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 <lustre_dlm.h>
111 struct cl_page_slice;
113 struct cl_lock_slice;
115 struct cl_lock_operations;
116 struct cl_page_operations;
124 * Device in the client stack.
126 * \see vvp_device, lov_device, lovsub_device, osc_device
130 struct lu_device cd_lu_dev;
133 /** \addtogroup cl_object cl_object
136 * "Data attributes" of cl_object. Data attributes can be updated
137 * independently for a sub-object, and top-object's attributes are calculated
138 * from sub-objects' ones.
141 /** Object size, in bytes */
144 unsigned int cat_kms_valid:1;
146 * Known minimal size, in bytes.
148 * This is only valid when at least one DLM lock is held.
151 /** Modification time. Measured in seconds since epoch. */
153 /** Access time. Measured in seconds since epoch. */
155 /** Change time. Measured in seconds since epoch. */
158 * Blocks allocated to this cl_object on the server file system.
160 * \todo XXX An interface for block size is needed.
164 * User identifier for quota purposes.
168 * Group identifier for quota purposes.
172 /* nlink of the directory */
175 /* Project identifier for quota purpose. */
180 * Fields in cl_attr that are being set.
195 * Sub-class of lu_object with methods common for objects on the client
198 * cl_object: represents a regular file system object, both a file and a
199 * stripe. cl_object is based on lu_object: it is identified by a fid,
200 * layered, cached, hashed, and lrued. Important distinction with the server
201 * side, where md_object and dt_object are used, is that cl_object "fans out"
202 * at the lov/sns level: depending on the file layout, single file is
203 * represented as a set of "sub-objects" (stripes). At the implementation
204 * level, struct lov_object contains an array of cl_objects. Each sub-object
205 * is a full-fledged cl_object, having its fid, living in the lru and hash
208 * This leads to the next important difference with the server side: on the
209 * client, it's quite usual to have objects with the different sequence of
210 * layers. For example, typical top-object is composed of the following
216 * whereas its sub-objects are composed of
221 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
222 * track of the object-subobject relationship.
224 * Sub-objects are not cached independently: when top-object is about to
225 * be discarded from the memory, all its sub-objects are torn-down and
228 * \see vvp_object, lov_object, lovsub_object, osc_object
232 struct lu_object co_lu;
233 /** per-object-layer operations */
234 const struct cl_object_operations *co_ops;
235 /** offset of page slice in cl_page buffer */
240 * Description of the client object configuration. This is used for the
241 * creation of a new client object that is identified by a more state than
244 struct cl_object_conf {
246 struct lu_object_conf coc_lu;
249 * Object layout. This is consumed by lov.
251 struct lu_buf coc_layout;
253 * Description of particular stripe location in the
254 * cluster. This is consumed by osc.
256 struct lov_oinfo *coc_oinfo;
259 * VFS inode. This is consumed by vvp.
261 struct inode *coc_inode;
263 * Layout lock handle.
265 struct ldlm_lock *coc_lock;
267 * Operation to handle layout, OBJECT_CONF_XYZ.
273 /** configure layout, set up a new stripe, must be called while
274 * holding layout lock. */
276 /** invalidate the current stripe configuration due to losing
278 OBJECT_CONF_INVALIDATE = 1,
279 /** wait for old layout to go away so that new layout can be
285 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
286 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
290 /** the buffer to return the layout in lov_mds_md format. */
291 struct lu_buf cl_buf;
292 /** size of layout in lov_mds_md format. */
294 /** Layout generation. */
296 /** whether layout is a composite one */
297 bool cl_is_composite;
298 /** Whether layout is a HSM released one */
303 * Operations implemented for each cl object layer.
305 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
307 struct cl_object_operations {
309 * Initialize page slice for this layer. Called top-to-bottom through
310 * every object layer when a new cl_page is instantiated. Layer
311 * keeping private per-page data, or requiring its own page operations
312 * vector should allocate these data here, and attach then to the page
313 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
316 * \retval NULL success.
318 * \retval ERR_PTR(errno) failure code.
320 * \retval valid-pointer pointer to already existing referenced page
321 * to be used instead of newly created.
323 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
324 struct cl_page *page, pgoff_t index);
326 * Initialize lock slice for this layer. Called top-to-bottom through
327 * every object layer when a new cl_lock is instantiated. Layer
328 * keeping private per-lock data, or requiring its own lock operations
329 * vector should allocate these data here, and attach then to the lock
330 * by calling cl_lock_slice_add(). Mandatory.
332 int (*coo_lock_init)(const struct lu_env *env,
333 struct cl_object *obj, struct cl_lock *lock,
334 const struct cl_io *io);
336 * Initialize io state for a given layer.
338 * called top-to-bottom once per io existence to initialize io
339 * state. If layer wants to keep some state for this type of io, it
340 * has to embed struct cl_io_slice in lu_env::le_ses, and register
341 * slice with cl_io_slice_add(). It is guaranteed that all threads
342 * participating in this io share the same session.
344 int (*coo_io_init)(const struct lu_env *env,
345 struct cl_object *obj, struct cl_io *io);
347 * Fill portion of \a attr that this layer controls. This method is
348 * called top-to-bottom through all object layers.
350 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
352 * \return 0: to continue
353 * \return +ve: to stop iterating through layers (but 0 is returned
354 * from enclosing cl_object_attr_get())
355 * \return -ve: to signal error
357 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
358 struct cl_attr *attr);
362 * \a valid is a bitmask composed from enum #cl_attr_valid, and
363 * indicating what attributes are to be set.
365 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
367 * \return the same convention as for
368 * cl_object_operations::coo_attr_get() is used.
370 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
371 const struct cl_attr *attr, unsigned valid);
373 * Update object configuration. Called top-to-bottom to modify object
376 * XXX error conditions and handling.
378 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
379 const struct cl_object_conf *conf);
381 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
382 * object. Layers are supposed to fill parts of \a lvb that will be
383 * shipped to the glimpse originator as a glimpse result.
385 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
386 * \see osc_object_glimpse()
388 int (*coo_glimpse)(const struct lu_env *env,
389 const struct cl_object *obj, struct ost_lvb *lvb);
391 * Object prune method. Called when the layout is going to change on
392 * this object, therefore each layer has to clean up their cache,
393 * mainly pages and locks.
395 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
397 * Object getstripe method.
399 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
400 struct lov_user_md __user *lum, size_t size);
402 * Get FIEMAP mapping from the object.
404 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
405 struct ll_fiemap_info_key *fmkey,
406 struct fiemap *fiemap, size_t *buflen);
408 * Get layout and generation of the object.
410 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
411 struct cl_layout *layout);
413 * Get maximum size of the object.
415 loff_t (*coo_maxbytes)(struct cl_object *obj);
417 * Set request attributes.
419 void (*coo_req_attr_set)(const struct lu_env *env,
420 struct cl_object *obj,
421 struct cl_req_attr *attr);
423 * Flush \a obj data corresponding to \a lock. Used for DoM
424 * locks in llite's cancelling blocking ast callback.
426 int (*coo_object_flush)(const struct lu_env *env,
427 struct cl_object *obj,
428 struct ldlm_lock *lock);
432 * Extended header for client object.
434 struct cl_object_header {
435 /** Standard lu_object_header. cl_object::co_lu::lo_header points
437 struct lu_object_header coh_lu;
440 * Parent object. It is assumed that an object has a well-defined
441 * parent, but not a well-defined child (there may be multiple
442 * sub-objects, for the same top-object). cl_object_header::coh_parent
443 * field allows certain code to be written generically, without
444 * limiting possible cl_object layouts unduly.
446 struct cl_object_header *coh_parent;
448 * Protects consistency between cl_attr of parent object and
449 * attributes of sub-objects, that the former is calculated ("merged")
452 * \todo XXX this can be read/write lock if needed.
454 spinlock_t coh_attr_guard;
456 * Size of cl_page + page slices
458 unsigned short coh_page_bufsize;
460 * Number of objects above this one: 0 for a top-object, 1 for its
463 unsigned char coh_nesting;
467 * Helper macro: iterate over all layers of the object \a obj, assigning every
468 * layer top-to-bottom to \a slice.
470 #define cl_object_for_each(slice, obj) \
471 list_for_each_entry((slice), \
472 &(obj)->co_lu.lo_header->loh_layers,\
476 * Helper macro: iterate over all layers of the object \a obj, assigning every
477 * layer bottom-to-top to \a slice.
479 #define cl_object_for_each_reverse(slice, obj) \
480 list_for_each_entry_reverse((slice), \
481 &(obj)->co_lu.lo_header->loh_layers,\
486 #define CL_PAGE_EOF ((pgoff_t)~0ull)
488 /** \addtogroup cl_page cl_page
492 * Layered client page.
494 * cl_page: represents a portion of a file, cached in the memory. All pages
495 * of the given file are of the same size, and are kept in the radix tree
496 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
497 * of the top-level file object are first class cl_objects, they have their
498 * own radix trees of pages and hence page is implemented as a sequence of
499 * struct cl_pages's, linked into double-linked list through
500 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
501 * corresponding radix tree at the corresponding logical offset.
503 * cl_page is associated with VM page of the hosting environment (struct
504 * page in Linux kernel, for example), struct page. It is assumed, that this
505 * association is implemented by one of cl_page layers (top layer in the
506 * current design) that
508 * - intercepts per-VM-page call-backs made by the environment (e.g.,
511 * - translates state (page flag bits) and locking between lustre and
514 * The association between cl_page and struct page is immutable and
515 * established when cl_page is created.
517 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
518 * this io an exclusive access to this page w.r.t. other io attempts and
519 * various events changing page state (such as transfer completion, or
520 * eviction of the page from the memory). Note, that in general cl_io
521 * cannot be identified with a particular thread, and page ownership is not
522 * exactly equal to the current thread holding a lock on the page. Layer
523 * implementing association between cl_page and struct page has to implement
524 * ownership on top of available synchronization mechanisms.
526 * While lustre client maintains the notion of an page ownership by io,
527 * hosting MM/VM usually has its own page concurrency control
528 * mechanisms. For example, in Linux, page access is synchronized by the
529 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
530 * takes care to acquire and release such locks as necessary around the
531 * calls to the file system methods (->readpage(), ->prepare_write(),
532 * ->commit_write(), etc.). This leads to the situation when there are two
533 * different ways to own a page in the client:
535 * - client code explicitly and voluntary owns the page (cl_page_own());
537 * - VM locks a page and then calls the client, that has "to assume"
538 * the ownership from the VM (cl_page_assume()).
540 * Dual methods to release ownership are cl_page_disown() and
541 * cl_page_unassume().
543 * cl_page is reference counted (cl_page::cp_ref). When reference counter
544 * drops to 0, the page is returned to the cache, unless it is in
545 * cl_page_state::CPS_FREEING state, in which case it is immediately
548 * The general logic guaranteeing the absence of "existential races" for
549 * pages is the following:
551 * - there are fixed known ways for a thread to obtain a new reference
554 * - by doing a lookup in the cl_object radix tree, protected by the
557 * - by starting from VM-locked struct page and following some
558 * hosting environment method (e.g., following ->private pointer in
559 * the case of Linux kernel), see cl_vmpage_page();
561 * - when the page enters cl_page_state::CPS_FREEING state, all these
562 * ways are severed with the proper synchronization
563 * (cl_page_delete());
565 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
568 * - no new references to the page in cl_page_state::CPS_FREEING state
569 * are allowed (checked in cl_page_get()).
571 * Together this guarantees that when last reference to a
572 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
573 * page, as neither references to it can be acquired at that point, nor
576 * cl_page is a state machine. States are enumerated in enum
577 * cl_page_state. Possible state transitions are enumerated in
578 * cl_page_state_set(). State transition process (i.e., actual changing of
579 * cl_page::cp_state field) is protected by the lock on the underlying VM
582 * Linux Kernel implementation.
584 * Binding between cl_page and struct page (which is a typedef for
585 * struct page) is implemented in the vvp layer. cl_page is attached to the
586 * ->private pointer of the struct page, together with the setting of
587 * PG_private bit in page->flags, and acquiring additional reference on the
588 * struct page (much like struct buffer_head, or any similar file system
589 * private data structures).
591 * PG_locked lock is used to implement both ownership and transfer
592 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
593 * states. No additional references are acquired for the duration of the
596 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
597 * write-out is "protected" by the special PG_writeback bit.
601 * States of cl_page. cl_page.c assumes particular order here.
603 * The page state machine is rather crude, as it doesn't recognize finer page
604 * states like "dirty" or "up to date". This is because such states are not
605 * always well defined for the whole stack (see, for example, the
606 * implementation of the read-ahead, that hides page up-to-dateness to track
607 * cache hits accurately). Such sub-states are maintained by the layers that
608 * are interested in them.
612 * Page is in the cache, un-owned. Page leaves cached state in the
615 * - [cl_page_state::CPS_OWNED] io comes across the page and
618 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
619 * req-formation engine decides that it wants to include this page
620 * into an RPC being constructed, and yanks it from the cache;
622 * - [cl_page_state::CPS_FREEING] VM callback is executed to
623 * evict the page form the memory;
625 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
629 * Page is exclusively owned by some cl_io. Page may end up in this
630 * state as a result of
632 * - io creating new page and immediately owning it;
634 * - [cl_page_state::CPS_CACHED] io finding existing cached page
637 * - [cl_page_state::CPS_OWNED] io finding existing owned page
638 * and waiting for owner to release the page;
640 * Page leaves owned state in the following cases:
642 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
643 * the cache, doing nothing;
645 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
648 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
649 * transfer for this page;
651 * - [cl_page_state::CPS_FREEING] io decides to destroy this
652 * page (e.g., as part of truncate or extent lock cancellation).
654 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
658 * Page is being written out, as a part of a transfer. This state is
659 * entered when req-formation logic decided that it wants this page to
660 * be sent through the wire _now_. Specifically, it means that once
661 * this state is achieved, transfer completion handler (with either
662 * success or failure indication) is guaranteed to be executed against
663 * this page independently of any locks and any scheduling decisions
664 * made by the hosting environment (that effectively means that the
665 * page is never put into cl_page_state::CPS_PAGEOUT state "in
666 * advance". This property is mentioned, because it is important when
667 * reasoning about possible dead-locks in the system). The page can
668 * enter this state as a result of
670 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
671 * write-out of this page, or
673 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
674 * that it has enough dirty pages cached to issue a "good"
677 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
678 * is completed---it is moved into cl_page_state::CPS_CACHED state.
680 * Underlying VM page is locked for the duration of transfer.
682 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
686 * Page is being read in, as a part of a transfer. This is quite
687 * similar to the cl_page_state::CPS_PAGEOUT state, except that
688 * read-in is always "immediate"---there is no such thing a sudden
689 * construction of read request from cached, presumably not up to date,
692 * Underlying VM page is locked for the duration of transfer.
694 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
698 * Page is being destroyed. This state is entered when client decides
699 * that page has to be deleted from its host object, as, e.g., a part
702 * Once this state is reached, there is no way to escape it.
704 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
711 /** Host page, the page is from the host inode which the cl_page
715 /** Transient page, the transient cl_page is used to bind a cl_page
716 * to vmpage which is not belonging to the same object of cl_page.
717 * it is used in DirectIO and lockless IO. */
722 #define CP_STATE_BITS 4
723 #define CP_TYPE_BITS 2
724 #define CP_MAX_LAYER 2
727 * Fields are protected by the lock on struct page, except for atomics and
730 * \invariant Data type invariants are in cl_page_invariant(). Basically:
731 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
732 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
733 * cl_page::cp_owner (when set).
736 /** Reference counter. */
738 /** layout_entry + stripe index, composed using lov_comp_index() */
739 unsigned int cp_lov_index;
740 /** page->index of the page within the whole file */
741 pgoff_t cp_page_index;
742 /** An object this page is a part of. Immutable after creation. */
743 struct cl_object *cp_obj;
745 struct page *cp_vmpage;
747 * Assigned if doing direct IO, because in this case cp_vmpage is not
748 * a valid page cache page, hence the inode cannot be inferred from
749 * cp_vmpage->mapping->host.
751 struct inode *cp_inode;
752 /** Linkage of pages within group. Pages must be owned */
753 struct list_head cp_batch;
754 /** array of slices offset. Immutable after creation. */
755 unsigned char cp_layer_offset[CP_MAX_LAYER];
756 /** current slice index */
757 unsigned char cp_layer_count:2;
759 * Page state. This field is const to avoid accidental update, it is
760 * modified only internally within cl_page.c. Protected by a VM lock.
762 enum cl_page_state cp_state:CP_STATE_BITS;
764 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
767 enum cl_page_type cp_type:CP_TYPE_BITS;
768 unsigned cp_defer_uptodate:1,
771 /* which slab kmem index this memory allocated from */
772 short int cp_kmem_index;
775 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
776 * by sub-io. Protected by a VM lock.
778 struct cl_io *cp_owner;
779 /** List of references to this page, for debugging. */
780 struct lu_ref cp_reference;
781 /** Link to an object, for debugging. */
782 struct lu_ref_link cp_obj_ref;
783 /** Link to a queue, for debugging. */
784 struct lu_ref_link cp_queue_ref;
785 /** Assigned if doing a sync_io */
786 struct cl_sync_io *cp_sync_io;
790 * Per-layer part of cl_page.
792 * \see vvp_page, lov_page, osc_page
794 struct cl_page_slice {
795 struct cl_page *cpl_page;
797 * Object slice corresponding to this page slice. Immutable after
800 struct cl_object *cpl_obj;
801 const struct cl_page_operations *cpl_ops;
805 * Lock mode. For the client extent locks.
817 * Requested transfer type.
826 * Per-layer page operations.
828 * Methods taking an \a io argument are for the activity happening in the
829 * context of given \a io. Page is assumed to be owned by that io, except for
832 * \see vvp_page_ops, lov_page_ops, osc_page_ops
834 struct cl_page_operations {
836 * cl_page<->struct page methods. Only one layer in the stack has to
837 * implement these. Current code assumes that this functionality is
838 * provided by the topmost layer, see cl_page_disown0() as an example.
842 * Update file attributes when all we have is this page. Used for tiny
843 * writes to update attributes when we don't have a full cl_io.
845 void (*cpo_page_touch)(const struct lu_env *env,
846 const struct cl_page_slice *slice, size_t to);
852 * Called when page is truncated from the object. Optional.
854 * \see cl_page_discard()
855 * \see vvp_page_discard(), osc_page_discard()
857 void (*cpo_discard)(const struct lu_env *env,
858 const struct cl_page_slice *slice,
861 * Called when page is removed from the cache, and is about to being
862 * destroyed. Optional.
864 * \see cl_page_delete()
865 * \see vvp_page_delete(), osc_page_delete()
867 void (*cpo_delete)(const struct lu_env *env,
868 const struct cl_page_slice *slice);
870 * Optional debugging helper. Prints given page slice.
872 * \see cl_page_print()
874 int (*cpo_print)(const struct lu_env *env,
875 const struct cl_page_slice *slice,
876 void *cookie, lu_printer_t p);
885 * Request type dependent vector of operations.
887 * Transfer operations depend on transfer mode (cl_req_type). To avoid
888 * passing transfer mode to each and every of these methods, and to
889 * avoid branching on request type inside of the methods, separate
890 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
891 * provided. That is, method invocation usually looks like
893 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
897 * Completion handler. This is guaranteed to be eventually
898 * fired after cl_page_prep() or cl_page_make_ready() call.
900 * This method can be called in a non-blocking context. It is
901 * guaranteed however, that the page involved and its object
902 * are pinned in memory (and, hence, calling cl_page_put() is
905 * \see cl_page_completion()
907 void (*cpo_completion)(const struct lu_env *env,
908 const struct cl_page_slice *slice,
912 * Tell transfer engine that only [to, from] part of a page should be
915 * This is used for immediate transfers.
917 * \todo XXX this is not very good interface. It would be much better
918 * if all transfer parameters were supplied as arguments to
919 * cl_io_operations::cio_submit() call, but it is not clear how to do
920 * this for page queues.
922 * \see cl_page_clip()
924 void (*cpo_clip)(const struct lu_env *env,
925 const struct cl_page_slice *slice,
928 * Write out a page by kernel. This is only called by ll_writepage
931 * \see cl_page_flush()
933 int (*cpo_flush)(const struct lu_env *env,
934 const struct cl_page_slice *slice,
940 * Helper macro, dumping detailed information about \a page into a log.
942 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
944 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
945 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
946 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
947 CDEBUG(mask, format , ## __VA_ARGS__); \
952 * Helper macro, dumping shorter information about \a page into a log.
954 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
956 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
957 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
958 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
959 CDEBUG(mask, format , ## __VA_ARGS__); \
963 static inline struct page *cl_page_vmpage(const struct cl_page *page)
965 LASSERT(page->cp_vmpage != NULL);
966 return page->cp_vmpage;
969 static inline pgoff_t cl_page_index(const struct cl_page *cp)
971 return cl_page_vmpage(cp)->index;
975 * Check if a cl_page is in use.
977 * Client cache holds a refcount, this refcount will be dropped when
978 * the page is taken out of cache, see vvp_page_delete().
980 static inline bool __page_in_use(const struct cl_page *page, int refc)
982 return (atomic_read(&page->cp_ref) > refc + 1);
986 * Caller itself holds a refcount of cl_page.
988 #define cl_page_in_use(pg) __page_in_use(pg, 1)
990 * Caller doesn't hold a refcount.
992 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
994 /* references: cl_page, page cache, optional + refcount for caller reference
995 * (always 0 or 1 currently)
997 static inline int vmpage_in_use(struct page *vmpage, int refcount)
999 return (page_count(vmpage) - page_mapcount(vmpage) > 2 + refcount);
1004 /** \addtogroup cl_lock cl_lock
1008 * Extent locking on the client.
1012 * The locking model of the new client code is built around
1016 * data-type representing an extent lock on a regular file. cl_lock is a
1017 * layered object (much like cl_object and cl_page), it consists of a header
1018 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1019 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1021 * Typical cl_lock consists of one layer:
1023 * - lov_lock (lov specific data).
1025 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1026 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1030 * Each sub-lock is associated with a cl_object (representing stripe
1031 * sub-object or the file to which top-level cl_lock is associated to), and is
1032 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1033 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1034 * is different from cl_page, that doesn't fan out (there is usually exactly
1035 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1036 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1040 * cl_lock is a cacheless data container for the requirements of locks to
1041 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1044 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1045 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1047 * INTERFACE AND USAGE
1049 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1050 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1051 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1052 * consists of multiple sub cl_locks, each sub locks will be enqueued
1053 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1054 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1057 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1058 * method will be called for each layer to release the resource held by this
1059 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1060 * clo_enqueue time, is released.
1062 * LDLM lock can only be canceled if there is no cl_lock using it.
1064 * Overall process of the locking during IO operation is as following:
1066 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1067 * is called on each layer. Responsibility of this method is to add locks,
1068 * needed by a given layer into cl_io.ci_lockset.
1070 * - once locks for all layers were collected, they are sorted to avoid
1071 * dead-locks (cl_io_locks_sort()), and enqueued.
1073 * - when all locks are acquired, IO is performed;
1075 * - locks are released after IO is complete.
1077 * Striping introduces major additional complexity into locking. The
1078 * fundamental problem is that it is generally unsafe to actively use (hold)
1079 * two locks on the different OST servers at the same time, as this introduces
1080 * inter-server dependency and can lead to cascading evictions.
1082 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1083 * that no multi-stripe locks are taken (note that this design abandons POSIX
1084 * read/write semantics). Such pieces ideally can be executed concurrently. At
1085 * the same time, certain types of IO cannot be sub-divived, without
1086 * sacrificing correctness. This includes:
1088 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1091 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1093 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1094 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1095 * has to be held together with the usual lock on [offset, offset + count].
1097 * Interaction with DLM
1099 * In the expected setup, cl_lock is ultimately backed up by a collection of
1100 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1101 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1102 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1103 * description of interaction with DLM.
1109 struct cl_lock_descr {
1110 /** Object this lock is granted for. */
1111 struct cl_object *cld_obj;
1112 /** Index of the first page protected by this lock. */
1114 /** Index of the last page (inclusive) protected by this lock. */
1116 /** Group ID, for group lock */
1119 enum cl_lock_mode cld_mode;
1121 * flags to enqueue lock. A combination of bit-flags from
1122 * enum cl_enq_flags.
1124 __u32 cld_enq_flags;
1127 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1128 #define PDESCR(descr) \
1129 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1130 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1132 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1135 * Layered client lock.
1138 /** List of slices. Immutable after creation. */
1139 struct list_head cll_layers;
1140 /** lock attribute, extent, cl_object, etc. */
1141 struct cl_lock_descr cll_descr;
1145 * Per-layer part of cl_lock
1147 * \see lov_lock, osc_lock
1149 struct cl_lock_slice {
1150 struct cl_lock *cls_lock;
1151 /** Object slice corresponding to this lock slice. Immutable after
1153 struct cl_object *cls_obj;
1154 const struct cl_lock_operations *cls_ops;
1155 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1156 struct list_head cls_linkage;
1161 * \see lov_lock_ops, osc_lock_ops
1163 struct cl_lock_operations {
1166 * Attempts to enqueue the lock. Called top-to-bottom.
1168 * \retval 0 this layer has enqueued the lock successfully
1169 * \retval >0 this layer has enqueued the lock, but need to wait on
1170 * @anchor for resources
1171 * \retval -ve failure
1173 * \see lov_lock_enqueue(), osc_lock_enqueue()
1175 int (*clo_enqueue)(const struct lu_env *env,
1176 const struct cl_lock_slice *slice,
1177 struct cl_io *io, struct cl_sync_io *anchor);
1179 * Cancel a lock, release its DLM lock ref, while does not cancel the
1182 void (*clo_cancel)(const struct lu_env *env,
1183 const struct cl_lock_slice *slice);
1186 * Destructor. Frees resources and the slice.
1188 * \see lov_lock_fini(), osc_lock_fini()
1190 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1192 * Optional debugging helper. Prints given lock slice.
1194 int (*clo_print)(const struct lu_env *env,
1195 void *cookie, lu_printer_t p,
1196 const struct cl_lock_slice *slice);
1199 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1201 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1202 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1203 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1204 CDEBUG(mask, format , ## __VA_ARGS__); \
1208 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1212 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1218 /** \addtogroup cl_page_list cl_page_list
1219 * Page list used to perform collective operations on a group of pages.
1221 * Pages are added to the list one by one. cl_page_list acquires a reference
1222 * for every page in it. Page list is used to perform collective operations on
1225 * - submit pages for an immediate transfer,
1227 * - own pages on behalf of certain io (waiting for each page in turn),
1231 * When list is finalized, it releases references on all pages it still has.
1233 * \todo XXX concurrency control.
1237 struct cl_page_list {
1239 struct list_head pl_pages;
1243 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1244 * contains an incoming page list and an outgoing page list.
1247 struct cl_page_list c2_qin;
1248 struct cl_page_list c2_qout;
1251 /** @} cl_page_list */
1253 /** \addtogroup cl_io cl_io
1258 * cl_io represents a high level I/O activity like
1259 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1262 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1263 * important distinction. We want to minimize number of calls to the allocator
1264 * in the fast path, e.g., in the case of read(2) when everything is cached:
1265 * client already owns the lock over region being read, and data are cached
1266 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1267 * per-layer io state is stored in the session, associated with the io, see
1268 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1269 * by using free-lists, see cl_env_get().
1271 * There is a small predefined number of possible io types, enumerated in enum
1274 * cl_io is a state machine, that can be advanced concurrently by the multiple
1275 * threads. It is up to these threads to control the concurrency and,
1276 * specifically, to detect when io is done, and its state can be safely
1279 * For read/write io overall execution plan is as following:
1281 * (0) initialize io state through all layers;
1283 * (1) loop: prepare chunk of work to do
1285 * (2) call all layers to collect locks they need to process current chunk
1287 * (3) sort all locks to avoid dead-locks, and acquire them
1289 * (4) process the chunk: call per-page methods
1290 * cl_io_operations::cio_prepare_write(),
1291 * cl_io_operations::cio_commit_write() for write)
1297 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1298 * address allocation efficiency issues mentioned above), and returns with the
1299 * special error condition from per-page method when current sub-io has to
1300 * block. This causes io loop to be repeated, and lov switches to the next
1301 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1306 /** read system call */
1308 /** write system call */
1310 /** truncate, utime system calls */
1312 /** get data version */
1315 * page fault handling
1319 * fsync system call handling
1320 * To write out a range of file
1324 * glimpse. An io context to acquire glimpse lock.
1328 * Miscellaneous io. This is used for occasional io activity that
1329 * doesn't fit into other types. Currently this is used for:
1331 * - cancellation of an extent lock. This io exists as a context
1332 * to write dirty pages from under the lock being canceled back
1335 * - VM induced page write-out. An io context for writing page out
1336 * for memory cleansing;
1338 * - grouplock. An io context to acquire group lock.
1340 * CIT_MISC io is used simply as a context in which locks and pages
1341 * are manipulated. Such io has no internal "process", that is,
1342 * cl_io_loop() is never called for it.
1347 * To give advice about access of a file
1351 * SEEK_HOLE/SEEK_DATA handling to search holes or data
1352 * across all file objects
1359 * States of cl_io state machine
1362 /** Not initialized. */
1366 /** IO iteration started. */
1370 /** Actual IO is in progress. */
1372 /** IO for the current iteration finished. */
1374 /** Locks released. */
1376 /** Iteration completed. */
1378 /** cl_io finalized. */
1383 * IO state private for a layer.
1385 * This is usually embedded into layer session data, rather than allocated
1388 * \see vvp_io, lov_io, osc_io
1390 struct cl_io_slice {
1391 struct cl_io *cis_io;
1392 /** corresponding object slice. Immutable after creation. */
1393 struct cl_object *cis_obj;
1394 /** io operations. Immutable after creation. */
1395 const struct cl_io_operations *cis_iop;
1397 * linkage into a list of all slices for a given cl_io, hanging off
1398 * cl_io::ci_layers. Immutable after creation.
1400 struct list_head cis_linkage;
1403 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1406 struct cl_read_ahead {
1407 /* Maximum page index the readahead window will end.
1408 * This is determined DLM lock coverage, RPC and stripe boundary.
1409 * cra_end is included. */
1410 pgoff_t cra_end_idx;
1411 /* optimal RPC size for this read, by pages */
1412 unsigned long cra_rpc_pages;
1413 /* Release callback. If readahead holds resources underneath, this
1414 * function should be called to release it. */
1415 void (*cra_release)(const struct lu_env *env,
1416 struct cl_read_ahead *ra);
1418 /* Callback data for cra_release routine */
1422 /* whether lock is in contention */
1423 bool cra_contention;
1426 static inline void cl_read_ahead_release(const struct lu_env *env,
1427 struct cl_read_ahead *ra)
1429 if (ra->cra_release != NULL)
1430 ra->cra_release(env, ra);
1431 memset(ra, 0, sizeof(*ra));
1436 * Per-layer io operations.
1437 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1439 struct cl_io_operations {
1441 * Vector of io state transition methods for every io type.
1443 * \see cl_page_operations::io
1447 * Prepare io iteration at a given layer.
1449 * Called top-to-bottom at the beginning of each iteration of
1450 * "io loop" (if it makes sense for this type of io). Here
1451 * layer selects what work it will do during this iteration.
1453 * \see cl_io_operations::cio_iter_fini()
1455 int (*cio_iter_init) (const struct lu_env *env,
1456 const struct cl_io_slice *slice);
1458 * Finalize io iteration.
1460 * Called bottom-to-top at the end of each iteration of "io
1461 * loop". Here layers can decide whether IO has to be
1464 * \see cl_io_operations::cio_iter_init()
1466 void (*cio_iter_fini) (const struct lu_env *env,
1467 const struct cl_io_slice *slice);
1469 * Collect locks for the current iteration of io.
1471 * Called top-to-bottom to collect all locks necessary for
1472 * this iteration. This methods shouldn't actually enqueue
1473 * anything, instead it should post a lock through
1474 * cl_io_lock_add(). Once all locks are collected, they are
1475 * sorted and enqueued in the proper order.
1477 int (*cio_lock) (const struct lu_env *env,
1478 const struct cl_io_slice *slice);
1480 * Finalize unlocking.
1482 * Called bottom-to-top to finish layer specific unlocking
1483 * functionality, after generic code released all locks
1484 * acquired by cl_io_operations::cio_lock().
1486 void (*cio_unlock)(const struct lu_env *env,
1487 const struct cl_io_slice *slice);
1489 * Start io iteration.
1491 * Once all locks are acquired, called top-to-bottom to
1492 * commence actual IO. In the current implementation,
1493 * top-level vvp_io_{read,write}_start() does all the work
1494 * synchronously by calling generic_file_*(), so other layers
1495 * are called when everything is done.
1497 int (*cio_start)(const struct lu_env *env,
1498 const struct cl_io_slice *slice);
1500 * Called top-to-bottom at the end of io loop. Here layer
1501 * might wait for an unfinished asynchronous io.
1503 void (*cio_end) (const struct lu_env *env,
1504 const struct cl_io_slice *slice);
1506 * Called bottom-to-top to notify layers that read/write IO
1507 * iteration finished, with \a nob bytes transferred.
1509 void (*cio_advance)(const struct lu_env *env,
1510 const struct cl_io_slice *slice,
1513 * Called once per io, bottom-to-top to release io resources.
1515 void (*cio_fini) (const struct lu_env *env,
1516 const struct cl_io_slice *slice);
1520 * Submit pages from \a queue->c2_qin for IO, and move
1521 * successfully submitted pages into \a queue->c2_qout. Return
1522 * non-zero if failed to submit even the single page. If
1523 * submission failed after some pages were moved into \a
1524 * queue->c2_qout, completion callback with non-zero ioret is
1527 int (*cio_submit)(const struct lu_env *env,
1528 const struct cl_io_slice *slice,
1529 enum cl_req_type crt,
1530 struct cl_2queue *queue);
1532 * Queue async page for write.
1533 * The difference between cio_submit and cio_queue is that
1534 * cio_submit is for urgent request.
1536 int (*cio_commit_async)(const struct lu_env *env,
1537 const struct cl_io_slice *slice,
1538 struct cl_page_list *queue, int from, int to,
1541 * Release active extent.
1543 void (*cio_extent_release)(const struct lu_env *env,
1544 const struct cl_io_slice *slice);
1546 * Decide maximum read ahead extent
1548 * \pre io->ci_type == CIT_READ
1550 int (*cio_read_ahead)(const struct lu_env *env,
1551 const struct cl_io_slice *slice,
1552 pgoff_t start, struct cl_read_ahead *ra);
1555 * Reserve LRU slots before IO.
1557 int (*cio_lru_reserve) (const struct lu_env *env,
1558 const struct cl_io_slice *slice,
1559 loff_t pos, size_t bytes);
1561 * Optional debugging helper. Print given io slice.
1563 int (*cio_print)(const struct lu_env *env, void *cookie,
1564 lu_printer_t p, const struct cl_io_slice *slice);
1568 * Flags to lock enqueue procedure.
1573 * instruct server to not block, if conflicting lock is found. Instead
1574 * -EAGAIN is returned immediately.
1576 CEF_NONBLOCK = 0x00000001,
1578 * Tell lower layers this is a glimpse request, translated to
1579 * LDLM_FL_HAS_INTENT at LDLM layer.
1581 * Also, because glimpse locks never block other locks, we count this
1582 * as automatically compatible with other osc locks.
1583 * (see osc_lock_compatible)
1585 CEF_GLIMPSE = 0x00000002,
1587 * tell the server to instruct (though a flag in the blocking ast) an
1588 * owner of the conflicting lock, that it can drop dirty pages
1589 * protected by this lock, without sending them to the server.
1591 CEF_DISCARD_DATA = 0x00000004,
1593 * tell the sub layers that it must be a `real' lock. This is used for
1594 * mmapped-buffer locks, glimpse locks, manually requested locks
1595 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1598 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1600 CEF_MUST = 0x00000008,
1602 * tell the sub layers that never request a `real' lock. This flag is
1603 * not used currently.
1605 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1606 * conversion policy: ci_lockreq describes generic information of lock
1607 * requirement for this IO, especially for locks which belong to the
1608 * object doing IO; however, lock itself may have precise requirements
1609 * that are described by the enqueue flags.
1611 CEF_NEVER = 0x00000010,
1613 * tell the dlm layer this is a speculative lock request
1614 * speculative lock requests are locks which are not requested as part
1615 * of an I/O operation. Instead, they are requested because we expect
1616 * to use them in the future. They are requested asynchronously at the
1619 * Currently used for asynchronous glimpse locks and manually requested
1620 * locks (LU_LADVISE_LOCKAHEAD).
1622 CEF_SPECULATIVE = 0x00000020,
1624 * enqueue a lock to test DLM lock existence.
1626 CEF_PEEK = 0x00000040,
1628 * Lock match only. Used by group lock in I/O as group lock
1629 * is known to exist.
1631 CEF_LOCK_MATCH = 0x00000080,
1633 * tell the DLM layer to lock only the requested range
1635 CEF_LOCK_NO_EXPAND = 0x00000100,
1637 * mask of enq_flags.
1639 CEF_MASK = 0x000001ff,
1643 * Link between lock and io. Intermediate structure is needed, because the
1644 * same lock can be part of multiple io's simultaneously.
1646 struct cl_io_lock_link {
1647 /** linkage into one of cl_lockset lists. */
1648 struct list_head cill_linkage;
1649 struct cl_lock cill_lock;
1650 /** optional destructor */
1651 void (*cill_fini)(const struct lu_env *env,
1652 struct cl_io_lock_link *link);
1654 #define cill_descr cill_lock.cll_descr
1657 * Lock-set represents a collection of locks, that io needs at a
1658 * time. Generally speaking, client tries to avoid holding multiple locks when
1661 * - holding extent locks over multiple ost's introduces the danger of
1662 * "cascading timeouts";
1664 * - holding multiple locks over the same ost is still dead-lock prone,
1665 * see comment in osc_lock_enqueue(),
1667 * but there are certain situations where this is unavoidable:
1669 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1671 * - truncate has to take [new-size, EOF] lock for correctness;
1673 * - SNS has to take locks across full stripe for correctness;
1675 * - in the case when user level buffer, supplied to {read,write}(file0),
1676 * is a part of a memory mapped lustre file, client has to take a dlm
1677 * locks on file0, and all files that back up the buffer (or a part of
1678 * the buffer, that is being processed in the current chunk, in any
1679 * case, there are situations where at least 2 locks are necessary).
1681 * In such cases we at least try to take locks in the same consistent
1682 * order. To this end, all locks are first collected, then sorted, and then
1686 /** locks to be acquired. */
1687 struct list_head cls_todo;
1688 /** locks acquired. */
1689 struct list_head cls_done;
1693 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1694 * but 'req' is always to be thought as 'request' :-)
1696 enum cl_io_lock_dmd {
1697 /** Always lock data (e.g., O_APPEND). */
1699 /** Layers are free to decide between local and global locking. */
1701 /** Never lock: there is no cache (e.g., liblustre). */
1705 enum cl_fsync_mode {
1706 /** start writeback, do not wait for them to finish */
1708 /** start writeback and wait for them to finish */
1710 /** discard all of dirty pages in a specific file range */
1711 CL_FSYNC_DISCARD = 2,
1712 /** start writeback and make sure they have reached storage before
1713 * return. OST_SYNC RPC must be issued and finished */
1717 struct cl_io_rw_common {
1722 enum cl_setattr_subtype {
1723 /** regular setattr **/
1727 /** fallocate(2) - mode preallocate **/
1728 CL_SETATTR_FALLOCATE
1731 struct cl_io_range {
1737 struct cl_io_pt *cip_next;
1738 struct kiocb cip_iocb;
1739 struct iov_iter cip_iter;
1740 struct file *cip_file;
1741 enum cl_io_type cip_iot;
1742 unsigned int cip_need_restart:1;
1751 * cl_io is shared by all threads participating in this IO (in current
1752 * implementation only one thread advances IO, but parallel IO design and
1753 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1754 * is up to these threads to serialize their activities, including updates to
1755 * mutable cl_io fields.
1758 /** type of this IO. Immutable after creation. */
1759 enum cl_io_type ci_type;
1760 /** current state of cl_io state machine. */
1761 enum cl_io_state ci_state;
1762 /** main object this io is against. Immutable after creation. */
1763 struct cl_object *ci_obj;
1764 /** top level dio_aio */
1765 struct cl_dio_aio *ci_dio_aio;
1767 * Upper layer io, of which this io is a part of. Immutable after
1770 struct cl_io *ci_parent;
1771 /** List of slices. Immutable after creation. */
1772 struct list_head ci_layers;
1773 /** list of locks (to be) acquired by this io. */
1774 struct cl_lockset ci_lockset;
1775 /** lock requirements, this is just a help info for sublayers. */
1776 enum cl_io_lock_dmd ci_lockreq;
1777 /** layout version when this IO occurs */
1778 __u32 ci_layout_version;
1781 struct cl_io_rw_common rd;
1784 struct cl_io_rw_common wr;
1788 struct cl_io_rw_common ci_rw;
1789 struct cl_setattr_io {
1790 struct ost_lvb sa_attr;
1791 unsigned int sa_attr_flags;
1792 unsigned int sa_avalid; /* ATTR_* */
1793 unsigned int sa_xvalid; /* OP_XVALID */
1794 int sa_stripe_index;
1795 struct ost_layout sa_layout;
1796 const struct lu_fid *sa_parent_fid;
1797 /* SETATTR interface is used for regular setattr, */
1798 /* truncate(2) and fallocate(2) subtypes */
1799 enum cl_setattr_subtype sa_subtype;
1800 /* The following are used for fallocate(2) */
1802 loff_t sa_falloc_offset;
1803 loff_t sa_falloc_end;
1804 uid_t sa_falloc_uid;
1805 gid_t sa_falloc_gid;
1806 __u32 sa_falloc_projid;
1808 struct cl_data_version_io {
1809 u64 dv_data_version;
1810 u32 dv_layout_version;
1813 struct cl_fault_io {
1814 /** page index within file. */
1816 /** bytes valid byte on a faulted page. */
1818 /** writable page? for nopage() only */
1820 /** page of an executable? */
1822 /** page_mkwrite() */
1824 /** resulting page */
1825 struct cl_page *ft_page;
1827 struct cl_fsync_io {
1830 /** file system level fid */
1831 struct lu_fid *fi_fid;
1832 enum cl_fsync_mode fi_mode;
1833 /* how many pages were written/discarded */
1834 unsigned int fi_nr_written;
1836 struct cl_ladvise_io {
1839 /** file system level fid */
1840 struct lu_fid *li_fid;
1841 enum lu_ladvise_type li_advice;
1844 struct cl_lseek_io {
1850 time64_t lm_next_rpc_time;
1853 struct cl_2queue ci_queue;
1856 unsigned int ci_continue:1,
1858 * This io has held grouplock, to inform sublayers that
1859 * don't do lockless i/o.
1863 * The whole IO need to be restarted because layout has been changed
1867 * to not refresh layout - the IO issuer knows that the layout won't
1868 * change(page operations, layout change causes all page to be
1869 * discarded), or it doesn't matter if it changes(sync).
1873 * Need MDS intervention to complete a write.
1874 * Write intent is required for the following cases:
1875 * 1. component being written is not initialized, or
1876 * 2. the mirrored files are NOT in WRITE_PENDING state.
1878 ci_need_write_intent:1,
1880 * Check if layout changed after the IO finishes. Mainly for HSM
1881 * requirement. If IO occurs to openning files, it doesn't need to
1882 * verify layout because HSM won't release openning files.
1883 * Right now, only two opertaions need to verify layout: glimpse
1888 * file is released, restore has to to be triggered by vvp layer
1890 ci_restore_needed:1,
1895 /* Tell sublayers not to expand LDLM locks requested for this IO */
1896 ci_lock_no_expand:1,
1898 * Set if non-delay RPC should be used for this IO.
1900 * If this file has multiple mirrors, and if the OSTs of the current
1901 * mirror is inaccessible, non-delay RPC would error out quickly so
1902 * that the upper layer can try to access the next mirror.
1906 * Set if IO is triggered by async workqueue readahead.
1908 ci_async_readahead:1,
1910 * Ignore lockless and do normal locking for this io.
1914 * Set if we've tried all mirrors for this read IO, if it's not set,
1915 * the read IO will check to-be-read OSCs' status, and make fast-switch
1916 * another mirror if some of the OSTs are not healthy.
1918 ci_tried_all_mirrors:1,
1920 * Random read hints, readahead will be disabled.
1924 * Sequential read hints.
1928 * Do parallel (async) submission of DIO RPCs. Note DIO is still sync
1929 * to userspace, only the RPCs are submitted async, then waited for at
1930 * the llite layer before returning.
1934 * Bypass quota check
1936 unsigned ci_noquota:1,
1938 * io_uring direct IO with flags IOCB_NOWAIT.
1942 * How many times the read has retried before this one.
1943 * Set by the top level and consumed by the LOV.
1945 unsigned ci_ndelay_tried;
1947 * Designated mirror index for this I/O.
1949 unsigned ci_designated_mirror;
1951 * Number of pages owned by this IO. For invariant checking.
1953 unsigned ci_owned_nr;
1955 * Range of write intent. Valid if ci_need_write_intent is set.
1957 struct lu_extent ci_write_intent;
1963 * Per-transfer attributes.
1965 struct cl_req_attr {
1966 enum cl_req_type cra_type;
1968 struct cl_page *cra_page;
1969 /** Generic attributes for the server consumption. */
1970 struct obdo *cra_oa;
1972 char cra_jobid[LUSTRE_JOBID_SIZE];
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 += cfs_size_round(size);
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,
2254 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2255 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2256 size_t cl_page_size(const struct cl_object *obj);
2258 void cl_lock_print(const struct lu_env *env, void *cookie,
2259 lu_printer_t printer, const struct cl_lock *lock);
2260 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2261 lu_printer_t printer,
2262 const struct cl_lock_descr *descr);
2266 * Data structure managing a client's cached pages. A count of
2267 * "unstable" pages is maintained, and an LRU of clean pages is
2268 * maintained. "unstable" pages are pages pinned by the ptlrpc
2269 * layer for recovery purposes.
2271 struct cl_client_cache {
2273 * # of client cache refcount
2274 * # of users (OSCs) + 2 (held by llite and lov)
2278 * # of threads are doing shrinking
2280 unsigned int ccc_lru_shrinkers;
2282 * # of LRU entries available
2284 atomic_long_t ccc_lru_left;
2286 * List of entities(OSCs) for this LRU cache
2288 struct list_head ccc_lru;
2290 * Max # of LRU entries
2292 unsigned long ccc_lru_max;
2294 * Lock to protect ccc_lru list
2296 spinlock_t ccc_lru_lock;
2298 * Set if unstable check is enabled
2300 unsigned int ccc_unstable_check:1;
2302 * # of unstable pages for this mount point
2304 atomic_long_t ccc_unstable_nr;
2306 * Waitq for awaiting unstable pages to reach zero.
2307 * Used at umounting time and signaled on BRW commit
2309 wait_queue_head_t ccc_unstable_waitq;
2311 * Serialize max_cache_mb write operation
2313 struct mutex ccc_max_cache_mb_lock;
2316 * cl_cache functions
2318 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2319 void cl_cache_incref(struct cl_client_cache *cache);
2320 void cl_cache_decref(struct cl_client_cache *cache);
2324 /** \defgroup cl_lock cl_lock
2326 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2327 struct cl_lock *lock);
2328 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2329 const struct cl_io *io);
2330 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2331 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2332 const struct lu_device_type *dtype);
2333 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2335 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2336 struct cl_lock *lock, struct cl_sync_io *anchor);
2337 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2341 /** \defgroup cl_io cl_io
2344 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2345 enum cl_io_type iot, struct cl_object *obj);
2346 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2347 enum cl_io_type iot, struct cl_object *obj);
2348 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2349 enum cl_io_type iot, loff_t pos, size_t count);
2350 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2352 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2353 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2354 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2355 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2356 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2357 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2358 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2359 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2360 struct cl_io_lock_link *link);
2361 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2362 struct cl_lock_descr *descr);
2363 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2364 enum cl_req_type iot, struct cl_2queue *queue);
2365 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2366 enum cl_req_type iot, struct cl_2queue *queue,
2368 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2369 struct cl_page_list *queue, int from, int to,
2371 void cl_io_extent_release (const struct lu_env *env, struct cl_io *io);
2372 int cl_io_lru_reserve(const struct lu_env *env, struct cl_io *io,
2373 loff_t pos, size_t bytes);
2374 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2375 pgoff_t start, struct cl_read_ahead *ra);
2376 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2380 * True, iff \a io is an O_APPEND write(2).
2382 static inline int cl_io_is_append(const struct cl_io *io)
2384 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2387 static inline int cl_io_is_sync_write(const struct cl_io *io)
2389 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2392 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2394 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2398 * True, iff \a io is a truncate(2).
2400 static inline int cl_io_is_trunc(const struct cl_io *io)
2402 return io->ci_type == CIT_SETATTR &&
2403 (io->u.ci_setattr.sa_avalid & ATTR_SIZE) &&
2404 (io->u.ci_setattr.sa_subtype != CL_SETATTR_FALLOCATE);
2407 static inline int cl_io_is_fallocate(const struct cl_io *io)
2409 return (io->ci_type == CIT_SETATTR) &&
2410 (io->u.ci_setattr.sa_subtype == CL_SETATTR_FALLOCATE);
2413 struct cl_io *cl_io_top(struct cl_io *io);
2415 void cl_io_print(const struct lu_env *env, void *cookie,
2416 lu_printer_t printer, const struct cl_io *io);
2418 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2420 typeof(foo_io) __foo_io = (foo_io); \
2422 memset(&__foo_io->base, 0, \
2423 sizeof(*__foo_io) - offsetof(typeof(*__foo_io), base)); \
2428 /** \defgroup cl_page_list cl_page_list
2432 * Last page in the page list.
2434 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2436 LASSERT(plist->pl_nr > 0);
2437 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2440 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2442 LASSERT(plist->pl_nr > 0);
2443 return list_first_entry(&plist->pl_pages, struct cl_page, cp_batch);
2447 * Iterate over pages in a page list.
2449 #define cl_page_list_for_each(page, list) \
2450 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2453 * Iterate over pages in a page list, taking possible removals into account.
2455 #define cl_page_list_for_each_safe(page, temp, list) \
2456 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2458 void cl_page_list_init(struct cl_page_list *plist);
2459 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page,
2461 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2462 struct cl_page *page);
2463 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2464 struct cl_page *page);
2465 void cl_page_list_splice(struct cl_page_list *list,
2466 struct cl_page_list *head);
2467 void cl_page_list_del(const struct lu_env *env,
2468 struct cl_page_list *plist, struct cl_page *page);
2469 void cl_page_list_disown(const struct lu_env *env,
2470 struct cl_page_list *plist);
2471 void cl_page_list_assume(const struct lu_env *env,
2472 struct cl_io *io, struct cl_page_list *plist);
2473 void cl_page_list_discard(const struct lu_env *env,
2474 struct cl_io *io, struct cl_page_list *plist);
2475 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2477 void cl_2queue_init(struct cl_2queue *queue);
2478 void cl_2queue_disown(const struct lu_env *env, struct cl_2queue *queue);
2479 void cl_2queue_assume(const struct lu_env *env, struct cl_io *io,
2480 struct cl_2queue *queue);
2481 void cl_2queue_discard(const struct lu_env *env, struct cl_io *io,
2482 struct cl_2queue *queue);
2483 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2484 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2486 /** @} cl_page_list */
2488 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2489 struct cl_req_attr *attr);
2491 /** \defgroup cl_sync_io cl_sync_io
2498 typedef void (cl_sync_io_end_t)(const struct lu_env *, struct cl_sync_io *);
2500 void cl_sync_io_init_notify(struct cl_sync_io *anchor, int nr, void *dio_aio,
2501 cl_sync_io_end_t *end);
2503 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2505 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2507 int cl_sync_io_wait_recycle(const struct lu_env *env, struct cl_sync_io *anchor,
2508 long timeout, int ioret);
2509 struct cl_dio_aio *cl_dio_aio_alloc(struct kiocb *iocb, struct cl_object *obj,
2511 struct cl_sub_dio *cl_sub_dio_alloc(struct cl_dio_aio *ll_aio, bool sync);
2512 void cl_dio_aio_free(const struct lu_env *env, struct cl_dio_aio *aio);
2513 void cl_sub_dio_free(struct cl_sub_dio *sdio);
2514 static inline void cl_sync_io_init(struct cl_sync_io *anchor, int nr)
2516 cl_sync_io_init_notify(anchor, nr, NULL, NULL);
2520 * Anchor for synchronous transfer. This is allocated on a stack by thread
2521 * doing synchronous transfer, and a pointer to this structure is set up in
2522 * every page submitted for transfer. Transfer completion routine updates
2523 * anchor and wakes up waiting thread when transfer is complete.
2526 /** number of pages yet to be transferred. */
2527 atomic_t csi_sync_nr;
2530 /** completion to be signaled when transfer is complete. */
2531 wait_queue_head_t csi_waitq;
2532 /** callback to invoke when this IO is finished */
2533 cl_sync_io_end_t *csi_end_io;
2534 /* private pointer for an associated DIO/AIO */
2538 /** direct IO pages */
2539 struct ll_dio_pages {
2541 * page array to be written. we don't support
2542 * partial pages except the last one.
2544 struct page **ldp_pages;
2545 /** # of pages in the array. */
2547 /* the file offset of the first page. */
2548 loff_t ldp_file_offset;
2551 /* Top level struct used for AIO and DIO */
2553 struct cl_sync_io cda_sync;
2554 struct cl_object *cda_obj;
2555 struct kiocb *cda_iocb;
2557 unsigned cda_no_aio_complete:1,
2561 /* Sub-dio used for splitting DIO (and AIO, because AIO is DIO) according to
2562 * the layout/striping, so we can do parallel submit of DIO RPCs
2565 struct cl_sync_io csd_sync;
2566 struct cl_page_list csd_pages;
2568 struct cl_dio_aio *csd_ll_aio;
2569 struct ll_dio_pages csd_dio_pages;
2570 unsigned csd_creator_free:1;
2572 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW) || \
2573 defined(HAVE_DIRECTIO_2ARGS)
2574 #define HAVE_DIO_ITER 1
2577 void ll_release_user_pages(struct page **pages, int npages);
2579 /** @} cl_sync_io */
2581 /** \defgroup cl_env cl_env
2583 * lu_env handling for a client.
2585 * lu_env is an environment within which lustre code executes. Its major part
2586 * is lu_context---a fast memory allocation mechanism that is used to conserve
2587 * precious kernel stack space. Originally lu_env was designed for a server,
2590 * - there is a (mostly) fixed number of threads, and
2592 * - call chains have no non-lustre portions inserted between lustre code.
2594 * On a client both these assumtpion fails, because every user thread can
2595 * potentially execute lustre code as part of a system call, and lustre calls
2596 * into VFS or MM that call back into lustre.
2598 * To deal with that, cl_env wrapper functions implement the following
2601 * - allocation and destruction of environment is amortized by caching no
2602 * longer used environments instead of destroying them;
2604 * \see lu_env, lu_context, lu_context_key
2607 struct lu_env *cl_env_get(__u16 *refcheck);
2608 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2609 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2610 unsigned cl_env_cache_purge(unsigned nr);
2611 struct lu_env *cl_env_percpu_get(void);
2612 void cl_env_percpu_put(struct lu_env *env);
2619 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2620 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2622 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2623 struct lu_device_type *ldt,
2624 struct lu_device *next);
2627 int cl_global_init(void);
2628 void cl_global_fini(void);
2630 #endif /* _LINUX_CL_OBJECT_H */