
810 int (*cpo_own)(const struct lu_env *env,
811 const struct cl_page_slice *slice,
812 struct cl_io *io, int nonblock);
813 /** Called when ownership it yielded. Optional.
815 * \see cl_page_disown()
816 * \see vvp_page_disown()
818 void (*cpo_disown)(const struct lu_env *env,
819 const struct cl_page_slice *slice, struct cl_io *io);
821 * Called for a page that is already "owned" by \a io from VM point of
824 * \see cl_page_assume()
825 * \see vvp_page_assume(), lov_page_assume()
827 void (*cpo_assume)(const struct lu_env *env,
828 const struct cl_page_slice *slice, struct cl_io *io);
829 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
830 * bottom-to-top when IO releases a page without actually unlocking
833 * \see cl_page_unassume()
834 * \see vvp_page_unassume()
836 void (*cpo_unassume)(const struct lu_env *env,
837 const struct cl_page_slice *slice,
840 * Announces whether the page contains valid data or not by \a uptodate.
842 * \see cl_page_export()
843 * \see vvp_page_export()
845 void (*cpo_export)(const struct lu_env *env,
846 const struct cl_page_slice *slice, int uptodate);
848 * Checks whether underlying VM page is locked (in the suitable
849 * sense). Used for assertions.
851 * \retval -EBUSY: page is protected by a lock of a given mode;
852 * \retval -ENODATA: page is not protected by a lock;
853 * \retval 0: this layer cannot decide. (Should never happen.)
855 int (*cpo_is_vmlocked)(const struct lu_env *env,
856 const struct cl_page_slice *slice);
862 * Called when page is truncated from the object. Optional.
864 * \see cl_page_discard()
865 * \see vvp_page_discard(), osc_page_discard()
867 void (*cpo_discard)(const struct lu_env *env,
868 const struct cl_page_slice *slice,
871 * Called when page is removed from the cache, and is about to being
872 * destroyed. Optional.
874 * \see cl_page_delete()
875 * \see vvp_page_delete(), osc_page_delete()
877 void (*cpo_delete)(const struct lu_env *env,
878 const struct cl_page_slice *slice);
879 /** Destructor. Frees resources and slice itself. */
880 void (*cpo_fini)(const struct lu_env *env,
881 struct cl_page_slice *slice);
884 * Checks whether the page is protected by a cl_lock. This is a
885 * per-layer method, because certain layers have ways to check for the
886 * lock much more efficiently than through the generic locks scan, or
887 * implement locking mechanisms separate from cl_lock, e.g.,
888 * LL_FILE_GROUP_LOCKED in vvp. If \a pending is true, check for locks
889 * being canceled, or scheduled for cancellation as soon as the last
890 * user goes away, too.
892 * \retval -EBUSY: page is protected by a lock of a given mode;
893 * \retval -ENODATA: page is not protected by a lock;
894 * \retval 0: this layer cannot decide.
896 * \see cl_page_is_under_lock()
898 int (*cpo_is_under_lock)(const struct lu_env *env,
899 const struct cl_page_slice *slice,
900 struct cl_io *io, pgoff_t *max);
903 * Optional debugging helper. Prints given page slice.
905 * \see cl_page_print()
907 int (*cpo_print)(const struct lu_env *env,
908 const struct cl_page_slice *slice,
909 void *cookie, lu_printer_t p);
913 * Transfer methods. See comment on cl_req for a description of
914 * transfer formation and life-cycle.
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 * cl_req 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(struct cl_page *page)
1042 LASSERT(page->cp_vmpage);
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
1073 * Extent locking on the client.
1077 * The locking model of the new client code is built around
1081 * data-type representing an extent lock on a regular file. cl_lock is a
1082 * layered object (much like cl_object and cl_page), it consists of a header
1083 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1084 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1086 * Typical cl_lock consists of the two layers:
1088 * - vvp_lock (vvp specific data), and
1089 * - lov_lock (lov specific data).
1091 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1092 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1094 * - lovsub_lock, and
1097 * Each sub-lock is associated with a cl_object (representing stripe
1098 * sub-object or the file to which top-level cl_lock is associated to), and is
1099 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1100 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1101 * is different from cl_page, that doesn't fan out (there is usually exactly
1102 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1103 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1107 * cl_lock is a cacheless data container for the requirements of locks to
1108 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1111 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1112 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1114 * INTERFACE AND USAGE
1116 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1117 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1118 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1119 * consists of multiple sub cl_locks, each sub locks will be enqueued
1120 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1121 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1124 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1125 * method will be called for each layer to release the resource held by this
1126 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1127 * clo_enqueue time, is released.
1129 * LDLM lock can only be canceled if there is no cl_lock using it.
1131 * Overall process of the locking during IO operation is as following:
1133 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1134 * is called on each layer. Responsibility of this method is to add locks,
1135 * needed by a given layer into cl_io.ci_lockset.
1137 * - once locks for all layers were collected, they are sorted to avoid
1138 * dead-locks (cl_io_locks_sort()), and enqueued.
1140 * - when all locks are acquired, IO is performed;
1142 * - locks are released after IO is complete.
1144 * Striping introduces major additional complexity into locking. The
1145 * fundamental problem is that it is generally unsafe to actively use (hold)
1146 * two locks on the different OST servers at the same time, as this introduces
1147 * inter-server dependency and can lead to cascading evictions.
1149 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1150 * that no multi-stripe locks are taken (note that this design abandons POSIX
1151 * read/write semantics). Such pieces ideally can be executed concurrently. At
1152 * the same time, certain types of IO cannot be sub-divived, without
1153 * sacrificing correctness. This includes:
1155 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1158 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1160 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1161 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1162 * has to be held together with the usual lock on [offset, offset + count].
1164 * Interaction with DLM
1166 * In the expected setup, cl_lock is ultimately backed up by a collection of
1167 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1168 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1169 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1170 * description of interaction with DLM.
1176 struct cl_lock_descr {
1177 /** Object this lock is granted for. */
1178 struct cl_object *cld_obj;
1179 /** Index of the first page protected by this lock. */
1181 /** Index of the last page (inclusive) protected by this lock. */
1183 /** Group ID, for group lock */
1186 enum cl_lock_mode cld_mode;
1188 * flags to enqueue lock. A combination of bit-flags from
1189 * enum cl_enq_flags.
1191 __u32 cld_enq_flags;
1194 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1195 #define PDESCR(descr) \
1196 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1197 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1199 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1202 * Layered client lock.
1205 /** List of slices. Immutable after creation. */
1206 struct list_head cll_layers;
1207 /** lock attribute, extent, cl_object, etc. */
1208 struct cl_lock_descr cll_descr;
1212 * Per-layer part of cl_lock
1214 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1216 struct cl_lock_slice {
1217 struct cl_lock *cls_lock;
1218 /** Object slice corresponding to this lock slice. Immutable after
1221 struct cl_object *cls_obj;
1222 const struct cl_lock_operations *cls_ops;
1223 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1224 struct list_head cls_linkage;
1229 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1231 struct cl_lock_operations {
1234 * Attempts to enqueue the lock. Called top-to-bottom.
1236 * \retval 0 this layer has enqueued the lock successfully
1237 * \retval >0 this layer has enqueued the lock, but need to wait on
1238 * @anchor for resources
1239 * \retval -ve failure
1241 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1242 * \see osc_lock_enqueue()
1244 int (*clo_enqueue)(const struct lu_env *env,
1245 const struct cl_lock_slice *slice,
1246 struct cl_io *io, struct cl_sync_io *anchor);
1248 * Cancel a lock, release its DLM lock ref, while does not cancel the
1251 void (*clo_cancel)(const struct lu_env *env,
1252 const struct cl_lock_slice *slice);
1255 * Destructor. Frees resources and the slice.
1257 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1258 * \see osc_lock_fini()
1260 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1262 * Optional debugging helper. Prints given lock slice.
1264 int (*clo_print)(const struct lu_env *env,
1265 void *cookie, lu_printer_t p,
1266 const struct cl_lock_slice *slice);
1269 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1271 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1273 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1274 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1275 CDEBUG(mask, format, ## __VA_ARGS__); \
1279 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1283 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1289 /** \addtogroup cl_page_list cl_page_list
1290 * Page list used to perform collective operations on a group of pages.
1292 * Pages are added to the list one by one. cl_page_list acquires a reference
1293 * for every page in it. Page list is used to perform collective operations on
1296 * - submit pages for an immediate transfer,
1298 * - own pages on behalf of certain io (waiting for each page in turn),
1302 * When list is finalized, it releases references on all pages it still has.
1304 * \todo XXX concurrency control.
1308 struct cl_page_list {
1310 struct list_head pl_pages;
1311 struct task_struct *pl_owner;
1315 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1316 * contains an incoming page list and an outgoing page list.
1319 struct cl_page_list c2_qin;
1320 struct cl_page_list c2_qout;
1323 /** @} cl_page_list */
1325 /** \addtogroup cl_io cl_io
1331 * cl_io represents a high level I/O activity like
1332 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1335 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1336 * important distinction. We want to minimize number of calls to the allocator
1337 * in the fast path, e.g., in the case of read(2) when everything is cached:
1338 * client already owns the lock over region being read, and data are cached
1339 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1340 * per-layer io state is stored in the session, associated with the io, see
1341 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1342 * by using free-lists, see cl_env_get().
1344 * There is a small predefined number of possible io types, enumerated in enum
1347 * cl_io is a state machine, that can be advanced concurrently by the multiple
1348 * threads. It is up to these threads to control the concurrency and,
1349 * specifically, to detect when io is done, and its state can be safely
1352 * For read/write io overall execution plan is as following:
1354 * (0) initialize io state through all layers;
1356 * (1) loop: prepare chunk of work to do
1358 * (2) call all layers to collect locks they need to process current chunk
1360 * (3) sort all locks to avoid dead-locks, and acquire them
1362 * (4) process the chunk: call per-page methods
1363 * (cl_io_operations::cio_read_page() for read,
1364 * cl_io_operations::cio_prepare_write(),
1365 * cl_io_operations::cio_commit_write() for write)
1371 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1372 * address allocation efficiency issues mentioned above), and returns with the
1373 * special error condition from per-page method when current sub-io has to
1374 * block. This causes io loop to be repeated, and lov switches to the next
1375 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1380 /** read system call */
1382 /** write system call */
1384 /** truncate, utime system calls */
1387 * page fault handling
1391 * fsync system call handling
1392 * To write out a range of file
1396 * Miscellaneous io. This is used for occasional io activity that
1397 * doesn't fit into other types. Currently this is used for:
1399 * - cancellation of an extent lock. This io exists as a context
1400 * to write dirty pages from under the lock being canceled back
1403 * - VM induced page write-out. An io context for writing page out
1404 * for memory cleansing;
1406 * - glimpse. An io context to acquire glimpse lock.
1408 * - grouplock. An io context to acquire group lock.
1410 * CIT_MISC io is used simply as a context in which locks and pages
1411 * are manipulated. Such io has no internal "process", that is,
1412 * cl_io_loop() is never called for it.
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 * Per-layer io operations.
1467 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1469 struct cl_io_operations {
1471 * Vector of io state transition methods for every io type.
1473 * \see cl_page_operations::io
1477 * Prepare io iteration at a given layer.
1479 * Called top-to-bottom at the beginning of each iteration of
1480 * "io loop" (if it makes sense for this type of io). Here
1481 * layer selects what work it will do during this iteration.
1483 * \see cl_io_operations::cio_iter_fini()
1485 int (*cio_iter_init)(const struct lu_env *env,
1486 const struct cl_io_slice *slice);
1488 * Finalize io iteration.
1490 * Called bottom-to-top at the end of each iteration of "io
1491 * loop". Here layers can decide whether IO has to be
1494 * \see cl_io_operations::cio_iter_init()
1496 void (*cio_iter_fini)(const struct lu_env *env,
1497 const struct cl_io_slice *slice);
1499 * Collect locks for the current iteration of io.
1501 * Called top-to-bottom to collect all locks necessary for
1502 * this iteration. This methods shouldn't actually enqueue
1503 * anything, instead it should post a lock through
1504 * cl_io_lock_add(). Once all locks are collected, they are
1505 * sorted and enqueued in the proper order.
1507 int (*cio_lock)(const struct lu_env *env,
1508 const struct cl_io_slice *slice);
1510 * Finalize unlocking.
1512 * Called bottom-to-top to finish layer specific unlocking
1513 * functionality, after generic code released all locks
1514 * acquired by cl_io_operations::cio_lock().
1516 void (*cio_unlock)(const struct lu_env *env,
1517 const struct cl_io_slice *slice);
1519 * Start io iteration.
1521 * Once all locks are acquired, called top-to-bottom to
1522 * commence actual IO. In the current implementation,
1523 * top-level vvp_io_{read,write}_start() does all the work
1524 * synchronously by calling generic_file_*(), so other layers
1525 * are called when everything is done.
1527 int (*cio_start)(const struct lu_env *env,
1528 const struct cl_io_slice *slice);
1530 * Called top-to-bottom at the end of io loop. Here layer
1531 * might wait for an unfinished asynchronous io.
1533 void (*cio_end)(const struct lu_env *env,
1534 const struct cl_io_slice *slice);
1536 * Called bottom-to-top to notify layers that read/write IO
1537 * iteration finished, with \a nob bytes transferred.
1539 void (*cio_advance)(const struct lu_env *env,
1540 const struct cl_io_slice *slice,
1543 * Called once per io, bottom-to-top to release io resources.
1545 void (*cio_fini)(const struct lu_env *env,
1546 const struct cl_io_slice *slice);
1550 * Submit pages from \a queue->c2_qin for IO, and move
1551 * successfully submitted pages into \a queue->c2_qout. Return
1552 * non-zero if failed to submit even the single page. If
1553 * submission failed after some pages were moved into \a
1554 * queue->c2_qout, completion callback with non-zero ioret is
1557 int (*cio_submit)(const struct lu_env *env,
1558 const struct cl_io_slice *slice,
1559 enum cl_req_type crt,
1560 struct cl_2queue *queue);
1562 * Queue async page for write.
1563 * The difference between cio_submit and cio_queue is that
1564 * cio_submit is for urgent request.
1566 int (*cio_commit_async)(const struct lu_env *env,
1567 const struct cl_io_slice *slice,
1568 struct cl_page_list *queue, int from, int to,
1571 * Read missing page.
1573 * Called by a top-level cl_io_operations::op[CIT_READ]::cio_start()
1574 * method, when it hits not-up-to-date page in the range. Optional.
1576 * \pre io->ci_type == CIT_READ
1578 int (*cio_read_page)(const struct lu_env *env,
1579 const struct cl_io_slice *slice,
1580 const struct cl_page_slice *page);
1582 * Optional debugging helper. Print given io slice.
1584 int (*cio_print)(const struct lu_env *env, void *cookie,
1585 lu_printer_t p, const struct cl_io_slice *slice);
1589 * Flags to lock enqueue procedure.
1594 * instruct server to not block, if conflicting lock is found. Instead
1595 * -EWOULDBLOCK is returned immediately.
1597 CEF_NONBLOCK = 0x00000001,
1599 * take lock asynchronously (out of order), as it cannot
1600 * deadlock. This is for LDLM_FL_HAS_INTENT locks used for glimpsing.
1602 CEF_ASYNC = 0x00000002,
1604 * tell the server to instruct (though a flag in the blocking ast) an
1605 * owner of the conflicting lock, that it can drop dirty pages
1606 * protected by this lock, without sending them to the server.
1608 CEF_DISCARD_DATA = 0x00000004,
1610 * tell the sub layers that it must be a `real' lock. This is used for
1611 * mmapped-buffer locks and glimpse locks that must be never converted
1612 * into lockless mode.
1614 * \see vvp_mmap_locks(), cl_glimpse_lock().
1616 CEF_MUST = 0x00000008,
1618 * tell the sub layers that never request a `real' lock. This flag is
1619 * not used currently.
1621 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1622 * conversion policy: ci_lockreq describes generic information of lock
1623 * requirement for this IO, especially for locks which belong to the
1624 * object doing IO; however, lock itself may have precise requirements
1625 * that are described by the enqueue flags.
1627 CEF_NEVER = 0x00000010,
1629 * for async glimpse lock.
1631 CEF_AGL = 0x00000020,
1633 * enqueue a lock to test DLM lock existence.
1635 CEF_PEEK = 0x00000040,
1637 * mask of enq_flags.
1639 CEF_MASK = 0x0000007f,
1643 * Link between lock and io. Intermediate structure is needed, because the
1644 * same lock can be part of multiple io's simultaneously.
1646 struct cl_io_lock_link {
1647 /** linkage into one of cl_lockset lists. */
1648 struct list_head cill_linkage;
1649 struct cl_lock cill_lock;
1650 /** optional destructor */
1651 void (*cill_fini)(const struct lu_env *env,
1652 struct cl_io_lock_link *link);
1654 #define cill_descr cill_lock.cll_descr
1657 * Lock-set represents a collection of locks, that io needs at a
1658 * time. Generally speaking, client tries to avoid holding multiple locks when
1661 * - holding extent locks over multiple ost's introduces the danger of
1662 * "cascading timeouts";
1664 * - holding multiple locks over the same ost is still dead-lock prone,
1665 * see comment in osc_lock_enqueue(),
1667 * but there are certain situations where this is unavoidable:
1669 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1671 * - truncate has to take [new-size, EOF] lock for correctness;
1673 * - SNS has to take locks across full stripe for correctness;
1675 * - in the case when user level buffer, supplied to {read,write}(file0),
1676 * is a part of a memory mapped lustre file, client has to take a dlm
1677 * locks on file0, and all files that back up the buffer (or a part of
1678 * the buffer, that is being processed in the current chunk, in any
1679 * case, there are situations where at least 2 locks are necessary).
1681 * In such cases we at least try to take locks in the same consistent
1682 * order. To this end, all locks are first collected, then sorted, and then
1686 /** locks to be acquired. */
1687 struct list_head cls_todo;
1688 /** locks acquired. */
1689 struct list_head cls_done;
1693 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1694 * but 'req' is always to be thought as 'request' :-)
1696 enum cl_io_lock_dmd {
1697 /** Always lock data (e.g., O_APPEND). */
1699 /** Layers are free to decide between local and global locking. */
1701 /** Never lock: there is no cache (e.g., lockless IO). */
1705 enum cl_fsync_mode {
1706 /** start writeback, do not wait for them to finish */
1708 /** start writeback and wait for them to finish */
1710 /** discard all of dirty pages in a specific file range */
1711 CL_FSYNC_DISCARD = 2,
1712 /** start writeback and make sure they have reached storage before
1713 * return. OST_SYNC RPC must be issued and finished
1718 struct cl_io_rw_common {
1727 * cl_io is shared by all threads participating in this IO (in current
1728 * implementation only one thread advances IO, but parallel IO design and
1729 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1730 * is up to these threads to serialize their activities, including updates to
1731 * mutable cl_io fields.
1734 /** type of this IO. Immutable after creation. */
1735 enum cl_io_type ci_type;
1736 /** current state of cl_io state machine. */
1737 enum cl_io_state ci_state;
1738 /** main object this io is against. Immutable after creation. */
1739 struct cl_object *ci_obj;
1741 * Upper layer io, of which this io is a part of. Immutable after
1744 struct cl_io *ci_parent;
1745 /** List of slices. Immutable after creation. */
1746 struct list_head ci_layers;
1747 /** list of locks (to be) acquired by this io. */
1748 struct cl_lockset ci_lockset;
1749 /** lock requirements, this is just a help info for sublayers. */
1750 enum cl_io_lock_dmd ci_lockreq;
1753 struct cl_io_rw_common rd;
1756 struct cl_io_rw_common wr;
1760 struct cl_io_rw_common ci_rw;
1761 struct cl_setattr_io {
1762 struct ost_lvb sa_attr;
1763 unsigned int sa_valid;
1765 struct cl_fault_io {
1766 /** page index within file. */
1768 /** bytes valid byte on a faulted page. */
1770 /** writable page? for nopage() only */
1772 /** page of an executable? */
1774 /** page_mkwrite() */
1776 /** resulting page */
1777 struct cl_page *ft_page;
1779 struct cl_fsync_io {
1782 /** file system level fid */
1783 struct lu_fid *fi_fid;
1784 enum cl_fsync_mode fi_mode;
1785 /* how many pages were written/discarded */
1786 unsigned int fi_nr_written;
1789 struct cl_2queue ci_queue;
1792 unsigned int ci_continue:1,
1794 * This io has held grouplock, to inform sublayers that
1795 * don't do lockless i/o.
1799 * The whole IO need to be restarted because layout has been changed
1803 * to not refresh layout - the IO issuer knows that the layout won't
1804 * change(page operations, layout change causes all page to be
1805 * discarded), or it doesn't matter if it changes(sync).
1809 * Check if layout changed after the IO finishes. Mainly for HSM
1810 * requirement. If IO occurs to openning files, it doesn't need to
1811 * verify layout because HSM won't release openning files.
1812 * Right now, only two operations need to verify layout: glimpse
1817 * file is released, restore has to to be triggered by vvp layer
1819 ci_restore_needed:1,
1825 * Number of pages owned by this IO. For invariant checking.
1827 unsigned ci_owned_nr;
1832 /** \addtogroup cl_req cl_req
1838 * There are two possible modes of transfer initiation on the client:
1840 * - immediate transfer: this is started when a high level io wants a page
1841 * or a collection of pages to be transferred right away. Examples:
1842 * read-ahead, synchronous read in the case of non-page aligned write,
1843 * page write-out as a part of extent lock cancellation, page write-out
1844 * as a part of memory cleansing. Immediate transfer can be both
1845 * cl_req_type::CRT_READ and cl_req_type::CRT_WRITE;
1847 * - opportunistic transfer (cl_req_type::CRT_WRITE only), that happens
1848 * when io wants to transfer a page to the server some time later, when
1849 * it can be done efficiently. Example: pages dirtied by the write(2)
1852 * In any case, transfer takes place in the form of a cl_req, which is a
1853 * representation for a network RPC.
1855 * Pages queued for an opportunistic transfer are cached until it is decided
1856 * that efficient RPC can be composed of them. This decision is made by "a
1857 * req-formation engine", currently implemented as a part of osc
1858 * layer. Req-formation depends on many factors: the size of the resulting
1859 * RPC, whether or not multi-object RPCs are supported by the server,
1860 * max-rpc-in-flight limitations, size of the dirty cache, etc.
1862 * For the immediate transfer io submits a cl_page_list, that req-formation
1863 * engine slices into cl_req's, possibly adding cached pages to some of
1864 * the resulting req's.
1866 * Whenever a page from cl_page_list is added to a newly constructed req, its
1867 * cl_page_operations::cpo_prep() layer methods are called. At that moment,
1868 * page state is atomically changed from cl_page_state::CPS_OWNED to
1869 * cl_page_state::CPS_PAGEOUT or cl_page_state::CPS_PAGEIN, cl_page::cp_owner
1870 * is zeroed, and cl_page::cp_req is set to the
1871 * req. cl_page_operations::cpo_prep() method at the particular layer might
1872 * return -EALREADY to indicate that it does not need to submit this page
1873 * at all. This is possible, for example, if page, submitted for read,
1874 * became up-to-date in the meantime; and for write, the page don't have
1875 * dirty bit marked. \see cl_io_submit_rw()
1877 * Whenever a cached page is added to a newly constructed req, its
1878 * cl_page_operations::cpo_make_ready() layer methods are called. At that
1879 * moment, page state is atomically changed from cl_page_state::CPS_CACHED to
1880 * cl_page_state::CPS_PAGEOUT, and cl_page::cp_req is set to
1881 * req. cl_page_operations::cpo_make_ready() method at the particular layer
1882 * might return -EAGAIN to indicate that this page is not eligible for the
1883 * transfer right now.
1887 * Plan is to divide transfers into "priority bands" (indicated when
1888 * submitting cl_page_list, and queuing a page for the opportunistic transfer)
1889 * and allow glueing of cached pages to immediate transfers only within single
1890 * band. This would make high priority transfers (like lock cancellation or
1891 * memory pressure induced write-out) really high priority.
1896 * Per-transfer attributes.
1898 struct cl_req_attr {
1899 /** Generic attributes for the server consumption. */
1900 struct obdo *cra_oa;
1902 char cra_jobid[LUSTRE_JOBID_SIZE];
1906 * Transfer request operations definable at every layer.
1908 * Concurrency: transfer formation engine synchronizes calls to all transfer
1911 struct cl_req_operations {
1913 * Invoked top-to-bottom by cl_req_prep() when transfer formation is
1914 * complete (all pages are added).
1916 * \see osc_req_prep()
1918 int (*cro_prep)(const struct lu_env *env,
1919 const struct cl_req_slice *slice);
1921 * Called top-to-bottom to fill in \a oa fields. This is called twice
1922 * with different flags, see bug 10150 and osc_build_req().
1924 * \param obj an object from cl_req which attributes are to be set in
1927 * \param oa struct obdo where attributes are placed
1929 * \param flags \a oa fields to be filled.
1931 void (*cro_attr_set)(const struct lu_env *env,
1932 const struct cl_req_slice *slice,
1933 const struct cl_object *obj,
1934 struct cl_req_attr *attr, u64 flags);
1936 * Called top-to-bottom from cl_req_completion() to notify layers that
1937 * transfer completed. Has to free all state allocated by
1938 * cl_device_operations::cdo_req_init().
1940 void (*cro_completion)(const struct lu_env *env,
1941 const struct cl_req_slice *slice, int ioret);
1945 * A per-object state that (potentially multi-object) transfer request keeps.
1948 /** object itself */
1949 struct cl_object *ro_obj;
1950 /** reference to cl_req_obj::ro_obj. For debugging. */
1951 struct lu_ref_link ro_obj_ref;
1952 /* something else? Number of pages for a given object? */
1958 * Transfer requests are not reference counted, because IO sub-system owns
1959 * them exclusively and knows when to free them.
1963 * cl_req is created by cl_req_alloc() that calls
1964 * cl_device_operations::cdo_req_init() device methods to allocate per-req
1965 * state in every layer.
1967 * Then pages are added (cl_req_page_add()), req keeps track of all objects it
1968 * contains pages for.
1970 * Once all pages were collected, cl_page_operations::cpo_prep() method is
1971 * called top-to-bottom. At that point layers can modify req, let it pass, or
1972 * deny it completely. This is to support things like SNS that have transfer
1973 * ordering requirements invisible to the individual req-formation engine.
1975 * On transfer completion (or transfer timeout, or failure to initiate the
1976 * transfer of an allocated req), cl_req_operations::cro_completion() method
1977 * is called, after execution of cl_page_operations::cpo_completion() of all
1981 enum cl_req_type crq_type;
1982 /** A list of pages being transferred */
1983 struct list_head crq_pages;
1984 /** Number of pages in cl_req::crq_pages */
1985 unsigned crq_nrpages;
1986 /** An array of objects which pages are in ->crq_pages */
1987 struct cl_req_obj *crq_o;
1988 /** Number of elements in cl_req::crq_objs[] */
1989 unsigned crq_nrobjs;
1990 struct list_head crq_layers;
1994 * Per-layer state for request.
1996 struct cl_req_slice {
1997 struct cl_req *crs_req;
1998 struct cl_device *crs_dev;
1999 struct list_head crs_linkage;
2000 const struct cl_req_operations *crs_ops;
2005 enum cache_stats_item {
2006 /** how many cache lookups were performed */
2008 /** how many times cache lookup resulted in a hit */
2010 /** how many entities are in the cache right now */
2012 /** how many entities in the cache are actively used (and cannot be
2013 * evicted) right now
2016 /** how many entities were created at all */
2021 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2024 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2026 struct cache_stats {
2027 const char *cs_name;
2028 atomic_t cs_stats[CS_NR];
2031 /** These are not exported so far */
2032 void cache_stats_init(struct cache_stats *cs, const char *name);
2035 * Client-side site. This represents particular client stack. "Global"
2036 * variables should (directly or indirectly) be added here to allow multiple
2037 * clients to co-exist in the single address space.
2040 struct lu_site cs_lu;
2042 * Statistical counters. Atomics do not scale, something better like
2043 * per-cpu counters is needed.
2045 * These are exported as /sys/kernel/debug/lustre/llite/.../site
2047 * When interpreting keep in mind that both sub-locks (and sub-pages)
2048 * and top-locks (and top-pages) are accounted here.
2050 struct cache_stats cs_pages;
2051 atomic_t cs_pages_state[CPS_NR];
2054 int cl_site_init(struct cl_site *s, struct cl_device *top);
2055 void cl_site_fini(struct cl_site *s);
2056 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2059 * Output client site statistical counters into a buffer. Suitable for
2060 * ll_rd_*()-style functions.
2062 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2067 * Type conversion and accessory functions.
2071 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2073 return container_of(site, struct cl_site, cs_lu);
2076 static inline int lu_device_is_cl(const struct lu_device *d)
2078 return d->ld_type->ldt_tags & LU_DEVICE_CL;
2081 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2083 LASSERT(!d || IS_ERR(d) || lu_device_is_cl(d));
2084 return container_of0(d, struct cl_device, cd_lu_dev);
2087 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2089 return &d->cd_lu_dev;
2092 static inline struct cl_object *lu2cl(const struct lu_object *o)
2094 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2095 return container_of0(o, struct cl_object, co_lu);
2098 static inline const struct cl_object_conf *
2099 lu2cl_conf(const struct lu_object_conf *conf)
2101 return container_of0(conf, struct cl_object_conf, coc_lu);
2104 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2106 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2109 static inline struct cl_device *cl_object_device(const struct cl_object *o)
2111 LASSERT(!o || IS_ERR(o) || lu_device_is_cl(o->co_lu.lo_dev));
2112 return container_of0(o->co_lu.lo_dev, struct cl_device, cd_lu_dev);
2115 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2117 return container_of0(h, struct cl_object_header, coh_lu);
2120 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2122 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2126 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2128 return luh2coh(obj->co_lu.lo_header);
2131 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2133 return lu_device_init(&d->cd_lu_dev, t);
2136 static inline void cl_device_fini(struct cl_device *d)
2138 lu_device_fini(&d->cd_lu_dev);
2141 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2142 struct cl_object *obj, pgoff_t index,
2143 const struct cl_page_operations *ops);
2144 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2145 struct cl_object *obj,
2146 const struct cl_lock_operations *ops);
2147 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2148 struct cl_object *obj, const struct cl_io_operations *ops);
2149 void cl_req_slice_add(struct cl_req *req, struct cl_req_slice *slice,
2150 struct cl_device *dev,
2151 const struct cl_req_operations *ops);
2154 /** \defgroup cl_object cl_object
2157 struct cl_object *cl_object_top(struct cl_object *o);
2158 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2159 const struct lu_fid *fid,
2160 const struct cl_object_conf *c);
2162 int cl_object_header_init(struct cl_object_header *h);
2163 void cl_object_put(const struct lu_env *env, struct cl_object *o);
2164 void cl_object_get(struct cl_object *o);
2165 void cl_object_attr_lock(struct cl_object *o);
2166 void cl_object_attr_unlock(struct cl_object *o);
2167 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2168 struct cl_attr *attr);
2169 int cl_object_attr_set(const struct lu_env *env, struct cl_object *obj,
2170 const struct cl_attr *attr, unsigned valid);
2171 int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj,
2172 struct ost_lvb *lvb);
2173 int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
2174 const struct cl_object_conf *conf);
2175 int cl_object_prune(const struct lu_env *env, struct cl_object *obj);
2176 void cl_object_kill(const struct lu_env *env, struct cl_object *obj);
2179 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2181 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2183 return cl_object_header(o0) == cl_object_header(o1);
2186 static inline void cl_object_page_init(struct cl_object *clob, int size)
2188 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2189 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2190 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2193 static inline void *cl_object_page_slice(struct cl_object *clob,
2194 struct cl_page *page)
2196 return (void *)((char *)page + clob->co_slice_off);
2200 * Return refcount of cl_object.
2202 static inline int cl_object_refc(struct cl_object *clob)
2204 struct lu_object_header *header = clob->co_lu.lo_header;
2206 return atomic_read(&header->loh_ref);
2211 /** \defgroup cl_page cl_page
2221 /* callback of cl_page_gang_lookup() */
2222 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *obj,
2223 pgoff_t idx, struct page *vmpage,
2224 enum cl_page_type type);
2225 struct cl_page *cl_page_alloc(const struct lu_env *env,
2226 struct cl_object *o, pgoff_t ind,
2227 struct page *vmpage,
2228 enum cl_page_type type);
2229 void cl_page_get(struct cl_page *page);
2230 void cl_page_put(const struct lu_env *env, struct cl_page *page);
2231 void cl_page_print(const struct lu_env *env, void *cookie, lu_printer_t printer,
2232 const struct cl_page *pg);
2233 void cl_page_header_print(const struct lu_env *env, void *cookie,
2234 lu_printer_t printer, const struct cl_page *pg);
2235 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj);
2237 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2238 const struct lu_device_type *dtype);
2243 * Functions dealing with the ownership of page by io.
2247 int cl_page_own(const struct lu_env *env,
2248 struct cl_io *io, struct cl_page *page);
2249 int cl_page_own_try(const struct lu_env *env,
2250 struct cl_io *io, struct cl_page *page);
2251 void cl_page_assume(const struct lu_env *env,
2252 struct cl_io *io, struct cl_page *page);
2253 void cl_page_unassume(const struct lu_env *env,
2254 struct cl_io *io, struct cl_page *pg);
2255 void cl_page_disown(const struct lu_env *env,
2256 struct cl_io *io, struct cl_page *page);
2257 void cl_page_disown0(const struct lu_env *env,
2258 struct cl_io *io, struct cl_page *pg);
2259 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io);
2266 * Functions dealing with the preparation of a page for a transfer, and
2267 * tracking transfer state.
2270 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
2271 struct cl_page *pg, enum cl_req_type crt);
2272 void cl_page_completion(const struct lu_env *env,
2273 struct cl_page *pg, enum cl_req_type crt, int ioret);
2274 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
2275 enum cl_req_type crt);
2276 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
2277 struct cl_page *pg, enum cl_req_type crt);
2278 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
2280 int cl_page_cancel(const struct lu_env *env, struct cl_page *page);
2281 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
2282 struct cl_page *pg);
2287 * \name helper routines
2288 * Functions to discard, delete and export a cl_page.
2291 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2292 struct cl_page *pg);
2293 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2294 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg);
2295 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate);
2296 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
2297 struct cl_page *page, pgoff_t *max_index);
2298 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2299 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2300 int cl_page_size(const struct cl_object *obj);
2301 int cl_pages_prune(const struct lu_env *env, struct cl_object *obj);
2303 void cl_lock_print(const struct lu_env *env, void *cookie,
2304 lu_printer_t printer, const struct cl_lock *lock);
2305 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2306 lu_printer_t printer,
2307 const struct cl_lock_descr *descr);
2311 * Data structure managing a client's cached pages. A count of
2312 * "unstable" pages is maintained, and an LRU of clean pages is
2313 * maintained. "unstable" pages are pages pinned by the ptlrpc
2314 * layer for recovery purposes.
2316 struct cl_client_cache {
2318 * # of client cache refcount
2319 * # of users (OSCs) + 2 (held by llite and lov)
2323 * # of threads are doing shrinking
2325 unsigned int ccc_lru_shrinkers;
2327 * # of LRU entries available
2329 atomic_long_t ccc_lru_left;
2331 * List of entities(OSCs) for this LRU cache
2333 struct list_head ccc_lru;
2335 * Max # of LRU entries
2337 unsigned long ccc_lru_max;
2339 * Lock to protect ccc_lru list
2341 spinlock_t ccc_lru_lock;
2343 * Set if unstable check is enabled
2345 unsigned int ccc_unstable_check:1;
2347 * # of unstable pages for this mount point
2349 atomic_long_t ccc_unstable_nr;
2351 * Waitq for awaiting unstable pages to reach zero.
2352 * Used at umounting time and signaled on BRW commit
2354 wait_queue_head_t ccc_unstable_waitq;
2359 * cl_cache functions
2361 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2362 void cl_cache_incref(struct cl_client_cache *cache);
2363 void cl_cache_decref(struct cl_client_cache *cache);
2367 /** \defgroup cl_lock cl_lock
2371 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2372 struct cl_lock *lock);
2373 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2374 const struct cl_io *io);
2375 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2376 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2377 const struct lu_device_type *dtype);
2378 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2379 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2380 struct cl_lock *lock, struct cl_sync_io *anchor);
2381 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2385 /** \defgroup cl_io cl_io
2389 int cl_io_init(const struct lu_env *env, struct cl_io *io,
2390 enum cl_io_type iot, struct cl_object *obj);
2391 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
2392 enum cl_io_type iot, struct cl_object *obj);
2393 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
2394 enum cl_io_type iot, loff_t pos, size_t count);
2395 int cl_io_loop(const struct lu_env *env, struct cl_io *io);
2397 void cl_io_fini(const struct lu_env *env, struct cl_io *io);
2398 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io);
2399 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io);
2400 int cl_io_lock(const struct lu_env *env, struct cl_io *io);
2401 void cl_io_unlock(const struct lu_env *env, struct cl_io *io);
2402 int cl_io_start(const struct lu_env *env, struct cl_io *io);
2403 void cl_io_end(const struct lu_env *env, struct cl_io *io);
2404 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
2405 struct cl_io_lock_link *link);
2406 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2407 struct cl_lock_descr *descr);
2408 int cl_io_read_page(const struct lu_env *env, struct cl_io *io,
2409 struct cl_page *page);
2410 int cl_io_submit_rw(const struct lu_env *env, struct cl_io *io,
2411 enum cl_req_type iot, struct cl_2queue *queue);
2412 int cl_io_submit_sync(const struct lu_env *env, struct cl_io *io,
2413 enum cl_req_type iot, struct cl_2queue *queue,
2415 int cl_io_commit_async(const struct lu_env *env, struct cl_io *io,
2416 struct cl_page_list *queue, int from, int to,
2418 int cl_io_is_going(const struct lu_env *env);
2421 * True, iff \a io is an O_APPEND write(2).
2423 static inline int cl_io_is_append(const struct cl_io *io)
2425 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2428 static inline int cl_io_is_sync_write(const struct cl_io *io)
2430 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2433 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2435 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2439 * True, iff \a io is a truncate(2).
2441 static inline int cl_io_is_trunc(const struct cl_io *io)
2443 return io->ci_type == CIT_SETATTR &&
2444 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2447 struct cl_io *cl_io_top(struct cl_io *io);
2449 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2451 typeof(foo_io) __foo_io = (foo_io); \
2453 CLASSERT(offsetof(typeof(*__foo_io), base) == 0); \
2454 memset(&__foo_io->base + 1, 0, \
2455 sizeof(*__foo_io) - sizeof(__foo_io->base)); \
2460 /** \defgroup cl_page_list cl_page_list
2465 * Last page in the page list.
2467 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2469 LASSERT(plist->pl_nr > 0);
2470 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2473 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2475 LASSERT(plist->pl_nr > 0);
2476 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2480 * Iterate over pages in a page list.
2482 #define cl_page_list_for_each(page, list) \
2483 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2486 * Iterate over pages in a page list, taking possible removals into account.
2488 #define cl_page_list_for_each_safe(page, temp, list) \
2489 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2491 void cl_page_list_init(struct cl_page_list *plist);
2492 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page);
2493 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2494 struct cl_page *page);
2495 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2496 struct cl_page *page);
2497 void cl_page_list_splice(struct cl_page_list *list, struct cl_page_list *head);
2498 void cl_page_list_del(const struct lu_env *env, struct cl_page_list *plist,
2499 struct cl_page *page);
2500 void cl_page_list_disown(const struct lu_env *env,
2501 struct cl_io *io, struct cl_page_list *plist);
2502 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2504 void cl_2queue_init(struct cl_2queue *queue);
2505 void cl_2queue_disown(const struct lu_env *env,
2506 struct cl_io *io, struct cl_2queue *queue);
2507 void cl_2queue_discard(const struct lu_env *env,
2508 struct cl_io *io, struct cl_2queue *queue);
2509 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2510 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2512 /** @} cl_page_list */
2514 /** \defgroup cl_req cl_req
2517 struct cl_req *cl_req_alloc(const struct lu_env *env, struct cl_page *page,
2518 enum cl_req_type crt, int nr_objects);
2520 void cl_req_page_add(const struct lu_env *env, struct cl_req *req,
2521 struct cl_page *page);
2522 void cl_req_page_done(const struct lu_env *env, struct cl_page *page);
2523 int cl_req_prep(const struct lu_env *env, struct cl_req *req);
2524 void cl_req_attr_set(const struct lu_env *env, struct cl_req *req,
2525 struct cl_req_attr *attr, u64 flags);
2526 void cl_req_completion(const struct lu_env *env, struct cl_req *req, int ioret);
2528 /** \defgroup cl_sync_io cl_sync_io
2533 * Anchor for synchronous transfer. This is allocated on a stack by thread
2534 * doing synchronous transfer, and a pointer to this structure is set up in
2535 * every page submitted for transfer. Transfer completion routine updates
2536 * anchor and wakes up waiting thread when transfer is complete.
2539 /** number of pages yet to be transferred. */
2540 atomic_t csi_sync_nr;
2543 /** barrier of destroy this structure */
2544 atomic_t csi_barrier;
2545 /** completion to be signaled when transfer is complete. */
2546 wait_queue_head_t csi_waitq;
2547 /** callback to invoke when this IO is finished */
2548 void (*csi_end_io)(const struct lu_env *,
2549 struct cl_sync_io *);
2552 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2553 void (*end)(const struct lu_env *, struct cl_sync_io *));
2554 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2556 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2558 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2560 /** @} cl_sync_io */
2564 /** \defgroup cl_env cl_env
2566 * lu_env handling for a client.
2568 * lu_env is an environment within which lustre code executes. Its major part
2569 * is lu_context---a fast memory allocation mechanism that is used to conserve
2570 * precious kernel stack space. Originally lu_env was designed for a server,
2573 * - there is a (mostly) fixed number of threads, and
2575 * - call chains have no non-lustre portions inserted between lustre code.
2577 * On a client both these assumption fails, because every user thread can
2578 * potentially execute lustre code as part of a system call, and lustre calls
2579 * into VFS or MM that call back into lustre.
2581 * To deal with that, cl_env wrapper functions implement the following
2584 * - allocation and destruction of environment is amortized by caching no
2585 * longer used environments instead of destroying them;
2587 * - there is a notion of "current" environment, attached to the kernel
2588 * data structure representing current thread Top-level lustre code
2589 * allocates an environment and makes it current, then calls into
2590 * non-lustre code, that in turn calls lustre back. Low-level lustre
2591 * code thus called can fetch environment created by the top-level code
2592 * and reuse it, avoiding additional environment allocation.
2593 * Right now, three interfaces can attach the cl_env to running thread:
2596 * - cl_env_reexit(cl_env_reenter had to be called priorly)
2598 * \see lu_env, lu_context, lu_context_key
2602 struct cl_env_nest {
2607 struct lu_env *cl_env_get(int *refcheck);
2608 struct lu_env *cl_env_alloc(int *refcheck, __u32 tags);
2609 struct lu_env *cl_env_nested_get(struct cl_env_nest *nest);
2610 void cl_env_put(struct lu_env *env, int *refcheck);
2611 void cl_env_nested_put(struct cl_env_nest *nest, struct lu_env *env);
2612 void *cl_env_reenter(void);
2613 void cl_env_reexit(void *cookie);
2614 void cl_env_implant(struct lu_env *env, int *refcheck);
2615 void cl_env_unplant(struct lu_env *env, int *refcheck);
2616 unsigned int cl_env_cache_purge(unsigned int nr);
2617 struct lu_env *cl_env_percpu_get(void);
2618 void cl_env_percpu_put(struct lu_env *env);
2625 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2627 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2628 struct lu_device_type *ldt,
2629 struct lu_device *next);
2632 int cl_global_init(void);
2633 void cl_global_fini(void);
2635 #endif /* _LINUX_CL_OBJECT_H */