4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 #ifndef _LUSTRE_CL_OBJECT_H
33 #define _LUSTRE_CL_OBJECT_H
35 /** \defgroup clio clio
37 * Client objects implement io operations and cache pages.
39 * Examples: lov and osc are implementations of cl interface.
41 * Big Theory Statement.
45 * Client implementation is based on the following data-types:
51 * - cl_lock represents an extent lock on an object.
53 * - cl_io represents high-level i/o activity such as whole read/write
54 * system call, or write-out of pages from under the lock being
55 * canceled. cl_io has sub-ios that can be stopped and resumed
56 * independently, thus achieving high degree of transfer
57 * parallelism. Single cl_io can be advanced forward by
58 * the multiple threads (although in the most usual case of
59 * read/write system call it is associated with the single user
60 * thread, that issued the system call).
64 * - to avoid confusion high-level I/O operation like read or write system
65 * call is referred to as "an io", whereas low-level I/O operation, like
66 * RPC, is referred to as "a transfer"
68 * - "generic code" means generic (not file system specific) code in the
69 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
70 * is not layer specific.
76 * - cl_object_header::coh_page_guard
79 * See the top comment in cl_object.c for the description of overall locking and
80 * reference-counting design.
82 * See comments below for the description of i/o, page, and dlm-locking
89 * super-class definitions.
91 #include <libcfs/libcfs.h>
92 #include <libcfs/libcfs_ptask.h>
93 #include <lu_object.h>
94 #include <linux/atomic.h>
95 #include <linux/mutex.h>
96 #include <linux/radix-tree.h>
97 #include <linux/spinlock.h>
98 #include <linux/wait.h>
99 #include <lustre_dlm.h>
109 struct cl_page_slice;
111 struct cl_lock_slice;
113 struct cl_lock_operations;
114 struct cl_page_operations;
121 extern struct cfs_ptask_engine *cl_io_engine;
124 * Device in the client stack.
126 * \see vvp_device, lov_device, lovsub_device, osc_device
130 struct lu_device cd_lu_dev;
133 /** \addtogroup cl_object cl_object
136 * "Data attributes" of cl_object. Data attributes can be updated
137 * independently for a sub-object, and top-object's attributes are calculated
138 * from sub-objects' ones.
141 /** Object size, in bytes */
144 * Known minimal size, in bytes.
146 * This is only valid when at least one DLM lock is held.
149 /** Modification time. Measured in seconds since epoch. */
151 /** Access time. Measured in seconds since epoch. */
153 /** Change time. Measured in seconds since epoch. */
156 * Blocks allocated to this cl_object on the server file system.
158 * \todo XXX An interface for block size is needed.
162 * User identifier for quota purposes.
166 * Group identifier for quota purposes.
170 /* nlink of the directory */
173 /* Project identifier for quota purpose. */
178 * Fields in cl_attr that are being set.
193 * Sub-class of lu_object with methods common for objects on the client
196 * cl_object: represents a regular file system object, both a file and a
197 * stripe. cl_object is based on lu_object: it is identified by a fid,
198 * layered, cached, hashed, and lrued. Important distinction with the server
199 * side, where md_object and dt_object are used, is that cl_object "fans out"
200 * at the lov/sns level: depending on the file layout, single file is
201 * represented as a set of "sub-objects" (stripes). At the implementation
202 * level, struct lov_object contains an array of cl_objects. Each sub-object
203 * is a full-fledged cl_object, having its fid, living in the lru and hash
206 * This leads to the next important difference with the server side: on the
207 * client, it's quite usual to have objects with the different sequence of
208 * layers. For example, typical top-object is composed of the following
214 * whereas its sub-objects are composed of
219 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
220 * track of the object-subobject relationship.
222 * Sub-objects are not cached independently: when top-object is about to
223 * be discarded from the memory, all its sub-objects are torn-down and
226 * \see vvp_object, lov_object, lovsub_object, osc_object
230 struct lu_object co_lu;
231 /** per-object-layer operations */
232 const struct cl_object_operations *co_ops;
233 /** offset of page slice in cl_page buffer */
238 * Description of the client object configuration. This is used for the
239 * creation of a new client object that is identified by a more state than
242 struct cl_object_conf {
244 struct lu_object_conf coc_lu;
247 * Object layout. This is consumed by lov.
249 struct lu_buf coc_layout;
251 * Description of particular stripe location in the
252 * cluster. This is consumed by osc.
254 struct lov_oinfo *coc_oinfo;
257 * VFS inode. This is consumed by vvp.
259 struct inode *coc_inode;
261 * Layout lock handle.
263 struct ldlm_lock *coc_lock;
265 * Operation to handle layout, OBJECT_CONF_XYZ.
271 /** configure layout, set up a new stripe, must be called while
272 * holding layout lock. */
274 /** invalidate the current stripe configuration due to losing
276 OBJECT_CONF_INVALIDATE = 1,
277 /** wait for old layout to go away so that new layout can be
283 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
284 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
288 /** the buffer to return the layout in lov_mds_md format. */
289 struct lu_buf cl_buf;
290 /** size of layout in lov_mds_md format. */
292 /** size of DoM component if exists or zero otherwise */
293 u64 cl_dom_comp_size;
294 /** Layout generation. */
296 /** whether layout is a composite one */
297 bool cl_is_composite;
301 * Operations implemented for each cl object layer.
303 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
305 struct cl_object_operations {
307 * Initialize page slice for this layer. Called top-to-bottom through
308 * every object layer when a new cl_page is instantiated. Layer
309 * keeping private per-page data, or requiring its own page operations
310 * vector should allocate these data here, and attach then to the page
311 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
314 * \retval NULL success.
316 * \retval ERR_PTR(errno) failure code.
318 * \retval valid-pointer pointer to already existing referenced page
319 * to be used instead of newly created.
321 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
322 struct cl_page *page, pgoff_t index);
324 * Initialize lock slice for this layer. Called top-to-bottom through
325 * every object layer when a new cl_lock is instantiated. Layer
326 * keeping private per-lock data, or requiring its own lock operations
327 * vector should allocate these data here, and attach then to the lock
328 * by calling cl_lock_slice_add(). Mandatory.
330 int (*coo_lock_init)(const struct lu_env *env,
331 struct cl_object *obj, struct cl_lock *lock,
332 const struct cl_io *io);
334 * Initialize io state for a given layer.
336 * called top-to-bottom once per io existence to initialize io
337 * state. If layer wants to keep some state for this type of io, it
338 * has to embed struct cl_io_slice in lu_env::le_ses, and register
339 * slice with cl_io_slice_add(). It is guaranteed that all threads
340 * participating in this io share the same session.
342 int (*coo_io_init)(const struct lu_env *env,
343 struct cl_object *obj, struct cl_io *io);
345 * Fill portion of \a attr that this layer controls. This method is
346 * called top-to-bottom through all object layers.
348 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
350 * \return 0: to continue
351 * \return +ve: to stop iterating through layers (but 0 is returned
352 * from enclosing cl_object_attr_get())
353 * \return -ve: to signal error
355 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
356 struct cl_attr *attr);
360 * \a valid is a bitmask composed from enum #cl_attr_valid, and
361 * indicating what attributes are to be set.
363 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
365 * \return the same convention as for
366 * cl_object_operations::coo_attr_get() is used.
368 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
369 const struct cl_attr *attr, unsigned valid);
371 * Update object configuration. Called top-to-bottom to modify object
374 * XXX error conditions and handling.
376 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
377 const struct cl_object_conf *conf);
379 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
380 * object. Layers are supposed to fill parts of \a lvb that will be
381 * shipped to the glimpse originator as a glimpse result.
383 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
384 * \see osc_object_glimpse()
386 int (*coo_glimpse)(const struct lu_env *env,
387 const struct cl_object *obj, struct ost_lvb *lvb);
389 * Object prune method. Called when the layout is going to change on
390 * this object, therefore each layer has to clean up their cache,
391 * mainly pages and locks.
393 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
395 * Object getstripe method.
397 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
398 struct lov_user_md __user *lum, size_t size);
400 * Get FIEMAP mapping from the object.
402 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
403 struct ll_fiemap_info_key *fmkey,
404 struct fiemap *fiemap, size_t *buflen);
406 * Get layout and generation of the object.
408 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
409 struct cl_layout *layout);
411 * Get maximum size of the object.
413 loff_t (*coo_maxbytes)(struct cl_object *obj);
415 * Set request attributes.
417 void (*coo_req_attr_set)(const struct lu_env *env,
418 struct cl_object *obj,
419 struct cl_req_attr *attr);
423 * Extended header for client object.
425 struct cl_object_header {
426 /** Standard lu_object_header. cl_object::co_lu::lo_header points
428 struct lu_object_header coh_lu;
431 * Parent object. It is assumed that an object has a well-defined
432 * parent, but not a well-defined child (there may be multiple
433 * sub-objects, for the same top-object). cl_object_header::coh_parent
434 * field allows certain code to be written generically, without
435 * limiting possible cl_object layouts unduly.
437 struct cl_object_header *coh_parent;
439 * Protects consistency between cl_attr of parent object and
440 * attributes of sub-objects, that the former is calculated ("merged")
443 * \todo XXX this can be read/write lock if needed.
445 spinlock_t coh_attr_guard;
447 * Size of cl_page + page slices
449 unsigned short coh_page_bufsize;
451 * Number of objects above this one: 0 for a top-object, 1 for its
454 unsigned char coh_nesting;
458 * Helper macro: iterate over all layers of the object \a obj, assigning every
459 * layer top-to-bottom to \a slice.
461 #define cl_object_for_each(slice, obj) \
462 list_for_each_entry((slice), \
463 &(obj)->co_lu.lo_header->loh_layers,\
467 * Helper macro: iterate over all layers of the object \a obj, assigning every
468 * layer bottom-to-top to \a slice.
470 #define cl_object_for_each_reverse(slice, obj) \
471 list_for_each_entry_reverse((slice), \
472 &(obj)->co_lu.lo_header->loh_layers,\
477 #define CL_PAGE_EOF ((pgoff_t)~0ull)
479 /** \addtogroup cl_page cl_page
483 * Layered client page.
485 * cl_page: represents a portion of a file, cached in the memory. All pages
486 * of the given file are of the same size, and are kept in the radix tree
487 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
488 * of the top-level file object are first class cl_objects, they have their
489 * own radix trees of pages and hence page is implemented as a sequence of
490 * struct cl_pages's, linked into double-linked list through
491 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
492 * corresponding radix tree at the corresponding logical offset.
494 * cl_page is associated with VM page of the hosting environment (struct
495 * page in Linux kernel, for example), struct page. It is assumed, that this
496 * association is implemented by one of cl_page layers (top layer in the
497 * current design) that
499 * - intercepts per-VM-page call-backs made by the environment (e.g.,
502 * - translates state (page flag bits) and locking between lustre and
505 * The association between cl_page and struct page is immutable and
506 * established when cl_page is created.
508 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
509 * this io an exclusive access to this page w.r.t. other io attempts and
510 * various events changing page state (such as transfer completion, or
511 * eviction of the page from the memory). Note, that in general cl_io
512 * cannot be identified with a particular thread, and page ownership is not
513 * exactly equal to the current thread holding a lock on the page. Layer
514 * implementing association between cl_page and struct page has to implement
515 * ownership on top of available synchronization mechanisms.
517 * While lustre client maintains the notion of an page ownership by io,
518 * hosting MM/VM usually has its own page concurrency control
519 * mechanisms. For example, in Linux, page access is synchronized by the
520 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
521 * takes care to acquire and release such locks as necessary around the
522 * calls to the file system methods (->readpage(), ->prepare_write(),
523 * ->commit_write(), etc.). This leads to the situation when there are two
524 * different ways to own a page in the client:
526 * - client code explicitly and voluntary owns the page (cl_page_own());
528 * - VM locks a page and then calls the client, that has "to assume"
529 * the ownership from the VM (cl_page_assume()).
531 * Dual methods to release ownership are cl_page_disown() and
532 * cl_page_unassume().
534 * cl_page is reference counted (cl_page::cp_ref). When reference counter
535 * drops to 0, the page is returned to the cache, unless it is in
536 * cl_page_state::CPS_FREEING state, in which case it is immediately
539 * The general logic guaranteeing the absence of "existential races" for
540 * pages is the following:
542 * - there are fixed known ways for a thread to obtain a new reference
545 * - by doing a lookup in the cl_object radix tree, protected by the
548 * - by starting from VM-locked struct page and following some
549 * hosting environment method (e.g., following ->private pointer in
550 * the case of Linux kernel), see cl_vmpage_page();
552 * - when the page enters cl_page_state::CPS_FREEING state, all these
553 * ways are severed with the proper synchronization
554 * (cl_page_delete());
556 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
559 * - no new references to the page in cl_page_state::CPS_FREEING state
560 * are allowed (checked in cl_page_get()).
562 * Together this guarantees that when last reference to a
563 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
564 * page, as neither references to it can be acquired at that point, nor
567 * cl_page is a state machine. States are enumerated in enum
568 * cl_page_state. Possible state transitions are enumerated in
569 * cl_page_state_set(). State transition process (i.e., actual changing of
570 * cl_page::cp_state field) is protected by the lock on the underlying VM
573 * Linux Kernel implementation.
575 * Binding between cl_page and struct page (which is a typedef for
576 * struct page) is implemented in the vvp layer. cl_page is attached to the
577 * ->private pointer of the struct page, together with the setting of
578 * PG_private bit in page->flags, and acquiring additional reference on the
579 * struct page (much like struct buffer_head, or any similar file system
580 * private data structures).
582 * PG_locked lock is used to implement both ownership and transfer
583 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
584 * states. No additional references are acquired for the duration of the
587 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
588 * write-out is "protected" by the special PG_writeback bit.
592 * States of cl_page. cl_page.c assumes particular order here.
594 * The page state machine is rather crude, as it doesn't recognize finer page
595 * states like "dirty" or "up to date". This is because such states are not
596 * always well defined for the whole stack (see, for example, the
597 * implementation of the read-ahead, that hides page up-to-dateness to track
598 * cache hits accurately). Such sub-states are maintained by the layers that
599 * are interested in them.
603 * Page is in the cache, un-owned. Page leaves cached state in the
606 * - [cl_page_state::CPS_OWNED] io comes across the page and
609 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
610 * req-formation engine decides that it wants to include this page
611 * into an RPC being constructed, and yanks it from the cache;
613 * - [cl_page_state::CPS_FREEING] VM callback is executed to
614 * evict the page form the memory;
616 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
620 * Page is exclusively owned by some cl_io. Page may end up in this
621 * state as a result of
623 * - io creating new page and immediately owning it;
625 * - [cl_page_state::CPS_CACHED] io finding existing cached page
628 * - [cl_page_state::CPS_OWNED] io finding existing owned page
629 * and waiting for owner to release the page;
631 * Page leaves owned state in the following cases:
633 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
634 * the cache, doing nothing;
636 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
639 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
640 * transfer for this page;
642 * - [cl_page_state::CPS_FREEING] io decides to destroy this
643 * page (e.g., as part of truncate or extent lock cancellation).
645 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
649 * Page is being written out, as a part of a transfer. This state is
650 * entered when req-formation logic decided that it wants this page to
651 * be sent through the wire _now_. Specifically, it means that once
652 * this state is achieved, transfer completion handler (with either
653 * success or failure indication) is guaranteed to be executed against
654 * this page independently of any locks and any scheduling decisions
655 * made by the hosting environment (that effectively means that the
656 * page is never put into cl_page_state::CPS_PAGEOUT state "in
657 * advance". This property is mentioned, because it is important when
658 * reasoning about possible dead-locks in the system). The page can
659 * enter this state as a result of
661 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
662 * write-out of this page, or
664 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
665 * that it has enough dirty pages cached to issue a "good"
668 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
669 * is completed---it is moved into cl_page_state::CPS_CACHED state.
671 * Underlying VM page is locked for the duration of transfer.
673 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
677 * Page is being read in, as a part of a transfer. This is quite
678 * similar to the cl_page_state::CPS_PAGEOUT state, except that
679 * read-in is always "immediate"---there is no such thing a sudden
680 * construction of read request from cached, presumably not up to date,
683 * Underlying VM page is locked for the duration of transfer.
685 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
689 * Page is being destroyed. This state is entered when client decides
690 * that page has to be deleted from its host object, as, e.g., a part
693 * Once this state is reached, there is no way to escape it.
695 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
702 /** Host page, the page is from the host inode which the cl_page
706 /** Transient page, the transient cl_page is used to bind a cl_page
707 * to vmpage which is not belonging to the same object of cl_page.
708 * it is used in DirectIO and lockless IO. */
713 * Fields are protected by the lock on struct page, except for atomics and
716 * \invariant Data type invariants are in cl_page_invariant(). Basically:
717 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
718 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
719 * cl_page::cp_owner (when set).
722 /** Reference counter. */
724 /** An object this page is a part of. Immutable after creation. */
725 struct cl_object *cp_obj;
727 struct page *cp_vmpage;
728 /** Linkage of pages within group. Pages must be owned */
729 struct list_head cp_batch;
730 /** List of slices. Immutable after creation. */
731 struct list_head cp_layers;
733 * Page state. This field is const to avoid accidental update, it is
734 * modified only internally within cl_page.c. Protected by a VM lock.
736 const enum cl_page_state cp_state;
738 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
741 enum cl_page_type cp_type;
744 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
745 * by sub-io. Protected by a VM lock.
747 struct cl_io *cp_owner;
748 /** List of references to this page, for debugging. */
749 struct lu_ref cp_reference;
750 /** Link to an object, for debugging. */
751 struct lu_ref_link cp_obj_ref;
752 /** Link to a queue, for debugging. */
753 struct lu_ref_link cp_queue_ref;
754 /** Assigned if doing a sync_io */
755 struct cl_sync_io *cp_sync_io;
759 * Per-layer part of cl_page.
761 * \see vvp_page, lov_page, osc_page
763 struct cl_page_slice {
764 struct cl_page *cpl_page;
767 * Object slice corresponding to this page slice. Immutable after
770 struct cl_object *cpl_obj;
771 const struct cl_page_operations *cpl_ops;
772 /** Linkage into cl_page::cp_layers. Immutable after creation. */
773 struct list_head cpl_linkage;
777 * Lock mode. For the client extent locks.
789 * Requested transfer type.
798 * Per-layer page operations.
800 * Methods taking an \a io argument are for the activity happening in the
801 * context of given \a io. Page is assumed to be owned by that io, except for
802 * the obvious cases (like cl_page_operations::cpo_own()).
804 * \see vvp_page_ops, lov_page_ops, osc_page_ops
806 struct cl_page_operations {
808 * cl_page<->struct page methods. Only one layer in the stack has to
809 * implement these. Current code assumes that this functionality is
810 * provided by the topmost layer, see cl_page_disown0() as an example.
814 * Called when \a io acquires this page into the exclusive
815 * ownership. When this method returns, it is guaranteed that the is
816 * not owned by other io, and no transfer is going on against
820 * \see vvp_page_own(), lov_page_own()
822 int (*cpo_own)(const struct lu_env *env,
823 const struct cl_page_slice *slice,
824 struct cl_io *io, int nonblock);
825 /** Called when ownership it yielded. Optional.
827 * \see cl_page_disown()
828 * \see vvp_page_disown()
830 void (*cpo_disown)(const struct lu_env *env,
831 const struct cl_page_slice *slice, struct cl_io *io);
833 * Called for a page that is already "owned" by \a io from VM point of
836 * \see cl_page_assume()
837 * \see vvp_page_assume(), lov_page_assume()
839 void (*cpo_assume)(const struct lu_env *env,
840 const struct cl_page_slice *slice, struct cl_io *io);
841 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
842 * bottom-to-top when IO releases a page without actually unlocking
845 * \see cl_page_unassume()
846 * \see vvp_page_unassume()
848 void (*cpo_unassume)(const struct lu_env *env,
849 const struct cl_page_slice *slice,
852 * Announces whether the page contains valid data or not by \a uptodate.
854 * \see cl_page_export()
855 * \see vvp_page_export()
857 void (*cpo_export)(const struct lu_env *env,
858 const struct cl_page_slice *slice, int uptodate);
860 * Checks whether underlying VM page is locked (in the suitable
861 * sense). Used for assertions.
863 * \retval -EBUSY: page is protected by a lock of a given mode;
864 * \retval -ENODATA: page is not protected by a lock;
865 * \retval 0: this layer cannot decide. (Should never happen.)
867 int (*cpo_is_vmlocked)(const struct lu_env *env,
868 const struct cl_page_slice *slice);
874 * Called when page is truncated from the object. Optional.
876 * \see cl_page_discard()
877 * \see vvp_page_discard(), osc_page_discard()
879 void (*cpo_discard)(const struct lu_env *env,
880 const struct cl_page_slice *slice,
883 * Called when page is removed from the cache, and is about to being
884 * destroyed. Optional.
886 * \see cl_page_delete()
887 * \see vvp_page_delete(), osc_page_delete()
889 void (*cpo_delete)(const struct lu_env *env,
890 const struct cl_page_slice *slice);
891 /** Destructor. Frees resources and slice itself. */
892 void (*cpo_fini)(const struct lu_env *env,
893 struct cl_page_slice *slice);
895 * Optional debugging helper. Prints given page slice.
897 * \see cl_page_print()
899 int (*cpo_print)(const struct lu_env *env,
900 const struct cl_page_slice *slice,
901 void *cookie, lu_printer_t p);
910 * Request type dependent vector of operations.
912 * Transfer operations depend on transfer mode (cl_req_type). To avoid
913 * passing transfer mode to each and every of these methods, and to
914 * avoid branching on request type inside of the methods, separate
915 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
916 * provided. That is, method invocation usually looks like
918 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
922 * Called when a page is submitted for a transfer as a part of
925 * \return 0 : page is eligible for submission;
926 * \return -EALREADY : skip this page;
927 * \return -ve : error.
929 * \see cl_page_prep()
931 int (*cpo_prep)(const struct lu_env *env,
932 const struct cl_page_slice *slice,
935 * Completion handler. This is guaranteed to be eventually
936 * fired after cl_page_operations::cpo_prep() or
937 * cl_page_operations::cpo_make_ready() call.
939 * This method can be called in a non-blocking context. It is
940 * guaranteed however, that the page involved and its object
941 * are pinned in memory (and, hence, calling cl_page_put() is
944 * \see cl_page_completion()
946 void (*cpo_completion)(const struct lu_env *env,
947 const struct cl_page_slice *slice,
950 * Called when cached page is about to be added to the
951 * ptlrpc request as a part of req formation.
953 * \return 0 : proceed with this page;
954 * \return -EAGAIN : skip this page;
955 * \return -ve : error.
957 * \see cl_page_make_ready()
959 int (*cpo_make_ready)(const struct lu_env *env,
960 const struct cl_page_slice *slice);
963 * Tell transfer engine that only [to, from] part of a page should be
966 * This is used for immediate transfers.
968 * \todo XXX this is not very good interface. It would be much better
969 * if all transfer parameters were supplied as arguments to
970 * cl_io_operations::cio_submit() call, but it is not clear how to do
971 * this for page queues.
973 * \see cl_page_clip()
975 void (*cpo_clip)(const struct lu_env *env,
976 const struct cl_page_slice *slice,
979 * \pre the page was queued for transferring.
980 * \post page is removed from client's pending list, or -EBUSY
981 * is returned if it has already been in transferring.
983 * This is one of seldom page operation which is:
984 * 0. called from top level;
985 * 1. don't have vmpage locked;
986 * 2. every layer should synchronize execution of its ->cpo_cancel()
987 * with completion handlers. Osc uses client obd lock for this
988 * purpose. Based on there is no vvp_page_cancel and
989 * lov_page_cancel(), cpo_cancel is defacto protected by client lock.
991 * \see osc_page_cancel().
993 int (*cpo_cancel)(const struct lu_env *env,
994 const struct cl_page_slice *slice);
996 * Write out a page by kernel. This is only called by ll_writepage
999 * \see cl_page_flush()
1001 int (*cpo_flush)(const struct lu_env *env,
1002 const struct cl_page_slice *slice,
1008 * Helper macro, dumping detailed information about \a page into a log.
1010 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1012 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1013 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1014 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1015 CDEBUG(mask, format , ## __VA_ARGS__); \
1020 * Helper macro, dumping shorter information about \a page into a log.
1022 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1024 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1025 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1026 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1027 CDEBUG(mask, format , ## __VA_ARGS__); \
1031 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1033 LASSERT(page->cp_vmpage != NULL);
1034 return page->cp_vmpage;
1038 * Check if a cl_page is in use.
1040 * Client cache holds a refcount, this refcount will be dropped when
1041 * the page is taken out of cache, see vvp_page_delete().
1043 static inline bool __page_in_use(const struct cl_page *page, int refc)
1045 return (atomic_read(&page->cp_ref) > refc + 1);
1049 * Caller itself holds a refcount of cl_page.
1051 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1053 * Caller doesn't hold a refcount.
1055 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1059 /** \addtogroup cl_lock cl_lock
1063 * Extent locking on the client.
1067 * The locking model of the new client code is built around
1071 * data-type representing an extent lock on a regular file. cl_lock is a
1072 * layered object (much like cl_object and cl_page), it consists of a header
1073 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1074 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1076 * Typical cl_lock consists of the two layers:
1078 * - vvp_lock (vvp specific data), and
1079 * - lov_lock (lov specific data).
1081 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1082 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1084 * - lovsub_lock, and
1087 * Each sub-lock is associated with a cl_object (representing stripe
1088 * sub-object or the file to which top-level cl_lock is associated to), and is
1089 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1090 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1091 * is different from cl_page, that doesn't fan out (there is usually exactly
1092 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1093 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1097 * cl_lock is a cacheless data container for the requirements of locks to
1098 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1101 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1102 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1104 * INTERFACE AND USAGE
1106 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1107 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1108 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1109 * consists of multiple sub cl_locks, each sub locks will be enqueued
1110 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1111 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1114 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1115 * method will be called for each layer to release the resource held by this
1116 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1117 * clo_enqueue time, is released.
1119 * LDLM lock can only be canceled if there is no cl_lock using it.
1121 * Overall process of the locking during IO operation is as following:
1123 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1124 * is called on each layer. Responsibility of this method is to add locks,
1125 * needed by a given layer into cl_io.ci_lockset.
1127 * - once locks for all layers were collected, they are sorted to avoid
1128 * dead-locks (cl_io_locks_sort()), and enqueued.
1130 * - when all locks are acquired, IO is performed;
1132 * - locks are released after IO is complete.
1134 * Striping introduces major additional complexity into locking. The
1135 * fundamental problem is that it is generally unsafe to actively use (hold)
1136 * two locks on the different OST servers at the same time, as this introduces
1137 * inter-server dependency and can lead to cascading evictions.
1139 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1140 * that no multi-stripe locks are taken (note that this design abandons POSIX
1141 * read/write semantics). Such pieces ideally can be executed concurrently. At
1142 * the same time, certain types of IO cannot be sub-divived, without
1143 * sacrificing correctness. This includes:
1145 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1148 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1150 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1151 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1152 * has to be held together with the usual lock on [offset, offset + count].
1154 * Interaction with DLM
1156 * In the expected setup, cl_lock is ultimately backed up by a collection of
1157 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1158 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1159 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1160 * description of interaction with DLM.
1166 struct cl_lock_descr {
1167 /** Object this lock is granted for. */
1168 struct cl_object *cld_obj;
1169 /** Index of the first page protected by this lock. */
1171 /** Index of the last page (inclusive) protected by this lock. */
1173 /** Group ID, for group lock */
1176 enum cl_lock_mode cld_mode;
1178 * flags to enqueue lock. A combination of bit-flags from
1179 * enum cl_enq_flags.
1181 __u32 cld_enq_flags;
1184 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1185 #define PDESCR(descr) \
1186 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1187 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1189 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1192 * Layered client lock.
1195 /** List of slices. Immutable after creation. */
1196 struct list_head cll_layers;
1197 /** lock attribute, extent, cl_object, etc. */
1198 struct cl_lock_descr cll_descr;
1202 * Per-layer part of cl_lock
1204 * \see vvp_lock, lov_lock, lovsub_lock, osc_lock
1206 struct cl_lock_slice {
1207 struct cl_lock *cls_lock;
1208 /** Object slice corresponding to this lock slice. Immutable after
1210 struct cl_object *cls_obj;
1211 const struct cl_lock_operations *cls_ops;
1212 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1213 struct list_head cls_linkage;
1218 * \see vvp_lock_ops, lov_lock_ops, lovsub_lock_ops, osc_lock_ops
1220 struct cl_lock_operations {
1223 * Attempts to enqueue the lock. Called top-to-bottom.
1225 * \retval 0 this layer has enqueued the lock successfully
1226 * \retval >0 this layer has enqueued the lock, but need to wait on
1227 * @anchor for resources
1228 * \retval -ve failure
1230 * \see vvp_lock_enqueue(), lov_lock_enqueue(), lovsub_lock_enqueue(),
1231 * \see osc_lock_enqueue()
1233 int (*clo_enqueue)(const struct lu_env *env,
1234 const struct cl_lock_slice *slice,
1235 struct cl_io *io, struct cl_sync_io *anchor);
1237 * Cancel a lock, release its DLM lock ref, while does not cancel the
1240 void (*clo_cancel)(const struct lu_env *env,
1241 const struct cl_lock_slice *slice);
1244 * Destructor. Frees resources and the slice.
1246 * \see vvp_lock_fini(), lov_lock_fini(), lovsub_lock_fini(),
1247 * \see osc_lock_fini()
1249 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1251 * Optional debugging helper. Prints given lock slice.
1253 int (*clo_print)(const struct lu_env *env,
1254 void *cookie, lu_printer_t p,
1255 const struct cl_lock_slice *slice);
1258 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1260 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1261 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1262 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1263 CDEBUG(mask, format , ## __VA_ARGS__); \
1267 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1271 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1277 /** \addtogroup cl_page_list cl_page_list
1278 * Page list used to perform collective operations on a group of pages.
1280 * Pages are added to the list one by one. cl_page_list acquires a reference
1281 * for every page in it. Page list is used to perform collective operations on
1284 * - submit pages for an immediate transfer,
1286 * - own pages on behalf of certain io (waiting for each page in turn),
1290 * When list is finalized, it releases references on all pages it still has.
1292 * \todo XXX concurrency control.
1296 struct cl_page_list {
1298 struct list_head pl_pages;
1299 struct task_struct *pl_owner;
1303 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1304 * contains an incoming page list and an outgoing page list.
1307 struct cl_page_list c2_qin;
1308 struct cl_page_list c2_qout;
1311 /** @} cl_page_list */
1313 /** \addtogroup cl_io cl_io
1318 * cl_io represents a high level I/O activity like
1319 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1322 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1323 * important distinction. We want to minimize number of calls to the allocator
1324 * in the fast path, e.g., in the case of read(2) when everything is cached:
1325 * client already owns the lock over region being read, and data are cached
1326 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1327 * per-layer io state is stored in the session, associated with the io, see
1328 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1329 * by using free-lists, see cl_env_get().
1331 * There is a small predefined number of possible io types, enumerated in enum
1334 * cl_io is a state machine, that can be advanced concurrently by the multiple
1335 * threads. It is up to these threads to control the concurrency and,
1336 * specifically, to detect when io is done, and its state can be safely
1339 * For read/write io overall execution plan is as following:
1341 * (0) initialize io state through all layers;
1343 * (1) loop: prepare chunk of work to do
1345 * (2) call all layers to collect locks they need to process current chunk
1347 * (3) sort all locks to avoid dead-locks, and acquire them
1349 * (4) process the chunk: call per-page methods
1350 * cl_io_operations::cio_prepare_write(),
1351 * cl_io_operations::cio_commit_write() for write)
1357 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1358 * address allocation efficiency issues mentioned above), and returns with the
1359 * special error condition from per-page method when current sub-io has to
1360 * block. This causes io loop to be repeated, and lov switches to the next
1361 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1366 /** read system call */
1368 /** write system call */
1370 /** truncate, utime system calls */
1372 /** get data version */
1375 * page fault handling
1379 * fsync system call handling
1380 * To write out a range of file
1384 * glimpse. An io context to acquire glimpse lock.
1388 * Miscellaneous io. This is used for occasional io activity that
1389 * doesn't fit into other types. Currently this is used for:
1391 * - cancellation of an extent lock. This io exists as a context
1392 * to write dirty pages from under the lock being canceled back
1395 * - VM induced page write-out. An io context for writing page out
1396 * for memory cleansing;
1398 * - grouplock. An io context to acquire group lock.
1400 * CIT_MISC io is used simply as a context in which locks and pages
1401 * are manipulated. Such io has no internal "process", that is,
1402 * cl_io_loop() is never called for it.
1407 * To give advice about access of a file
1414 * States of cl_io state machine
1417 /** Not initialized. */
1421 /** IO iteration started. */
1425 /** Actual IO is in progress. */
1427 /** IO for the current iteration finished. */
1429 /** Locks released. */
1431 /** Iteration completed. */
1433 /** cl_io finalized. */
1438 * IO state private for a layer.
1440 * This is usually embedded into layer session data, rather than allocated
1443 * \see vvp_io, lov_io, osc_io
1445 struct cl_io_slice {
1446 struct cl_io *cis_io;
1447 /** corresponding object slice. Immutable after creation. */
1448 struct cl_object *cis_obj;
1449 /** io operations. Immutable after creation. */
1450 const struct cl_io_operations *cis_iop;
1452 * linkage into a list of all slices for a given cl_io, hanging off
1453 * cl_io::ci_layers. Immutable after creation.
1455 struct list_head cis_linkage;
1458 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1461 struct cl_read_ahead {
1462 /* Maximum page index the readahead window will end.
1463 * This is determined DLM lock coverage, RPC and stripe boundary.
1464 * cra_end is included. */
1466 /* optimal RPC size for this read, by pages */
1467 unsigned long cra_rpc_size;
1468 /* Release callback. If readahead holds resources underneath, this
1469 * function should be called to release it. */
1470 void (*cra_release)(const struct lu_env *env, void *cbdata);
1471 /* Callback data for cra_release routine */
1475 static inline void cl_read_ahead_release(const struct lu_env *env,
1476 struct cl_read_ahead *ra)
1478 if (ra->cra_release != NULL)
1479 ra->cra_release(env, ra->cra_cbdata);
1480 memset(ra, 0, sizeof(*ra));
1485 * Per-layer io operations.
1486 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1488 struct cl_io_operations {
1490 * Vector of io state transition methods for every io type.
1492 * \see cl_page_operations::io
1496 * Prepare io iteration at a given layer.
1498 * Called top-to-bottom at the beginning of each iteration of
1499 * "io loop" (if it makes sense for this type of io). Here
1500 * layer selects what work it will do during this iteration.
1502 * \see cl_io_operations::cio_iter_fini()
1504 int (*cio_iter_init) (const struct lu_env *env,
1505 const struct cl_io_slice *slice);
1507 * Finalize io iteration.
1509 * Called bottom-to-top at the end of each iteration of "io
1510 * loop". Here layers can decide whether IO has to be
1513 * \see cl_io_operations::cio_iter_init()
1515 void (*cio_iter_fini) (const struct lu_env *env,
1516 const struct cl_io_slice *slice);
1518 * Collect locks for the current iteration of io.
1520 * Called top-to-bottom to collect all locks necessary for
1521 * this iteration. This methods shouldn't actually enqueue
1522 * anything, instead it should post a lock through
1523 * cl_io_lock_add(). Once all locks are collected, they are
1524 * sorted and enqueued in the proper order.
1526 int (*cio_lock) (const struct lu_env *env,
1527 const struct cl_io_slice *slice);
1529 * Finalize unlocking.
1531 * Called bottom-to-top to finish layer specific unlocking
1532 * functionality, after generic code released all locks
1533 * acquired by cl_io_operations::cio_lock().
1535 void (*cio_unlock)(const struct lu_env *env,
1536 const struct cl_io_slice *slice);
1538 * Start io iteration.
1540 * Once all locks are acquired, called top-to-bottom to
1541 * commence actual IO. In the current implementation,
1542 * top-level vvp_io_{read,write}_start() does all the work
1543 * synchronously by calling generic_file_*(), so other layers
1544 * are called when everything is done.
1546 int (*cio_start)(const struct lu_env *env,
1547 const struct cl_io_slice *slice);
1549 * Called top-to-bottom at the end of io loop. Here layer
1550 * might wait for an unfinished asynchronous io.
1552 void (*cio_end) (const struct lu_env *env,
1553 const struct cl_io_slice *slice);
1555 * Called bottom-to-top to notify layers that read/write IO
1556 * iteration finished, with \a nob bytes transferred.
1558 void (*cio_advance)(const struct lu_env *env,
1559 const struct cl_io_slice *slice,
1562 * Called once per io, bottom-to-top to release io resources.
1564 void (*cio_fini) (const struct lu_env *env,
1565 const struct cl_io_slice *slice);
1569 * Submit pages from \a queue->c2_qin for IO, and move
1570 * successfully submitted pages into \a queue->c2_qout. Return
1571 * non-zero if failed to submit even the single page. If
1572 * submission failed after some pages were moved into \a
1573 * queue->c2_qout, completion callback with non-zero ioret is
1576 int (*cio_submit)(const struct lu_env *env,
1577 const struct cl_io_slice *slice,
1578 enum cl_req_type crt,
1579 struct cl_2queue *queue);
1581 * Queue async page for write.
1582 * The difference between cio_submit and cio_queue is that
1583 * cio_submit is for urgent request.
1585 int (*cio_commit_async)(const struct lu_env *env,
1586 const struct cl_io_slice *slice,
1587 struct cl_page_list *queue, int from, int to,
1590 * Decide maximum read ahead extent
1592 * \pre io->ci_type == CIT_READ
1594 int (*cio_read_ahead)(const struct lu_env *env,
1595 const struct cl_io_slice *slice,
1596 pgoff_t start, struct cl_read_ahead *ra);
1598 * Optional debugging helper. Print given io slice.
1600 int (*cio_print)(const struct lu_env *env, void *cookie,
1601 lu_printer_t p, const struct cl_io_slice *slice);
1605 * Flags to lock enqueue procedure.
1610 * instruct server to not block, if conflicting lock is found. Instead
1611 * -EWOULDBLOCK is returned immediately.
1613 CEF_NONBLOCK = 0x00000001,
1615 * Tell lower layers this is a glimpse request, translated to
1616 * LDLM_FL_HAS_INTENT at LDLM layer.
1618 * Also, because glimpse locks never block other locks, we count this
1619 * as automatically compatible with other osc locks.
1620 * (see osc_lock_compatible)
1622 CEF_GLIMPSE = 0x00000002,
1624 * tell the server to instruct (though a flag in the blocking ast) an
1625 * owner of the conflicting lock, that it can drop dirty pages
1626 * protected by this lock, without sending them to the server.
1628 CEF_DISCARD_DATA = 0x00000004,
1630 * tell the sub layers that it must be a `real' lock. This is used for
1631 * mmapped-buffer locks, glimpse locks, manually requested locks
1632 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1635 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1637 CEF_MUST = 0x00000008,
1639 * tell the sub layers that never request a `real' lock. This flag is
1640 * not used currently.
1642 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1643 * conversion policy: ci_lockreq describes generic information of lock
1644 * requirement for this IO, especially for locks which belong to the
1645 * object doing IO; however, lock itself may have precise requirements
1646 * that are described by the enqueue flags.
1648 CEF_NEVER = 0x00000010,
1650 * tell the dlm layer this is a speculative lock request
1651 * speculative lock requests are locks which are not requested as part
1652 * of an I/O operation. Instead, they are requested because we expect
1653 * to use them in the future. They are requested asynchronously at the
1656 * Currently used for asynchronous glimpse locks and manually requested
1657 * locks (LU_LADVISE_LOCKAHEAD).
1659 CEF_SPECULATIVE = 0x00000020,
1661 * enqueue a lock to test DLM lock existence.
1663 CEF_PEEK = 0x00000040,
1665 * Lock match only. Used by group lock in I/O as group lock
1666 * is known to exist.
1668 CEF_LOCK_MATCH = 0x00000080,
1670 * tell the DLM layer to lock only the requested range
1672 CEF_LOCK_NO_EXPAND = 0x00000100,
1674 * mask of enq_flags.
1676 CEF_MASK = 0x000001ff,
1680 * Link between lock and io. Intermediate structure is needed, because the
1681 * same lock can be part of multiple io's simultaneously.
1683 struct cl_io_lock_link {
1684 /** linkage into one of cl_lockset lists. */
1685 struct list_head cill_linkage;
1686 struct cl_lock cill_lock;
1687 /** optional destructor */
1688 void (*cill_fini)(const struct lu_env *env,
1689 struct cl_io_lock_link *link);
1691 #define cill_descr cill_lock.cll_descr
1694 * Lock-set represents a collection of locks, that io needs at a
1695 * time. Generally speaking, client tries to avoid holding multiple locks when
1698 * - holding extent locks over multiple ost's introduces the danger of
1699 * "cascading timeouts";
1701 * - holding multiple locks over the same ost is still dead-lock prone,
1702 * see comment in osc_lock_enqueue(),
1704 * but there are certain situations where this is unavoidable:
1706 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1708 * - truncate has to take [new-size, EOF] lock for correctness;
1710 * - SNS has to take locks across full stripe for correctness;
1712 * - in the case when user level buffer, supplied to {read,write}(file0),
1713 * is a part of a memory mapped lustre file, client has to take a dlm
1714 * locks on file0, and all files that back up the buffer (or a part of
1715 * the buffer, that is being processed in the current chunk, in any
1716 * case, there are situations where at least 2 locks are necessary).
1718 * In such cases we at least try to take locks in the same consistent
1719 * order. To this end, all locks are first collected, then sorted, and then
1723 /** locks to be acquired. */
1724 struct list_head cls_todo;
1725 /** locks acquired. */
1726 struct list_head cls_done;
1730 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1731 * but 'req' is always to be thought as 'request' :-)
1733 enum cl_io_lock_dmd {
1734 /** Always lock data (e.g., O_APPEND). */
1736 /** Layers are free to decide between local and global locking. */
1738 /** Never lock: there is no cache (e.g., liblustre). */
1742 enum cl_fsync_mode {
1743 /** start writeback, do not wait for them to finish */
1745 /** start writeback and wait for them to finish */
1747 /** discard all of dirty pages in a specific file range */
1748 CL_FSYNC_DISCARD = 2,
1749 /** start writeback and make sure they have reached storage before
1750 * return. OST_SYNC RPC must be issued and finished */
1754 struct cl_io_range {
1760 struct cl_io_pt *cip_next;
1761 struct cfs_ptask cip_task;
1762 struct kiocb cip_iocb;
1763 struct iov_iter cip_iter;
1764 struct file *cip_file;
1765 enum cl_io_type cip_iot;
1766 unsigned int cip_need_restart:1;
1775 * cl_io is shared by all threads participating in this IO (in current
1776 * implementation only one thread advances IO, but parallel IO design and
1777 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1778 * is up to these threads to serialize their activities, including updates to
1779 * mutable cl_io fields.
1782 /** type of this IO. Immutable after creation. */
1783 enum cl_io_type ci_type;
1784 /** current state of cl_io state machine. */
1785 enum cl_io_state ci_state;
1786 /** main object this io is against. Immutable after creation. */
1787 struct cl_object *ci_obj;
1789 * Upper layer io, of which this io is a part of. Immutable after
1792 struct cl_io *ci_parent;
1793 /** List of slices. Immutable after creation. */
1794 struct list_head ci_layers;
1795 /** list of locks (to be) acquired by this io. */
1796 struct cl_lockset ci_lockset;
1797 /** lock requirements, this is just a help info for sublayers. */
1798 enum cl_io_lock_dmd ci_lockreq;
1799 /** layout version when this IO occurs */
1800 __u32 ci_layout_version;
1803 struct iov_iter rw_iter;
1804 struct kiocb rw_iocb;
1805 struct cl_io_range rw_range;
1806 struct file *rw_file;
1807 unsigned int rw_nonblock:1,
1810 int (*rw_ptask)(struct cfs_ptask *ptask);
1812 struct cl_setattr_io {
1813 struct ost_lvb sa_attr;
1814 unsigned int sa_attr_flags;
1815 unsigned int sa_valid;
1816 int sa_stripe_index;
1817 struct ost_layout sa_layout;
1818 const struct lu_fid *sa_parent_fid;
1820 struct cl_data_version_io {
1821 u64 dv_data_version;
1824 struct cl_fault_io {
1825 /** page index within file. */
1827 /** bytes valid byte on a faulted page. */
1829 /** writable page? for nopage() only */
1831 /** page of an executable? */
1833 /** page_mkwrite() */
1835 /** resulting page */
1836 struct cl_page *ft_page;
1838 struct cl_fsync_io {
1841 /** file system level fid */
1842 struct lu_fid *fi_fid;
1843 enum cl_fsync_mode fi_mode;
1844 /* how many pages were written/discarded */
1845 unsigned int fi_nr_written;
1847 struct cl_ladvise_io {
1850 /** file system level fid */
1851 struct lu_fid *li_fid;
1852 enum lu_ladvise_type li_advice;
1856 struct cl_2queue ci_queue;
1859 unsigned int ci_continue:1,
1861 * This io has held grouplock, to inform sublayers that
1862 * don't do lockless i/o.
1866 * The whole IO need to be restarted because layout has been changed
1870 * to not refresh layout - the IO issuer knows that the layout won't
1871 * change(page operations, layout change causes all page to be
1872 * discarded), or it doesn't matter if it changes(sync).
1876 * Need MDS intervention to complete a write.
1877 * Write intent is required for the following cases:
1878 * 1. component being written is not initialized, or
1879 * 2. the mirrored files are NOT in WRITE_PENDING state.
1881 ci_need_write_intent:1,
1883 * Check if layout changed after the IO finishes. Mainly for HSM
1884 * requirement. If IO occurs to openning files, it doesn't need to
1885 * verify layout because HSM won't release openning files.
1886 * Right now, only two opertaions need to verify layout: glimpse
1891 * file is released, restore has to to be triggered by vvp layer
1893 ci_restore_needed:1,
1898 /** Set to 1 if parallel execution is allowed for current I/O? */
1900 /* Tell sublayers not to expand LDLM locks requested for this IO */
1901 ci_lock_no_expand:1,
1903 * Set if non-delay RPC should be used for this IO.
1905 * If this file has multiple mirrors, and if the OSTs of the current
1906 * mirror is inaccessible, non-delay RPC would error out quickly so
1907 * that the upper layer can try to access the next mirror.
1911 * How many times the read has retried before this one.
1912 * Set by the top level and consumed by the LOV.
1914 unsigned ci_ndelay_tried;
1916 * Number of pages owned by this IO. For invariant checking.
1918 unsigned ci_owned_nr;
1920 * Range of write intent. Valid if ci_need_write_intent is set.
1922 struct lu_extent ci_write_intent;
1928 * Per-transfer attributes.
1930 struct cl_req_attr {
1931 enum cl_req_type cra_type;
1933 struct cl_page *cra_page;
1934 /** Generic attributes for the server consumption. */
1935 struct obdo *cra_oa;
1937 char cra_jobid[LUSTRE_JOBID_SIZE];
1940 enum cache_stats_item {
1941 /** how many cache lookups were performed */
1943 /** how many times cache lookup resulted in a hit */
1945 /** how many entities are in the cache right now */
1947 /** how many entities in the cache are actively used (and cannot be
1948 * evicted) right now */
1950 /** how many entities were created at all */
1955 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
1958 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
1960 struct cache_stats {
1961 const char *cs_name;
1962 atomic_t cs_stats[CS_NR];
1965 /** These are not exported so far */
1966 void cache_stats_init (struct cache_stats *cs, const char *name);
1969 * Client-side site. This represents particular client stack. "Global"
1970 * variables should (directly or indirectly) be added here to allow multiple
1971 * clients to co-exist in the single address space.
1974 struct lu_site cs_lu;
1976 * Statistical counters. Atomics do not scale, something better like
1977 * per-cpu counters is needed.
1979 * These are exported as /proc/fs/lustre/llite/.../site
1981 * When interpreting keep in mind that both sub-locks (and sub-pages)
1982 * and top-locks (and top-pages) are accounted here.
1984 struct cache_stats cs_pages;
1985 atomic_t cs_pages_state[CPS_NR];
1988 int cl_site_init(struct cl_site *s, struct cl_device *top);
1989 void cl_site_fini(struct cl_site *s);
1990 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
1993 * Output client site statistical counters into a buffer. Suitable for
1994 * ll_rd_*()-style functions.
1996 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2001 * Type conversion and accessory functions.
2005 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2007 return container_of(site, struct cl_site, cs_lu);
2010 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2012 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2013 return container_of0(d, struct cl_device, cd_lu_dev);
2016 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2018 return &d->cd_lu_dev;
2021 static inline struct cl_object *lu2cl(const struct lu_object *o)
2023 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2024 return container_of0(o, struct cl_object, co_lu);
2027 static inline const struct cl_object_conf *
2028 lu2cl_conf(const struct lu_object_conf *conf)
2030 return container_of0(conf, struct cl_object_conf, coc_lu);
2033 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2035 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2038 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2040 return container_of0(h, struct cl_object_header, coh_lu);
2043 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2045 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2049 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2051 return luh2coh(obj->co_lu.lo_header);
2054 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2056 return lu_device_init(&d->cd_lu_dev, t);
2059 static inline void cl_device_fini(struct cl_device *d)
2061 lu_device_fini(&d->cd_lu_dev);
2064 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2065 struct cl_object *obj, pgoff_t index,
2066 const struct cl_page_operations *ops);
2067 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2068 struct cl_object *obj,
2069 const struct cl_lock_operations *ops);
2070 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2071 struct cl_object *obj, const struct cl_io_operations *ops);
2074 /** \defgroup cl_object cl_object
2076 struct cl_object *cl_object_top (struct cl_object *o);
2077 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2078 const struct lu_fid *fid,
2079 const struct cl_object_conf *c);
2081 int cl_object_header_init(struct cl_object_header *h);
2082 void cl_object_header_fini(struct cl_object_header *h);
2083 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2084 void cl_object_get (struct cl_object *o);
2085 void cl_object_attr_lock (struct cl_object *o);
2086 void cl_object_attr_unlock(struct cl_object *o);
2087 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2088 struct cl_attr *attr);
2089 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2090 const struct cl_attr *attr, unsigned valid);
2091 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2092 struct ost_lvb *lvb);
2093 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2094 const struct cl_object_conf *conf);
2095 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2096 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2097 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2098 struct lov_user_md __user *lum, size_t size);
2099 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2100 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2102 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2103 struct cl_layout *cl);
2104 loff_t cl_object_maxbytes(struct cl_object *obj);
2107 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2109 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2111 return cl_object_header(o0) == cl_object_header(o1);
2114 static inline void cl_object_page_init(struct cl_object *clob, int size)
2116 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2117 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2118 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2121 static inline void *cl_object_page_slice(struct cl_object *clob,
2122 struct cl_page *page)
2124 return (void *)((char *)page + clob->co_slice_off);
2128 * Return refcount of cl_object.
2130 static inline int cl_object_refc(struct cl_object *clob)
2132 struct lu_object_header *header = clob->co_lu.lo_header;
2133 return atomic_read(&header->loh_ref);
2138 /** \defgroup cl_page cl_page
2146 /* callback of cl_page_gang_lookup() */
2148 struct cl_page *cl_page_find (const struct lu_env *env,
2149 struct cl_object *obj,
2150 pgoff_t idx, struct page *vmpage,
2151 enum cl_page_type type);
2152 struct cl_page *cl_page_alloc (const struct lu_env *env,
2153 struct cl_object *o, pgoff_t ind,
2154 struct page *vmpage,
2155 enum cl_page_type type);
2156 void cl_page_get (struct cl_page *page);
2157 void cl_page_put (const struct lu_env *env,
2158 struct cl_page *page);
2159 void cl_page_print (const struct lu_env *env, void *cookie,
2160 lu_printer_t printer,
2161 const struct cl_page *pg);
2162 void cl_page_header_print(const struct lu_env *env, void *cookie,
2163 lu_printer_t printer,
2164 const struct cl_page *pg);
2165 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2166 struct cl_page *cl_page_top (struct cl_page *page);
2168 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2169 const struct lu_device_type *dtype);
2174 * Functions dealing with the ownership of page by io.
2178 int cl_page_own (const struct lu_env *env,
2179 struct cl_io *io, struct cl_page *page);
2180 int cl_page_own_try (const struct lu_env *env,
2181 struct cl_io *io, struct cl_page *page);
2182 void cl_page_assume (const struct lu_env *env,
2183 struct cl_io *io, struct cl_page *page);
2184 void cl_page_unassume (const struct lu_env *env,
2185 struct cl_io *io, struct cl_page *pg);
2186 void cl_page_disown (const struct lu_env *env,
2187 struct cl_io *io, struct cl_page *page);
2188 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2195 * Functions dealing with the preparation of a page for a transfer, and
2196 * tracking transfer state.
2199 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2200 struct cl_page *pg, enum cl_req_type crt);
2201 void cl_page_completion (const struct lu_env *env,
2202 struct cl_page *pg, enum cl_req_type crt, int ioret);
2203 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2204 enum cl_req_type crt);
2205 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2206 struct cl_page *pg, enum cl_req_type crt);
2207 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2209 int cl_page_cancel (const struct lu_env *env, struct cl_page *page);
2210 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2211 struct cl_page *pg);
2217 * \name helper routines
2218 * Functions to discard, delete and export a cl_page.
2221 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2222 struct cl_page *pg);
2223 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2224 int cl_page_is_vmlocked(const struct lu_env *env,
2225 const struct cl_page *pg);
2226 void cl_page_export(const struct lu_env *env,
2227 struct cl_page *pg, int uptodate);
2228 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2229 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2230 size_t cl_page_size(const struct cl_object *obj);
2232 void cl_lock_print(const struct lu_env *env, void *cookie,
2233 lu_printer_t printer, const struct cl_lock *lock);
2234 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2235 lu_printer_t printer,
2236 const struct cl_lock_descr *descr);
2240 * Data structure managing a client's cached pages. A count of
2241 * "unstable" pages is maintained, and an LRU of clean pages is
2242 * maintained. "unstable" pages are pages pinned by the ptlrpc
2243 * layer for recovery purposes.
2245 struct cl_client_cache {
2247 * # of client cache refcount
2248 * # of users (OSCs) + 2 (held by llite and lov)
2252 * # of threads are doing shrinking
2254 unsigned int ccc_lru_shrinkers;
2256 * # of LRU entries available
2258 atomic_long_t ccc_lru_left;
2260 * List of entities(OSCs) for this LRU cache
2262 struct list_head ccc_lru;
2264 * Max # of LRU entries
2266 unsigned long ccc_lru_max;
2268 * Lock to protect ccc_lru list
2270 spinlock_t ccc_lru_lock;
2272 * Set if unstable check is enabled
2274 unsigned int ccc_unstable_check:1;
2276 * # of unstable pages for this mount point
2278 atomic_long_t ccc_unstable_nr;
2280 * Waitq for awaiting unstable pages to reach zero.
2281 * Used at umounting time and signaled on BRW commit
2283 wait_queue_head_t ccc_unstable_waitq;
2286 * cl_cache functions
2288 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2289 void cl_cache_incref(struct cl_client_cache *cache);
2290 void cl_cache_decref(struct cl_client_cache *cache);
2294 /** \defgroup cl_lock cl_lock
2296 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2297 struct cl_lock *lock);
2298 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2299 const struct cl_io *io);
2300 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2301 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2302 const struct lu_device_type *dtype);
2303 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2305 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2306 struct cl_lock *lock, struct cl_sync_io *anchor);
2307 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2311 /** \defgroup cl_io cl_io
2314 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2315 enum cl_io_type iot, struct cl_object *obj);
2316 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2317 enum cl_io_type iot, struct cl_object *obj);
2318 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2319 enum cl_io_type iot, loff_t pos, size_t count);
2320 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2322 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2323 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2324 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2325 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2326 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2327 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2328 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2329 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2330 struct cl_io_lock_link *link);
2331 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2332 struct cl_lock_descr *descr);
2333 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2334 enum cl_req_type iot, struct cl_2queue *queue);
2335 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2336 enum cl_req_type iot, struct cl_2queue *queue,
2338 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2339 struct cl_page_list *queue, int from, int to,
2341 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2342 pgoff_t start, struct cl_read_ahead *ra);
2343 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2345 int cl_io_cancel (const struct lu_env *env, struct cl_io *io,
2346 struct cl_page_list *queue);
2349 * True, iff \a io is an O_APPEND write(2).
2351 static inline int cl_io_is_append(const struct cl_io *io)
2353 return io->ci_type == CIT_WRITE && io->u.ci_rw.rw_append;
2356 static inline int cl_io_is_sync_write(const struct cl_io *io)
2358 return io->ci_type == CIT_WRITE && io->u.ci_rw.rw_sync;
2361 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2363 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2367 * True, iff \a io is a truncate(2).
2369 static inline int cl_io_is_trunc(const struct cl_io *io)
2371 return io->ci_type == CIT_SETATTR &&
2372 (io->u.ci_setattr.sa_valid & ATTR_SIZE);
2375 struct cl_io *cl_io_top(struct cl_io *io);
2377 void cl_io_print(const struct lu_env *env, void *cookie,
2378 lu_printer_t printer, const struct cl_io *io);
2380 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2382 typeof(foo_io) __foo_io = (foo_io); \
2384 memset(&__foo_io->base, 0, \
2385 sizeof(*__foo_io) - offsetof(typeof(*__foo_io), base)); \
2390 /** \defgroup cl_page_list cl_page_list
2394 * Last page in the page list.
2396 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2398 LASSERT(plist->pl_nr > 0);
2399 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2402 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2404 LASSERT(plist->pl_nr > 0);
2405 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2409 * Iterate over pages in a page list.
2411 #define cl_page_list_for_each(page, list) \
2412 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2415 * Iterate over pages in a page list, taking possible removals into account.
2417 #define cl_page_list_for_each_safe(page, temp, list) \
2418 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2420 void cl_page_list_init (struct cl_page_list *plist);
2421 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2422 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2423 struct cl_page *page);
2424 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2425 struct cl_page *page);
2426 void cl_page_list_splice (struct cl_page_list *list,
2427 struct cl_page_list *head);
2428 void cl_page_list_del (const struct lu_env *env,
2429 struct cl_page_list *plist, struct cl_page *page);
2430 void cl_page_list_disown (const struct lu_env *env,
2431 struct cl_io *io, struct cl_page_list *plist);
2432 void cl_page_list_assume (const struct lu_env *env,
2433 struct cl_io *io, struct cl_page_list *plist);
2434 void cl_page_list_discard(const struct lu_env *env,
2435 struct cl_io *io, struct cl_page_list *plist);
2436 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2438 void cl_2queue_init (struct cl_2queue *queue);
2439 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2440 void cl_2queue_disown (const struct lu_env *env,
2441 struct cl_io *io, struct cl_2queue *queue);
2442 void cl_2queue_assume (const struct lu_env *env,
2443 struct cl_io *io, struct cl_2queue *queue);
2444 void cl_2queue_discard (const struct lu_env *env,
2445 struct cl_io *io, struct cl_2queue *queue);
2446 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2447 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2449 /** @} cl_page_list */
2451 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2452 struct cl_req_attr *attr);
2454 /** \defgroup cl_sync_io cl_sync_io
2458 * Anchor for synchronous transfer. This is allocated on a stack by thread
2459 * doing synchronous transfer, and a pointer to this structure is set up in
2460 * every page submitted for transfer. Transfer completion routine updates
2461 * anchor and wakes up waiting thread when transfer is complete.
2464 /** number of pages yet to be transferred. */
2465 atomic_t csi_sync_nr;
2468 /** barrier of destroy this structure */
2469 atomic_t csi_barrier;
2470 /** completion to be signaled when transfer is complete. */
2471 wait_queue_head_t csi_waitq;
2472 /** callback to invoke when this IO is finished */
2473 void (*csi_end_io)(const struct lu_env *,
2474 struct cl_sync_io *);
2477 void cl_sync_io_init(struct cl_sync_io *anchor, int nr,
2478 void (*end)(const struct lu_env *, struct cl_sync_io *));
2479 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2481 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2483 void cl_sync_io_end(const struct lu_env *env, struct cl_sync_io *anchor);
2485 /** @} cl_sync_io */
2487 /** \defgroup cl_env cl_env
2489 * lu_env handling for a client.
2491 * lu_env is an environment within which lustre code executes. Its major part
2492 * is lu_context---a fast memory allocation mechanism that is used to conserve
2493 * precious kernel stack space. Originally lu_env was designed for a server,
2496 * - there is a (mostly) fixed number of threads, and
2498 * - call chains have no non-lustre portions inserted between lustre code.
2500 * On a client both these assumtpion fails, because every user thread can
2501 * potentially execute lustre code as part of a system call, and lustre calls
2502 * into VFS or MM that call back into lustre.
2504 * To deal with that, cl_env wrapper functions implement the following
2507 * - allocation and destruction of environment is amortized by caching no
2508 * longer used environments instead of destroying them;
2510 * \see lu_env, lu_context, lu_context_key
2513 struct lu_env *cl_env_get(__u16 *refcheck);
2514 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2515 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2516 unsigned cl_env_cache_purge(unsigned nr);
2517 struct lu_env *cl_env_percpu_get(void);
2518 void cl_env_percpu_put(struct lu_env *env);
2525 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2526 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2528 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2529 struct lu_device_type *ldt,
2530 struct lu_device *next);
2533 int cl_global_init(void);
2534 void cl_global_fini(void);
2536 #endif /* _LINUX_CL_OBJECT_H */