4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 #ifndef _LUSTRE_CL_OBJECT_H
32 #define _LUSTRE_CL_OBJECT_H
34 /** \defgroup clio clio
36 * Client objects implement io operations and cache pages.
38 * Examples: lov and osc are implementations of cl interface.
40 * Big Theory Statement.
44 * Client implementation is based on the following data-types:
50 * - cl_lock represents an extent lock on an object.
52 * - cl_io represents high-level i/o activity such as whole read/write
53 * system call, or write-out of pages from under the lock being
54 * canceled. cl_io has sub-ios that can be stopped and resumed
55 * independently, thus achieving high degree of transfer
56 * parallelism. Single cl_io can be advanced forward by
57 * the multiple threads (although in the most usual case of
58 * read/write system call it is associated with the single user
59 * thread, that issued the system call).
63 * - to avoid confusion high-level I/O operation like read or write system
64 * call is referred to as "an io", whereas low-level I/O operation, like
65 * RPC, is referred to as "a transfer"
67 * - "generic code" means generic (not file system specific) code in the
68 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
69 * is not layer specific.
75 * - cl_object_header::coh_page_guard
78 * See the top comment in cl_object.c for the description of overall locking and
79 * reference-counting design.
81 * See comments below for the description of i/o, page, and dlm-locking
88 * super-class definitions.
90 #include <linux/aio.h>
93 #include <libcfs/libcfs.h>
94 #include <lu_object.h>
95 #include <linux/atomic.h>
96 #include <linux/mutex.h>
97 #include <linux/radix-tree.h>
98 #include <linux/spinlock.h>
99 #include <linux/wait.h>
100 #include <linux/pagevec.h>
101 #include <libcfs/linux/linux-misc.h>
102 #include <lustre_dlm.h>
112 struct cl_page_slice;
114 struct cl_lock_slice;
116 struct cl_lock_operations;
117 struct cl_page_operations;
125 * Device in the client stack.
127 * \see vvp_device, lov_device, lovsub_device, osc_device
131 struct lu_device cd_lu_dev;
134 /** \addtogroup cl_object cl_object
137 * "Data attributes" of cl_object. Data attributes can be updated
138 * independently for a sub-object, and top-object's attributes are calculated
139 * from sub-objects' ones.
142 /** Object size, in bytes */
145 unsigned int cat_kms_valid:1;
147 * Known minimal size, in bytes.
149 * This is only valid when at least one DLM lock is held.
152 /** Modification time. Measured in seconds since epoch. */
154 /** Access time. Measured in seconds since epoch. */
156 /** Change time. Measured in seconds since epoch. */
159 * Blocks allocated to this cl_object on the server file system.
161 * \todo XXX An interface for block size is needed.
165 * User identifier for quota purposes.
169 * Group identifier for quota purposes.
173 /* nlink of the directory */
176 /* Project identifier for quota purpose. */
181 * Fields in cl_attr that are being set.
196 * Sub-class of lu_object with methods common for objects on the client
199 * cl_object: represents a regular file system object, both a file and a
200 * stripe. cl_object is based on lu_object: it is identified by a fid,
201 * layered, cached, hashed, and lrued. Important distinction with the server
202 * side, where md_object and dt_object are used, is that cl_object "fans out"
203 * at the lov/sns level: depending on the file layout, single file is
204 * represented as a set of "sub-objects" (stripes). At the implementation
205 * level, struct lov_object contains an array of cl_objects. Each sub-object
206 * is a full-fledged cl_object, having its fid, living in the lru and hash
209 * This leads to the next important difference with the server side: on the
210 * client, it's quite usual to have objects with the different sequence of
211 * layers. For example, typical top-object is composed of the following
217 * whereas its sub-objects are composed of
222 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
223 * track of the object-subobject relationship.
225 * Sub-objects are not cached independently: when top-object is about to
226 * be discarded from the memory, all its sub-objects are torn-down and
229 * \see vvp_object, lov_object, lovsub_object, osc_object
233 struct lu_object co_lu;
234 /** per-object-layer operations */
235 const struct cl_object_operations *co_ops;
236 /** offset of page slice in cl_page buffer */
241 * Description of the client object configuration. This is used for the
242 * creation of a new client object that is identified by a more state than
245 struct cl_object_conf {
247 struct lu_object_conf coc_lu;
250 * Object layout. This is consumed by lov.
252 struct lu_buf coc_layout;
254 * Description of particular stripe location in the
255 * cluster. This is consumed by osc.
257 struct lov_oinfo *coc_oinfo;
260 * VFS inode. This is consumed by vvp.
262 struct inode *coc_inode;
264 * Layout lock handle.
266 struct ldlm_lock *coc_lock;
269 * Operation to handle layout, OBJECT_CONF_XYZ.
275 /** configure layout, set up a new stripe, must be called while
276 * holding layout lock. */
278 /** invalidate the current stripe configuration due to losing
280 OBJECT_CONF_INVALIDATE = 1,
281 /** wait for old layout to go away so that new layout can be
287 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
288 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
292 /** the buffer to return the layout in lov_mds_md format. */
293 struct lu_buf cl_buf;
294 /** size of layout in lov_mds_md format. */
296 /** Layout generation. */
298 /** whether layout is a composite one */
299 bool cl_is_composite;
300 /** Whether layout is a HSM released one */
314 * Operations implemented for each cl object layer.
316 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
318 struct cl_object_operations {
320 * Initialize page slice for this layer. Called top-to-bottom through
321 * every object layer when a new cl_page is instantiated. Layer
322 * keeping private per-page data, or requiring its own page operations
323 * vector should allocate these data here, and attach then to the page
324 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
327 * \retval NULL success.
329 * \retval ERR_PTR(errno) failure code.
331 * \retval valid-pointer pointer to already existing referenced page
332 * to be used instead of newly created.
334 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
335 struct cl_page *page, pgoff_t index);
337 * Initialize lock slice for this layer. Called top-to-bottom through
338 * every object layer when a new cl_lock is instantiated. Layer
339 * keeping private per-lock data, or requiring its own lock operations
340 * vector should allocate these data here, and attach then to the lock
341 * by calling cl_lock_slice_add(). Mandatory.
343 int (*coo_lock_init)(const struct lu_env *env,
344 struct cl_object *obj, struct cl_lock *lock,
345 const struct cl_io *io);
347 * Initialize io state for a given layer.
349 * called top-to-bottom once per io existence to initialize io
350 * state. If layer wants to keep some state for this type of io, it
351 * has to embed struct cl_io_slice in lu_env::le_ses, and register
352 * slice with cl_io_slice_add(). It is guaranteed that all threads
353 * participating in this io share the same session.
355 int (*coo_io_init)(const struct lu_env *env,
356 struct cl_object *obj, struct cl_io *io);
358 * Fill portion of \a attr that this layer controls. This method is
359 * called top-to-bottom through all object layers.
361 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
363 * \return 0: to continue
364 * \return +ve: to stop iterating through layers (but 0 is returned
365 * from enclosing cl_object_attr_get())
366 * \return -ve: to signal error
368 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
369 struct cl_attr *attr);
373 * \a valid is a bitmask composed from enum #cl_attr_valid, and
374 * indicating what attributes are to be set.
376 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
378 * \return the same convention as for
379 * cl_object_operations::coo_attr_get() is used.
381 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
382 const struct cl_attr *attr, unsigned valid);
384 * Update object configuration. Called top-to-bottom to modify object
387 * XXX error conditions and handling.
389 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
390 const struct cl_object_conf *conf);
392 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
393 * object. Layers are supposed to fill parts of \a lvb that will be
394 * shipped to the glimpse originator as a glimpse result.
396 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
397 * \see osc_object_glimpse()
399 int (*coo_glimpse)(const struct lu_env *env,
400 const struct cl_object *obj, struct ost_lvb *lvb);
402 * Object prune method. Called when the layout is going to change on
403 * this object, therefore each layer has to clean up their cache,
404 * mainly pages and locks.
406 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
408 * Object getstripe method.
410 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
411 struct lov_user_md __user *lum, size_t size);
413 * Get FIEMAP mapping from the object.
415 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
416 struct ll_fiemap_info_key *fmkey,
417 struct fiemap *fiemap, size_t *buflen);
419 * Get layout and generation of the object.
421 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
422 struct cl_layout *layout);
424 * Get maximum size of the object.
426 loff_t (*coo_maxbytes)(struct cl_object *obj);
428 * Set request attributes.
430 void (*coo_req_attr_set)(const struct lu_env *env,
431 struct cl_object *obj,
432 struct cl_req_attr *attr);
434 * Flush \a obj data corresponding to \a lock. Used for DoM
435 * locks in llite's cancelling blocking ast callback.
437 int (*coo_object_flush)(const struct lu_env *env,
438 struct cl_object *obj,
439 struct ldlm_lock *lock);
441 * operate upon inode. Used in LOV to lock/unlock inode from vvp layer.
443 int (*coo_inode_ops)(const struct lu_env *env, struct cl_object *obj,
444 enum coo_inode_opc opc, void *data);
448 * Extended header for client object.
450 struct cl_object_header {
451 /** Standard lu_object_header. cl_object::co_lu::lo_header points
453 struct lu_object_header coh_lu;
456 * Parent object. It is assumed that an object has a well-defined
457 * parent, but not a well-defined child (there may be multiple
458 * sub-objects, for the same top-object). cl_object_header::coh_parent
459 * field allows certain code to be written generically, without
460 * limiting possible cl_object layouts unduly.
462 struct cl_object_header *coh_parent;
464 * Protects consistency between cl_attr of parent object and
465 * attributes of sub-objects, that the former is calculated ("merged")
468 * \todo XXX this can be read/write lock if needed.
470 spinlock_t coh_attr_guard;
472 * Size of cl_page + page slices
474 unsigned short coh_page_bufsize;
476 * Number of objects above this one: 0 for a top-object, 1 for its
479 unsigned char coh_nesting;
483 * Helper macro: iterate over all layers of the object \a obj, assigning every
484 * layer top-to-bottom to \a slice.
486 #define cl_object_for_each(slice, obj) \
487 list_for_each_entry((slice), \
488 &(obj)->co_lu.lo_header->loh_layers,\
492 * Helper macro: iterate over all layers of the object \a obj, assigning every
493 * layer bottom-to-top to \a slice.
495 #define cl_object_for_each_reverse(slice, obj) \
496 list_for_each_entry_reverse((slice), \
497 &(obj)->co_lu.lo_header->loh_layers,\
502 #define CL_PAGE_EOF ((pgoff_t)~0ull)
504 /** \addtogroup cl_page cl_page
508 * Layered client page.
510 * cl_page: represents a portion of a file, cached in the memory. All pages
511 * of the given file are of the same size, and are kept in the radix tree
512 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
513 * of the top-level file object are first class cl_objects, they have their
514 * own radix trees of pages and hence page is implemented as a sequence of
515 * struct cl_pages's, linked into double-linked list through
516 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
517 * corresponding radix tree at the corresponding logical offset.
519 * cl_page is associated with VM page of the hosting environment (struct
520 * page in Linux kernel, for example), struct page. It is assumed, that this
521 * association is implemented by one of cl_page layers (top layer in the
522 * current design) that
524 * - intercepts per-VM-page call-backs made by the environment (e.g.,
527 * - translates state (page flag bits) and locking between lustre and
530 * The association between cl_page and struct page is immutable and
531 * established when cl_page is created.
533 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
534 * this io an exclusive access to this page w.r.t. other io attempts and
535 * various events changing page state (such as transfer completion, or
536 * eviction of the page from the memory). Note, that in general cl_io
537 * cannot be identified with a particular thread, and page ownership is not
538 * exactly equal to the current thread holding a lock on the page. Layer
539 * implementing association between cl_page and struct page has to implement
540 * ownership on top of available synchronization mechanisms.
542 * While lustre client maintains the notion of an page ownership by io,
543 * hosting MM/VM usually has its own page concurrency control
544 * mechanisms. For example, in Linux, page access is synchronized by the
545 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
546 * takes care to acquire and release such locks as necessary around the
547 * calls to the file system methods (->readpage(), ->prepare_write(),
548 * ->commit_write(), etc.). This leads to the situation when there are two
549 * different ways to own a page in the client:
551 * - client code explicitly and voluntary owns the page (cl_page_own());
553 * - VM locks a page and then calls the client, that has "to assume"
554 * the ownership from the VM (cl_page_assume()).
556 * Dual methods to release ownership are cl_page_disown() and
557 * cl_page_unassume().
559 * cl_page is reference counted (cl_page::cp_ref). When reference counter
560 * drops to 0, the page is returned to the cache, unless it is in
561 * cl_page_state::CPS_FREEING state, in which case it is immediately
564 * The general logic guaranteeing the absence of "existential races" for
565 * pages is the following:
567 * - there are fixed known ways for a thread to obtain a new reference
570 * - by doing a lookup in the cl_object radix tree, protected by the
573 * - by starting from VM-locked struct page and following some
574 * hosting environment method (e.g., following ->private pointer in
575 * the case of Linux kernel), see cl_vmpage_page();
577 * - when the page enters cl_page_state::CPS_FREEING state, all these
578 * ways are severed with the proper synchronization
579 * (cl_page_delete());
581 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
584 * - no new references to the page in cl_page_state::CPS_FREEING state
585 * are allowed (checked in cl_page_get()).
587 * Together this guarantees that when last reference to a
588 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
589 * page, as neither references to it can be acquired at that point, nor
592 * cl_page is a state machine. States are enumerated in enum
593 * cl_page_state. Possible state transitions are enumerated in
594 * cl_page_state_set(). State transition process (i.e., actual changing of
595 * cl_page::cp_state field) is protected by the lock on the underlying VM
598 * Linux Kernel implementation.
600 * Binding between cl_page and struct page (which is a typedef for
601 * struct page) is implemented in the vvp layer. cl_page is attached to the
602 * ->private pointer of the struct page, together with the setting of
603 * PG_private bit in page->flags, and acquiring additional reference on the
604 * struct page (much like struct buffer_head, or any similar file system
605 * private data structures).
607 * PG_locked lock is used to implement both ownership and transfer
608 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
609 * states. No additional references are acquired for the duration of the
612 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
613 * write-out is "protected" by the special PG_writeback bit.
617 * States of cl_page. cl_page.c assumes particular order here.
619 * The page state machine is rather crude, as it doesn't recognize finer page
620 * states like "dirty" or "up to date". This is because such states are not
621 * always well defined for the whole stack (see, for example, the
622 * implementation of the read-ahead, that hides page up-to-dateness to track
623 * cache hits accurately). Such sub-states are maintained by the layers that
624 * are interested in them.
628 * Page is in the cache, un-owned. Page leaves cached state in the
631 * - [cl_page_state::CPS_OWNED] io comes across the page and
634 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
635 * req-formation engine decides that it wants to include this page
636 * into an RPC being constructed, and yanks it from the cache;
638 * - [cl_page_state::CPS_FREEING] VM callback is executed to
639 * evict the page form the memory;
641 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
645 * Page is exclusively owned by some cl_io. Page may end up in this
646 * state as a result of
648 * - io creating new page and immediately owning it;
650 * - [cl_page_state::CPS_CACHED] io finding existing cached page
653 * - [cl_page_state::CPS_OWNED] io finding existing owned page
654 * and waiting for owner to release the page;
656 * Page leaves owned state in the following cases:
658 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
659 * the cache, doing nothing;
661 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
664 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
665 * transfer for this page;
667 * - [cl_page_state::CPS_FREEING] io decides to destroy this
668 * page (e.g., as part of truncate or extent lock cancellation).
670 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
674 * Page is being written out, as a part of a transfer. This state is
675 * entered when req-formation logic decided that it wants this page to
676 * be sent through the wire _now_. Specifically, it means that once
677 * this state is achieved, transfer completion handler (with either
678 * success or failure indication) is guaranteed to be executed against
679 * this page independently of any locks and any scheduling decisions
680 * made by the hosting environment (that effectively means that the
681 * page is never put into cl_page_state::CPS_PAGEOUT state "in
682 * advance". This property is mentioned, because it is important when
683 * reasoning about possible dead-locks in the system). The page can
684 * enter this state as a result of
686 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
687 * write-out of this page, or
689 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
690 * that it has enough dirty pages cached to issue a "good"
693 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
694 * is completed---it is moved into cl_page_state::CPS_CACHED state.
696 * Underlying VM page is locked for the duration of transfer.
698 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
702 * Page is being read in, as a part of a transfer. This is quite
703 * similar to the cl_page_state::CPS_PAGEOUT state, except that
704 * read-in is always "immediate"---there is no such thing a sudden
705 * construction of read request from cached, presumably not up to date,
708 * Underlying VM page is locked for the duration of transfer.
710 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
714 * Page is being destroyed. This state is entered when client decides
715 * that page has to be deleted from its host object, as, e.g., a part
718 * Once this state is reached, there is no way to escape it.
720 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
727 /** Host page, the page is from the host inode which the cl_page
731 /** Transient page, the transient cl_page is used to bind a cl_page
732 * to vmpage which is not belonging to the same object of cl_page.
733 * it is used in DirectIO and lockless IO. */
738 #define CP_STATE_BITS 4
739 #define CP_TYPE_BITS 2
740 #define CP_MAX_LAYER 2
743 * Fields are protected by the lock on struct page, except for atomics and
746 * \invariant Data type invariants are in cl_page_invariant(). Basically:
747 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
748 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
749 * cl_page::cp_owner (when set).
752 /** Reference counter. */
754 /** layout_entry + stripe index, composed using lov_comp_index() */
755 unsigned int cp_lov_index;
756 /** page->index of the page within the whole file */
757 pgoff_t cp_page_index;
758 /** An object this page is a part of. Immutable after creation. */
759 struct cl_object *cp_obj;
761 struct page *cp_vmpage;
763 * Assigned if doing direct IO, because in this case cp_vmpage is not
764 * a valid page cache page, hence the inode cannot be inferred from
765 * cp_vmpage->mapping->host.
767 struct inode *cp_inode;
768 /** Linkage of pages within group. Pages must be owned */
769 struct list_head cp_batch;
770 /** array of slices offset. Immutable after creation. */
771 unsigned char cp_layer_offset[CP_MAX_LAYER];
772 /** current slice index */
773 unsigned char cp_layer_count:2;
775 * Page state. This field is const to avoid accidental update, it is
776 * modified only internally within cl_page.c. Protected by a VM lock.
778 enum cl_page_state cp_state:CP_STATE_BITS;
780 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
783 enum cl_page_type cp_type:CP_TYPE_BITS;
784 unsigned cp_defer_uptodate:1,
787 /* which slab kmem index this memory allocated from */
788 short int cp_kmem_index;
791 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
792 * by sub-io. Protected by a VM lock.
794 struct cl_io *cp_owner;
795 /** List of references to this page, for debugging. */
796 struct lu_ref cp_reference;
797 /** Link to an object, for debugging. */
798 struct lu_ref_link cp_obj_ref;
799 /** Link to a queue, for debugging. */
800 struct lu_ref_link cp_queue_ref;
801 /** Assigned if doing a sync_io */
802 struct cl_sync_io *cp_sync_io;
806 * Per-layer part of cl_page.
808 * \see vvp_page, lov_page, osc_page
810 struct cl_page_slice {
811 struct cl_page *cpl_page;
812 const struct cl_page_operations *cpl_ops;
816 * Lock mode. For the client extent locks.
828 * Requested transfer type.
837 * Per-layer page operations.
839 * Methods taking an \a io argument are for the activity happening in the
840 * context of given \a io. Page is assumed to be owned by that io, except for
843 * \see vvp_page_ops, lov_page_ops, osc_page_ops
845 struct cl_page_operations {
847 * cl_page<->struct page methods. Only one layer in the stack has to
848 * implement these. Current code assumes that this functionality is
849 * provided by the topmost layer, see __cl_page_disown() as an example.
853 * Update file attributes when all we have is this page. Used for tiny
854 * writes to update attributes when we don't have a full cl_io.
856 void (*cpo_page_touch)(const struct lu_env *env,
857 const struct cl_page_slice *slice, size_t to);
863 * Called when page is truncated from the object. Optional.
865 * \see cl_page_discard()
866 * \see vvp_page_discard(), osc_page_discard()
868 void (*cpo_discard)(const struct lu_env *env,
869 const struct cl_page_slice *slice,
872 * Called when page is removed from the cache, and is about to being
873 * destroyed. Optional.
875 * \see cl_page_delete()
876 * \see vvp_page_delete(), osc_page_delete()
878 void (*cpo_delete)(const struct lu_env *env,
879 const struct cl_page_slice *slice);
881 * Optional debugging helper. Prints given page slice.
883 * \see cl_page_print()
885 int (*cpo_print)(const struct lu_env *env,
886 const struct cl_page_slice *slice,
887 void *cookie, lu_printer_t p);
896 * Request type dependent vector of operations.
898 * Transfer operations depend on transfer mode (cl_req_type). To avoid
899 * passing transfer mode to each and every of these methods, and to
900 * avoid branching on request type inside of the methods, separate
901 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
902 * provided. That is, method invocation usually looks like
904 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
908 * Completion handler. This is guaranteed to be eventually
909 * fired after cl_page_prep() or cl_page_make_ready() call.
911 * This method can be called in a non-blocking context. It is
912 * guaranteed however, that the page involved and its object
913 * are pinned in memory (and, hence, calling cl_page_put() is
916 * \see cl_page_completion()
918 void (*cpo_completion)(const struct lu_env *env,
919 const struct cl_page_slice *slice,
923 * Tell transfer engine that only [to, from] part of a page should be
926 * This is used for immediate transfers.
928 * \todo XXX this is not very good interface. It would be much better
929 * if all transfer parameters were supplied as arguments to
930 * cl_io_operations::cio_submit() call, but it is not clear how to do
931 * this for page queues.
933 * \see cl_page_clip()
935 void (*cpo_clip)(const struct lu_env *env,
936 const struct cl_page_slice *slice,
939 * Write out a page by kernel. This is only called by ll_writepage
942 * \see cl_page_flush()
944 int (*cpo_flush)(const struct lu_env *env,
945 const struct cl_page_slice *slice,
951 * Helper macro, dumping detailed information about \a page into a log.
953 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
955 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
956 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
957 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
958 CDEBUG(mask, format , ## __VA_ARGS__); \
963 * Helper macro, dumping shorter information about \a page into a log.
965 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
967 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
968 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
969 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
970 CDEBUG(mask, format , ## __VA_ARGS__); \
974 static inline struct page *cl_page_vmpage(const struct cl_page *page)
976 LASSERT(page->cp_vmpage != NULL);
977 return page->cp_vmpage;
980 static inline pgoff_t cl_page_index(const struct cl_page *cp)
982 return cl_page_vmpage(cp)->index;
986 * Check if a cl_page is in use.
988 * Client cache holds a refcount, this refcount will be dropped when
989 * the page is taken out of cache, see vvp_page_delete().
991 static inline bool __page_in_use(const struct cl_page *page, int refc)
993 return (refcount_read(&page->cp_ref) > refc + 1);
997 * Caller itself holds a refcount of cl_page.
999 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1001 * Caller doesn't hold a refcount.
1003 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1007 /** \addtogroup cl_lock cl_lock
1011 * Extent locking on the client.
1015 * The locking model of the new client code is built around
1019 * data-type representing an extent lock on a regular file. cl_lock is a
1020 * layered object (much like cl_object and cl_page), it consists of a header
1021 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1022 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1024 * Typical cl_lock consists of one layer:
1026 * - lov_lock (lov specific data).
1028 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1029 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1033 * Each sub-lock is associated with a cl_object (representing stripe
1034 * sub-object or the file to which top-level cl_lock is associated to), and is
1035 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1036 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1037 * is different from cl_page, that doesn't fan out (there is usually exactly
1038 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1039 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1043 * cl_lock is a cacheless data container for the requirements of locks to
1044 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1047 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1048 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1050 * INTERFACE AND USAGE
1052 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1053 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1054 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1055 * consists of multiple sub cl_locks, each sub locks will be enqueued
1056 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1057 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1060 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1061 * method will be called for each layer to release the resource held by this
1062 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1063 * clo_enqueue time, is released.
1065 * LDLM lock can only be canceled if there is no cl_lock using it.
1067 * Overall process of the locking during IO operation is as following:
1069 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1070 * is called on each layer. Responsibility of this method is to add locks,
1071 * needed by a given layer into cl_io.ci_lockset.
1073 * - once locks for all layers were collected, they are sorted to avoid
1074 * dead-locks (cl_io_locks_sort()), and enqueued.
1076 * - when all locks are acquired, IO is performed;
1078 * - locks are released after IO is complete.
1080 * Striping introduces major additional complexity into locking. The
1081 * fundamental problem is that it is generally unsafe to actively use (hold)
1082 * two locks on the different OST servers at the same time, as this introduces
1083 * inter-server dependency and can lead to cascading evictions.
1085 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1086 * that no multi-stripe locks are taken (note that this design abandons POSIX
1087 * read/write semantics). Such pieces ideally can be executed concurrently. At
1088 * the same time, certain types of IO cannot be sub-divived, without
1089 * sacrificing correctness. This includes:
1091 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1094 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1096 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1097 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1098 * has to be held together with the usual lock on [offset, offset + count].
1100 * Interaction with DLM
1102 * In the expected setup, cl_lock is ultimately backed up by a collection of
1103 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1104 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1105 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1106 * description of interaction with DLM.
1112 struct cl_lock_descr {
1113 /** Object this lock is granted for. */
1114 struct cl_object *cld_obj;
1115 /** Index of the first page protected by this lock. */
1117 /** Index of the last page (inclusive) protected by this lock. */
1119 /** Group ID, for group lock */
1122 enum cl_lock_mode cld_mode;
1124 * flags to enqueue lock. A combination of bit-flags from
1125 * enum cl_enq_flags.
1127 __u32 cld_enq_flags;
1130 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1131 #define PDESCR(descr) \
1132 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1133 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1135 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1138 * Layered client lock.
1141 /** List of slices. Immutable after creation. */
1142 struct list_head cll_layers;
1143 /** lock attribute, extent, cl_object, etc. */
1144 struct cl_lock_descr cll_descr;
1148 * Per-layer part of cl_lock
1150 * \see lov_lock, osc_lock
1152 struct cl_lock_slice {
1153 struct cl_lock *cls_lock;
1154 /** Object slice corresponding to this lock slice. Immutable after
1156 struct cl_object *cls_obj;
1157 const struct cl_lock_operations *cls_ops;
1158 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1159 struct list_head cls_linkage;
1164 * \see lov_lock_ops, osc_lock_ops
1166 struct cl_lock_operations {
1169 * Attempts to enqueue the lock. Called top-to-bottom.
1171 * \retval 0 this layer has enqueued the lock successfully
1172 * \retval >0 this layer has enqueued the lock, but need to wait on
1173 * @anchor for resources
1174 * \retval -ve failure
1176 * \see lov_lock_enqueue(), osc_lock_enqueue()
1178 int (*clo_enqueue)(const struct lu_env *env,
1179 const struct cl_lock_slice *slice,
1180 struct cl_io *io, struct cl_sync_io *anchor);
1182 * Cancel a lock, release its DLM lock ref, while does not cancel the
1185 void (*clo_cancel)(const struct lu_env *env,
1186 const struct cl_lock_slice *slice);
1189 * Destructor. Frees resources and the slice.
1191 * \see lov_lock_fini(), osc_lock_fini()
1193 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1195 * Optional debugging helper. Prints given lock slice.
1197 int (*clo_print)(const struct lu_env *env,
1198 void *cookie, lu_printer_t p,
1199 const struct cl_lock_slice *slice);
1202 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1204 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1205 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1206 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1207 CDEBUG(mask, format , ## __VA_ARGS__); \
1211 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1215 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1221 /** \addtogroup cl_page_list cl_page_list
1222 * Page list used to perform collective operations on a group of pages.
1224 * Pages are added to the list one by one. cl_page_list acquires a reference
1225 * for every page in it. Page list is used to perform collective operations on
1228 * - submit pages for an immediate transfer,
1230 * - own pages on behalf of certain io (waiting for each page in turn),
1234 * When list is finalized, it releases references on all pages it still has.
1236 * \todo XXX concurrency control.
1240 struct cl_page_list {
1242 struct list_head pl_pages;
1246 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1247 * contains an incoming page list and an outgoing page list.
1250 struct cl_page_list c2_qin;
1251 struct cl_page_list c2_qout;
1254 /** @} cl_page_list */
1256 /** \addtogroup cl_io cl_io
1261 * cl_io represents a high level I/O activity like
1262 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1265 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1266 * important distinction. We want to minimize number of calls to the allocator
1267 * in the fast path, e.g., in the case of read(2) when everything is cached:
1268 * client already owns the lock over region being read, and data are cached
1269 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1270 * per-layer io state is stored in the session, associated with the io, see
1271 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1272 * by using free-lists, see cl_env_get().
1274 * There is a small predefined number of possible io types, enumerated in enum
1277 * cl_io is a state machine, that can be advanced concurrently by the multiple
1278 * threads. It is up to these threads to control the concurrency and,
1279 * specifically, to detect when io is done, and its state can be safely
1282 * For read/write io overall execution plan is as following:
1284 * (0) initialize io state through all layers;
1286 * (1) loop: prepare chunk of work to do
1288 * (2) call all layers to collect locks they need to process current chunk
1290 * (3) sort all locks to avoid dead-locks, and acquire them
1292 * (4) process the chunk: call per-page methods
1293 * cl_io_operations::cio_prepare_write(),
1294 * cl_io_operations::cio_commit_write() for write)
1300 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1301 * address allocation efficiency issues mentioned above), and returns with the
1302 * special error condition from per-page method when current sub-io has to
1303 * block. This causes io loop to be repeated, and lov switches to the next
1304 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1309 /** read system call */
1311 /** write system call */
1313 /** truncate, utime system calls */
1315 /** get data version */
1318 * page fault handling
1322 * fsync system call handling
1323 * To write out a range of file
1327 * glimpse. An io context to acquire glimpse lock.
1331 * Miscellaneous io. This is used for occasional io activity that
1332 * doesn't fit into other types. Currently this is used for:
1334 * - cancellation of an extent lock. This io exists as a context
1335 * to write dirty pages from under the lock being canceled back
1338 * - VM induced page write-out. An io context for writing page out
1339 * for memory cleansing;
1341 * - grouplock. An io context to acquire group lock.
1343 * CIT_MISC io is used simply as a context in which locks and pages
1344 * are manipulated. Such io has no internal "process", that is,
1345 * cl_io_loop() is never called for it.
1350 * To give advice about access of a file
1354 * SEEK_HOLE/SEEK_DATA handling to search holes or data
1355 * across all file objects
1362 * States of cl_io state machine
1365 /** Not initialized. */
1369 /** IO iteration started. */
1373 /** Actual IO is in progress. */
1375 /** IO for the current iteration finished. */
1377 /** Locks released. */
1379 /** Iteration completed. */
1381 /** cl_io finalized. */
1386 * IO state private for a layer.
1388 * This is usually embedded into layer session data, rather than allocated
1391 * \see vvp_io, lov_io, osc_io
1393 struct cl_io_slice {
1394 struct cl_io *cis_io;
1395 /** corresponding object slice. Immutable after creation. */
1396 struct cl_object *cis_obj;
1397 /** io operations. Immutable after creation. */
1398 const struct cl_io_operations *cis_iop;
1400 * linkage into a list of all slices for a given cl_io, hanging off
1401 * cl_io::ci_layers. Immutable after creation.
1403 struct list_head cis_linkage;
1406 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1409 struct cl_read_ahead {
1410 /* Maximum page index the readahead window will end.
1411 * This is determined DLM lock coverage, RPC and stripe boundary.
1412 * cra_end is included. */
1413 pgoff_t cra_end_idx;
1414 /* optimal RPC size for this read, by pages */
1415 unsigned long cra_rpc_pages;
1416 /* Release callback. If readahead holds resources underneath, this
1417 * function should be called to release it. */
1418 void (*cra_release)(const struct lu_env *env,
1419 struct cl_read_ahead *ra);
1421 /* Callback data for cra_release routine */
1425 /* whether lock is in contention */
1426 bool cra_contention;
1429 static inline void cl_read_ahead_release(const struct lu_env *env,
1430 struct cl_read_ahead *ra)
1432 if (ra->cra_release != NULL)
1433 ra->cra_release(env, ra);
1434 memset(ra, 0, sizeof(*ra));
1439 * Per-layer io operations.
1440 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1442 struct cl_io_operations {
1444 * Vector of io state transition methods for every io type.
1446 * \see cl_page_operations::io
1450 * Prepare io iteration at a given layer.
1452 * Called top-to-bottom at the beginning of each iteration of
1453 * "io loop" (if it makes sense for this type of io). Here
1454 * layer selects what work it will do during this iteration.
1456 * \see cl_io_operations::cio_iter_fini()
1458 int (*cio_iter_init) (const struct lu_env *env,
1459 const struct cl_io_slice *slice);
1461 * Finalize io iteration.
1463 * Called bottom-to-top at the end of each iteration of "io
1464 * loop". Here layers can decide whether IO has to be
1467 * \see cl_io_operations::cio_iter_init()
1469 void (*cio_iter_fini) (const struct lu_env *env,
1470 const struct cl_io_slice *slice);
1472 * Collect locks for the current iteration of io.
1474 * Called top-to-bottom to collect all locks necessary for
1475 * this iteration. This methods shouldn't actually enqueue
1476 * anything, instead it should post a lock through
1477 * cl_io_lock_add(). Once all locks are collected, they are
1478 * sorted and enqueued in the proper order.
1480 int (*cio_lock) (const struct lu_env *env,
1481 const struct cl_io_slice *slice);
1483 * Finalize unlocking.
1485 * Called bottom-to-top to finish layer specific unlocking
1486 * functionality, after generic code released all locks
1487 * acquired by cl_io_operations::cio_lock().
1489 void (*cio_unlock)(const struct lu_env *env,
1490 const struct cl_io_slice *slice);
1492 * Start io iteration.
1494 * Once all locks are acquired, called top-to-bottom to
1495 * commence actual IO. In the current implementation,
1496 * top-level vvp_io_{read,write}_start() does all the work
1497 * synchronously by calling generic_file_*(), so other layers
1498 * are called when everything is done.
1500 int (*cio_start)(const struct lu_env *env,
1501 const struct cl_io_slice *slice);
1503 * Called top-to-bottom at the end of io loop. Here layer
1504 * might wait for an unfinished asynchronous io.
1506 void (*cio_end) (const struct lu_env *env,
1507 const struct cl_io_slice *slice);
1509 * Called bottom-to-top to notify layers that read/write IO
1510 * iteration finished, with \a nob bytes transferred.
1512 void (*cio_advance)(const struct lu_env *env,
1513 const struct cl_io_slice *slice,
1516 * Called once per io, bottom-to-top to release io resources.
1518 void (*cio_fini) (const struct lu_env *env,
1519 const struct cl_io_slice *slice);
1523 * Submit pages from \a queue->c2_qin for IO, and move
1524 * successfully submitted pages into \a queue->c2_qout. Return
1525 * non-zero if failed to submit even the single page. If
1526 * submission failed after some pages were moved into \a
1527 * queue->c2_qout, completion callback with non-zero ioret is
1530 int (*cio_submit)(const struct lu_env *env,
1531 const struct cl_io_slice *slice,
1532 enum cl_req_type crt,
1533 struct cl_2queue *queue);
1535 * Queue async page for write.
1536 * The difference between cio_submit and cio_queue is that
1537 * cio_submit is for urgent request.
1539 int (*cio_commit_async)(const struct lu_env *env,
1540 const struct cl_io_slice *slice,
1541 struct cl_page_list *queue, int from, int to,
1544 * Release active extent.
1546 void (*cio_extent_release)(const struct lu_env *env,
1547 const struct cl_io_slice *slice);
1549 * Decide maximum read ahead extent
1551 * \pre io->ci_type == CIT_READ
1553 int (*cio_read_ahead)(const struct lu_env *env,
1554 const struct cl_io_slice *slice,
1555 pgoff_t start, struct cl_read_ahead *ra);
1558 * Reserve LRU slots before IO.
1560 int (*cio_lru_reserve) (const struct lu_env *env,
1561 const struct cl_io_slice *slice,
1562 loff_t pos, size_t bytes);
1564 * Optional debugging helper. Print given io slice.
1566 int (*cio_print)(const struct lu_env *env, void *cookie,
1567 lu_printer_t p, const struct cl_io_slice *slice);
1571 * Flags to lock enqueue procedure.
1576 * instruct server to not block, if conflicting lock is found. Instead
1577 * -EAGAIN is returned immediately.
1579 CEF_NONBLOCK = 0x00000001,
1581 * Tell lower layers this is a glimpse request, translated to
1582 * LDLM_FL_HAS_INTENT at LDLM layer.
1584 * Also, because glimpse locks never block other locks, we count this
1585 * as automatically compatible with other osc locks.
1586 * (see osc_lock_compatible)
1588 CEF_GLIMPSE = 0x00000002,
1590 * tell the server to instruct (though a flag in the blocking ast) an
1591 * owner of the conflicting lock, that it can drop dirty pages
1592 * protected by this lock, without sending them to the server.
1594 CEF_DISCARD_DATA = 0x00000004,
1596 * tell the sub layers that it must be a `real' lock. This is used for
1597 * mmapped-buffer locks, glimpse locks, manually requested locks
1598 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1601 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1603 CEF_MUST = 0x00000008,
1605 * tell the sub layers that never request a `real' lock. This flag is
1606 * not used currently.
1608 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1609 * conversion policy: ci_lockreq describes generic information of lock
1610 * requirement for this IO, especially for locks which belong to the
1611 * object doing IO; however, lock itself may have precise requirements
1612 * that are described by the enqueue flags.
1614 CEF_NEVER = 0x00000010,
1616 * tell the dlm layer this is a speculative lock request
1617 * speculative lock requests are locks which are not requested as part
1618 * of an I/O operation. Instead, they are requested because we expect
1619 * to use them in the future. They are requested asynchronously at the
1622 * Currently used for asynchronous glimpse locks and manually requested
1623 * locks (LU_LADVISE_LOCKAHEAD).
1625 CEF_SPECULATIVE = 0x00000020,
1627 * enqueue a lock to test DLM lock existence.
1629 CEF_PEEK = 0x00000040,
1631 * Lock match only. Used by group lock in I/O as group lock
1632 * is known to exist.
1634 CEF_LOCK_MATCH = 0x00000080,
1636 * tell the DLM layer to lock only the requested range
1638 CEF_LOCK_NO_EXPAND = 0x00000100,
1640 * mask of enq_flags.
1642 CEF_MASK = 0x000001ff,
1646 * Link between lock and io. Intermediate structure is needed, because the
1647 * same lock can be part of multiple io's simultaneously.
1649 struct cl_io_lock_link {
1650 /** linkage into one of cl_lockset lists. */
1651 struct list_head cill_linkage;
1652 struct cl_lock cill_lock;
1653 /** optional destructor */
1654 void (*cill_fini)(const struct lu_env *env,
1655 struct cl_io_lock_link *link);
1657 #define cill_descr cill_lock.cll_descr
1660 * Lock-set represents a collection of locks, that io needs at a
1661 * time. Generally speaking, client tries to avoid holding multiple locks when
1664 * - holding extent locks over multiple ost's introduces the danger of
1665 * "cascading timeouts";
1667 * - holding multiple locks over the same ost is still dead-lock prone,
1668 * see comment in osc_lock_enqueue(),
1670 * but there are certain situations where this is unavoidable:
1672 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1674 * - truncate has to take [new-size, EOF] lock for correctness;
1676 * - SNS has to take locks across full stripe for correctness;
1678 * - in the case when user level buffer, supplied to {read,write}(file0),
1679 * is a part of a memory mapped lustre file, client has to take a dlm
1680 * locks on file0, and all files that back up the buffer (or a part of
1681 * the buffer, that is being processed in the current chunk, in any
1682 * case, there are situations where at least 2 locks are necessary).
1684 * In such cases we at least try to take locks in the same consistent
1685 * order. To this end, all locks are first collected, then sorted, and then
1689 /** locks to be acquired. */
1690 struct list_head cls_todo;
1691 /** locks acquired. */
1692 struct list_head cls_done;
1696 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1697 * but 'req' is always to be thought as 'request' :-)
1699 enum cl_io_lock_dmd {
1700 /** Always lock data (e.g., O_APPEND). */
1702 /** Layers are free to decide between local and global locking. */
1704 /** Never lock: there is no cache (e.g., liblustre). */
1708 enum cl_fsync_mode {
1709 /** start writeback, do not wait for them to finish */
1711 /** start writeback and wait for them to finish */
1713 /** discard all of dirty pages in a specific file range */
1714 CL_FSYNC_DISCARD = 2,
1715 /** start writeback and make sure they have reached storage before
1716 * return. OST_SYNC RPC must be issued and finished */
1720 struct cl_io_rw_common {
1725 enum cl_setattr_subtype {
1726 /** regular setattr **/
1730 /** fallocate(2) - mode preallocate **/
1731 CL_SETATTR_FALLOCATE
1734 struct cl_io_range {
1740 struct cl_io_pt *cip_next;
1741 struct kiocb cip_iocb;
1742 struct iov_iter cip_iter;
1743 struct file *cip_file;
1744 enum cl_io_type cip_iot;
1745 unsigned int cip_need_restart:1;
1754 * cl_io is shared by all threads participating in this IO (in current
1755 * implementation only one thread advances IO, but parallel IO design and
1756 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1757 * is up to these threads to serialize their activities, including updates to
1758 * mutable cl_io fields.
1761 /** type of this IO. Immutable after creation. */
1762 enum cl_io_type ci_type;
1763 /** current state of cl_io state machine. */
1764 enum cl_io_state ci_state;
1765 /** main object this io is against. Immutable after creation. */
1766 struct cl_object *ci_obj;
1767 /** top level dio_aio */
1768 struct cl_dio_aio *ci_dio_aio;
1770 * Upper layer io, of which this io is a part of. Immutable after
1773 struct cl_io *ci_parent;
1774 /** List of slices. Immutable after creation. */
1775 struct list_head ci_layers;
1776 /** list of locks (to be) acquired by this io. */
1777 struct cl_lockset ci_lockset;
1778 /** lock requirements, this is just a help info for sublayers. */
1779 enum cl_io_lock_dmd ci_lockreq;
1780 /** layout version when this IO occurs */
1781 __u32 ci_layout_version;
1784 struct cl_io_rw_common rd;
1787 struct cl_io_rw_common wr;
1791 struct cl_io_rw_common ci_rw;
1792 struct cl_setattr_io {
1793 struct ost_lvb sa_attr;
1794 unsigned int sa_attr_flags;
1795 unsigned int sa_avalid; /* ATTR_* */
1796 unsigned int sa_xvalid; /* OP_XVALID */
1797 int sa_stripe_index;
1798 struct ost_layout sa_layout;
1799 const struct lu_fid *sa_parent_fid;
1800 /* SETATTR interface is used for regular setattr, */
1801 /* truncate(2) and fallocate(2) subtypes */
1802 enum cl_setattr_subtype sa_subtype;
1803 /* The following are used for fallocate(2) */
1805 loff_t sa_falloc_offset;
1806 loff_t sa_falloc_end;
1807 uid_t sa_falloc_uid;
1808 gid_t sa_falloc_gid;
1809 __u32 sa_falloc_projid;
1811 struct cl_data_version_io {
1812 u64 dv_data_version;
1813 u32 dv_layout_version;
1816 struct cl_fault_io {
1817 /** page index within file. */
1819 /** bytes valid byte on a faulted page. */
1821 /** writable page? for nopage() only */
1823 /** page of an executable? */
1825 /** page_mkwrite() */
1827 /** resulting page */
1828 struct cl_page *ft_page;
1830 struct cl_fsync_io {
1833 /** file system level fid */
1834 struct lu_fid *fi_fid;
1835 enum cl_fsync_mode fi_mode;
1836 /* how many pages were written/discarded */
1837 unsigned int fi_nr_written;
1839 struct cl_ladvise_io {
1842 /** file system level fid */
1843 struct lu_fid *li_fid;
1844 enum lu_ladvise_type li_advice;
1847 struct cl_lseek_io {
1853 time64_t lm_next_rpc_time;
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 /* Tell sublayers not to expand LDLM locks requested for this IO */
1899 ci_lock_no_expand:1,
1901 * Set if non-delay RPC should be used for this IO.
1903 * If this file has multiple mirrors, and if the OSTs of the current
1904 * mirror is inaccessible, non-delay RPC would error out quickly so
1905 * that the upper layer can try to access the next mirror.
1909 * Set if IO is triggered by async workqueue readahead.
1911 ci_async_readahead:1,
1913 * Ignore lockless and do normal locking for this io.
1917 * Set if we've tried all mirrors for this read IO, if it's not set,
1918 * the read IO will check to-be-read OSCs' status, and make fast-switch
1919 * another mirror if some of the OSTs are not healthy.
1921 ci_tried_all_mirrors:1,
1923 * Random read hints, readahead will be disabled.
1927 * Sequential read hints.
1931 * Do parallel (async) submission of DIO RPCs. Note DIO is still sync
1932 * to userspace, only the RPCs are submitted async, then waited for at
1933 * the llite layer before returning.
1937 * Bypass quota check
1939 unsigned ci_noquota:1,
1941 * io_uring direct IO with flags IOCB_NOWAIT.
1945 * The filesystem must exclusively acquire invalidate_lock before
1946 * invalidating page cache in truncate / hole punch / DLM extent
1947 * lock blocking AST path (and thus calling into ->invalidatepage)
1948 * to block races between page cache invalidation and page cache
1949 * filling functions (fault, read, ...)
1951 ci_invalidate_page_cache:1;
1954 * How many times the read has retried before this one.
1955 * Set by the top level and consumed by the LOV.
1957 unsigned ci_ndelay_tried;
1959 * Designated mirror index for this I/O.
1961 unsigned ci_designated_mirror;
1963 * Number of pages owned by this IO. For invariant checking.
1965 unsigned ci_owned_nr;
1967 * Range of write intent. Valid if ci_need_write_intent is set.
1969 struct lu_extent ci_write_intent;
1975 * Per-transfer attributes.
1977 struct cl_req_attr {
1978 enum cl_req_type cra_type;
1980 struct cl_page *cra_page;
1981 /** Generic attributes for the server consumption. */
1982 struct obdo *cra_oa;
1984 char cra_jobid[LUSTRE_JOBID_SIZE];
1985 /** uid/gid of the process doing an io */
1990 enum cache_stats_item {
1991 /** how many cache lookups were performed */
1993 /** how many times cache lookup resulted in a hit */
1995 /** how many entities are in the cache right now */
1997 /** how many entities in the cache are actively used (and cannot be
1998 * evicted) right now */
2000 /** how many entities were created at all */
2005 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2008 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2010 struct cache_stats {
2011 const char *cs_name;
2012 atomic_t cs_stats[CS_NR];
2015 /** These are not exported so far */
2016 void cache_stats_init (struct cache_stats *cs, const char *name);
2019 * Client-side site. This represents particular client stack. "Global"
2020 * variables should (directly or indirectly) be added here to allow multiple
2021 * clients to co-exist in the single address space.
2024 struct lu_site cs_lu;
2026 * Statistical counters. Atomics do not scale, something better like
2027 * per-cpu counters is needed.
2029 * These are exported as /proc/fs/lustre/llite/.../site
2031 * When interpreting keep in mind that both sub-locks (and sub-pages)
2032 * and top-locks (and top-pages) are accounted here.
2034 struct cache_stats cs_pages;
2035 atomic_t cs_pages_state[CPS_NR];
2038 int cl_site_init(struct cl_site *s, struct cl_device *top);
2039 void cl_site_fini(struct cl_site *s);
2040 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2043 * Output client site statistical counters into a buffer. Suitable for
2044 * ll_rd_*()-style functions.
2046 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2051 * Type conversion and accessory functions.
2055 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2057 return container_of(site, struct cl_site, cs_lu);
2060 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2062 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2063 return container_of_safe(d, struct cl_device, cd_lu_dev);
2066 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2068 return &d->cd_lu_dev;
2071 static inline struct cl_object *lu2cl(const struct lu_object *o)
2073 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2074 return container_of_safe(o, struct cl_object, co_lu);
2077 static inline const struct cl_object_conf *
2078 lu2cl_conf(const struct lu_object_conf *conf)
2080 return container_of_safe(conf, struct cl_object_conf, coc_lu);
2083 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2085 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2088 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2090 return container_of_safe(h, struct cl_object_header, coh_lu);
2093 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2095 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2099 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2101 return luh2coh(obj->co_lu.lo_header);
2104 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2106 return lu_device_init(&d->cd_lu_dev, t);
2109 static inline void cl_device_fini(struct cl_device *d)
2111 lu_device_fini(&d->cd_lu_dev);
2114 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2115 struct cl_object *obj,
2116 const struct cl_page_operations *ops);
2117 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2118 struct cl_object *obj,
2119 const struct cl_lock_operations *ops);
2120 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2121 struct cl_object *obj, const struct cl_io_operations *ops);
2124 /** \defgroup cl_object cl_object
2126 struct cl_object *cl_object_top (struct cl_object *o);
2127 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2128 const struct lu_fid *fid,
2129 const struct cl_object_conf *c);
2131 int cl_object_header_init(struct cl_object_header *h);
2132 void cl_object_header_fini(struct cl_object_header *h);
2133 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2134 void cl_object_get (struct cl_object *o);
2135 void cl_object_attr_lock (struct cl_object *o);
2136 void cl_object_attr_unlock(struct cl_object *o);
2137 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2138 struct cl_attr *attr);
2139 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2140 const struct cl_attr *attr, unsigned valid);
2141 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2142 struct ost_lvb *lvb);
2143 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2144 const struct cl_object_conf *conf);
2145 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2146 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2147 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2148 struct lov_user_md __user *lum, size_t size);
2149 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2150 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2152 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2153 struct cl_layout *cl);
2154 loff_t cl_object_maxbytes(struct cl_object *obj);
2155 int cl_object_flush(const struct lu_env *env, struct cl_object *obj,
2156 struct ldlm_lock *lock);
2157 int cl_object_inode_ops(const struct lu_env *env, struct cl_object *obj,
2158 enum coo_inode_opc opc, void *data);
2162 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2164 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2166 return cl_object_header(o0) == cl_object_header(o1);
2169 static inline void cl_object_page_init(struct cl_object *clob, int size)
2171 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2172 cl_object_header(clob)->coh_page_bufsize += round_up(size, 8);
2173 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2176 static inline void *cl_object_page_slice(struct cl_object *clob,
2177 struct cl_page *page)
2179 return (void *)((char *)page + clob->co_slice_off);
2183 * Return refcount of cl_object.
2185 static inline int cl_object_refc(struct cl_object *clob)
2187 struct lu_object_header *header = clob->co_lu.lo_header;
2188 return atomic_read(&header->loh_ref);
2193 /** \defgroup cl_page cl_page
2195 struct cl_page *cl_page_find (const struct lu_env *env,
2196 struct cl_object *obj,
2197 pgoff_t idx, struct page *vmpage,
2198 enum cl_page_type type);
2199 struct cl_page *cl_page_alloc (const struct lu_env *env,
2200 struct cl_object *o, pgoff_t ind,
2201 struct page *vmpage,
2202 enum cl_page_type type);
2203 void cl_page_get (struct cl_page *page);
2204 void cl_page_put (const struct lu_env *env,
2205 struct cl_page *page);
2206 void cl_pagevec_put (const struct lu_env *env,
2207 struct cl_page *page,
2208 struct pagevec *pvec);
2209 void cl_page_print (const struct lu_env *env, void *cookie,
2210 lu_printer_t printer,
2211 const struct cl_page *pg);
2212 void cl_page_header_print(const struct lu_env *env, void *cookie,
2213 lu_printer_t printer,
2214 const struct cl_page *pg);
2215 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2220 * Functions dealing with the ownership of page by io.
2224 int cl_page_own (const struct lu_env *env,
2225 struct cl_io *io, struct cl_page *page);
2226 int cl_page_own_try (const struct lu_env *env,
2227 struct cl_io *io, struct cl_page *page);
2228 void cl_page_assume (const struct lu_env *env,
2229 struct cl_io *io, struct cl_page *page);
2230 void cl_page_unassume (const struct lu_env *env,
2231 struct cl_io *io, struct cl_page *pg);
2232 void cl_page_disown (const struct lu_env *env,
2233 struct cl_io *io, struct cl_page *page);
2234 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2241 * Functions dealing with the preparation of a page for a transfer, and
2242 * tracking transfer state.
2245 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2246 struct cl_page *pg, enum cl_req_type crt);
2247 void cl_page_completion (const struct lu_env *env,
2248 struct cl_page *pg, enum cl_req_type crt, int ioret);
2249 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2250 enum cl_req_type crt);
2251 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2252 struct cl_page *pg, enum cl_req_type crt);
2253 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2255 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2256 struct cl_page *pg);
2262 * \name helper routines
2263 * Functions to discard, delete and export a cl_page.
2266 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2267 struct cl_page *pg);
2268 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2269 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
2272 void cl_lock_print(const struct lu_env *env, void *cookie,
2273 lu_printer_t printer, const struct cl_lock *lock);
2274 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2275 lu_printer_t printer,
2276 const struct cl_lock_descr *descr);
2280 * Data structure managing a client's cached pages. A count of
2281 * "unstable" pages is maintained, and an LRU of clean pages is
2282 * maintained. "unstable" pages are pages pinned by the ptlrpc
2283 * layer for recovery purposes.
2285 struct cl_client_cache {
2287 * # of client cache refcount
2288 * # of users (OSCs) + 2 (held by llite and lov)
2290 refcount_t ccc_users;
2292 * # of threads are doing shrinking
2294 unsigned int ccc_lru_shrinkers;
2296 * # of LRU entries available
2298 atomic_long_t ccc_lru_left;
2300 * List of entities(OSCs) for this LRU cache
2302 struct list_head ccc_lru;
2304 * Max # of LRU entries
2306 unsigned long ccc_lru_max;
2308 * Lock to protect ccc_lru list
2310 spinlock_t ccc_lru_lock;
2312 * Set if unstable check is enabled
2314 unsigned int ccc_unstable_check:1;
2316 * # of unstable pages for this mount point
2318 atomic_long_t ccc_unstable_nr;
2320 * Waitq for awaiting unstable pages to reach zero.
2321 * Used at umounting time and signaled on BRW commit
2323 wait_queue_head_t ccc_unstable_waitq;
2325 * Serialize max_cache_mb write operation
2327 struct mutex ccc_max_cache_mb_lock;
2330 * cl_cache functions
2332 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2333 void cl_cache_incref(struct cl_client_cache *cache);
2334 void cl_cache_decref(struct cl_client_cache *cache);
2338 /** \defgroup cl_lock cl_lock
2340 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2341 struct cl_lock *lock);
2342 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2343 const struct cl_io *io);
2344 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2345 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2346 const struct lu_device_type *dtype);
2347 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2349 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2350 struct cl_lock *lock, struct cl_sync_io *anchor);
2351 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2355 /** \defgroup cl_io cl_io
2358 int cl_io_init(const struct lu_env *env, struct cl_io *io,
2359 enum cl_io_type iot, struct cl_object *obj);
2360 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
2361 enum cl_io_type iot, struct cl_object *obj);
2362 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
2363 enum cl_io_type iot, loff_t pos, size_t bytes);
2364 int cl_io_loop(const struct lu_env *env, struct cl_io *io);
2366 void cl_io_fini(const struct lu_env *env, struct cl_io *io);
2367 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io);
2368 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io);
2369 int cl_io_lock(const struct lu_env *env, struct cl_io *io);
2370 void cl_io_unlock(const struct lu_env *env, struct cl_io *io);
2371 int cl_io_start(const struct lu_env *env, struct cl_io *io);
2372 void cl_io_end(const struct lu_env *env, struct cl_io *io);
2373 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
2374 struct cl_io_lock_link *link);
2375 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2376 struct cl_lock_descr *descr);
2377 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2378 enum cl_req_type iot, struct cl_2queue *queue);
2379 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2380 enum cl_req_type iot, struct cl_2queue *queue,
2382 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2383 struct cl_page_list *queue, int from, int to,
2385 void cl_io_extent_release (const struct lu_env *env, struct cl_io *io);
2386 int cl_io_lru_reserve(const struct lu_env *env, struct cl_io *io,
2387 loff_t pos, size_t bytes);
2388 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2389 pgoff_t start, struct cl_read_ahead *ra);
2390 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io,
2394 * True, iff \a io is an O_APPEND write(2).
2396 static inline int cl_io_is_append(const struct cl_io *io)
2398 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2401 static inline int cl_io_is_sync_write(const struct cl_io *io)
2403 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2406 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2408 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2412 * True, iff \a io is a truncate(2).
2414 static inline int cl_io_is_trunc(const struct cl_io *io)
2416 return io->ci_type == CIT_SETATTR &&
2417 (io->u.ci_setattr.sa_avalid & ATTR_SIZE) &&
2418 (io->u.ci_setattr.sa_subtype != CL_SETATTR_FALLOCATE);
2421 static inline int cl_io_is_fallocate(const struct cl_io *io)
2423 return (io->ci_type == CIT_SETATTR) &&
2424 (io->u.ci_setattr.sa_subtype == CL_SETATTR_FALLOCATE);
2427 struct cl_io *cl_io_top(struct cl_io *io);
2429 #define CL_IO_SLICE_CLEAN(obj, base) memset_startat(obj, 0, base)
2433 /** \defgroup cl_page_list cl_page_list
2437 * Last page in the page list.
2439 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2441 LASSERT(plist->pl_nr > 0);
2442 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2445 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2447 LASSERT(plist->pl_nr > 0);
2448 return list_first_entry(&plist->pl_pages, struct cl_page, cp_batch);
2452 * Iterate over pages in a page list.
2454 #define cl_page_list_for_each(page, list) \
2455 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2458 * Iterate over pages in a page list, taking possible removals into account.
2460 #define cl_page_list_for_each_safe(page, temp, list) \
2461 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2463 void cl_page_list_init(struct cl_page_list *plist);
2464 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page,
2466 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2467 struct cl_page *page);
2468 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2469 struct cl_page *page);
2470 void cl_page_list_splice(struct cl_page_list *list,
2471 struct cl_page_list *head);
2472 void cl_page_list_del(const struct lu_env *env,
2473 struct cl_page_list *plist, struct cl_page *page);
2474 void cl_page_list_disown(const struct lu_env *env,
2475 struct cl_page_list *plist);
2476 void cl_page_list_assume(const struct lu_env *env,
2477 struct cl_io *io, struct cl_page_list *plist);
2478 void cl_page_list_discard(const struct lu_env *env,
2479 struct cl_io *io, struct cl_page_list *plist);
2480 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2482 void cl_2queue_init(struct cl_2queue *queue);
2483 void cl_2queue_disown(const struct lu_env *env, struct cl_2queue *queue);
2484 void cl_2queue_assume(const struct lu_env *env, struct cl_io *io,
2485 struct cl_2queue *queue);
2486 void cl_2queue_discard(const struct lu_env *env, struct cl_io *io,
2487 struct cl_2queue *queue);
2488 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2489 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2491 /** @} cl_page_list */
2493 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2494 struct cl_req_attr *attr);
2496 /** \defgroup cl_sync_io cl_sync_io
2503 typedef void (cl_sync_io_end_t)(const struct lu_env *, struct cl_sync_io *);
2505 void cl_sync_io_init_notify(struct cl_sync_io *anchor, int nr, void *dio_aio,
2506 cl_sync_io_end_t *end);
2508 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2510 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2512 int cl_sync_io_wait_recycle(const struct lu_env *env, struct cl_sync_io *anchor,
2513 long timeout, int ioret);
2514 struct cl_dio_aio *cl_dio_aio_alloc(struct kiocb *iocb, struct cl_object *obj,
2516 struct cl_sub_dio *cl_sub_dio_alloc(struct cl_dio_aio *ll_aio,
2517 struct iov_iter *iter, bool write,
2519 void cl_dio_aio_free(const struct lu_env *env, struct cl_dio_aio *aio);
2520 void cl_sub_dio_free(struct cl_sub_dio *sdio);
2521 static inline void cl_sync_io_init(struct cl_sync_io *anchor, int nr)
2523 cl_sync_io_init_notify(anchor, nr, NULL, NULL);
2527 * Anchor for synchronous transfer. This is allocated on a stack by thread
2528 * doing synchronous transfer, and a pointer to this structure is set up in
2529 * every page submitted for transfer. Transfer completion routine updates
2530 * anchor and wakes up waiting thread when transfer is complete.
2533 /** number of pages yet to be transferred. */
2534 atomic_t csi_sync_nr;
2535 /** has this i/o completed? */
2536 atomic_t csi_complete;
2539 /** completion to be signaled when transfer is complete. */
2540 wait_queue_head_t csi_waitq;
2541 /** callback to invoke when this IO is finished */
2542 cl_sync_io_end_t *csi_end_io;
2543 /* private pointer for an associated DIO/AIO */
2547 /** direct IO pages */
2548 struct ll_dio_pages {
2550 * page array to be written. we don't support
2551 * partial pages except the last one.
2553 struct page **ldp_pages;
2554 /** # of pages in the array. */
2556 /* the file offset of the first page. */
2557 loff_t ldp_file_offset;
2560 /* Top level struct used for AIO and DIO */
2562 struct cl_sync_io cda_sync;
2563 struct cl_object *cda_obj;
2564 struct kiocb *cda_iocb;
2566 struct mm_struct *cda_mm;
2567 unsigned cda_no_aio_complete:1,
2571 /* Sub-dio used for splitting DIO (and AIO, because AIO is DIO) according to
2572 * the layout/striping, so we can do parallel submit of DIO RPCs
2575 struct cl_sync_io csd_sync;
2576 struct cl_page_list csd_pages;
2578 struct cl_dio_aio *csd_ll_aio;
2579 struct ll_dio_pages csd_dio_pages;
2580 struct iov_iter csd_iter;
2581 unsigned csd_creator_free:1,
2584 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW) || \
2585 defined(HAVE_DIRECTIO_2ARGS)
2586 #define HAVE_DIO_ITER 1
2589 void ll_release_user_pages(struct page **pages, int npages);
2591 #ifndef HAVE_KTHREAD_USE_MM
2592 #define kthread_use_mm(mm) use_mm(mm)
2593 #define kthread_unuse_mm(mm) unuse_mm(mm)
2596 /** @} cl_sync_io */
2598 /** \defgroup cl_env cl_env
2600 * lu_env handling for a client.
2602 * lu_env is an environment within which lustre code executes. Its major part
2603 * is lu_context---a fast memory allocation mechanism that is used to conserve
2604 * precious kernel stack space. Originally lu_env was designed for a server,
2607 * - there is a (mostly) fixed number of threads, and
2609 * - call chains have no non-lustre portions inserted between lustre code.
2611 * On a client both these assumtpion fails, because every user thread can
2612 * potentially execute lustre code as part of a system call, and lustre calls
2613 * into VFS or MM that call back into lustre.
2615 * To deal with that, cl_env wrapper functions implement the following
2618 * - allocation and destruction of environment is amortized by caching no
2619 * longer used environments instead of destroying them;
2621 * \see lu_env, lu_context, lu_context_key
2624 struct lu_env *cl_env_get(__u16 *refcheck);
2625 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2626 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2627 unsigned cl_env_cache_purge(unsigned nr);
2628 struct lu_env *cl_env_percpu_get(void);
2629 void cl_env_percpu_put(struct lu_env *env);
2636 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2637 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2639 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2640 struct lu_device_type *ldt,
2641 struct lu_device *next);
2644 int cl_global_init(void);
2645 void cl_global_fini(void);
2647 #endif /* _LINUX_CL_OBJECT_H */