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;
268 * Operation to handle layout, OBJECT_CONF_XYZ.
274 /** configure layout, set up a new stripe, must be called while
275 * holding layout lock. */
277 /** invalidate the current stripe configuration due to losing
279 OBJECT_CONF_INVALIDATE = 1,
280 /** wait for old layout to go away so that new layout can be
286 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
287 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
291 /** the buffer to return the layout in lov_mds_md format. */
292 struct lu_buf cl_buf;
293 /** size of layout in lov_mds_md format. */
295 /** Layout generation. */
297 /** whether layout is a composite one */
298 bool cl_is_composite;
299 /** Whether layout is a HSM released one */
304 * Operations implemented for each cl object layer.
306 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
308 struct cl_object_operations {
310 * Initialize page slice for this layer. Called top-to-bottom through
311 * every object layer when a new cl_page is instantiated. Layer
312 * keeping private per-page data, or requiring its own page operations
313 * vector should allocate these data here, and attach then to the page
314 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
317 * \retval NULL success.
319 * \retval ERR_PTR(errno) failure code.
321 * \retval valid-pointer pointer to already existing referenced page
322 * to be used instead of newly created.
324 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
325 struct cl_page *page, pgoff_t index);
327 * Initialize lock slice for this layer. Called top-to-bottom through
328 * every object layer when a new cl_lock is instantiated. Layer
329 * keeping private per-lock data, or requiring its own lock operations
330 * vector should allocate these data here, and attach then to the lock
331 * by calling cl_lock_slice_add(). Mandatory.
333 int (*coo_lock_init)(const struct lu_env *env,
334 struct cl_object *obj, struct cl_lock *lock,
335 const struct cl_io *io);
337 * Initialize io state for a given layer.
339 * called top-to-bottom once per io existence to initialize io
340 * state. If layer wants to keep some state for this type of io, it
341 * has to embed struct cl_io_slice in lu_env::le_ses, and register
342 * slice with cl_io_slice_add(). It is guaranteed that all threads
343 * participating in this io share the same session.
345 int (*coo_io_init)(const struct lu_env *env,
346 struct cl_object *obj, struct cl_io *io);
348 * Fill portion of \a attr that this layer controls. This method is
349 * called top-to-bottom through all object layers.
351 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
353 * \return 0: to continue
354 * \return +ve: to stop iterating through layers (but 0 is returned
355 * from enclosing cl_object_attr_get())
356 * \return -ve: to signal error
358 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
359 struct cl_attr *attr);
363 * \a valid is a bitmask composed from enum #cl_attr_valid, and
364 * indicating what attributes are to be set.
366 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
368 * \return the same convention as for
369 * cl_object_operations::coo_attr_get() is used.
371 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
372 const struct cl_attr *attr, unsigned valid);
374 * Update object configuration. Called top-to-bottom to modify object
377 * XXX error conditions and handling.
379 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
380 const struct cl_object_conf *conf);
382 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
383 * object. Layers are supposed to fill parts of \a lvb that will be
384 * shipped to the glimpse originator as a glimpse result.
386 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
387 * \see osc_object_glimpse()
389 int (*coo_glimpse)(const struct lu_env *env,
390 const struct cl_object *obj, struct ost_lvb *lvb);
392 * Object prune method. Called when the layout is going to change on
393 * this object, therefore each layer has to clean up their cache,
394 * mainly pages and locks.
396 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
398 * Object getstripe method.
400 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
401 struct lov_user_md __user *lum, size_t size);
403 * Get FIEMAP mapping from the object.
405 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
406 struct ll_fiemap_info_key *fmkey,
407 struct fiemap *fiemap, size_t *buflen);
409 * Get layout and generation of the object.
411 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
412 struct cl_layout *layout);
414 * Get maximum size of the object.
416 loff_t (*coo_maxbytes)(struct cl_object *obj);
418 * Set request attributes.
420 void (*coo_req_attr_set)(const struct lu_env *env,
421 struct cl_object *obj,
422 struct cl_req_attr *attr);
424 * Flush \a obj data corresponding to \a lock. Used for DoM
425 * locks in llite's cancelling blocking ast callback.
427 int (*coo_object_flush)(const struct lu_env *env,
428 struct cl_object *obj,
429 struct ldlm_lock *lock);
433 * Extended header for client object.
435 struct cl_object_header {
436 /** Standard lu_object_header. cl_object::co_lu::lo_header points
438 struct lu_object_header coh_lu;
441 * Parent object. It is assumed that an object has a well-defined
442 * parent, but not a well-defined child (there may be multiple
443 * sub-objects, for the same top-object). cl_object_header::coh_parent
444 * field allows certain code to be written generically, without
445 * limiting possible cl_object layouts unduly.
447 struct cl_object_header *coh_parent;
449 * Protects consistency between cl_attr of parent object and
450 * attributes of sub-objects, that the former is calculated ("merged")
453 * \todo XXX this can be read/write lock if needed.
455 spinlock_t coh_attr_guard;
457 * Size of cl_page + page slices
459 unsigned short coh_page_bufsize;
461 * Number of objects above this one: 0 for a top-object, 1 for its
464 unsigned char coh_nesting;
468 * Helper macro: iterate over all layers of the object \a obj, assigning every
469 * layer top-to-bottom to \a slice.
471 #define cl_object_for_each(slice, obj) \
472 list_for_each_entry((slice), \
473 &(obj)->co_lu.lo_header->loh_layers,\
477 * Helper macro: iterate over all layers of the object \a obj, assigning every
478 * layer bottom-to-top to \a slice.
480 #define cl_object_for_each_reverse(slice, obj) \
481 list_for_each_entry_reverse((slice), \
482 &(obj)->co_lu.lo_header->loh_layers,\
487 #define CL_PAGE_EOF ((pgoff_t)~0ull)
489 /** \addtogroup cl_page cl_page
493 * Layered client page.
495 * cl_page: represents a portion of a file, cached in the memory. All pages
496 * of the given file are of the same size, and are kept in the radix tree
497 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
498 * of the top-level file object are first class cl_objects, they have their
499 * own radix trees of pages and hence page is implemented as a sequence of
500 * struct cl_pages's, linked into double-linked list through
501 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
502 * corresponding radix tree at the corresponding logical offset.
504 * cl_page is associated with VM page of the hosting environment (struct
505 * page in Linux kernel, for example), struct page. It is assumed, that this
506 * association is implemented by one of cl_page layers (top layer in the
507 * current design) that
509 * - intercepts per-VM-page call-backs made by the environment (e.g.,
512 * - translates state (page flag bits) and locking between lustre and
515 * The association between cl_page and struct page is immutable and
516 * established when cl_page is created.
518 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
519 * this io an exclusive access to this page w.r.t. other io attempts and
520 * various events changing page state (such as transfer completion, or
521 * eviction of the page from the memory). Note, that in general cl_io
522 * cannot be identified with a particular thread, and page ownership is not
523 * exactly equal to the current thread holding a lock on the page. Layer
524 * implementing association between cl_page and struct page has to implement
525 * ownership on top of available synchronization mechanisms.
527 * While lustre client maintains the notion of an page ownership by io,
528 * hosting MM/VM usually has its own page concurrency control
529 * mechanisms. For example, in Linux, page access is synchronized by the
530 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
531 * takes care to acquire and release such locks as necessary around the
532 * calls to the file system methods (->readpage(), ->prepare_write(),
533 * ->commit_write(), etc.). This leads to the situation when there are two
534 * different ways to own a page in the client:
536 * - client code explicitly and voluntary owns the page (cl_page_own());
538 * - VM locks a page and then calls the client, that has "to assume"
539 * the ownership from the VM (cl_page_assume()).
541 * Dual methods to release ownership are cl_page_disown() and
542 * cl_page_unassume().
544 * cl_page is reference counted (cl_page::cp_ref). When reference counter
545 * drops to 0, the page is returned to the cache, unless it is in
546 * cl_page_state::CPS_FREEING state, in which case it is immediately
549 * The general logic guaranteeing the absence of "existential races" for
550 * pages is the following:
552 * - there are fixed known ways for a thread to obtain a new reference
555 * - by doing a lookup in the cl_object radix tree, protected by the
558 * - by starting from VM-locked struct page and following some
559 * hosting environment method (e.g., following ->private pointer in
560 * the case of Linux kernel), see cl_vmpage_page();
562 * - when the page enters cl_page_state::CPS_FREEING state, all these
563 * ways are severed with the proper synchronization
564 * (cl_page_delete());
566 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
569 * - no new references to the page in cl_page_state::CPS_FREEING state
570 * are allowed (checked in cl_page_get()).
572 * Together this guarantees that when last reference to a
573 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
574 * page, as neither references to it can be acquired at that point, nor
577 * cl_page is a state machine. States are enumerated in enum
578 * cl_page_state. Possible state transitions are enumerated in
579 * cl_page_state_set(). State transition process (i.e., actual changing of
580 * cl_page::cp_state field) is protected by the lock on the underlying VM
583 * Linux Kernel implementation.
585 * Binding between cl_page and struct page (which is a typedef for
586 * struct page) is implemented in the vvp layer. cl_page is attached to the
587 * ->private pointer of the struct page, together with the setting of
588 * PG_private bit in page->flags, and acquiring additional reference on the
589 * struct page (much like struct buffer_head, or any similar file system
590 * private data structures).
592 * PG_locked lock is used to implement both ownership and transfer
593 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
594 * states. No additional references are acquired for the duration of the
597 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
598 * write-out is "protected" by the special PG_writeback bit.
602 * States of cl_page. cl_page.c assumes particular order here.
604 * The page state machine is rather crude, as it doesn't recognize finer page
605 * states like "dirty" or "up to date". This is because such states are not
606 * always well defined for the whole stack (see, for example, the
607 * implementation of the read-ahead, that hides page up-to-dateness to track
608 * cache hits accurately). Such sub-states are maintained by the layers that
609 * are interested in them.
613 * Page is in the cache, un-owned. Page leaves cached state in the
616 * - [cl_page_state::CPS_OWNED] io comes across the page and
619 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
620 * req-formation engine decides that it wants to include this page
621 * into an RPC being constructed, and yanks it from the cache;
623 * - [cl_page_state::CPS_FREEING] VM callback is executed to
624 * evict the page form the memory;
626 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
630 * Page is exclusively owned by some cl_io. Page may end up in this
631 * state as a result of
633 * - io creating new page and immediately owning it;
635 * - [cl_page_state::CPS_CACHED] io finding existing cached page
638 * - [cl_page_state::CPS_OWNED] io finding existing owned page
639 * and waiting for owner to release the page;
641 * Page leaves owned state in the following cases:
643 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
644 * the cache, doing nothing;
646 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
649 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
650 * transfer for this page;
652 * - [cl_page_state::CPS_FREEING] io decides to destroy this
653 * page (e.g., as part of truncate or extent lock cancellation).
655 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
659 * Page is being written out, as a part of a transfer. This state is
660 * entered when req-formation logic decided that it wants this page to
661 * be sent through the wire _now_. Specifically, it means that once
662 * this state is achieved, transfer completion handler (with either
663 * success or failure indication) is guaranteed to be executed against
664 * this page independently of any locks and any scheduling decisions
665 * made by the hosting environment (that effectively means that the
666 * page is never put into cl_page_state::CPS_PAGEOUT state "in
667 * advance". This property is mentioned, because it is important when
668 * reasoning about possible dead-locks in the system). The page can
669 * enter this state as a result of
671 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
672 * write-out of this page, or
674 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
675 * that it has enough dirty pages cached to issue a "good"
678 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
679 * is completed---it is moved into cl_page_state::CPS_CACHED state.
681 * Underlying VM page is locked for the duration of transfer.
683 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
687 * Page is being read in, as a part of a transfer. This is quite
688 * similar to the cl_page_state::CPS_PAGEOUT state, except that
689 * read-in is always "immediate"---there is no such thing a sudden
690 * construction of read request from cached, presumably not up to date,
693 * Underlying VM page is locked for the duration of transfer.
695 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
699 * Page is being destroyed. This state is entered when client decides
700 * that page has to be deleted from its host object, as, e.g., a part
703 * Once this state is reached, there is no way to escape it.
705 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
712 /** Host page, the page is from the host inode which the cl_page
716 /** Transient page, the transient cl_page is used to bind a cl_page
717 * to vmpage which is not belonging to the same object of cl_page.
718 * it is used in DirectIO and lockless IO. */
723 #define CP_STATE_BITS 4
724 #define CP_TYPE_BITS 2
725 #define CP_MAX_LAYER 2
728 * Fields are protected by the lock on struct page, except for atomics and
731 * \invariant Data type invariants are in cl_page_invariant(). Basically:
732 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
733 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
734 * cl_page::cp_owner (when set).
737 /** Reference counter. */
739 /** layout_entry + stripe index, composed using lov_comp_index() */
740 unsigned int cp_lov_index;
741 /** page->index of the page within the whole file */
742 pgoff_t cp_page_index;
743 /** An object this page is a part of. Immutable after creation. */
744 struct cl_object *cp_obj;
746 struct page *cp_vmpage;
748 * Assigned if doing direct IO, because in this case cp_vmpage is not
749 * a valid page cache page, hence the inode cannot be inferred from
750 * cp_vmpage->mapping->host.
752 struct inode *cp_inode;
753 /** Linkage of pages within group. Pages must be owned */
754 struct list_head cp_batch;
755 /** array of slices offset. Immutable after creation. */
756 unsigned char cp_layer_offset[CP_MAX_LAYER];
757 /** current slice index */
758 unsigned char cp_layer_count:2;
760 * Page state. This field is const to avoid accidental update, it is
761 * modified only internally within cl_page.c. Protected by a VM lock.
763 enum cl_page_state cp_state:CP_STATE_BITS;
765 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
768 enum cl_page_type cp_type:CP_TYPE_BITS;
769 unsigned cp_defer_uptodate:1,
772 /* which slab kmem index this memory allocated from */
773 short int cp_kmem_index;
776 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
777 * by sub-io. Protected by a VM lock.
779 struct cl_io *cp_owner;
780 /** List of references to this page, for debugging. */
781 struct lu_ref cp_reference;
782 /** Link to an object, for debugging. */
783 struct lu_ref_link cp_obj_ref;
784 /** Link to a queue, for debugging. */
785 struct lu_ref_link cp_queue_ref;
786 /** Assigned if doing a sync_io */
787 struct cl_sync_io *cp_sync_io;
791 * Per-layer part of cl_page.
793 * \see vvp_page, lov_page, osc_page
795 struct cl_page_slice {
796 struct cl_page *cpl_page;
797 const struct cl_page_operations *cpl_ops;
801 * Lock mode. For the client extent locks.
813 * Requested transfer type.
822 * Per-layer page operations.
824 * Methods taking an \a io argument are for the activity happening in the
825 * context of given \a io. Page is assumed to be owned by that io, except for
828 * \see vvp_page_ops, lov_page_ops, osc_page_ops
830 struct cl_page_operations {
832 * cl_page<->struct page methods. Only one layer in the stack has to
833 * implement these. Current code assumes that this functionality is
834 * provided by the topmost layer, see __cl_page_disown() as an example.
838 * Update file attributes when all we have is this page. Used for tiny
839 * writes to update attributes when we don't have a full cl_io.
841 void (*cpo_page_touch)(const struct lu_env *env,
842 const struct cl_page_slice *slice, size_t to);
848 * Called when page is truncated from the object. Optional.
850 * \see cl_page_discard()
851 * \see vvp_page_discard(), osc_page_discard()
853 void (*cpo_discard)(const struct lu_env *env,
854 const struct cl_page_slice *slice,
857 * Called when page is removed from the cache, and is about to being
858 * destroyed. Optional.
860 * \see cl_page_delete()
861 * \see vvp_page_delete(), osc_page_delete()
863 void (*cpo_delete)(const struct lu_env *env,
864 const struct cl_page_slice *slice);
866 * Optional debugging helper. Prints given page slice.
868 * \see cl_page_print()
870 int (*cpo_print)(const struct lu_env *env,
871 const struct cl_page_slice *slice,
872 void *cookie, lu_printer_t p);
881 * Request type dependent vector of operations.
883 * Transfer operations depend on transfer mode (cl_req_type). To avoid
884 * passing transfer mode to each and every of these methods, and to
885 * avoid branching on request type inside of the methods, separate
886 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
887 * provided. That is, method invocation usually looks like
889 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
893 * Completion handler. This is guaranteed to be eventually
894 * fired after cl_page_prep() or cl_page_make_ready() call.
896 * This method can be called in a non-blocking context. It is
897 * guaranteed however, that the page involved and its object
898 * are pinned in memory (and, hence, calling cl_page_put() is
901 * \see cl_page_completion()
903 void (*cpo_completion)(const struct lu_env *env,
904 const struct cl_page_slice *slice,
908 * Tell transfer engine that only [to, from] part of a page should be
911 * This is used for immediate transfers.
913 * \todo XXX this is not very good interface. It would be much better
914 * if all transfer parameters were supplied as arguments to
915 * cl_io_operations::cio_submit() call, but it is not clear how to do
916 * this for page queues.
918 * \see cl_page_clip()
920 void (*cpo_clip)(const struct lu_env *env,
921 const struct cl_page_slice *slice,
924 * Write out a page by kernel. This is only called by ll_writepage
927 * \see cl_page_flush()
929 int (*cpo_flush)(const struct lu_env *env,
930 const struct cl_page_slice *slice,
936 * Helper macro, dumping detailed information about \a page into a log.
938 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
940 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
941 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
942 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
943 CDEBUG(mask, format , ## __VA_ARGS__); \
948 * Helper macro, dumping shorter information about \a page into a log.
950 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
952 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
953 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
954 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
955 CDEBUG(mask, format , ## __VA_ARGS__); \
959 static inline struct page *cl_page_vmpage(const struct cl_page *page)
961 LASSERT(page->cp_vmpage != NULL);
962 return page->cp_vmpage;
965 static inline pgoff_t cl_page_index(const struct cl_page *cp)
967 return cl_page_vmpage(cp)->index;
971 * Check if a cl_page is in use.
973 * Client cache holds a refcount, this refcount will be dropped when
974 * the page is taken out of cache, see vvp_page_delete().
976 static inline bool __page_in_use(const struct cl_page *page, int refc)
978 return (refcount_read(&page->cp_ref) > refc + 1);
982 * Caller itself holds a refcount of cl_page.
984 #define cl_page_in_use(pg) __page_in_use(pg, 1)
986 * Caller doesn't hold a refcount.
988 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
992 /** \addtogroup cl_lock cl_lock
996 * Extent locking on the client.
1000 * The locking model of the new client code is built around
1004 * data-type representing an extent lock on a regular file. cl_lock is a
1005 * layered object (much like cl_object and cl_page), it consists of a header
1006 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1007 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1009 * Typical cl_lock consists of one layer:
1011 * - lov_lock (lov specific data).
1013 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1014 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1018 * Each sub-lock is associated with a cl_object (representing stripe
1019 * sub-object or the file to which top-level cl_lock is associated to), and is
1020 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1021 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1022 * is different from cl_page, that doesn't fan out (there is usually exactly
1023 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1024 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1028 * cl_lock is a cacheless data container for the requirements of locks to
1029 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1032 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1033 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1035 * INTERFACE AND USAGE
1037 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1038 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1039 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1040 * consists of multiple sub cl_locks, each sub locks will be enqueued
1041 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1042 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1045 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1046 * method will be called for each layer to release the resource held by this
1047 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1048 * clo_enqueue time, is released.
1050 * LDLM lock can only be canceled if there is no cl_lock using it.
1052 * Overall process of the locking during IO operation is as following:
1054 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1055 * is called on each layer. Responsibility of this method is to add locks,
1056 * needed by a given layer into cl_io.ci_lockset.
1058 * - once locks for all layers were collected, they are sorted to avoid
1059 * dead-locks (cl_io_locks_sort()), and enqueued.
1061 * - when all locks are acquired, IO is performed;
1063 * - locks are released after IO is complete.
1065 * Striping introduces major additional complexity into locking. The
1066 * fundamental problem is that it is generally unsafe to actively use (hold)
1067 * two locks on the different OST servers at the same time, as this introduces
1068 * inter-server dependency and can lead to cascading evictions.
1070 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1071 * that no multi-stripe locks are taken (note that this design abandons POSIX
1072 * read/write semantics). Such pieces ideally can be executed concurrently. At
1073 * the same time, certain types of IO cannot be sub-divived, without
1074 * sacrificing correctness. This includes:
1076 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1079 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1081 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1082 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1083 * has to be held together with the usual lock on [offset, offset + count].
1085 * Interaction with DLM
1087 * In the expected setup, cl_lock is ultimately backed up by a collection of
1088 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1089 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1090 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1091 * description of interaction with DLM.
1097 struct cl_lock_descr {
1098 /** Object this lock is granted for. */
1099 struct cl_object *cld_obj;
1100 /** Index of the first page protected by this lock. */
1102 /** Index of the last page (inclusive) protected by this lock. */
1104 /** Group ID, for group lock */
1107 enum cl_lock_mode cld_mode;
1109 * flags to enqueue lock. A combination of bit-flags from
1110 * enum cl_enq_flags.
1112 __u32 cld_enq_flags;
1115 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1116 #define PDESCR(descr) \
1117 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1118 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1120 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1123 * Layered client lock.
1126 /** List of slices. Immutable after creation. */
1127 struct list_head cll_layers;
1128 /** lock attribute, extent, cl_object, etc. */
1129 struct cl_lock_descr cll_descr;
1133 * Per-layer part of cl_lock
1135 * \see lov_lock, osc_lock
1137 struct cl_lock_slice {
1138 struct cl_lock *cls_lock;
1139 /** Object slice corresponding to this lock slice. Immutable after
1141 struct cl_object *cls_obj;
1142 const struct cl_lock_operations *cls_ops;
1143 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1144 struct list_head cls_linkage;
1149 * \see lov_lock_ops, osc_lock_ops
1151 struct cl_lock_operations {
1154 * Attempts to enqueue the lock. Called top-to-bottom.
1156 * \retval 0 this layer has enqueued the lock successfully
1157 * \retval >0 this layer has enqueued the lock, but need to wait on
1158 * @anchor for resources
1159 * \retval -ve failure
1161 * \see lov_lock_enqueue(), osc_lock_enqueue()
1163 int (*clo_enqueue)(const struct lu_env *env,
1164 const struct cl_lock_slice *slice,
1165 struct cl_io *io, struct cl_sync_io *anchor);
1167 * Cancel a lock, release its DLM lock ref, while does not cancel the
1170 void (*clo_cancel)(const struct lu_env *env,
1171 const struct cl_lock_slice *slice);
1174 * Destructor. Frees resources and the slice.
1176 * \see lov_lock_fini(), osc_lock_fini()
1178 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1180 * Optional debugging helper. Prints given lock slice.
1182 int (*clo_print)(const struct lu_env *env,
1183 void *cookie, lu_printer_t p,
1184 const struct cl_lock_slice *slice);
1187 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1189 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1190 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1191 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1192 CDEBUG(mask, format , ## __VA_ARGS__); \
1196 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1200 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1206 /** \addtogroup cl_page_list cl_page_list
1207 * Page list used to perform collective operations on a group of pages.
1209 * Pages are added to the list one by one. cl_page_list acquires a reference
1210 * for every page in it. Page list is used to perform collective operations on
1213 * - submit pages for an immediate transfer,
1215 * - own pages on behalf of certain io (waiting for each page in turn),
1219 * When list is finalized, it releases references on all pages it still has.
1221 * \todo XXX concurrency control.
1225 struct cl_page_list {
1227 struct list_head pl_pages;
1231 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1232 * contains an incoming page list and an outgoing page list.
1235 struct cl_page_list c2_qin;
1236 struct cl_page_list c2_qout;
1239 /** @} cl_page_list */
1241 /** \addtogroup cl_io cl_io
1246 * cl_io represents a high level I/O activity like
1247 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1250 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1251 * important distinction. We want to minimize number of calls to the allocator
1252 * in the fast path, e.g., in the case of read(2) when everything is cached:
1253 * client already owns the lock over region being read, and data are cached
1254 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1255 * per-layer io state is stored in the session, associated with the io, see
1256 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1257 * by using free-lists, see cl_env_get().
1259 * There is a small predefined number of possible io types, enumerated in enum
1262 * cl_io is a state machine, that can be advanced concurrently by the multiple
1263 * threads. It is up to these threads to control the concurrency and,
1264 * specifically, to detect when io is done, and its state can be safely
1267 * For read/write io overall execution plan is as following:
1269 * (0) initialize io state through all layers;
1271 * (1) loop: prepare chunk of work to do
1273 * (2) call all layers to collect locks they need to process current chunk
1275 * (3) sort all locks to avoid dead-locks, and acquire them
1277 * (4) process the chunk: call per-page methods
1278 * cl_io_operations::cio_prepare_write(),
1279 * cl_io_operations::cio_commit_write() for write)
1285 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1286 * address allocation efficiency issues mentioned above), and returns with the
1287 * special error condition from per-page method when current sub-io has to
1288 * block. This causes io loop to be repeated, and lov switches to the next
1289 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1294 /** read system call */
1296 /** write system call */
1298 /** truncate, utime system calls */
1300 /** get data version */
1303 * page fault handling
1307 * fsync system call handling
1308 * To write out a range of file
1312 * glimpse. An io context to acquire glimpse lock.
1316 * Miscellaneous io. This is used for occasional io activity that
1317 * doesn't fit into other types. Currently this is used for:
1319 * - cancellation of an extent lock. This io exists as a context
1320 * to write dirty pages from under the lock being canceled back
1323 * - VM induced page write-out. An io context for writing page out
1324 * for memory cleansing;
1326 * - grouplock. An io context to acquire group lock.
1328 * CIT_MISC io is used simply as a context in which locks and pages
1329 * are manipulated. Such io has no internal "process", that is,
1330 * cl_io_loop() is never called for it.
1335 * To give advice about access of a file
1339 * SEEK_HOLE/SEEK_DATA handling to search holes or data
1340 * across all file objects
1347 * States of cl_io state machine
1350 /** Not initialized. */
1354 /** IO iteration started. */
1358 /** Actual IO is in progress. */
1360 /** IO for the current iteration finished. */
1362 /** Locks released. */
1364 /** Iteration completed. */
1366 /** cl_io finalized. */
1371 * IO state private for a layer.
1373 * This is usually embedded into layer session data, rather than allocated
1376 * \see vvp_io, lov_io, osc_io
1378 struct cl_io_slice {
1379 struct cl_io *cis_io;
1380 /** corresponding object slice. Immutable after creation. */
1381 struct cl_object *cis_obj;
1382 /** io operations. Immutable after creation. */
1383 const struct cl_io_operations *cis_iop;
1385 * linkage into a list of all slices for a given cl_io, hanging off
1386 * cl_io::ci_layers. Immutable after creation.
1388 struct list_head cis_linkage;
1391 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1394 struct cl_read_ahead {
1395 /* Maximum page index the readahead window will end.
1396 * This is determined DLM lock coverage, RPC and stripe boundary.
1397 * cra_end is included. */
1398 pgoff_t cra_end_idx;
1399 /* optimal RPC size for this read, by pages */
1400 unsigned long cra_rpc_pages;
1401 /* Release callback. If readahead holds resources underneath, this
1402 * function should be called to release it. */
1403 void (*cra_release)(const struct lu_env *env,
1404 struct cl_read_ahead *ra);
1406 /* Callback data for cra_release routine */
1410 /* whether lock is in contention */
1411 bool cra_contention;
1414 static inline void cl_read_ahead_release(const struct lu_env *env,
1415 struct cl_read_ahead *ra)
1417 if (ra->cra_release != NULL)
1418 ra->cra_release(env, ra);
1419 memset(ra, 0, sizeof(*ra));
1424 * Per-layer io operations.
1425 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1427 struct cl_io_operations {
1429 * Vector of io state transition methods for every io type.
1431 * \see cl_page_operations::io
1435 * Prepare io iteration at a given layer.
1437 * Called top-to-bottom at the beginning of each iteration of
1438 * "io loop" (if it makes sense for this type of io). Here
1439 * layer selects what work it will do during this iteration.
1441 * \see cl_io_operations::cio_iter_fini()
1443 int (*cio_iter_init) (const struct lu_env *env,
1444 const struct cl_io_slice *slice);
1446 * Finalize io iteration.
1448 * Called bottom-to-top at the end of each iteration of "io
1449 * loop". Here layers can decide whether IO has to be
1452 * \see cl_io_operations::cio_iter_init()
1454 void (*cio_iter_fini) (const struct lu_env *env,
1455 const struct cl_io_slice *slice);
1457 * Collect locks for the current iteration of io.
1459 * Called top-to-bottom to collect all locks necessary for
1460 * this iteration. This methods shouldn't actually enqueue
1461 * anything, instead it should post a lock through
1462 * cl_io_lock_add(). Once all locks are collected, they are
1463 * sorted and enqueued in the proper order.
1465 int (*cio_lock) (const struct lu_env *env,
1466 const struct cl_io_slice *slice);
1468 * Finalize unlocking.
1470 * Called bottom-to-top to finish layer specific unlocking
1471 * functionality, after generic code released all locks
1472 * acquired by cl_io_operations::cio_lock().
1474 void (*cio_unlock)(const struct lu_env *env,
1475 const struct cl_io_slice *slice);
1477 * Start io iteration.
1479 * Once all locks are acquired, called top-to-bottom to
1480 * commence actual IO. In the current implementation,
1481 * top-level vvp_io_{read,write}_start() does all the work
1482 * synchronously by calling generic_file_*(), so other layers
1483 * are called when everything is done.
1485 int (*cio_start)(const struct lu_env *env,
1486 const struct cl_io_slice *slice);
1488 * Called top-to-bottom at the end of io loop. Here layer
1489 * might wait for an unfinished asynchronous io.
1491 void (*cio_end) (const struct lu_env *env,
1492 const struct cl_io_slice *slice);
1494 * Called bottom-to-top to notify layers that read/write IO
1495 * iteration finished, with \a nob bytes transferred.
1497 void (*cio_advance)(const struct lu_env *env,
1498 const struct cl_io_slice *slice,
1501 * Called once per io, bottom-to-top to release io resources.
1503 void (*cio_fini) (const struct lu_env *env,
1504 const struct cl_io_slice *slice);
1508 * Submit pages from \a queue->c2_qin for IO, and move
1509 * successfully submitted pages into \a queue->c2_qout. Return
1510 * non-zero if failed to submit even the single page. If
1511 * submission failed after some pages were moved into \a
1512 * queue->c2_qout, completion callback with non-zero ioret is
1515 int (*cio_submit)(const struct lu_env *env,
1516 const struct cl_io_slice *slice,
1517 enum cl_req_type crt,
1518 struct cl_2queue *queue);
1520 * Queue async page for write.
1521 * The difference between cio_submit and cio_queue is that
1522 * cio_submit is for urgent request.
1524 int (*cio_commit_async)(const struct lu_env *env,
1525 const struct cl_io_slice *slice,
1526 struct cl_page_list *queue, int from, int to,
1529 * Release active extent.
1531 void (*cio_extent_release)(const struct lu_env *env,
1532 const struct cl_io_slice *slice);
1534 * Decide maximum read ahead extent
1536 * \pre io->ci_type == CIT_READ
1538 int (*cio_read_ahead)(const struct lu_env *env,
1539 const struct cl_io_slice *slice,
1540 pgoff_t start, struct cl_read_ahead *ra);
1543 * Reserve LRU slots before IO.
1545 int (*cio_lru_reserve) (const struct lu_env *env,
1546 const struct cl_io_slice *slice,
1547 loff_t pos, size_t bytes);
1549 * Optional debugging helper. Print given io slice.
1551 int (*cio_print)(const struct lu_env *env, void *cookie,
1552 lu_printer_t p, const struct cl_io_slice *slice);
1556 * Flags to lock enqueue procedure.
1561 * instruct server to not block, if conflicting lock is found. Instead
1562 * -EAGAIN is returned immediately.
1564 CEF_NONBLOCK = 0x00000001,
1566 * Tell lower layers this is a glimpse request, translated to
1567 * LDLM_FL_HAS_INTENT at LDLM layer.
1569 * Also, because glimpse locks never block other locks, we count this
1570 * as automatically compatible with other osc locks.
1571 * (see osc_lock_compatible)
1573 CEF_GLIMPSE = 0x00000002,
1575 * tell the server to instruct (though a flag in the blocking ast) an
1576 * owner of the conflicting lock, that it can drop dirty pages
1577 * protected by this lock, without sending them to the server.
1579 CEF_DISCARD_DATA = 0x00000004,
1581 * tell the sub layers that it must be a `real' lock. This is used for
1582 * mmapped-buffer locks, glimpse locks, manually requested locks
1583 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1586 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1588 CEF_MUST = 0x00000008,
1590 * tell the sub layers that never request a `real' lock. This flag is
1591 * not used currently.
1593 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1594 * conversion policy: ci_lockreq describes generic information of lock
1595 * requirement for this IO, especially for locks which belong to the
1596 * object doing IO; however, lock itself may have precise requirements
1597 * that are described by the enqueue flags.
1599 CEF_NEVER = 0x00000010,
1601 * tell the dlm layer this is a speculative lock request
1602 * speculative lock requests are locks which are not requested as part
1603 * of an I/O operation. Instead, they are requested because we expect
1604 * to use them in the future. They are requested asynchronously at the
1607 * Currently used for asynchronous glimpse locks and manually requested
1608 * locks (LU_LADVISE_LOCKAHEAD).
1610 CEF_SPECULATIVE = 0x00000020,
1612 * enqueue a lock to test DLM lock existence.
1614 CEF_PEEK = 0x00000040,
1616 * Lock match only. Used by group lock in I/O as group lock
1617 * is known to exist.
1619 CEF_LOCK_MATCH = 0x00000080,
1621 * tell the DLM layer to lock only the requested range
1623 CEF_LOCK_NO_EXPAND = 0x00000100,
1625 * mask of enq_flags.
1627 CEF_MASK = 0x000001ff,
1631 * Link between lock and io. Intermediate structure is needed, because the
1632 * same lock can be part of multiple io's simultaneously.
1634 struct cl_io_lock_link {
1635 /** linkage into one of cl_lockset lists. */
1636 struct list_head cill_linkage;
1637 struct cl_lock cill_lock;
1638 /** optional destructor */
1639 void (*cill_fini)(const struct lu_env *env,
1640 struct cl_io_lock_link *link);
1642 #define cill_descr cill_lock.cll_descr
1645 * Lock-set represents a collection of locks, that io needs at a
1646 * time. Generally speaking, client tries to avoid holding multiple locks when
1649 * - holding extent locks over multiple ost's introduces the danger of
1650 * "cascading timeouts";
1652 * - holding multiple locks over the same ost is still dead-lock prone,
1653 * see comment in osc_lock_enqueue(),
1655 * but there are certain situations where this is unavoidable:
1657 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1659 * - truncate has to take [new-size, EOF] lock for correctness;
1661 * - SNS has to take locks across full stripe for correctness;
1663 * - in the case when user level buffer, supplied to {read,write}(file0),
1664 * is a part of a memory mapped lustre file, client has to take a dlm
1665 * locks on file0, and all files that back up the buffer (or a part of
1666 * the buffer, that is being processed in the current chunk, in any
1667 * case, there are situations where at least 2 locks are necessary).
1669 * In such cases we at least try to take locks in the same consistent
1670 * order. To this end, all locks are first collected, then sorted, and then
1674 /** locks to be acquired. */
1675 struct list_head cls_todo;
1676 /** locks acquired. */
1677 struct list_head cls_done;
1681 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1682 * but 'req' is always to be thought as 'request' :-)
1684 enum cl_io_lock_dmd {
1685 /** Always lock data (e.g., O_APPEND). */
1687 /** Layers are free to decide between local and global locking. */
1689 /** Never lock: there is no cache (e.g., liblustre). */
1693 enum cl_fsync_mode {
1694 /** start writeback, do not wait for them to finish */
1696 /** start writeback and wait for them to finish */
1698 /** discard all of dirty pages in a specific file range */
1699 CL_FSYNC_DISCARD = 2,
1700 /** start writeback and make sure they have reached storage before
1701 * return. OST_SYNC RPC must be issued and finished */
1705 struct cl_io_rw_common {
1710 enum cl_setattr_subtype {
1711 /** regular setattr **/
1715 /** fallocate(2) - mode preallocate **/
1716 CL_SETATTR_FALLOCATE
1719 struct cl_io_range {
1725 struct cl_io_pt *cip_next;
1726 struct kiocb cip_iocb;
1727 struct iov_iter cip_iter;
1728 struct file *cip_file;
1729 enum cl_io_type cip_iot;
1730 unsigned int cip_need_restart:1;
1739 * cl_io is shared by all threads participating in this IO (in current
1740 * implementation only one thread advances IO, but parallel IO design and
1741 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1742 * is up to these threads to serialize their activities, including updates to
1743 * mutable cl_io fields.
1746 /** type of this IO. Immutable after creation. */
1747 enum cl_io_type ci_type;
1748 /** current state of cl_io state machine. */
1749 enum cl_io_state ci_state;
1750 /** main object this io is against. Immutable after creation. */
1751 struct cl_object *ci_obj;
1752 /** top level dio_aio */
1753 struct cl_dio_aio *ci_dio_aio;
1755 * Upper layer io, of which this io is a part of. Immutable after
1758 struct cl_io *ci_parent;
1759 /** List of slices. Immutable after creation. */
1760 struct list_head ci_layers;
1761 /** list of locks (to be) acquired by this io. */
1762 struct cl_lockset ci_lockset;
1763 /** lock requirements, this is just a help info for sublayers. */
1764 enum cl_io_lock_dmd ci_lockreq;
1765 /** layout version when this IO occurs */
1766 __u32 ci_layout_version;
1769 struct cl_io_rw_common rd;
1772 struct cl_io_rw_common wr;
1776 struct cl_io_rw_common ci_rw;
1777 struct cl_setattr_io {
1778 struct ost_lvb sa_attr;
1779 unsigned int sa_attr_flags;
1780 unsigned int sa_avalid; /* ATTR_* */
1781 unsigned int sa_xvalid; /* OP_XVALID */
1782 int sa_stripe_index;
1783 struct ost_layout sa_layout;
1784 const struct lu_fid *sa_parent_fid;
1785 /* SETATTR interface is used for regular setattr, */
1786 /* truncate(2) and fallocate(2) subtypes */
1787 enum cl_setattr_subtype sa_subtype;
1788 /* The following are used for fallocate(2) */
1790 loff_t sa_falloc_offset;
1791 loff_t sa_falloc_end;
1792 uid_t sa_falloc_uid;
1793 gid_t sa_falloc_gid;
1794 __u32 sa_falloc_projid;
1796 struct cl_data_version_io {
1797 u64 dv_data_version;
1798 u32 dv_layout_version;
1801 struct cl_fault_io {
1802 /** page index within file. */
1804 /** bytes valid byte on a faulted page. */
1806 /** writable page? for nopage() only */
1808 /** page of an executable? */
1810 /** page_mkwrite() */
1812 /** resulting page */
1813 struct cl_page *ft_page;
1815 struct cl_fsync_io {
1818 /** file system level fid */
1819 struct lu_fid *fi_fid;
1820 enum cl_fsync_mode fi_mode;
1821 /* how many pages were written/discarded */
1822 unsigned int fi_nr_written;
1824 struct cl_ladvise_io {
1827 /** file system level fid */
1828 struct lu_fid *li_fid;
1829 enum lu_ladvise_type li_advice;
1832 struct cl_lseek_io {
1838 time64_t lm_next_rpc_time;
1841 struct cl_2queue ci_queue;
1844 unsigned int ci_continue:1,
1846 * This io has held grouplock, to inform sublayers that
1847 * don't do lockless i/o.
1851 * The whole IO need to be restarted because layout has been changed
1855 * to not refresh layout - the IO issuer knows that the layout won't
1856 * change(page operations, layout change causes all page to be
1857 * discarded), or it doesn't matter if it changes(sync).
1861 * Need MDS intervention to complete a write.
1862 * Write intent is required for the following cases:
1863 * 1. component being written is not initialized, or
1864 * 2. the mirrored files are NOT in WRITE_PENDING state.
1866 ci_need_write_intent:1,
1868 * Check if layout changed after the IO finishes. Mainly for HSM
1869 * requirement. If IO occurs to openning files, it doesn't need to
1870 * verify layout because HSM won't release openning files.
1871 * Right now, only two opertaions need to verify layout: glimpse
1876 * file is released, restore has to to be triggered by vvp layer
1878 ci_restore_needed:1,
1883 /* Tell sublayers not to expand LDLM locks requested for this IO */
1884 ci_lock_no_expand:1,
1886 * Set if non-delay RPC should be used for this IO.
1888 * If this file has multiple mirrors, and if the OSTs of the current
1889 * mirror is inaccessible, non-delay RPC would error out quickly so
1890 * that the upper layer can try to access the next mirror.
1894 * Set if IO is triggered by async workqueue readahead.
1896 ci_async_readahead:1,
1898 * Ignore lockless and do normal locking for this io.
1902 * Set if we've tried all mirrors for this read IO, if it's not set,
1903 * the read IO will check to-be-read OSCs' status, and make fast-switch
1904 * another mirror if some of the OSTs are not healthy.
1906 ci_tried_all_mirrors:1,
1908 * Random read hints, readahead will be disabled.
1912 * Sequential read hints.
1916 * Do parallel (async) submission of DIO RPCs. Note DIO is still sync
1917 * to userspace, only the RPCs are submitted async, then waited for at
1918 * the llite layer before returning.
1922 * Bypass quota check
1924 unsigned ci_noquota:1,
1926 * io_uring direct IO with flags IOCB_NOWAIT.
1930 * The filesystem must exclusively acquire invalidate_lock before
1931 * invalidating page cache in truncate / hole punch / DLM extent
1932 * lock blocking AST path (and thus calling into ->invalidatepage)
1933 * to block races between page cache invalidation and page cache
1934 * filling functions (fault, read, ...)
1936 ci_invalidate_page_cache:1;
1939 * How many times the read has retried before this one.
1940 * Set by the top level and consumed by the LOV.
1942 unsigned ci_ndelay_tried;
1944 * Designated mirror index for this I/O.
1946 unsigned ci_designated_mirror;
1948 * Number of pages owned by this IO. For invariant checking.
1950 unsigned ci_owned_nr;
1952 * Range of write intent. Valid if ci_need_write_intent is set.
1954 struct lu_extent ci_write_intent;
1960 * Per-transfer attributes.
1962 struct cl_req_attr {
1963 enum cl_req_type cra_type;
1965 struct cl_page *cra_page;
1966 /** Generic attributes for the server consumption. */
1967 struct obdo *cra_oa;
1969 char cra_jobid[LUSTRE_JOBID_SIZE];
1970 /** uid/gid of the process doing an io */
1975 enum cache_stats_item {
1976 /** how many cache lookups were performed */
1978 /** how many times cache lookup resulted in a hit */
1980 /** how many entities are in the cache right now */
1982 /** how many entities in the cache are actively used (and cannot be
1983 * evicted) right now */
1985 /** how many entities were created at all */
1990 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
1993 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
1995 struct cache_stats {
1996 const char *cs_name;
1997 atomic_t cs_stats[CS_NR];
2000 /** These are not exported so far */
2001 void cache_stats_init (struct cache_stats *cs, const char *name);
2004 * Client-side site. This represents particular client stack. "Global"
2005 * variables should (directly or indirectly) be added here to allow multiple
2006 * clients to co-exist in the single address space.
2009 struct lu_site cs_lu;
2011 * Statistical counters. Atomics do not scale, something better like
2012 * per-cpu counters is needed.
2014 * These are exported as /proc/fs/lustre/llite/.../site
2016 * When interpreting keep in mind that both sub-locks (and sub-pages)
2017 * and top-locks (and top-pages) are accounted here.
2019 struct cache_stats cs_pages;
2020 atomic_t cs_pages_state[CPS_NR];
2023 int cl_site_init(struct cl_site *s, struct cl_device *top);
2024 void cl_site_fini(struct cl_site *s);
2025 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2028 * Output client site statistical counters into a buffer. Suitable for
2029 * ll_rd_*()-style functions.
2031 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2036 * Type conversion and accessory functions.
2040 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2042 return container_of(site, struct cl_site, cs_lu);
2045 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2047 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2048 return container_of_safe(d, struct cl_device, cd_lu_dev);
2051 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2053 return &d->cd_lu_dev;
2056 static inline struct cl_object *lu2cl(const struct lu_object *o)
2058 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2059 return container_of_safe(o, struct cl_object, co_lu);
2062 static inline const struct cl_object_conf *
2063 lu2cl_conf(const struct lu_object_conf *conf)
2065 return container_of_safe(conf, struct cl_object_conf, coc_lu);
2068 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2070 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2073 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2075 return container_of_safe(h, struct cl_object_header, coh_lu);
2078 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2080 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2084 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2086 return luh2coh(obj->co_lu.lo_header);
2089 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2091 return lu_device_init(&d->cd_lu_dev, t);
2094 static inline void cl_device_fini(struct cl_device *d)
2096 lu_device_fini(&d->cd_lu_dev);
2099 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2100 struct cl_object *obj,
2101 const struct cl_page_operations *ops);
2102 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2103 struct cl_object *obj,
2104 const struct cl_lock_operations *ops);
2105 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2106 struct cl_object *obj, const struct cl_io_operations *ops);
2109 /** \defgroup cl_object cl_object
2111 struct cl_object *cl_object_top (struct cl_object *o);
2112 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2113 const struct lu_fid *fid,
2114 const struct cl_object_conf *c);
2116 int cl_object_header_init(struct cl_object_header *h);
2117 void cl_object_header_fini(struct cl_object_header *h);
2118 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2119 void cl_object_get (struct cl_object *o);
2120 void cl_object_attr_lock (struct cl_object *o);
2121 void cl_object_attr_unlock(struct cl_object *o);
2122 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2123 struct cl_attr *attr);
2124 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2125 const struct cl_attr *attr, unsigned valid);
2126 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2127 struct ost_lvb *lvb);
2128 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2129 const struct cl_object_conf *conf);
2130 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2131 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2132 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2133 struct lov_user_md __user *lum, size_t size);
2134 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2135 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2137 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2138 struct cl_layout *cl);
2139 loff_t cl_object_maxbytes(struct cl_object *obj);
2140 int cl_object_flush(const struct lu_env *env, struct cl_object *obj,
2141 struct ldlm_lock *lock);
2145 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2147 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2149 return cl_object_header(o0) == cl_object_header(o1);
2152 static inline void cl_object_page_init(struct cl_object *clob, int size)
2154 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2155 cl_object_header(clob)->coh_page_bufsize += round_up(size, 8);
2156 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2159 static inline void *cl_object_page_slice(struct cl_object *clob,
2160 struct cl_page *page)
2162 return (void *)((char *)page + clob->co_slice_off);
2166 * Return refcount of cl_object.
2168 static inline int cl_object_refc(struct cl_object *clob)
2170 struct lu_object_header *header = clob->co_lu.lo_header;
2171 return atomic_read(&header->loh_ref);
2176 /** \defgroup cl_page cl_page
2178 struct cl_page *cl_page_find (const struct lu_env *env,
2179 struct cl_object *obj,
2180 pgoff_t idx, struct page *vmpage,
2181 enum cl_page_type type);
2182 struct cl_page *cl_page_alloc (const struct lu_env *env,
2183 struct cl_object *o, pgoff_t ind,
2184 struct page *vmpage,
2185 enum cl_page_type type);
2186 void cl_page_get (struct cl_page *page);
2187 void cl_page_put (const struct lu_env *env,
2188 struct cl_page *page);
2189 void cl_pagevec_put (const struct lu_env *env,
2190 struct cl_page *page,
2191 struct pagevec *pvec);
2192 void cl_page_print (const struct lu_env *env, void *cookie,
2193 lu_printer_t printer,
2194 const struct cl_page *pg);
2195 void cl_page_header_print(const struct lu_env *env, void *cookie,
2196 lu_printer_t printer,
2197 const struct cl_page *pg);
2198 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2203 * Functions dealing with the ownership of page by io.
2207 int cl_page_own (const struct lu_env *env,
2208 struct cl_io *io, struct cl_page *page);
2209 int cl_page_own_try (const struct lu_env *env,
2210 struct cl_io *io, struct cl_page *page);
2211 void cl_page_assume (const struct lu_env *env,
2212 struct cl_io *io, struct cl_page *page);
2213 void cl_page_unassume (const struct lu_env *env,
2214 struct cl_io *io, struct cl_page *pg);
2215 void cl_page_disown (const struct lu_env *env,
2216 struct cl_io *io, struct cl_page *page);
2217 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2224 * Functions dealing with the preparation of a page for a transfer, and
2225 * tracking transfer state.
2228 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2229 struct cl_page *pg, enum cl_req_type crt);
2230 void cl_page_completion (const struct lu_env *env,
2231 struct cl_page *pg, enum cl_req_type crt, int ioret);
2232 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2233 enum cl_req_type crt);
2234 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2235 struct cl_page *pg, enum cl_req_type crt);
2236 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2238 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2239 struct cl_page *pg);
2245 * \name helper routines
2246 * Functions to discard, delete and export a cl_page.
2249 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2250 struct cl_page *pg);
2251 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2252 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
2255 void cl_lock_print(const struct lu_env *env, void *cookie,
2256 lu_printer_t printer, const struct cl_lock *lock);
2257 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2258 lu_printer_t printer,
2259 const struct cl_lock_descr *descr);
2263 * Data structure managing a client's cached pages. A count of
2264 * "unstable" pages is maintained, and an LRU of clean pages is
2265 * maintained. "unstable" pages are pages pinned by the ptlrpc
2266 * layer for recovery purposes.
2268 struct cl_client_cache {
2270 * # of client cache refcount
2271 * # of users (OSCs) + 2 (held by llite and lov)
2273 refcount_t ccc_users;
2275 * # of threads are doing shrinking
2277 unsigned int ccc_lru_shrinkers;
2279 * # of LRU entries available
2281 atomic_long_t ccc_lru_left;
2283 * List of entities(OSCs) for this LRU cache
2285 struct list_head ccc_lru;
2287 * Max # of LRU entries
2289 unsigned long ccc_lru_max;
2291 * Lock to protect ccc_lru list
2293 spinlock_t ccc_lru_lock;
2295 * Set if unstable check is enabled
2297 unsigned int ccc_unstable_check:1;
2299 * # of unstable pages for this mount point
2301 atomic_long_t ccc_unstable_nr;
2303 * Waitq for awaiting unstable pages to reach zero.
2304 * Used at umounting time and signaled on BRW commit
2306 wait_queue_head_t ccc_unstable_waitq;
2308 * Serialize max_cache_mb write operation
2310 struct mutex ccc_max_cache_mb_lock;
2313 * cl_cache functions
2315 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2316 void cl_cache_incref(struct cl_client_cache *cache);
2317 void cl_cache_decref(struct cl_client_cache *cache);
2321 /** \defgroup cl_lock cl_lock
2323 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2324 struct cl_lock *lock);
2325 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2326 const struct cl_io *io);
2327 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2328 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2329 const struct lu_device_type *dtype);
2330 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2332 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2333 struct cl_lock *lock, struct cl_sync_io *anchor);
2334 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2338 /** \defgroup cl_io cl_io
2341 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2342 enum cl_io_type iot, struct cl_object *obj);
2343 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2344 enum cl_io_type iot, struct cl_object *obj);
2345 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2346 enum cl_io_type iot, loff_t pos, size_t count);
2347 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2349 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2350 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2351 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2352 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2353 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2354 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2355 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2356 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2357 struct cl_io_lock_link *link);
2358 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2359 struct cl_lock_descr *descr);
2360 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2361 enum cl_req_type iot, struct cl_2queue *queue);
2362 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2363 enum cl_req_type iot, struct cl_2queue *queue,
2365 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2366 struct cl_page_list *queue, int from, int to,
2368 void cl_io_extent_release (const struct lu_env *env, struct cl_io *io);
2369 int cl_io_lru_reserve(const struct lu_env *env, struct cl_io *io,
2370 loff_t pos, size_t bytes);
2371 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2372 pgoff_t start, struct cl_read_ahead *ra);
2373 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2377 * True, iff \a io is an O_APPEND write(2).
2379 static inline int cl_io_is_append(const struct cl_io *io)
2381 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2384 static inline int cl_io_is_sync_write(const struct cl_io *io)
2386 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2389 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2391 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2395 * True, iff \a io is a truncate(2).
2397 static inline int cl_io_is_trunc(const struct cl_io *io)
2399 return io->ci_type == CIT_SETATTR &&
2400 (io->u.ci_setattr.sa_avalid & ATTR_SIZE) &&
2401 (io->u.ci_setattr.sa_subtype != CL_SETATTR_FALLOCATE);
2404 static inline int cl_io_is_fallocate(const struct cl_io *io)
2406 return (io->ci_type == CIT_SETATTR) &&
2407 (io->u.ci_setattr.sa_subtype == CL_SETATTR_FALLOCATE);
2410 struct cl_io *cl_io_top(struct cl_io *io);
2412 void cl_io_print(const struct lu_env *env, void *cookie,
2413 lu_printer_t printer, const struct cl_io *io);
2415 #define CL_IO_SLICE_CLEAN(obj, base) memset_startat(obj, 0, base)
2419 /** \defgroup cl_page_list cl_page_list
2423 * Last page in the page list.
2425 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2427 LASSERT(plist->pl_nr > 0);
2428 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2431 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2433 LASSERT(plist->pl_nr > 0);
2434 return list_first_entry(&plist->pl_pages, struct cl_page, cp_batch);
2438 * Iterate over pages in a page list.
2440 #define cl_page_list_for_each(page, list) \
2441 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2444 * Iterate over pages in a page list, taking possible removals into account.
2446 #define cl_page_list_for_each_safe(page, temp, list) \
2447 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2449 void cl_page_list_init(struct cl_page_list *plist);
2450 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page,
2452 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2453 struct cl_page *page);
2454 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2455 struct cl_page *page);
2456 void cl_page_list_splice(struct cl_page_list *list,
2457 struct cl_page_list *head);
2458 void cl_page_list_del(const struct lu_env *env,
2459 struct cl_page_list *plist, struct cl_page *page);
2460 void cl_page_list_disown(const struct lu_env *env,
2461 struct cl_page_list *plist);
2462 void cl_page_list_assume(const struct lu_env *env,
2463 struct cl_io *io, struct cl_page_list *plist);
2464 void cl_page_list_discard(const struct lu_env *env,
2465 struct cl_io *io, struct cl_page_list *plist);
2466 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2468 void cl_2queue_init(struct cl_2queue *queue);
2469 void cl_2queue_disown(const struct lu_env *env, struct cl_2queue *queue);
2470 void cl_2queue_assume(const struct lu_env *env, struct cl_io *io,
2471 struct cl_2queue *queue);
2472 void cl_2queue_discard(const struct lu_env *env, struct cl_io *io,
2473 struct cl_2queue *queue);
2474 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2475 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2477 /** @} cl_page_list */
2479 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2480 struct cl_req_attr *attr);
2482 /** \defgroup cl_sync_io cl_sync_io
2489 typedef void (cl_sync_io_end_t)(const struct lu_env *, struct cl_sync_io *);
2491 void cl_sync_io_init_notify(struct cl_sync_io *anchor, int nr, void *dio_aio,
2492 cl_sync_io_end_t *end);
2494 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2496 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2498 int cl_sync_io_wait_recycle(const struct lu_env *env, struct cl_sync_io *anchor,
2499 long timeout, int ioret);
2500 struct cl_dio_aio *cl_dio_aio_alloc(struct kiocb *iocb, struct cl_object *obj,
2502 struct cl_sub_dio *cl_sub_dio_alloc(struct cl_dio_aio *ll_aio, bool sync);
2503 void cl_dio_aio_free(const struct lu_env *env, struct cl_dio_aio *aio);
2504 void cl_sub_dio_free(struct cl_sub_dio *sdio);
2505 static inline void cl_sync_io_init(struct cl_sync_io *anchor, int nr)
2507 cl_sync_io_init_notify(anchor, nr, NULL, NULL);
2511 * Anchor for synchronous transfer. This is allocated on a stack by thread
2512 * doing synchronous transfer, and a pointer to this structure is set up in
2513 * every page submitted for transfer. Transfer completion routine updates
2514 * anchor and wakes up waiting thread when transfer is complete.
2517 /** number of pages yet to be transferred. */
2518 atomic_t csi_sync_nr;
2521 /** completion to be signaled when transfer is complete. */
2522 wait_queue_head_t csi_waitq;
2523 /** callback to invoke when this IO is finished */
2524 cl_sync_io_end_t *csi_end_io;
2525 /* private pointer for an associated DIO/AIO */
2529 /** direct IO pages */
2530 struct ll_dio_pages {
2532 * page array to be written. we don't support
2533 * partial pages except the last one.
2535 struct page **ldp_pages;
2536 /** # of pages in the array. */
2538 /* the file offset of the first page. */
2539 loff_t ldp_file_offset;
2542 /* Top level struct used for AIO and DIO */
2544 struct cl_sync_io cda_sync;
2545 struct cl_object *cda_obj;
2546 struct kiocb *cda_iocb;
2548 unsigned cda_no_aio_complete:1,
2552 /* Sub-dio used for splitting DIO (and AIO, because AIO is DIO) according to
2553 * the layout/striping, so we can do parallel submit of DIO RPCs
2556 struct cl_sync_io csd_sync;
2557 struct cl_page_list csd_pages;
2559 struct cl_dio_aio *csd_ll_aio;
2560 struct ll_dio_pages csd_dio_pages;
2561 unsigned csd_creator_free:1;
2563 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW) || \
2564 defined(HAVE_DIRECTIO_2ARGS)
2565 #define HAVE_DIO_ITER 1
2568 void ll_release_user_pages(struct page **pages, int npages);
2570 /** @} cl_sync_io */
2572 /** \defgroup cl_env cl_env
2574 * lu_env handling for a client.
2576 * lu_env is an environment within which lustre code executes. Its major part
2577 * is lu_context---a fast memory allocation mechanism that is used to conserve
2578 * precious kernel stack space. Originally lu_env was designed for a server,
2581 * - there is a (mostly) fixed number of threads, and
2583 * - call chains have no non-lustre portions inserted between lustre code.
2585 * On a client both these assumtpion fails, because every user thread can
2586 * potentially execute lustre code as part of a system call, and lustre calls
2587 * into VFS or MM that call back into lustre.
2589 * To deal with that, cl_env wrapper functions implement the following
2592 * - allocation and destruction of environment is amortized by caching no
2593 * longer used environments instead of destroying them;
2595 * \see lu_env, lu_context, lu_context_key
2598 struct lu_env *cl_env_get(__u16 *refcheck);
2599 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2600 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2601 unsigned cl_env_cache_purge(unsigned nr);
2602 struct lu_env *cl_env_percpu_get(void);
2603 void cl_env_percpu_put(struct lu_env *env);
2610 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2611 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2613 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2614 struct lu_device_type *ldt,
2615 struct lu_device *next);
2618 int cl_global_init(void);
2619 void cl_global_fini(void);
2621 #endif /* _LINUX_CL_OBJECT_H */