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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 #ifndef _LUSTRE_CL_OBJECT_H
33 #define _LUSTRE_CL_OBJECT_H
35 /** \defgroup clio clio
37 * Client objects implement io operations and cache pages.
39 * Examples: lov and osc are implementations of cl interface.
41 * Big Theory Statement.
45 * Client implementation is based on the following data-types:
51 * - cl_lock represents an extent lock on an object.
53 * - cl_io represents high-level i/o activity such as whole read/write
54 * system call, or write-out of pages from under the lock being
55 * canceled. cl_io has sub-ios that can be stopped and resumed
56 * independently, thus achieving high degree of transfer
57 * parallelism. Single cl_io can be advanced forward by
58 * the multiple threads (although in the most usual case of
59 * read/write system call it is associated with the single user
60 * thread, that issued the system call).
64 * - to avoid confusion high-level I/O operation like read or write system
65 * call is referred to as "an io", whereas low-level I/O operation, like
66 * RPC, is referred to as "a transfer"
68 * - "generic code" means generic (not file system specific) code in the
69 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
70 * is not layer specific.
76 * - cl_object_header::coh_page_guard
79 * See the top comment in cl_object.c for the description of overall locking and
80 * reference-counting design.
82 * See comments below for the description of i/o, page, and dlm-locking
89 * super-class definitions.
91 #include <linux/aio.h>
94 #include <libcfs/libcfs.h>
95 #include <lu_object.h>
96 #include <linux/atomic.h>
97 #include <linux/mutex.h>
98 #include <linux/radix-tree.h>
99 #include <linux/spinlock.h>
100 #include <linux/wait.h>
101 #include <linux/pagevec.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 * Known minimal size, in bytes.
147 * This is only valid when at least one DLM lock is held.
150 /** Modification time. Measured in seconds since epoch. */
152 /** Access time. Measured in seconds since epoch. */
154 /** Change time. Measured in seconds since epoch. */
157 * Blocks allocated to this cl_object on the server file system.
159 * \todo XXX An interface for block size is needed.
163 * User identifier for quota purposes.
167 * Group identifier for quota purposes.
171 /* nlink of the directory */
174 /* Project identifier for quota purpose. */
179 * Fields in cl_attr that are being set.
194 * Sub-class of lu_object with methods common for objects on the client
197 * cl_object: represents a regular file system object, both a file and a
198 * stripe. cl_object is based on lu_object: it is identified by a fid,
199 * layered, cached, hashed, and lrued. Important distinction with the server
200 * side, where md_object and dt_object are used, is that cl_object "fans out"
201 * at the lov/sns level: depending on the file layout, single file is
202 * represented as a set of "sub-objects" (stripes). At the implementation
203 * level, struct lov_object contains an array of cl_objects. Each sub-object
204 * is a full-fledged cl_object, having its fid, living in the lru and hash
207 * This leads to the next important difference with the server side: on the
208 * client, it's quite usual to have objects with the different sequence of
209 * layers. For example, typical top-object is composed of the following
215 * whereas its sub-objects are composed of
220 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
221 * track of the object-subobject relationship.
223 * Sub-objects are not cached independently: when top-object is about to
224 * be discarded from the memory, all its sub-objects are torn-down and
227 * \see vvp_object, lov_object, lovsub_object, osc_object
231 struct lu_object co_lu;
232 /** per-object-layer operations */
233 const struct cl_object_operations *co_ops;
234 /** offset of page slice in cl_page buffer */
239 * Description of the client object configuration. This is used for the
240 * creation of a new client object that is identified by a more state than
243 struct cl_object_conf {
245 struct lu_object_conf coc_lu;
248 * Object layout. This is consumed by lov.
250 struct lu_buf coc_layout;
252 * Description of particular stripe location in the
253 * cluster. This is consumed by osc.
255 struct lov_oinfo *coc_oinfo;
258 * VFS inode. This is consumed by vvp.
260 struct inode *coc_inode;
262 * Layout lock handle.
264 struct ldlm_lock *coc_lock;
266 * Operation to handle layout, OBJECT_CONF_XYZ.
272 /** configure layout, set up a new stripe, must be called while
273 * holding layout lock. */
275 /** invalidate the current stripe configuration due to losing
277 OBJECT_CONF_INVALIDATE = 1,
278 /** wait for old layout to go away so that new layout can be
284 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
285 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
289 /** the buffer to return the layout in lov_mds_md format. */
290 struct lu_buf cl_buf;
291 /** size of layout in lov_mds_md format. */
293 /** Layout generation. */
295 /** whether layout is a composite one */
296 bool cl_is_composite;
297 /** Whether layout is a HSM released one */
302 * Operations implemented for each cl object layer.
304 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
306 struct cl_object_operations {
308 * Initialize page slice for this layer. Called top-to-bottom through
309 * every object layer when a new cl_page is instantiated. Layer
310 * keeping private per-page data, or requiring its own page operations
311 * vector should allocate these data here, and attach then to the page
312 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
315 * \retval NULL success.
317 * \retval ERR_PTR(errno) failure code.
319 * \retval valid-pointer pointer to already existing referenced page
320 * to be used instead of newly created.
322 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
323 struct cl_page *page, pgoff_t index);
325 * Initialize lock slice for this layer. Called top-to-bottom through
326 * every object layer when a new cl_lock is instantiated. Layer
327 * keeping private per-lock data, or requiring its own lock operations
328 * vector should allocate these data here, and attach then to the lock
329 * by calling cl_lock_slice_add(). Mandatory.
331 int (*coo_lock_init)(const struct lu_env *env,
332 struct cl_object *obj, struct cl_lock *lock,
333 const struct cl_io *io);
335 * Initialize io state for a given layer.
337 * called top-to-bottom once per io existence to initialize io
338 * state. If layer wants to keep some state for this type of io, it
339 * has to embed struct cl_io_slice in lu_env::le_ses, and register
340 * slice with cl_io_slice_add(). It is guaranteed that all threads
341 * participating in this io share the same session.
343 int (*coo_io_init)(const struct lu_env *env,
344 struct cl_object *obj, struct cl_io *io);
346 * Fill portion of \a attr that this layer controls. This method is
347 * called top-to-bottom through all object layers.
349 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
351 * \return 0: to continue
352 * \return +ve: to stop iterating through layers (but 0 is returned
353 * from enclosing cl_object_attr_get())
354 * \return -ve: to signal error
356 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
357 struct cl_attr *attr);
361 * \a valid is a bitmask composed from enum #cl_attr_valid, and
362 * indicating what attributes are to be set.
364 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
366 * \return the same convention as for
367 * cl_object_operations::coo_attr_get() is used.
369 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
370 const struct cl_attr *attr, unsigned valid);
372 * Update object configuration. Called top-to-bottom to modify object
375 * XXX error conditions and handling.
377 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
378 const struct cl_object_conf *conf);
380 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
381 * object. Layers are supposed to fill parts of \a lvb that will be
382 * shipped to the glimpse originator as a glimpse result.
384 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
385 * \see osc_object_glimpse()
387 int (*coo_glimpse)(const struct lu_env *env,
388 const struct cl_object *obj, struct ost_lvb *lvb);
390 * Object prune method. Called when the layout is going to change on
391 * this object, therefore each layer has to clean up their cache,
392 * mainly pages and locks.
394 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
396 * Object getstripe method.
398 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
399 struct lov_user_md __user *lum, size_t size);
401 * Get FIEMAP mapping from the object.
403 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
404 struct ll_fiemap_info_key *fmkey,
405 struct fiemap *fiemap, size_t *buflen);
407 * Get layout and generation of the object.
409 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
410 struct cl_layout *layout);
412 * Get maximum size of the object.
414 loff_t (*coo_maxbytes)(struct cl_object *obj);
416 * Set request attributes.
418 void (*coo_req_attr_set)(const struct lu_env *env,
419 struct cl_object *obj,
420 struct cl_req_attr *attr);
422 * Flush \a obj data corresponding to \a lock. Used for DoM
423 * locks in llite's cancelling blocking ast callback.
425 int (*coo_object_flush)(const struct lu_env *env,
426 struct cl_object *obj,
427 struct ldlm_lock *lock);
431 * Extended header for client object.
433 struct cl_object_header {
434 /** Standard lu_object_header. cl_object::co_lu::lo_header points
436 struct lu_object_header coh_lu;
439 * Parent object. It is assumed that an object has a well-defined
440 * parent, but not a well-defined child (there may be multiple
441 * sub-objects, for the same top-object). cl_object_header::coh_parent
442 * field allows certain code to be written generically, without
443 * limiting possible cl_object layouts unduly.
445 struct cl_object_header *coh_parent;
447 * Protects consistency between cl_attr of parent object and
448 * attributes of sub-objects, that the former is calculated ("merged")
451 * \todo XXX this can be read/write lock if needed.
453 spinlock_t coh_attr_guard;
455 * Size of cl_page + page slices
457 unsigned short coh_page_bufsize;
459 * Number of objects above this one: 0 for a top-object, 1 for its
462 unsigned char coh_nesting;
466 * Helper macro: iterate over all layers of the object \a obj, assigning every
467 * layer top-to-bottom to \a slice.
469 #define cl_object_for_each(slice, obj) \
470 list_for_each_entry((slice), \
471 &(obj)->co_lu.lo_header->loh_layers,\
475 * Helper macro: iterate over all layers of the object \a obj, assigning every
476 * layer bottom-to-top to \a slice.
478 #define cl_object_for_each_reverse(slice, obj) \
479 list_for_each_entry_reverse((slice), \
480 &(obj)->co_lu.lo_header->loh_layers,\
485 #define CL_PAGE_EOF ((pgoff_t)~0ull)
487 /** \addtogroup cl_page cl_page
491 * Layered client page.
493 * cl_page: represents a portion of a file, cached in the memory. All pages
494 * of the given file are of the same size, and are kept in the radix tree
495 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
496 * of the top-level file object are first class cl_objects, they have their
497 * own radix trees of pages and hence page is implemented as a sequence of
498 * struct cl_pages's, linked into double-linked list through
499 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
500 * corresponding radix tree at the corresponding logical offset.
502 * cl_page is associated with VM page of the hosting environment (struct
503 * page in Linux kernel, for example), struct page. It is assumed, that this
504 * association is implemented by one of cl_page layers (top layer in the
505 * current design) that
507 * - intercepts per-VM-page call-backs made by the environment (e.g.,
510 * - translates state (page flag bits) and locking between lustre and
513 * The association between cl_page and struct page is immutable and
514 * established when cl_page is created.
516 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
517 * this io an exclusive access to this page w.r.t. other io attempts and
518 * various events changing page state (such as transfer completion, or
519 * eviction of the page from the memory). Note, that in general cl_io
520 * cannot be identified with a particular thread, and page ownership is not
521 * exactly equal to the current thread holding a lock on the page. Layer
522 * implementing association between cl_page and struct page has to implement
523 * ownership on top of available synchronization mechanisms.
525 * While lustre client maintains the notion of an page ownership by io,
526 * hosting MM/VM usually has its own page concurrency control
527 * mechanisms. For example, in Linux, page access is synchronized by the
528 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
529 * takes care to acquire and release such locks as necessary around the
530 * calls to the file system methods (->readpage(), ->prepare_write(),
531 * ->commit_write(), etc.). This leads to the situation when there are two
532 * different ways to own a page in the client:
534 * - client code explicitly and voluntary owns the page (cl_page_own());
536 * - VM locks a page and then calls the client, that has "to assume"
537 * the ownership from the VM (cl_page_assume()).
539 * Dual methods to release ownership are cl_page_disown() and
540 * cl_page_unassume().
542 * cl_page is reference counted (cl_page::cp_ref). When reference counter
543 * drops to 0, the page is returned to the cache, unless it is in
544 * cl_page_state::CPS_FREEING state, in which case it is immediately
547 * The general logic guaranteeing the absence of "existential races" for
548 * pages is the following:
550 * - there are fixed known ways for a thread to obtain a new reference
553 * - by doing a lookup in the cl_object radix tree, protected by the
556 * - by starting from VM-locked struct page and following some
557 * hosting environment method (e.g., following ->private pointer in
558 * the case of Linux kernel), see cl_vmpage_page();
560 * - when the page enters cl_page_state::CPS_FREEING state, all these
561 * ways are severed with the proper synchronization
562 * (cl_page_delete());
564 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
567 * - no new references to the page in cl_page_state::CPS_FREEING state
568 * are allowed (checked in cl_page_get()).
570 * Together this guarantees that when last reference to a
571 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
572 * page, as neither references to it can be acquired at that point, nor
575 * cl_page is a state machine. States are enumerated in enum
576 * cl_page_state. Possible state transitions are enumerated in
577 * cl_page_state_set(). State transition process (i.e., actual changing of
578 * cl_page::cp_state field) is protected by the lock on the underlying VM
581 * Linux Kernel implementation.
583 * Binding between cl_page and struct page (which is a typedef for
584 * struct page) is implemented in the vvp layer. cl_page is attached to the
585 * ->private pointer of the struct page, together with the setting of
586 * PG_private bit in page->flags, and acquiring additional reference on the
587 * struct page (much like struct buffer_head, or any similar file system
588 * private data structures).
590 * PG_locked lock is used to implement both ownership and transfer
591 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
592 * states. No additional references are acquired for the duration of the
595 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
596 * write-out is "protected" by the special PG_writeback bit.
600 * States of cl_page. cl_page.c assumes particular order here.
602 * The page state machine is rather crude, as it doesn't recognize finer page
603 * states like "dirty" or "up to date". This is because such states are not
604 * always well defined for the whole stack (see, for example, the
605 * implementation of the read-ahead, that hides page up-to-dateness to track
606 * cache hits accurately). Such sub-states are maintained by the layers that
607 * are interested in them.
611 * Page is in the cache, un-owned. Page leaves cached state in the
614 * - [cl_page_state::CPS_OWNED] io comes across the page and
617 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
618 * req-formation engine decides that it wants to include this page
619 * into an RPC being constructed, and yanks it from the cache;
621 * - [cl_page_state::CPS_FREEING] VM callback is executed to
622 * evict the page form the memory;
624 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
628 * Page is exclusively owned by some cl_io. Page may end up in this
629 * state as a result of
631 * - io creating new page and immediately owning it;
633 * - [cl_page_state::CPS_CACHED] io finding existing cached page
636 * - [cl_page_state::CPS_OWNED] io finding existing owned page
637 * and waiting for owner to release the page;
639 * Page leaves owned state in the following cases:
641 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
642 * the cache, doing nothing;
644 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
647 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
648 * transfer for this page;
650 * - [cl_page_state::CPS_FREEING] io decides to destroy this
651 * page (e.g., as part of truncate or extent lock cancellation).
653 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
657 * Page is being written out, as a part of a transfer. This state is
658 * entered when req-formation logic decided that it wants this page to
659 * be sent through the wire _now_. Specifically, it means that once
660 * this state is achieved, transfer completion handler (with either
661 * success or failure indication) is guaranteed to be executed against
662 * this page independently of any locks and any scheduling decisions
663 * made by the hosting environment (that effectively means that the
664 * page is never put into cl_page_state::CPS_PAGEOUT state "in
665 * advance". This property is mentioned, because it is important when
666 * reasoning about possible dead-locks in the system). The page can
667 * enter this state as a result of
669 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
670 * write-out of this page, or
672 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
673 * that it has enough dirty pages cached to issue a "good"
676 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
677 * is completed---it is moved into cl_page_state::CPS_CACHED state.
679 * Underlying VM page is locked for the duration of transfer.
681 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
685 * Page is being read in, as a part of a transfer. This is quite
686 * similar to the cl_page_state::CPS_PAGEOUT state, except that
687 * read-in is always "immediate"---there is no such thing a sudden
688 * construction of read request from cached, presumably not up to date,
691 * Underlying VM page is locked for the duration of transfer.
693 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
697 * Page is being destroyed. This state is entered when client decides
698 * that page has to be deleted from its host object, as, e.g., a part
701 * Once this state is reached, there is no way to escape it.
703 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
710 /** Host page, the page is from the host inode which the cl_page
714 /** Transient page, the transient cl_page is used to bind a cl_page
715 * to vmpage which is not belonging to the same object of cl_page.
716 * it is used in DirectIO and lockless IO. */
721 #define CP_STATE_BITS 4
722 #define CP_TYPE_BITS 2
723 #define CP_MAX_LAYER 3
726 * Fields are protected by the lock on struct page, except for atomics and
729 * \invariant Data type invariants are in cl_page_invariant(). Basically:
730 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
731 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
732 * cl_page::cp_owner (when set).
735 /** Reference counter. */
737 /** layout_entry + stripe index, composed using lov_comp_index() */
738 unsigned int cp_lov_index;
739 pgoff_t cp_osc_index;
740 /** An object this page is a part of. Immutable after creation. */
741 struct cl_object *cp_obj;
743 struct page *cp_vmpage;
745 * Assigned if doing direct IO, because in this case cp_vmpage is not
746 * a valid page cache page, hence the inode cannot be inferred from
747 * cp_vmpage->mapping->host.
749 struct inode *cp_inode;
750 /** Linkage of pages within group. Pages must be owned */
751 struct list_head cp_batch;
752 /** array of slices offset. Immutable after creation. */
753 unsigned char cp_layer_offset[CP_MAX_LAYER]; /* 24 bits */
754 /** current slice index */
755 unsigned char cp_layer_count:2; /* 26 bits */
757 * Page state. This field is const to avoid accidental update, it is
758 * modified only internally within cl_page.c. Protected by a VM lock.
760 enum cl_page_state cp_state:CP_STATE_BITS; /* 30 bits */
762 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
765 enum cl_page_type cp_type:CP_TYPE_BITS; /* 32 bits */
766 /* which slab kmem index this memory allocated from */
767 short int cp_kmem_index; /* 48 bits */
768 unsigned int cp_unused1:16; /* 64 bits */
771 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
772 * by sub-io. Protected by a VM lock.
774 struct cl_io *cp_owner;
775 /** List of references to this page, for debugging. */
776 struct lu_ref cp_reference;
777 /** Link to an object, for debugging. */
778 struct lu_ref_link cp_obj_ref;
779 /** Link to a queue, for debugging. */
780 struct lu_ref_link cp_queue_ref;
781 /** Assigned if doing a sync_io */
782 struct cl_sync_io *cp_sync_io;
786 * Per-layer part of cl_page.
788 * \see vvp_page, lov_page, osc_page
790 struct cl_page_slice {
791 struct cl_page *cpl_page;
793 * Object slice corresponding to this page slice. Immutable after
796 struct cl_object *cpl_obj;
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
826 * the obvious cases (like cl_page_operations::cpo_own()).
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_disown0() as an example.
838 * Called when \a io acquires this page into the exclusive
839 * ownership. When this method returns, it is guaranteed that the is
840 * not owned by other io, and no transfer is going on against
844 * \see vvp_page_own(), lov_page_own()
846 int (*cpo_own)(const struct lu_env *env,
847 const struct cl_page_slice *slice,
848 struct cl_io *io, int nonblock);
849 /** Called when ownership it yielded. Optional.
851 * \see cl_page_disown()
852 * \see vvp_page_disown()
854 void (*cpo_disown)(const struct lu_env *env,
855 const struct cl_page_slice *slice, struct cl_io *io);
857 * Called for a page that is already "owned" by \a io from VM point of
860 * \see cl_page_assume()
861 * \see vvp_page_assume(), lov_page_assume()
863 void (*cpo_assume)(const struct lu_env *env,
864 const struct cl_page_slice *slice, struct cl_io *io);
865 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
866 * bottom-to-top when IO releases a page without actually unlocking
869 * \see cl_page_unassume()
870 * \see vvp_page_unassume()
872 void (*cpo_unassume)(const struct lu_env *env,
873 const struct cl_page_slice *slice,
876 * Announces whether the page contains valid data or not by \a uptodate.
878 * \see cl_page_export()
879 * \see vvp_page_export()
881 void (*cpo_export)(const struct lu_env *env,
882 const struct cl_page_slice *slice, int uptodate);
884 * Checks whether underlying VM page is locked (in the suitable
885 * sense). Used for assertions.
887 * \retval -EBUSY: page is protected by a lock of a given mode;
888 * \retval -ENODATA: page is not protected by a lock;
889 * \retval 0: this layer cannot decide. (Should never happen.)
891 int (*cpo_is_vmlocked)(const struct lu_env *env,
892 const struct cl_page_slice *slice);
895 * Update file attributes when all we have is this page. Used for tiny
896 * writes to update attributes when we don't have a full cl_io.
898 void (*cpo_page_touch)(const struct lu_env *env,
899 const struct cl_page_slice *slice, size_t to);
905 * Called when page is truncated from the object. Optional.
907 * \see cl_page_discard()
908 * \see vvp_page_discard(), osc_page_discard()
910 void (*cpo_discard)(const struct lu_env *env,
911 const struct cl_page_slice *slice,
914 * Called when page is removed from the cache, and is about to being
915 * destroyed. Optional.
917 * \see cl_page_delete()
918 * \see vvp_page_delete(), osc_page_delete()
920 void (*cpo_delete)(const struct lu_env *env,
921 const struct cl_page_slice *slice);
922 /** Destructor. Frees resources and slice itself. */
923 void (*cpo_fini)(const struct lu_env *env,
924 struct cl_page_slice *slice,
925 struct pagevec *pvec);
927 * Optional debugging helper. Prints given page slice.
929 * \see cl_page_print()
931 int (*cpo_print)(const struct lu_env *env,
932 const struct cl_page_slice *slice,
933 void *cookie, lu_printer_t p);
942 * Request type dependent vector of operations.
944 * Transfer operations depend on transfer mode (cl_req_type). To avoid
945 * passing transfer mode to each and every of these methods, and to
946 * avoid branching on request type inside of the methods, separate
947 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
948 * provided. That is, method invocation usually looks like
950 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
954 * Called when a page is submitted for a transfer as a part of
957 * \return 0 : page is eligible for submission;
958 * \return -EALREADY : skip this page;
959 * \return -ve : error.
961 * \see cl_page_prep()
963 int (*cpo_prep)(const struct lu_env *env,
964 const struct cl_page_slice *slice,
967 * Completion handler. This is guaranteed to be eventually
968 * fired after cl_page_operations::cpo_prep() or
969 * cl_page_operations::cpo_make_ready() call.
971 * This method can be called in a non-blocking context. It is
972 * guaranteed however, that the page involved and its object
973 * are pinned in memory (and, hence, calling cl_page_put() is
976 * \see cl_page_completion()
978 void (*cpo_completion)(const struct lu_env *env,
979 const struct cl_page_slice *slice,
982 * Called when cached page is about to be added to the
983 * ptlrpc request as a part of req formation.
985 * \return 0 : proceed with this page;
986 * \return -EAGAIN : skip this page;
987 * \return -ve : error.
989 * \see cl_page_make_ready()
991 int (*cpo_make_ready)(const struct lu_env *env,
992 const struct cl_page_slice *slice);
995 * Tell transfer engine that only [to, from] part of a page should be
998 * This is used for immediate transfers.
1000 * \todo XXX this is not very good interface. It would be much better
1001 * if all transfer parameters were supplied as arguments to
1002 * cl_io_operations::cio_submit() call, but it is not clear how to do
1003 * this for page queues.
1005 * \see cl_page_clip()
1007 void (*cpo_clip)(const struct lu_env *env,
1008 const struct cl_page_slice *slice,
1011 * Write out a page by kernel. This is only called by ll_writepage
1014 * \see cl_page_flush()
1016 int (*cpo_flush)(const struct lu_env *env,
1017 const struct cl_page_slice *slice,
1023 * Helper macro, dumping detailed information about \a page into a log.
1025 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1027 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1028 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1029 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1030 CDEBUG(mask, format , ## __VA_ARGS__); \
1035 * Helper macro, dumping shorter information about \a page into a log.
1037 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1039 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1040 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1041 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1042 CDEBUG(mask, format , ## __VA_ARGS__); \
1046 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1048 LASSERT(page->cp_vmpage != NULL);
1049 return page->cp_vmpage;
1053 * Check if a cl_page is in use.
1055 * Client cache holds a refcount, this refcount will be dropped when
1056 * the page is taken out of cache, see vvp_page_delete().
1058 static inline bool __page_in_use(const struct cl_page *page, int refc)
1060 return (atomic_read(&page->cp_ref) > refc + 1);
1064 * Caller itself holds a refcount of cl_page.
1066 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1068 * Caller doesn't hold a refcount.
1070 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1074 /** \addtogroup cl_lock cl_lock
1078 * Extent locking on the client.
1082 * The locking model of the new client code is built around
1086 * data-type representing an extent lock on a regular file. cl_lock is a
1087 * layered object (much like cl_object and cl_page), it consists of a header
1088 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1089 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1091 * Typical cl_lock consists of one layer:
1093 * - lov_lock (lov specific data).
1095 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1096 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1100 * Each sub-lock is associated with a cl_object (representing stripe
1101 * sub-object or the file to which top-level cl_lock is associated to), and is
1102 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1103 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1104 * is different from cl_page, that doesn't fan out (there is usually exactly
1105 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1106 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1110 * cl_lock is a cacheless data container for the requirements of locks to
1111 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1114 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1115 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1117 * INTERFACE AND USAGE
1119 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1120 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1121 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1122 * consists of multiple sub cl_locks, each sub locks will be enqueued
1123 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1124 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1127 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1128 * method will be called for each layer to release the resource held by this
1129 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1130 * clo_enqueue time, is released.
1132 * LDLM lock can only be canceled if there is no cl_lock using it.
1134 * Overall process of the locking during IO operation is as following:
1136 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1137 * is called on each layer. Responsibility of this method is to add locks,
1138 * needed by a given layer into cl_io.ci_lockset.
1140 * - once locks for all layers were collected, they are sorted to avoid
1141 * dead-locks (cl_io_locks_sort()), and enqueued.
1143 * - when all locks are acquired, IO is performed;
1145 * - locks are released after IO is complete.
1147 * Striping introduces major additional complexity into locking. The
1148 * fundamental problem is that it is generally unsafe to actively use (hold)
1149 * two locks on the different OST servers at the same time, as this introduces
1150 * inter-server dependency and can lead to cascading evictions.
1152 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1153 * that no multi-stripe locks are taken (note that this design abandons POSIX
1154 * read/write semantics). Such pieces ideally can be executed concurrently. At
1155 * the same time, certain types of IO cannot be sub-divived, without
1156 * sacrificing correctness. This includes:
1158 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1161 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1163 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1164 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1165 * has to be held together with the usual lock on [offset, offset + count].
1167 * Interaction with DLM
1169 * In the expected setup, cl_lock is ultimately backed up by a collection of
1170 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1171 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1172 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1173 * description of interaction with DLM.
1179 struct cl_lock_descr {
1180 /** Object this lock is granted for. */
1181 struct cl_object *cld_obj;
1182 /** Index of the first page protected by this lock. */
1184 /** Index of the last page (inclusive) protected by this lock. */
1186 /** Group ID, for group lock */
1189 enum cl_lock_mode cld_mode;
1191 * flags to enqueue lock. A combination of bit-flags from
1192 * enum cl_enq_flags.
1194 __u32 cld_enq_flags;
1197 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1198 #define PDESCR(descr) \
1199 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1200 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1202 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1205 * Layered client lock.
1208 /** List of slices. Immutable after creation. */
1209 struct list_head cll_layers;
1210 /** lock attribute, extent, cl_object, etc. */
1211 struct cl_lock_descr cll_descr;
1215 * Per-layer part of cl_lock
1217 * \see lov_lock, osc_lock
1219 struct cl_lock_slice {
1220 struct cl_lock *cls_lock;
1221 /** Object slice corresponding to this lock slice. Immutable after
1223 struct cl_object *cls_obj;
1224 const struct cl_lock_operations *cls_ops;
1225 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1226 struct list_head cls_linkage;
1231 * \see lov_lock_ops, osc_lock_ops
1233 struct cl_lock_operations {
1236 * Attempts to enqueue the lock. Called top-to-bottom.
1238 * \retval 0 this layer has enqueued the lock successfully
1239 * \retval >0 this layer has enqueued the lock, but need to wait on
1240 * @anchor for resources
1241 * \retval -ve failure
1243 * \see lov_lock_enqueue(), osc_lock_enqueue()
1245 int (*clo_enqueue)(const struct lu_env *env,
1246 const struct cl_lock_slice *slice,
1247 struct cl_io *io, struct cl_sync_io *anchor);
1249 * Cancel a lock, release its DLM lock ref, while does not cancel the
1252 void (*clo_cancel)(const struct lu_env *env,
1253 const struct cl_lock_slice *slice);
1256 * Destructor. Frees resources and the slice.
1258 * \see lov_lock_fini(), osc_lock_fini()
1260 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1262 * Optional debugging helper. Prints given lock slice.
1264 int (*clo_print)(const struct lu_env *env,
1265 void *cookie, lu_printer_t p,
1266 const struct cl_lock_slice *slice);
1269 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1271 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1272 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1273 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1274 CDEBUG(mask, format , ## __VA_ARGS__); \
1278 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1282 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1288 /** \addtogroup cl_page_list cl_page_list
1289 * Page list used to perform collective operations on a group of pages.
1291 * Pages are added to the list one by one. cl_page_list acquires a reference
1292 * for every page in it. Page list is used to perform collective operations on
1295 * - submit pages for an immediate transfer,
1297 * - own pages on behalf of certain io (waiting for each page in turn),
1301 * When list is finalized, it releases references on all pages it still has.
1303 * \todo XXX concurrency control.
1307 struct cl_page_list {
1309 struct list_head pl_pages;
1313 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1314 * contains an incoming page list and an outgoing page list.
1317 struct cl_page_list c2_qin;
1318 struct cl_page_list c2_qout;
1321 /** @} cl_page_list */
1323 /** \addtogroup cl_io cl_io
1328 * cl_io represents a high level I/O activity like
1329 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1332 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1333 * important distinction. We want to minimize number of calls to the allocator
1334 * in the fast path, e.g., in the case of read(2) when everything is cached:
1335 * client already owns the lock over region being read, and data are cached
1336 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1337 * per-layer io state is stored in the session, associated with the io, see
1338 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1339 * by using free-lists, see cl_env_get().
1341 * There is a small predefined number of possible io types, enumerated in enum
1344 * cl_io is a state machine, that can be advanced concurrently by the multiple
1345 * threads. It is up to these threads to control the concurrency and,
1346 * specifically, to detect when io is done, and its state can be safely
1349 * For read/write io overall execution plan is as following:
1351 * (0) initialize io state through all layers;
1353 * (1) loop: prepare chunk of work to do
1355 * (2) call all layers to collect locks they need to process current chunk
1357 * (3) sort all locks to avoid dead-locks, and acquire them
1359 * (4) process the chunk: call per-page methods
1360 * cl_io_operations::cio_prepare_write(),
1361 * cl_io_operations::cio_commit_write() for write)
1367 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1368 * address allocation efficiency issues mentioned above), and returns with the
1369 * special error condition from per-page method when current sub-io has to
1370 * block. This causes io loop to be repeated, and lov switches to the next
1371 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1376 /** read system call */
1378 /** write system call */
1380 /** truncate, utime system calls */
1382 /** get data version */
1385 * page fault handling
1389 * fsync system call handling
1390 * To write out a range of file
1394 * glimpse. An io context to acquire glimpse lock.
1398 * Miscellaneous io. This is used for occasional io activity that
1399 * doesn't fit into other types. Currently this is used for:
1401 * - cancellation of an extent lock. This io exists as a context
1402 * to write dirty pages from under the lock being canceled back
1405 * - VM induced page write-out. An io context for writing page out
1406 * for memory cleansing;
1408 * - grouplock. An io context to acquire group lock.
1410 * CIT_MISC io is used simply as a context in which locks and pages
1411 * are manipulated. Such io has no internal "process", that is,
1412 * cl_io_loop() is never called for it.
1417 * To give advice about access of a file
1421 * SEEK_HOLE/SEEK_DATA handling to search holes or data
1422 * across all file objects
1429 * States of cl_io state machine
1432 /** Not initialized. */
1436 /** IO iteration started. */
1440 /** Actual IO is in progress. */
1442 /** IO for the current iteration finished. */
1444 /** Locks released. */
1446 /** Iteration completed. */
1448 /** cl_io finalized. */
1453 * IO state private for a layer.
1455 * This is usually embedded into layer session data, rather than allocated
1458 * \see vvp_io, lov_io, osc_io
1460 struct cl_io_slice {
1461 struct cl_io *cis_io;
1462 /** corresponding object slice. Immutable after creation. */
1463 struct cl_object *cis_obj;
1464 /** io operations. Immutable after creation. */
1465 const struct cl_io_operations *cis_iop;
1467 * linkage into a list of all slices for a given cl_io, hanging off
1468 * cl_io::ci_layers. Immutable after creation.
1470 struct list_head cis_linkage;
1473 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1476 struct cl_read_ahead {
1477 /* Maximum page index the readahead window will end.
1478 * This is determined DLM lock coverage, RPC and stripe boundary.
1479 * cra_end is included. */
1480 pgoff_t cra_end_idx;
1481 /* optimal RPC size for this read, by pages */
1482 unsigned long cra_rpc_pages;
1483 /* Release callback. If readahead holds resources underneath, this
1484 * function should be called to release it. */
1485 void (*cra_release)(const struct lu_env *env, void *cbdata);
1486 /* Callback data for cra_release routine */
1488 /* whether lock is in contention */
1489 bool cra_contention;
1492 static inline void cl_read_ahead_release(const struct lu_env *env,
1493 struct cl_read_ahead *ra)
1495 if (ra->cra_release != NULL)
1496 ra->cra_release(env, ra->cra_cbdata);
1497 memset(ra, 0, sizeof(*ra));
1502 * Per-layer io operations.
1503 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1505 struct cl_io_operations {
1507 * Vector of io state transition methods for every io type.
1509 * \see cl_page_operations::io
1513 * Prepare io iteration at a given layer.
1515 * Called top-to-bottom at the beginning of each iteration of
1516 * "io loop" (if it makes sense for this type of io). Here
1517 * layer selects what work it will do during this iteration.
1519 * \see cl_io_operations::cio_iter_fini()
1521 int (*cio_iter_init) (const struct lu_env *env,
1522 const struct cl_io_slice *slice);
1524 * Finalize io iteration.
1526 * Called bottom-to-top at the end of each iteration of "io
1527 * loop". Here layers can decide whether IO has to be
1530 * \see cl_io_operations::cio_iter_init()
1532 void (*cio_iter_fini) (const struct lu_env *env,
1533 const struct cl_io_slice *slice);
1535 * Collect locks for the current iteration of io.
1537 * Called top-to-bottom to collect all locks necessary for
1538 * this iteration. This methods shouldn't actually enqueue
1539 * anything, instead it should post a lock through
1540 * cl_io_lock_add(). Once all locks are collected, they are
1541 * sorted and enqueued in the proper order.
1543 int (*cio_lock) (const struct lu_env *env,
1544 const struct cl_io_slice *slice);
1546 * Finalize unlocking.
1548 * Called bottom-to-top to finish layer specific unlocking
1549 * functionality, after generic code released all locks
1550 * acquired by cl_io_operations::cio_lock().
1552 void (*cio_unlock)(const struct lu_env *env,
1553 const struct cl_io_slice *slice);
1555 * Start io iteration.
1557 * Once all locks are acquired, called top-to-bottom to
1558 * commence actual IO. In the current implementation,
1559 * top-level vvp_io_{read,write}_start() does all the work
1560 * synchronously by calling generic_file_*(), so other layers
1561 * are called when everything is done.
1563 int (*cio_start)(const struct lu_env *env,
1564 const struct cl_io_slice *slice);
1566 * Called top-to-bottom at the end of io loop. Here layer
1567 * might wait for an unfinished asynchronous io.
1569 void (*cio_end) (const struct lu_env *env,
1570 const struct cl_io_slice *slice);
1572 * Called bottom-to-top to notify layers that read/write IO
1573 * iteration finished, with \a nob bytes transferred.
1575 void (*cio_advance)(const struct lu_env *env,
1576 const struct cl_io_slice *slice,
1579 * Called once per io, bottom-to-top to release io resources.
1581 void (*cio_fini) (const struct lu_env *env,
1582 const struct cl_io_slice *slice);
1586 * Submit pages from \a queue->c2_qin for IO, and move
1587 * successfully submitted pages into \a queue->c2_qout. Return
1588 * non-zero if failed to submit even the single page. If
1589 * submission failed after some pages were moved into \a
1590 * queue->c2_qout, completion callback with non-zero ioret is
1593 int (*cio_submit)(const struct lu_env *env,
1594 const struct cl_io_slice *slice,
1595 enum cl_req_type crt,
1596 struct cl_2queue *queue);
1598 * Queue async page for write.
1599 * The difference between cio_submit and cio_queue is that
1600 * cio_submit is for urgent request.
1602 int (*cio_commit_async)(const struct lu_env *env,
1603 const struct cl_io_slice *slice,
1604 struct cl_page_list *queue, int from, int to,
1607 * Decide maximum read ahead extent
1609 * \pre io->ci_type == CIT_READ
1611 int (*cio_read_ahead)(const struct lu_env *env,
1612 const struct cl_io_slice *slice,
1613 pgoff_t start, struct cl_read_ahead *ra);
1615 * Optional debugging helper. Print given io slice.
1617 int (*cio_print)(const struct lu_env *env, void *cookie,
1618 lu_printer_t p, const struct cl_io_slice *slice);
1622 * Flags to lock enqueue procedure.
1627 * instruct server to not block, if conflicting lock is found. Instead
1628 * -EWOULDBLOCK is returned immediately.
1630 CEF_NONBLOCK = 0x00000001,
1632 * Tell lower layers this is a glimpse request, translated to
1633 * LDLM_FL_HAS_INTENT at LDLM layer.
1635 * Also, because glimpse locks never block other locks, we count this
1636 * as automatically compatible with other osc locks.
1637 * (see osc_lock_compatible)
1639 CEF_GLIMPSE = 0x00000002,
1641 * tell the server to instruct (though a flag in the blocking ast) an
1642 * owner of the conflicting lock, that it can drop dirty pages
1643 * protected by this lock, without sending them to the server.
1645 CEF_DISCARD_DATA = 0x00000004,
1647 * tell the sub layers that it must be a `real' lock. This is used for
1648 * mmapped-buffer locks, glimpse locks, manually requested locks
1649 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1652 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1654 CEF_MUST = 0x00000008,
1656 * tell the sub layers that never request a `real' lock. This flag is
1657 * not used currently.
1659 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1660 * conversion policy: ci_lockreq describes generic information of lock
1661 * requirement for this IO, especially for locks which belong to the
1662 * object doing IO; however, lock itself may have precise requirements
1663 * that are described by the enqueue flags.
1665 CEF_NEVER = 0x00000010,
1667 * tell the dlm layer this is a speculative lock request
1668 * speculative lock requests are locks which are not requested as part
1669 * of an I/O operation. Instead, they are requested because we expect
1670 * to use them in the future. They are requested asynchronously at the
1673 * Currently used for asynchronous glimpse locks and manually requested
1674 * locks (LU_LADVISE_LOCKAHEAD).
1676 CEF_SPECULATIVE = 0x00000020,
1678 * enqueue a lock to test DLM lock existence.
1680 CEF_PEEK = 0x00000040,
1682 * Lock match only. Used by group lock in I/O as group lock
1683 * is known to exist.
1685 CEF_LOCK_MATCH = 0x00000080,
1687 * tell the DLM layer to lock only the requested range
1689 CEF_LOCK_NO_EXPAND = 0x00000100,
1691 * mask of enq_flags.
1693 CEF_MASK = 0x000001ff,
1697 * Link between lock and io. Intermediate structure is needed, because the
1698 * same lock can be part of multiple io's simultaneously.
1700 struct cl_io_lock_link {
1701 /** linkage into one of cl_lockset lists. */
1702 struct list_head cill_linkage;
1703 struct cl_lock cill_lock;
1704 /** optional destructor */
1705 void (*cill_fini)(const struct lu_env *env,
1706 struct cl_io_lock_link *link);
1708 #define cill_descr cill_lock.cll_descr
1711 * Lock-set represents a collection of locks, that io needs at a
1712 * time. Generally speaking, client tries to avoid holding multiple locks when
1715 * - holding extent locks over multiple ost's introduces the danger of
1716 * "cascading timeouts";
1718 * - holding multiple locks over the same ost is still dead-lock prone,
1719 * see comment in osc_lock_enqueue(),
1721 * but there are certain situations where this is unavoidable:
1723 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1725 * - truncate has to take [new-size, EOF] lock for correctness;
1727 * - SNS has to take locks across full stripe for correctness;
1729 * - in the case when user level buffer, supplied to {read,write}(file0),
1730 * is a part of a memory mapped lustre file, client has to take a dlm
1731 * locks on file0, and all files that back up the buffer (or a part of
1732 * the buffer, that is being processed in the current chunk, in any
1733 * case, there are situations where at least 2 locks are necessary).
1735 * In such cases we at least try to take locks in the same consistent
1736 * order. To this end, all locks are first collected, then sorted, and then
1740 /** locks to be acquired. */
1741 struct list_head cls_todo;
1742 /** locks acquired. */
1743 struct list_head cls_done;
1747 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1748 * but 'req' is always to be thought as 'request' :-)
1750 enum cl_io_lock_dmd {
1751 /** Always lock data (e.g., O_APPEND). */
1753 /** Layers are free to decide between local and global locking. */
1755 /** Never lock: there is no cache (e.g., liblustre). */
1759 enum cl_fsync_mode {
1760 /** start writeback, do not wait for them to finish */
1762 /** start writeback and wait for them to finish */
1764 /** discard all of dirty pages in a specific file range */
1765 CL_FSYNC_DISCARD = 2,
1766 /** start writeback and make sure they have reached storage before
1767 * return. OST_SYNC RPC must be issued and finished */
1771 struct cl_io_rw_common {
1776 enum cl_setattr_subtype {
1777 /** regular setattr **/
1781 /** fallocate(2) - mode preallocate **/
1782 CL_SETATTR_FALLOCATE
1785 struct cl_io_range {
1791 struct cl_io_pt *cip_next;
1792 struct kiocb cip_iocb;
1793 struct iov_iter cip_iter;
1794 struct file *cip_file;
1795 enum cl_io_type cip_iot;
1796 unsigned int cip_need_restart:1;
1805 * cl_io is shared by all threads participating in this IO (in current
1806 * implementation only one thread advances IO, but parallel IO design and
1807 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1808 * is up to these threads to serialize their activities, including updates to
1809 * mutable cl_io fields.
1812 /** type of this IO. Immutable after creation. */
1813 enum cl_io_type ci_type;
1814 /** current state of cl_io state machine. */
1815 enum cl_io_state ci_state;
1816 /** main object this io is against. Immutable after creation. */
1817 struct cl_object *ci_obj;
1818 /** one AIO request might be split in cl_io_loop */
1819 struct cl_dio_aio *ci_aio;
1821 * Upper layer io, of which this io is a part of. Immutable after
1824 struct cl_io *ci_parent;
1825 /** List of slices. Immutable after creation. */
1826 struct list_head ci_layers;
1827 /** list of locks (to be) acquired by this io. */
1828 struct cl_lockset ci_lockset;
1829 /** lock requirements, this is just a help info for sublayers. */
1830 enum cl_io_lock_dmd ci_lockreq;
1831 /** layout version when this IO occurs */
1832 __u32 ci_layout_version;
1835 struct cl_io_rw_common rd;
1838 struct cl_io_rw_common wr;
1842 struct cl_io_rw_common ci_rw;
1843 struct cl_setattr_io {
1844 struct ost_lvb sa_attr;
1845 unsigned int sa_attr_flags;
1846 unsigned int sa_avalid; /* ATTR_* */
1847 unsigned int sa_xvalid; /* OP_XVALID */
1848 int sa_stripe_index;
1849 struct ost_layout sa_layout;
1850 const struct lu_fid *sa_parent_fid;
1851 /* SETATTR interface is used for regular setattr, */
1852 /* truncate(2) and fallocate(2) subtypes */
1853 enum cl_setattr_subtype sa_subtype;
1854 /* The following are used for fallocate(2) */
1856 loff_t sa_falloc_offset;
1857 loff_t sa_falloc_len;
1858 loff_t sa_falloc_end;
1860 struct cl_data_version_io {
1861 u64 dv_data_version;
1862 u32 dv_layout_version;
1865 struct cl_fault_io {
1866 /** page index within file. */
1868 /** bytes valid byte on a faulted page. */
1870 /** writable page? for nopage() only */
1872 /** page of an executable? */
1874 /** page_mkwrite() */
1876 /** resulting page */
1877 struct cl_page *ft_page;
1879 struct cl_fsync_io {
1882 /** file system level fid */
1883 struct lu_fid *fi_fid;
1884 enum cl_fsync_mode fi_mode;
1885 /* how many pages were written/discarded */
1886 unsigned int fi_nr_written;
1888 struct cl_ladvise_io {
1891 /** file system level fid */
1892 struct lu_fid *li_fid;
1893 enum lu_ladvise_type li_advice;
1896 struct cl_lseek_io {
1902 struct cl_2queue ci_queue;
1905 unsigned int ci_continue:1,
1907 * This io has held grouplock, to inform sublayers that
1908 * don't do lockless i/o.
1912 * The whole IO need to be restarted because layout has been changed
1916 * to not refresh layout - the IO issuer knows that the layout won't
1917 * change(page operations, layout change causes all page to be
1918 * discarded), or it doesn't matter if it changes(sync).
1922 * Need MDS intervention to complete a write.
1923 * Write intent is required for the following cases:
1924 * 1. component being written is not initialized, or
1925 * 2. the mirrored files are NOT in WRITE_PENDING state.
1927 ci_need_write_intent:1,
1929 * Check if layout changed after the IO finishes. Mainly for HSM
1930 * requirement. If IO occurs to openning files, it doesn't need to
1931 * verify layout because HSM won't release openning files.
1932 * Right now, only two opertaions need to verify layout: glimpse
1937 * file is released, restore has to to be triggered by vvp layer
1939 ci_restore_needed:1,
1944 /* Tell sublayers not to expand LDLM locks requested for this IO */
1945 ci_lock_no_expand:1,
1947 * Set if non-delay RPC should be used for this IO.
1949 * If this file has multiple mirrors, and if the OSTs of the current
1950 * mirror is inaccessible, non-delay RPC would error out quickly so
1951 * that the upper layer can try to access the next mirror.
1955 * Set if IO is triggered by async workqueue readahead.
1957 ci_async_readahead:1,
1959 * Ignore lockless and do normal locking for this io.
1963 * Set if we've tried all mirrors for this read IO, if it's not set,
1964 * the read IO will check to-be-read OSCs' status, and make fast-switch
1965 * another mirror if some of the OSTs are not healthy.
1967 ci_tried_all_mirrors:1,
1969 * Random read hints, readahead will be disabled.
1973 * Sequential read hints.
1977 * Bypass quota check
1979 unsigned ci_noquota:1;
1981 * How many times the read has retried before this one.
1982 * Set by the top level and consumed by the LOV.
1984 unsigned ci_ndelay_tried;
1986 * Designated mirror index for this I/O.
1988 unsigned ci_designated_mirror;
1990 * Number of pages owned by this IO. For invariant checking.
1992 unsigned ci_owned_nr;
1994 * Range of write intent. Valid if ci_need_write_intent is set.
1996 struct lu_extent ci_write_intent;
2002 * Per-transfer attributes.
2004 struct cl_req_attr {
2005 enum cl_req_type cra_type;
2007 struct cl_page *cra_page;
2008 /** Generic attributes for the server consumption. */
2009 struct obdo *cra_oa;
2011 char cra_jobid[LUSTRE_JOBID_SIZE];
2014 enum cache_stats_item {
2015 /** how many cache lookups were performed */
2017 /** how many times cache lookup resulted in a hit */
2019 /** how many entities are in the cache right now */
2021 /** how many entities in the cache are actively used (and cannot be
2022 * evicted) right now */
2024 /** how many entities were created at all */
2029 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2032 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2034 struct cache_stats {
2035 const char *cs_name;
2036 atomic_t cs_stats[CS_NR];
2039 /** These are not exported so far */
2040 void cache_stats_init (struct cache_stats *cs, const char *name);
2043 * Client-side site. This represents particular client stack. "Global"
2044 * variables should (directly or indirectly) be added here to allow multiple
2045 * clients to co-exist in the single address space.
2048 struct lu_site cs_lu;
2050 * Statistical counters. Atomics do not scale, something better like
2051 * per-cpu counters is needed.
2053 * These are exported as /proc/fs/lustre/llite/.../site
2055 * When interpreting keep in mind that both sub-locks (and sub-pages)
2056 * and top-locks (and top-pages) are accounted here.
2058 struct cache_stats cs_pages;
2059 atomic_t cs_pages_state[CPS_NR];
2062 int cl_site_init(struct cl_site *s, struct cl_device *top);
2063 void cl_site_fini(struct cl_site *s);
2064 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2067 * Output client site statistical counters into a buffer. Suitable for
2068 * ll_rd_*()-style functions.
2070 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2075 * Type conversion and accessory functions.
2079 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2081 return container_of(site, struct cl_site, cs_lu);
2084 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2086 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2087 return container_of_safe(d, struct cl_device, cd_lu_dev);
2090 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2092 return &d->cd_lu_dev;
2095 static inline struct cl_object *lu2cl(const struct lu_object *o)
2097 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2098 return container_of_safe(o, struct cl_object, co_lu);
2101 static inline const struct cl_object_conf *
2102 lu2cl_conf(const struct lu_object_conf *conf)
2104 return container_of_safe(conf, struct cl_object_conf, coc_lu);
2107 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2109 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2112 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2114 return container_of_safe(h, struct cl_object_header, coh_lu);
2117 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2119 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2123 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2125 return luh2coh(obj->co_lu.lo_header);
2128 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2130 return lu_device_init(&d->cd_lu_dev, t);
2133 static inline void cl_device_fini(struct cl_device *d)
2135 lu_device_fini(&d->cd_lu_dev);
2138 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2139 struct cl_object *obj,
2140 const struct cl_page_operations *ops);
2141 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2142 struct cl_object *obj,
2143 const struct cl_lock_operations *ops);
2144 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2145 struct cl_object *obj, const struct cl_io_operations *ops);
2148 /** \defgroup cl_object cl_object
2150 struct cl_object *cl_object_top (struct cl_object *o);
2151 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2152 const struct lu_fid *fid,
2153 const struct cl_object_conf *c);
2155 int cl_object_header_init(struct cl_object_header *h);
2156 void cl_object_header_fini(struct cl_object_header *h);
2157 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2158 void cl_object_get (struct cl_object *o);
2159 void cl_object_attr_lock (struct cl_object *o);
2160 void cl_object_attr_unlock(struct cl_object *o);
2161 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2162 struct cl_attr *attr);
2163 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2164 const struct cl_attr *attr, unsigned valid);
2165 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2166 struct ost_lvb *lvb);
2167 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2168 const struct cl_object_conf *conf);
2169 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2170 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2171 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2172 struct lov_user_md __user *lum, size_t size);
2173 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2174 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2176 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2177 struct cl_layout *cl);
2178 loff_t cl_object_maxbytes(struct cl_object *obj);
2179 int cl_object_flush(const struct lu_env *env, struct cl_object *obj,
2180 struct ldlm_lock *lock);
2184 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2186 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2188 return cl_object_header(o0) == cl_object_header(o1);
2191 static inline void cl_object_page_init(struct cl_object *clob, int size)
2193 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2194 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2195 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2198 static inline void *cl_object_page_slice(struct cl_object *clob,
2199 struct cl_page *page)
2201 return (void *)((char *)page + clob->co_slice_off);
2205 * Return refcount of cl_object.
2207 static inline int cl_object_refc(struct cl_object *clob)
2209 struct lu_object_header *header = clob->co_lu.lo_header;
2210 return atomic_read(&header->loh_ref);
2215 /** \defgroup cl_page cl_page
2217 struct cl_page *cl_page_find (const struct lu_env *env,
2218 struct cl_object *obj,
2219 pgoff_t idx, struct page *vmpage,
2220 enum cl_page_type type);
2221 struct cl_page *cl_page_alloc (const struct lu_env *env,
2222 struct cl_object *o, pgoff_t ind,
2223 struct page *vmpage,
2224 enum cl_page_type type);
2225 void cl_page_get (struct cl_page *page);
2226 void cl_page_put (const struct lu_env *env,
2227 struct cl_page *page);
2228 void cl_pagevec_put (const struct lu_env *env,
2229 struct cl_page *page,
2230 struct pagevec *pvec);
2231 void cl_page_print (const struct lu_env *env, void *cookie,
2232 lu_printer_t printer,
2233 const struct cl_page *pg);
2234 void cl_page_header_print(const struct lu_env *env, void *cookie,
2235 lu_printer_t printer,
2236 const struct cl_page *pg);
2237 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2238 struct cl_page *cl_page_top (struct cl_page *page);
2240 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2241 const struct lu_device_type *dtype);
2246 * Functions dealing with the ownership of page by io.
2250 int cl_page_own (const struct lu_env *env,
2251 struct cl_io *io, struct cl_page *page);
2252 int cl_page_own_try (const struct lu_env *env,
2253 struct cl_io *io, struct cl_page *page);
2254 void cl_page_assume (const struct lu_env *env,
2255 struct cl_io *io, struct cl_page *page);
2256 void cl_page_unassume (const struct lu_env *env,
2257 struct cl_io *io, struct cl_page *pg);
2258 void cl_page_disown (const struct lu_env *env,
2259 struct cl_io *io, struct cl_page *page);
2260 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2267 * Functions dealing with the preparation of a page for a transfer, and
2268 * tracking transfer state.
2271 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2272 struct cl_page *pg, enum cl_req_type crt);
2273 void cl_page_completion (const struct lu_env *env,
2274 struct cl_page *pg, enum cl_req_type crt, int ioret);
2275 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2276 enum cl_req_type crt);
2277 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2278 struct cl_page *pg, enum cl_req_type crt);
2279 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2281 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2282 struct cl_page *pg);
2288 * \name helper routines
2289 * Functions to discard, delete and export a cl_page.
2292 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2293 struct cl_page *pg);
2294 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2295 int cl_page_is_vmlocked(const struct lu_env *env,
2296 const struct cl_page *pg);
2297 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
2299 void cl_page_export(const struct lu_env *env,
2300 struct cl_page *pg, int uptodate);
2301 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2302 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2303 size_t cl_page_size(const struct cl_object *obj);
2305 void cl_lock_print(const struct lu_env *env, void *cookie,
2306 lu_printer_t printer, const struct cl_lock *lock);
2307 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2308 lu_printer_t printer,
2309 const struct cl_lock_descr *descr);
2313 * Data structure managing a client's cached pages. A count of
2314 * "unstable" pages is maintained, and an LRU of clean pages is
2315 * maintained. "unstable" pages are pages pinned by the ptlrpc
2316 * layer for recovery purposes.
2318 struct cl_client_cache {
2320 * # of client cache refcount
2321 * # of users (OSCs) + 2 (held by llite and lov)
2325 * # of threads are doing shrinking
2327 unsigned int ccc_lru_shrinkers;
2329 * # of LRU entries available
2331 atomic_long_t ccc_lru_left;
2333 * List of entities(OSCs) for this LRU cache
2335 struct list_head ccc_lru;
2337 * Max # of LRU entries
2339 unsigned long ccc_lru_max;
2341 * Lock to protect ccc_lru list
2343 spinlock_t ccc_lru_lock;
2345 * Set if unstable check is enabled
2347 unsigned int ccc_unstable_check:1;
2349 * # of unstable pages for this mount point
2351 atomic_long_t ccc_unstable_nr;
2353 * Waitq for awaiting unstable pages to reach zero.
2354 * Used at umounting time and signaled on BRW commit
2356 wait_queue_head_t ccc_unstable_waitq;
2358 * Serialize max_cache_mb write operation
2360 struct mutex ccc_max_cache_mb_lock;
2363 * cl_cache functions
2365 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2366 void cl_cache_incref(struct cl_client_cache *cache);
2367 void cl_cache_decref(struct cl_client_cache *cache);
2371 /** \defgroup cl_lock cl_lock
2373 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2374 struct cl_lock *lock);
2375 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2376 const struct cl_io *io);
2377 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2378 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2379 const struct lu_device_type *dtype);
2380 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2382 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2383 struct cl_lock *lock, struct cl_sync_io *anchor);
2384 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2388 /** \defgroup cl_io cl_io
2391 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2392 enum cl_io_type iot, struct cl_object *obj);
2393 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2394 enum cl_io_type iot, struct cl_object *obj);
2395 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2396 enum cl_io_type iot, loff_t pos, size_t count);
2397 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2399 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2400 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2401 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2402 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2403 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2404 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2405 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2406 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2407 struct cl_io_lock_link *link);
2408 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2409 struct cl_lock_descr *descr);
2410 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2411 enum cl_req_type iot, struct cl_2queue *queue);
2412 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2413 enum cl_req_type iot, struct cl_2queue *queue,
2415 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2416 struct cl_page_list *queue, int from, int to,
2418 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2419 pgoff_t start, struct cl_read_ahead *ra);
2420 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2424 * True, iff \a io is an O_APPEND write(2).
2426 static inline int cl_io_is_append(const struct cl_io *io)
2428 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2431 static inline int cl_io_is_sync_write(const struct cl_io *io)
2433 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2436 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2438 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2442 * True, iff \a io is a truncate(2).
2444 static inline int cl_io_is_trunc(const struct cl_io *io)
2446 return io->ci_type == CIT_SETATTR &&
2447 (io->u.ci_setattr.sa_avalid & ATTR_SIZE) &&
2448 (io->u.ci_setattr.sa_subtype != CL_SETATTR_FALLOCATE);
2451 static inline int cl_io_is_fallocate(const struct cl_io *io)
2453 return (io->ci_type == CIT_SETATTR) &&
2454 (io->u.ci_setattr.sa_subtype == CL_SETATTR_FALLOCATE);
2457 struct cl_io *cl_io_top(struct cl_io *io);
2459 void cl_io_print(const struct lu_env *env, void *cookie,
2460 lu_printer_t printer, const struct cl_io *io);
2462 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2464 typeof(foo_io) __foo_io = (foo_io); \
2466 memset(&__foo_io->base, 0, \
2467 sizeof(*__foo_io) - offsetof(typeof(*__foo_io), base)); \
2472 /** \defgroup cl_page_list cl_page_list
2476 * Last page in the page list.
2478 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2480 LASSERT(plist->pl_nr > 0);
2481 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2484 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2486 LASSERT(plist->pl_nr > 0);
2487 return list_entry(plist->pl_pages.next, struct cl_page, cp_batch);
2491 * Iterate over pages in a page list.
2493 #define cl_page_list_for_each(page, list) \
2494 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2497 * Iterate over pages in a page list, taking possible removals into account.
2499 #define cl_page_list_for_each_safe(page, temp, list) \
2500 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2502 void cl_page_list_init (struct cl_page_list *plist);
2503 void cl_page_list_add (struct cl_page_list *plist, struct cl_page *page);
2504 void cl_page_list_move (struct cl_page_list *dst, struct cl_page_list *src,
2505 struct cl_page *page);
2506 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2507 struct cl_page *page);
2508 void cl_page_list_splice (struct cl_page_list *list,
2509 struct cl_page_list *head);
2510 void cl_page_list_del (const struct lu_env *env,
2511 struct cl_page_list *plist, struct cl_page *page);
2512 void cl_page_list_disown (const struct lu_env *env,
2513 struct cl_io *io, struct cl_page_list *plist);
2514 void cl_page_list_assume (const struct lu_env *env,
2515 struct cl_io *io, struct cl_page_list *plist);
2516 void cl_page_list_discard(const struct lu_env *env,
2517 struct cl_io *io, struct cl_page_list *plist);
2518 void cl_page_list_fini (const struct lu_env *env, struct cl_page_list *plist);
2520 void cl_2queue_init (struct cl_2queue *queue);
2521 void cl_2queue_add (struct cl_2queue *queue, struct cl_page *page);
2522 void cl_2queue_disown (const struct lu_env *env,
2523 struct cl_io *io, struct cl_2queue *queue);
2524 void cl_2queue_assume (const struct lu_env *env,
2525 struct cl_io *io, struct cl_2queue *queue);
2526 void cl_2queue_discard (const struct lu_env *env,
2527 struct cl_io *io, struct cl_2queue *queue);
2528 void cl_2queue_fini (const struct lu_env *env, struct cl_2queue *queue);
2529 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2531 /** @} cl_page_list */
2533 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2534 struct cl_req_attr *attr);
2536 /** \defgroup cl_sync_io cl_sync_io
2542 typedef void (cl_sync_io_end_t)(const struct lu_env *, struct cl_sync_io *);
2544 void cl_sync_io_init_notify(struct cl_sync_io *anchor, int nr,
2545 struct cl_dio_aio *aio, cl_sync_io_end_t *end);
2547 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2549 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2551 struct cl_dio_aio *cl_aio_alloc(struct kiocb *iocb);
2552 void cl_aio_free(struct cl_dio_aio *aio);
2553 static inline void cl_sync_io_init(struct cl_sync_io *anchor, int nr)
2555 cl_sync_io_init_notify(anchor, nr, NULL, NULL);
2559 * Anchor for synchronous transfer. This is allocated on a stack by thread
2560 * doing synchronous transfer, and a pointer to this structure is set up in
2561 * every page submitted for transfer. Transfer completion routine updates
2562 * anchor and wakes up waiting thread when transfer is complete.
2565 /** number of pages yet to be transferred. */
2566 atomic_t csi_sync_nr;
2569 /** completion to be signaled when transfer is complete. */
2570 wait_queue_head_t csi_waitq;
2571 /** callback to invoke when this IO is finished */
2572 cl_sync_io_end_t *csi_end_io;
2573 /** aio private data */
2574 struct cl_dio_aio *csi_aio;
2577 /** To support Direct AIO */
2579 struct cl_sync_io cda_sync;
2580 struct cl_page_list cda_pages;
2581 struct kiocb *cda_iocb;
2583 unsigned cda_no_aio_complete:1;
2586 /** @} cl_sync_io */
2588 /** \defgroup cl_env cl_env
2590 * lu_env handling for a client.
2592 * lu_env is an environment within which lustre code executes. Its major part
2593 * is lu_context---a fast memory allocation mechanism that is used to conserve
2594 * precious kernel stack space. Originally lu_env was designed for a server,
2597 * - there is a (mostly) fixed number of threads, and
2599 * - call chains have no non-lustre portions inserted between lustre code.
2601 * On a client both these assumtpion fails, because every user thread can
2602 * potentially execute lustre code as part of a system call, and lustre calls
2603 * into VFS or MM that call back into lustre.
2605 * To deal with that, cl_env wrapper functions implement the following
2608 * - allocation and destruction of environment is amortized by caching no
2609 * longer used environments instead of destroying them;
2611 * \see lu_env, lu_context, lu_context_key
2614 struct lu_env *cl_env_get(__u16 *refcheck);
2615 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2616 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2617 unsigned cl_env_cache_purge(unsigned nr);
2618 struct lu_env *cl_env_percpu_get(void);
2619 void cl_env_percpu_put(struct lu_env *env);
2626 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2627 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2629 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2630 struct lu_device_type *ldt,
2631 struct lu_device *next);
2634 int cl_global_init(void);
2635 void cl_global_fini(void);
2637 #endif /* _LINUX_CL_OBJECT_H */