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 * Known minimal size, in bytes.
146 * This is only valid when at least one DLM lock is held.
149 /** Modification time. Measured in seconds since epoch. */
151 /** Access time. Measured in seconds since epoch. */
153 /** Change time. Measured in seconds since epoch. */
156 * Blocks allocated to this cl_object on the server file system.
158 * \todo XXX An interface for block size is needed.
162 * User identifier for quota purposes.
166 * Group identifier for quota purposes.
170 /* nlink of the directory */
173 /* Project identifier for quota purpose. */
178 * Fields in cl_attr that are being set.
193 * Sub-class of lu_object with methods common for objects on the client
196 * cl_object: represents a regular file system object, both a file and a
197 * stripe. cl_object is based on lu_object: it is identified by a fid,
198 * layered, cached, hashed, and lrued. Important distinction with the server
199 * side, where md_object and dt_object are used, is that cl_object "fans out"
200 * at the lov/sns level: depending on the file layout, single file is
201 * represented as a set of "sub-objects" (stripes). At the implementation
202 * level, struct lov_object contains an array of cl_objects. Each sub-object
203 * is a full-fledged cl_object, having its fid, living in the lru and hash
206 * This leads to the next important difference with the server side: on the
207 * client, it's quite usual to have objects with the different sequence of
208 * layers. For example, typical top-object is composed of the following
214 * whereas its sub-objects are composed of
219 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
220 * track of the object-subobject relationship.
222 * Sub-objects are not cached independently: when top-object is about to
223 * be discarded from the memory, all its sub-objects are torn-down and
226 * \see vvp_object, lov_object, lovsub_object, osc_object
230 struct lu_object co_lu;
231 /** per-object-layer operations */
232 const struct cl_object_operations *co_ops;
233 /** offset of page slice in cl_page buffer */
238 * Description of the client object configuration. This is used for the
239 * creation of a new client object that is identified by a more state than
242 struct cl_object_conf {
244 struct lu_object_conf coc_lu;
247 * Object layout. This is consumed by lov.
249 struct lu_buf coc_layout;
251 * Description of particular stripe location in the
252 * cluster. This is consumed by osc.
254 struct lov_oinfo *coc_oinfo;
257 * VFS inode. This is consumed by vvp.
259 struct inode *coc_inode;
261 * Layout lock handle.
263 struct ldlm_lock *coc_lock;
265 * Operation to handle layout, OBJECT_CONF_XYZ.
271 /** configure layout, set up a new stripe, must be called while
272 * holding layout lock. */
274 /** invalidate the current stripe configuration due to losing
276 OBJECT_CONF_INVALIDATE = 1,
277 /** wait for old layout to go away so that new layout can be
283 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
284 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
288 /** the buffer to return the layout in lov_mds_md format. */
289 struct lu_buf cl_buf;
290 /** size of layout in lov_mds_md format. */
292 /** Layout generation. */
294 /** whether layout is a composite one */
295 bool cl_is_composite;
296 /** Whether layout is a HSM released one */
301 * Operations implemented for each cl object layer.
303 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
305 struct cl_object_operations {
307 * Initialize page slice for this layer. Called top-to-bottom through
308 * every object layer when a new cl_page is instantiated. Layer
309 * keeping private per-page data, or requiring its own page operations
310 * vector should allocate these data here, and attach then to the page
311 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
314 * \retval NULL success.
316 * \retval ERR_PTR(errno) failure code.
318 * \retval valid-pointer pointer to already existing referenced page
319 * to be used instead of newly created.
321 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
322 struct cl_page *page, pgoff_t index);
324 * Initialize lock slice for this layer. Called top-to-bottom through
325 * every object layer when a new cl_lock is instantiated. Layer
326 * keeping private per-lock data, or requiring its own lock operations
327 * vector should allocate these data here, and attach then to the lock
328 * by calling cl_lock_slice_add(). Mandatory.
330 int (*coo_lock_init)(const struct lu_env *env,
331 struct cl_object *obj, struct cl_lock *lock,
332 const struct cl_io *io);
334 * Initialize io state for a given layer.
336 * called top-to-bottom once per io existence to initialize io
337 * state. If layer wants to keep some state for this type of io, it
338 * has to embed struct cl_io_slice in lu_env::le_ses, and register
339 * slice with cl_io_slice_add(). It is guaranteed that all threads
340 * participating in this io share the same session.
342 int (*coo_io_init)(const struct lu_env *env,
343 struct cl_object *obj, struct cl_io *io);
345 * Fill portion of \a attr that this layer controls. This method is
346 * called top-to-bottom through all object layers.
348 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
350 * \return 0: to continue
351 * \return +ve: to stop iterating through layers (but 0 is returned
352 * from enclosing cl_object_attr_get())
353 * \return -ve: to signal error
355 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
356 struct cl_attr *attr);
360 * \a valid is a bitmask composed from enum #cl_attr_valid, and
361 * indicating what attributes are to be set.
363 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
365 * \return the same convention as for
366 * cl_object_operations::coo_attr_get() is used.
368 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
369 const struct cl_attr *attr, unsigned valid);
371 * Update object configuration. Called top-to-bottom to modify object
374 * XXX error conditions and handling.
376 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
377 const struct cl_object_conf *conf);
379 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
380 * object. Layers are supposed to fill parts of \a lvb that will be
381 * shipped to the glimpse originator as a glimpse result.
383 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
384 * \see osc_object_glimpse()
386 int (*coo_glimpse)(const struct lu_env *env,
387 const struct cl_object *obj, struct ost_lvb *lvb);
389 * Object prune method. Called when the layout is going to change on
390 * this object, therefore each layer has to clean up their cache,
391 * mainly pages and locks.
393 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
395 * Object getstripe method.
397 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
398 struct lov_user_md __user *lum, size_t size);
400 * Get FIEMAP mapping from the object.
402 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
403 struct ll_fiemap_info_key *fmkey,
404 struct fiemap *fiemap, size_t *buflen);
406 * Get layout and generation of the object.
408 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
409 struct cl_layout *layout);
411 * Get maximum size of the object.
413 loff_t (*coo_maxbytes)(struct cl_object *obj);
415 * Set request attributes.
417 void (*coo_req_attr_set)(const struct lu_env *env,
418 struct cl_object *obj,
419 struct cl_req_attr *attr);
421 * Flush \a obj data corresponding to \a lock. Used for DoM
422 * locks in llite's cancelling blocking ast callback.
424 int (*coo_object_flush)(const struct lu_env *env,
425 struct cl_object *obj,
426 struct ldlm_lock *lock);
430 * Extended header for client object.
432 struct cl_object_header {
433 /** Standard lu_object_header. cl_object::co_lu::lo_header points
435 struct lu_object_header coh_lu;
438 * Parent object. It is assumed that an object has a well-defined
439 * parent, but not a well-defined child (there may be multiple
440 * sub-objects, for the same top-object). cl_object_header::coh_parent
441 * field allows certain code to be written generically, without
442 * limiting possible cl_object layouts unduly.
444 struct cl_object_header *coh_parent;
446 * Protects consistency between cl_attr of parent object and
447 * attributes of sub-objects, that the former is calculated ("merged")
450 * \todo XXX this can be read/write lock if needed.
452 spinlock_t coh_attr_guard;
454 * Size of cl_page + page slices
456 unsigned short coh_page_bufsize;
458 * Number of objects above this one: 0 for a top-object, 1 for its
461 unsigned char coh_nesting;
465 * Helper macro: iterate over all layers of the object \a obj, assigning every
466 * layer top-to-bottom to \a slice.
468 #define cl_object_for_each(slice, obj) \
469 list_for_each_entry((slice), \
470 &(obj)->co_lu.lo_header->loh_layers,\
474 * Helper macro: iterate over all layers of the object \a obj, assigning every
475 * layer bottom-to-top to \a slice.
477 #define cl_object_for_each_reverse(slice, obj) \
478 list_for_each_entry_reverse((slice), \
479 &(obj)->co_lu.lo_header->loh_layers,\
484 #define CL_PAGE_EOF ((pgoff_t)~0ull)
486 /** \addtogroup cl_page cl_page
490 * Layered client page.
492 * cl_page: represents a portion of a file, cached in the memory. All pages
493 * of the given file are of the same size, and are kept in the radix tree
494 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
495 * of the top-level file object are first class cl_objects, they have their
496 * own radix trees of pages and hence page is implemented as a sequence of
497 * struct cl_pages's, linked into double-linked list through
498 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
499 * corresponding radix tree at the corresponding logical offset.
501 * cl_page is associated with VM page of the hosting environment (struct
502 * page in Linux kernel, for example), struct page. It is assumed, that this
503 * association is implemented by one of cl_page layers (top layer in the
504 * current design) that
506 * - intercepts per-VM-page call-backs made by the environment (e.g.,
509 * - translates state (page flag bits) and locking between lustre and
512 * The association between cl_page and struct page is immutable and
513 * established when cl_page is created.
515 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
516 * this io an exclusive access to this page w.r.t. other io attempts and
517 * various events changing page state (such as transfer completion, or
518 * eviction of the page from the memory). Note, that in general cl_io
519 * cannot be identified with a particular thread, and page ownership is not
520 * exactly equal to the current thread holding a lock on the page. Layer
521 * implementing association between cl_page and struct page has to implement
522 * ownership on top of available synchronization mechanisms.
524 * While lustre client maintains the notion of an page ownership by io,
525 * hosting MM/VM usually has its own page concurrency control
526 * mechanisms. For example, in Linux, page access is synchronized by the
527 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
528 * takes care to acquire and release such locks as necessary around the
529 * calls to the file system methods (->readpage(), ->prepare_write(),
530 * ->commit_write(), etc.). This leads to the situation when there are two
531 * different ways to own a page in the client:
533 * - client code explicitly and voluntary owns the page (cl_page_own());
535 * - VM locks a page and then calls the client, that has "to assume"
536 * the ownership from the VM (cl_page_assume()).
538 * Dual methods to release ownership are cl_page_disown() and
539 * cl_page_unassume().
541 * cl_page is reference counted (cl_page::cp_ref). When reference counter
542 * drops to 0, the page is returned to the cache, unless it is in
543 * cl_page_state::CPS_FREEING state, in which case it is immediately
546 * The general logic guaranteeing the absence of "existential races" for
547 * pages is the following:
549 * - there are fixed known ways for a thread to obtain a new reference
552 * - by doing a lookup in the cl_object radix tree, protected by the
555 * - by starting from VM-locked struct page and following some
556 * hosting environment method (e.g., following ->private pointer in
557 * the case of Linux kernel), see cl_vmpage_page();
559 * - when the page enters cl_page_state::CPS_FREEING state, all these
560 * ways are severed with the proper synchronization
561 * (cl_page_delete());
563 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
566 * - no new references to the page in cl_page_state::CPS_FREEING state
567 * are allowed (checked in cl_page_get()).
569 * Together this guarantees that when last reference to a
570 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
571 * page, as neither references to it can be acquired at that point, nor
574 * cl_page is a state machine. States are enumerated in enum
575 * cl_page_state. Possible state transitions are enumerated in
576 * cl_page_state_set(). State transition process (i.e., actual changing of
577 * cl_page::cp_state field) is protected by the lock on the underlying VM
580 * Linux Kernel implementation.
582 * Binding between cl_page and struct page (which is a typedef for
583 * struct page) is implemented in the vvp layer. cl_page is attached to the
584 * ->private pointer of the struct page, together with the setting of
585 * PG_private bit in page->flags, and acquiring additional reference on the
586 * struct page (much like struct buffer_head, or any similar file system
587 * private data structures).
589 * PG_locked lock is used to implement both ownership and transfer
590 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
591 * states. No additional references are acquired for the duration of the
594 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
595 * write-out is "protected" by the special PG_writeback bit.
599 * States of cl_page. cl_page.c assumes particular order here.
601 * The page state machine is rather crude, as it doesn't recognize finer page
602 * states like "dirty" or "up to date". This is because such states are not
603 * always well defined for the whole stack (see, for example, the
604 * implementation of the read-ahead, that hides page up-to-dateness to track
605 * cache hits accurately). Such sub-states are maintained by the layers that
606 * are interested in them.
610 * Page is in the cache, un-owned. Page leaves cached state in the
613 * - [cl_page_state::CPS_OWNED] io comes across the page and
616 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
617 * req-formation engine decides that it wants to include this page
618 * into an RPC being constructed, and yanks it from the cache;
620 * - [cl_page_state::CPS_FREEING] VM callback is executed to
621 * evict the page form the memory;
623 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
627 * Page is exclusively owned by some cl_io. Page may end up in this
628 * state as a result of
630 * - io creating new page and immediately owning it;
632 * - [cl_page_state::CPS_CACHED] io finding existing cached page
635 * - [cl_page_state::CPS_OWNED] io finding existing owned page
636 * and waiting for owner to release the page;
638 * Page leaves owned state in the following cases:
640 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
641 * the cache, doing nothing;
643 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
646 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
647 * transfer for this page;
649 * - [cl_page_state::CPS_FREEING] io decides to destroy this
650 * page (e.g., as part of truncate or extent lock cancellation).
652 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
656 * Page is being written out, as a part of a transfer. This state is
657 * entered when req-formation logic decided that it wants this page to
658 * be sent through the wire _now_. Specifically, it means that once
659 * this state is achieved, transfer completion handler (with either
660 * success or failure indication) is guaranteed to be executed against
661 * this page independently of any locks and any scheduling decisions
662 * made by the hosting environment (that effectively means that the
663 * page is never put into cl_page_state::CPS_PAGEOUT state "in
664 * advance". This property is mentioned, because it is important when
665 * reasoning about possible dead-locks in the system). The page can
666 * enter this state as a result of
668 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
669 * write-out of this page, or
671 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
672 * that it has enough dirty pages cached to issue a "good"
675 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
676 * is completed---it is moved into cl_page_state::CPS_CACHED state.
678 * Underlying VM page is locked for the duration of transfer.
680 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
684 * Page is being read in, as a part of a transfer. This is quite
685 * similar to the cl_page_state::CPS_PAGEOUT state, except that
686 * read-in is always "immediate"---there is no such thing a sudden
687 * construction of read request from cached, presumably not up to date,
690 * Underlying VM page is locked for the duration of transfer.
692 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
696 * Page is being destroyed. This state is entered when client decides
697 * that page has to be deleted from its host object, as, e.g., a part
700 * Once this state is reached, there is no way to escape it.
702 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
709 /** Host page, the page is from the host inode which the cl_page
713 /** Transient page, the transient cl_page is used to bind a cl_page
714 * to vmpage which is not belonging to the same object of cl_page.
715 * it is used in DirectIO and lockless IO. */
720 #define CP_STATE_BITS 4
721 #define CP_TYPE_BITS 2
722 #define CP_MAX_LAYER 2
725 * Fields are protected by the lock on struct page, except for atomics and
728 * \invariant Data type invariants are in cl_page_invariant(). Basically:
729 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
730 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
731 * cl_page::cp_owner (when set).
734 /** Reference counter. */
736 /** layout_entry + stripe index, composed using lov_comp_index() */
737 unsigned int cp_lov_index;
738 /** page->index of the page within the whole file */
739 pgoff_t cp_page_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];
754 /** current slice index */
755 unsigned char cp_layer_count:2;
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;
762 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
765 enum cl_page_type cp_type:CP_TYPE_BITS;
766 /* which slab kmem index this memory allocated from */
767 short int cp_kmem_index;
770 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
771 * by sub-io. Protected by a VM lock.
773 struct cl_io *cp_owner;
774 /** List of references to this page, for debugging. */
775 struct lu_ref cp_reference;
776 /** Link to an object, for debugging. */
777 struct lu_ref_link cp_obj_ref;
778 /** Link to a queue, for debugging. */
779 struct lu_ref_link cp_queue_ref;
780 /** Assigned if doing a sync_io */
781 struct cl_sync_io *cp_sync_io;
785 * Per-layer part of cl_page.
787 * \see vvp_page, lov_page, osc_page
789 struct cl_page_slice {
790 struct cl_page *cpl_page;
792 * Object slice corresponding to this page slice. Immutable after
795 struct cl_object *cpl_obj;
796 const struct cl_page_operations *cpl_ops;
800 * Lock mode. For the client extent locks.
812 * Requested transfer type.
821 * Per-layer page operations.
823 * Methods taking an \a io argument are for the activity happening in the
824 * context of given \a io. Page is assumed to be owned by that io, except for
825 * the obvious cases (like cl_page_operations::cpo_own()).
827 * \see vvp_page_ops, lov_page_ops, osc_page_ops
829 struct cl_page_operations {
831 * cl_page<->struct page methods. Only one layer in the stack has to
832 * implement these. Current code assumes that this functionality is
833 * provided by the topmost layer, see cl_page_disown0() as an example.
837 * Called when \a io acquires this page into the exclusive
838 * ownership. When this method returns, it is guaranteed that the is
839 * not owned by other io, and no transfer is going on against
843 * \see vvp_page_own(), lov_page_own()
845 int (*cpo_own)(const struct lu_env *env,
846 const struct cl_page_slice *slice,
847 struct cl_io *io, int nonblock);
848 /** Called when ownership it yielded. Optional.
850 * \see cl_page_disown()
851 * \see vvp_page_disown()
853 void (*cpo_disown)(const struct lu_env *env,
854 const struct cl_page_slice *slice, struct cl_io *io);
856 * Called for a page that is already "owned" by \a io from VM point of
859 * \see cl_page_assume()
860 * \see vvp_page_assume(), lov_page_assume()
862 void (*cpo_assume)(const struct lu_env *env,
863 const struct cl_page_slice *slice, struct cl_io *io);
864 /** Dual to cl_page_operations::cpo_assume(). Optional. Called
865 * bottom-to-top when IO releases a page without actually unlocking
868 * \see cl_page_unassume()
869 * \see vvp_page_unassume()
871 void (*cpo_unassume)(const struct lu_env *env,
872 const struct cl_page_slice *slice,
875 * Update file attributes when all we have is this page. Used for tiny
876 * writes to update attributes when we don't have a full cl_io.
878 void (*cpo_page_touch)(const struct lu_env *env,
879 const struct cl_page_slice *slice, size_t to);
885 * Called when page is truncated from the object. Optional.
887 * \see cl_page_discard()
888 * \see vvp_page_discard(), osc_page_discard()
890 void (*cpo_discard)(const struct lu_env *env,
891 const struct cl_page_slice *slice,
894 * Called when page is removed from the cache, and is about to being
895 * destroyed. Optional.
897 * \see cl_page_delete()
898 * \see vvp_page_delete(), osc_page_delete()
900 void (*cpo_delete)(const struct lu_env *env,
901 const struct cl_page_slice *slice);
902 /** Destructor. Frees resources and slice itself. */
903 void (*cpo_fini)(const struct lu_env *env,
904 struct cl_page_slice *slice,
905 struct pagevec *pvec);
907 * Optional debugging helper. Prints given page slice.
909 * \see cl_page_print()
911 int (*cpo_print)(const struct lu_env *env,
912 const struct cl_page_slice *slice,
913 void *cookie, lu_printer_t p);
922 * Request type dependent vector of operations.
924 * Transfer operations depend on transfer mode (cl_req_type). To avoid
925 * passing transfer mode to each and every of these methods, and to
926 * avoid branching on request type inside of the methods, separate
927 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
928 * provided. That is, method invocation usually looks like
930 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
934 * Called when a page is submitted for a transfer as a part of
937 * \return 0 : page is eligible for submission;
938 * \return -EALREADY : skip this page;
939 * \return -ve : error.
941 * \see cl_page_prep()
943 int (*cpo_prep)(const struct lu_env *env,
944 const struct cl_page_slice *slice,
947 * Completion handler. This is guaranteed to be eventually
948 * fired after cl_page_operations::cpo_prep() or
949 * cl_page_operations::cpo_make_ready() call.
951 * This method can be called in a non-blocking context. It is
952 * guaranteed however, that the page involved and its object
953 * are pinned in memory (and, hence, calling cl_page_put() is
956 * \see cl_page_completion()
958 void (*cpo_completion)(const struct lu_env *env,
959 const struct cl_page_slice *slice,
962 * Called when cached page is about to be added to the
963 * ptlrpc request as a part of req formation.
965 * \return 0 : proceed with this page;
966 * \return -EAGAIN : skip this page;
967 * \return -ve : error.
969 * \see cl_page_make_ready()
971 int (*cpo_make_ready)(const struct lu_env *env,
972 const struct cl_page_slice *slice);
975 * Tell transfer engine that only [to, from] part of a page should be
978 * This is used for immediate transfers.
980 * \todo XXX this is not very good interface. It would be much better
981 * if all transfer parameters were supplied as arguments to
982 * cl_io_operations::cio_submit() call, but it is not clear how to do
983 * this for page queues.
985 * \see cl_page_clip()
987 void (*cpo_clip)(const struct lu_env *env,
988 const struct cl_page_slice *slice,
991 * Write out a page by kernel. This is only called by ll_writepage
994 * \see cl_page_flush()
996 int (*cpo_flush)(const struct lu_env *env,
997 const struct cl_page_slice *slice,
1003 * Helper macro, dumping detailed information about \a page into a log.
1005 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
1007 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1008 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1009 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
1010 CDEBUG(mask, format , ## __VA_ARGS__); \
1015 * Helper macro, dumping shorter information about \a page into a log.
1017 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
1019 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1020 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1021 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
1022 CDEBUG(mask, format , ## __VA_ARGS__); \
1026 static inline struct page *cl_page_vmpage(const struct cl_page *page)
1028 LASSERT(page->cp_vmpage != NULL);
1029 return page->cp_vmpage;
1033 * Check if a cl_page is in use.
1035 * Client cache holds a refcount, this refcount will be dropped when
1036 * the page is taken out of cache, see vvp_page_delete().
1038 static inline bool __page_in_use(const struct cl_page *page, int refc)
1040 return (atomic_read(&page->cp_ref) > refc + 1);
1044 * Caller itself holds a refcount of cl_page.
1046 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1048 * Caller doesn't hold a refcount.
1050 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1052 /* references: cl_page, page cache, optional + refcount for caller reference
1053 * (always 0 or 1 currently)
1055 static inline int vmpage_in_use(struct page *vmpage, int refcount)
1057 return (page_count(vmpage) - page_mapcount(vmpage) > 2 + refcount);
1062 /** \addtogroup cl_lock cl_lock
1066 * Extent locking on the client.
1070 * The locking model of the new client code is built around
1074 * data-type representing an extent lock on a regular file. cl_lock is a
1075 * layered object (much like cl_object and cl_page), it consists of a header
1076 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1077 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1079 * Typical cl_lock consists of one layer:
1081 * - lov_lock (lov specific data).
1083 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1084 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1088 * Each sub-lock is associated with a cl_object (representing stripe
1089 * sub-object or the file to which top-level cl_lock is associated to), and is
1090 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1091 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1092 * is different from cl_page, that doesn't fan out (there is usually exactly
1093 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1094 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1098 * cl_lock is a cacheless data container for the requirements of locks to
1099 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1102 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1103 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1105 * INTERFACE AND USAGE
1107 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1108 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1109 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1110 * consists of multiple sub cl_locks, each sub locks will be enqueued
1111 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1112 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1115 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1116 * method will be called for each layer to release the resource held by this
1117 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1118 * clo_enqueue time, is released.
1120 * LDLM lock can only be canceled if there is no cl_lock using it.
1122 * Overall process of the locking during IO operation is as following:
1124 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1125 * is called on each layer. Responsibility of this method is to add locks,
1126 * needed by a given layer into cl_io.ci_lockset.
1128 * - once locks for all layers were collected, they are sorted to avoid
1129 * dead-locks (cl_io_locks_sort()), and enqueued.
1131 * - when all locks are acquired, IO is performed;
1133 * - locks are released after IO is complete.
1135 * Striping introduces major additional complexity into locking. The
1136 * fundamental problem is that it is generally unsafe to actively use (hold)
1137 * two locks on the different OST servers at the same time, as this introduces
1138 * inter-server dependency and can lead to cascading evictions.
1140 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1141 * that no multi-stripe locks are taken (note that this design abandons POSIX
1142 * read/write semantics). Such pieces ideally can be executed concurrently. At
1143 * the same time, certain types of IO cannot be sub-divived, without
1144 * sacrificing correctness. This includes:
1146 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1149 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1151 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1152 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1153 * has to be held together with the usual lock on [offset, offset + count].
1155 * Interaction with DLM
1157 * In the expected setup, cl_lock is ultimately backed up by a collection of
1158 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1159 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1160 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1161 * description of interaction with DLM.
1167 struct cl_lock_descr {
1168 /** Object this lock is granted for. */
1169 struct cl_object *cld_obj;
1170 /** Index of the first page protected by this lock. */
1172 /** Index of the last page (inclusive) protected by this lock. */
1174 /** Group ID, for group lock */
1177 enum cl_lock_mode cld_mode;
1179 * flags to enqueue lock. A combination of bit-flags from
1180 * enum cl_enq_flags.
1182 __u32 cld_enq_flags;
1185 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1186 #define PDESCR(descr) \
1187 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1188 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1190 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1193 * Layered client lock.
1196 /** List of slices. Immutable after creation. */
1197 struct list_head cll_layers;
1198 /** lock attribute, extent, cl_object, etc. */
1199 struct cl_lock_descr cll_descr;
1203 * Per-layer part of cl_lock
1205 * \see lov_lock, osc_lock
1207 struct cl_lock_slice {
1208 struct cl_lock *cls_lock;
1209 /** Object slice corresponding to this lock slice. Immutable after
1211 struct cl_object *cls_obj;
1212 const struct cl_lock_operations *cls_ops;
1213 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1214 struct list_head cls_linkage;
1219 * \see lov_lock_ops, osc_lock_ops
1221 struct cl_lock_operations {
1224 * Attempts to enqueue the lock. Called top-to-bottom.
1226 * \retval 0 this layer has enqueued the lock successfully
1227 * \retval >0 this layer has enqueued the lock, but need to wait on
1228 * @anchor for resources
1229 * \retval -ve failure
1231 * \see lov_lock_enqueue(), osc_lock_enqueue()
1233 int (*clo_enqueue)(const struct lu_env *env,
1234 const struct cl_lock_slice *slice,
1235 struct cl_io *io, struct cl_sync_io *anchor);
1237 * Cancel a lock, release its DLM lock ref, while does not cancel the
1240 void (*clo_cancel)(const struct lu_env *env,
1241 const struct cl_lock_slice *slice);
1244 * Destructor. Frees resources and the slice.
1246 * \see lov_lock_fini(), osc_lock_fini()
1248 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1250 * Optional debugging helper. Prints given lock slice.
1252 int (*clo_print)(const struct lu_env *env,
1253 void *cookie, lu_printer_t p,
1254 const struct cl_lock_slice *slice);
1257 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1259 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1260 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1261 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1262 CDEBUG(mask, format , ## __VA_ARGS__); \
1266 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1270 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1276 /** \addtogroup cl_page_list cl_page_list
1277 * Page list used to perform collective operations on a group of pages.
1279 * Pages are added to the list one by one. cl_page_list acquires a reference
1280 * for every page in it. Page list is used to perform collective operations on
1283 * - submit pages for an immediate transfer,
1285 * - own pages on behalf of certain io (waiting for each page in turn),
1289 * When list is finalized, it releases references on all pages it still has.
1291 * \todo XXX concurrency control.
1295 struct cl_page_list {
1297 struct list_head pl_pages;
1301 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1302 * contains an incoming page list and an outgoing page list.
1305 struct cl_page_list c2_qin;
1306 struct cl_page_list c2_qout;
1309 /** @} cl_page_list */
1311 /** \addtogroup cl_io cl_io
1316 * cl_io represents a high level I/O activity like
1317 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1320 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1321 * important distinction. We want to minimize number of calls to the allocator
1322 * in the fast path, e.g., in the case of read(2) when everything is cached:
1323 * client already owns the lock over region being read, and data are cached
1324 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1325 * per-layer io state is stored in the session, associated with the io, see
1326 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1327 * by using free-lists, see cl_env_get().
1329 * There is a small predefined number of possible io types, enumerated in enum
1332 * cl_io is a state machine, that can be advanced concurrently by the multiple
1333 * threads. It is up to these threads to control the concurrency and,
1334 * specifically, to detect when io is done, and its state can be safely
1337 * For read/write io overall execution plan is as following:
1339 * (0) initialize io state through all layers;
1341 * (1) loop: prepare chunk of work to do
1343 * (2) call all layers to collect locks they need to process current chunk
1345 * (3) sort all locks to avoid dead-locks, and acquire them
1347 * (4) process the chunk: call per-page methods
1348 * cl_io_operations::cio_prepare_write(),
1349 * cl_io_operations::cio_commit_write() for write)
1355 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1356 * address allocation efficiency issues mentioned above), and returns with the
1357 * special error condition from per-page method when current sub-io has to
1358 * block. This causes io loop to be repeated, and lov switches to the next
1359 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1364 /** read system call */
1366 /** write system call */
1368 /** truncate, utime system calls */
1370 /** get data version */
1373 * page fault handling
1377 * fsync system call handling
1378 * To write out a range of file
1382 * glimpse. An io context to acquire glimpse lock.
1386 * Miscellaneous io. This is used for occasional io activity that
1387 * doesn't fit into other types. Currently this is used for:
1389 * - cancellation of an extent lock. This io exists as a context
1390 * to write dirty pages from under the lock being canceled back
1393 * - VM induced page write-out. An io context for writing page out
1394 * for memory cleansing;
1396 * - grouplock. An io context to acquire group lock.
1398 * CIT_MISC io is used simply as a context in which locks and pages
1399 * are manipulated. Such io has no internal "process", that is,
1400 * cl_io_loop() is never called for it.
1405 * To give advice about access of a file
1409 * SEEK_HOLE/SEEK_DATA handling to search holes or data
1410 * across all file objects
1417 * States of cl_io state machine
1420 /** Not initialized. */
1424 /** IO iteration started. */
1428 /** Actual IO is in progress. */
1430 /** IO for the current iteration finished. */
1432 /** Locks released. */
1434 /** Iteration completed. */
1436 /** cl_io finalized. */
1441 * IO state private for a layer.
1443 * This is usually embedded into layer session data, rather than allocated
1446 * \see vvp_io, lov_io, osc_io
1448 struct cl_io_slice {
1449 struct cl_io *cis_io;
1450 /** corresponding object slice. Immutable after creation. */
1451 struct cl_object *cis_obj;
1452 /** io operations. Immutable after creation. */
1453 const struct cl_io_operations *cis_iop;
1455 * linkage into a list of all slices for a given cl_io, hanging off
1456 * cl_io::ci_layers. Immutable after creation.
1458 struct list_head cis_linkage;
1461 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1464 struct cl_read_ahead {
1465 /* Maximum page index the readahead window will end.
1466 * This is determined DLM lock coverage, RPC and stripe boundary.
1467 * cra_end is included. */
1468 pgoff_t cra_end_idx;
1469 /* optimal RPC size for this read, by pages */
1470 unsigned long cra_rpc_pages;
1471 /* Release callback. If readahead holds resources underneath, this
1472 * function should be called to release it. */
1473 void (*cra_release)(const struct lu_env *env,
1474 struct cl_read_ahead *ra);
1476 /* Callback data for cra_release routine */
1480 /* whether lock is in contention */
1481 bool cra_contention;
1484 static inline void cl_read_ahead_release(const struct lu_env *env,
1485 struct cl_read_ahead *ra)
1487 if (ra->cra_release != NULL)
1488 ra->cra_release(env, ra);
1489 memset(ra, 0, sizeof(*ra));
1494 * Per-layer io operations.
1495 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1497 struct cl_io_operations {
1499 * Vector of io state transition methods for every io type.
1501 * \see cl_page_operations::io
1505 * Prepare io iteration at a given layer.
1507 * Called top-to-bottom at the beginning of each iteration of
1508 * "io loop" (if it makes sense for this type of io). Here
1509 * layer selects what work it will do during this iteration.
1511 * \see cl_io_operations::cio_iter_fini()
1513 int (*cio_iter_init) (const struct lu_env *env,
1514 const struct cl_io_slice *slice);
1516 * Finalize io iteration.
1518 * Called bottom-to-top at the end of each iteration of "io
1519 * loop". Here layers can decide whether IO has to be
1522 * \see cl_io_operations::cio_iter_init()
1524 void (*cio_iter_fini) (const struct lu_env *env,
1525 const struct cl_io_slice *slice);
1527 * Collect locks for the current iteration of io.
1529 * Called top-to-bottom to collect all locks necessary for
1530 * this iteration. This methods shouldn't actually enqueue
1531 * anything, instead it should post a lock through
1532 * cl_io_lock_add(). Once all locks are collected, they are
1533 * sorted and enqueued in the proper order.
1535 int (*cio_lock) (const struct lu_env *env,
1536 const struct cl_io_slice *slice);
1538 * Finalize unlocking.
1540 * Called bottom-to-top to finish layer specific unlocking
1541 * functionality, after generic code released all locks
1542 * acquired by cl_io_operations::cio_lock().
1544 void (*cio_unlock)(const struct lu_env *env,
1545 const struct cl_io_slice *slice);
1547 * Start io iteration.
1549 * Once all locks are acquired, called top-to-bottom to
1550 * commence actual IO. In the current implementation,
1551 * top-level vvp_io_{read,write}_start() does all the work
1552 * synchronously by calling generic_file_*(), so other layers
1553 * are called when everything is done.
1555 int (*cio_start)(const struct lu_env *env,
1556 const struct cl_io_slice *slice);
1558 * Called top-to-bottom at the end of io loop. Here layer
1559 * might wait for an unfinished asynchronous io.
1561 void (*cio_end) (const struct lu_env *env,
1562 const struct cl_io_slice *slice);
1564 * Called bottom-to-top to notify layers that read/write IO
1565 * iteration finished, with \a nob bytes transferred.
1567 void (*cio_advance)(const struct lu_env *env,
1568 const struct cl_io_slice *slice,
1571 * Called once per io, bottom-to-top to release io resources.
1573 void (*cio_fini) (const struct lu_env *env,
1574 const struct cl_io_slice *slice);
1578 * Submit pages from \a queue->c2_qin for IO, and move
1579 * successfully submitted pages into \a queue->c2_qout. Return
1580 * non-zero if failed to submit even the single page. If
1581 * submission failed after some pages were moved into \a
1582 * queue->c2_qout, completion callback with non-zero ioret is
1585 int (*cio_submit)(const struct lu_env *env,
1586 const struct cl_io_slice *slice,
1587 enum cl_req_type crt,
1588 struct cl_2queue *queue);
1590 * Queue async page for write.
1591 * The difference between cio_submit and cio_queue is that
1592 * cio_submit is for urgent request.
1594 int (*cio_commit_async)(const struct lu_env *env,
1595 const struct cl_io_slice *slice,
1596 struct cl_page_list *queue, int from, int to,
1599 * Release active extent.
1601 void (*cio_extent_release)(const struct lu_env *env,
1602 const struct cl_io_slice *slice);
1604 * Decide maximum read ahead extent
1606 * \pre io->ci_type == CIT_READ
1608 int (*cio_read_ahead)(const struct lu_env *env,
1609 const struct cl_io_slice *slice,
1610 pgoff_t start, struct cl_read_ahead *ra);
1613 * Reserve LRU slots before IO.
1615 int (*cio_lru_reserve) (const struct lu_env *env,
1616 const struct cl_io_slice *slice,
1617 loff_t pos, size_t bytes);
1619 * Optional debugging helper. Print given io slice.
1621 int (*cio_print)(const struct lu_env *env, void *cookie,
1622 lu_printer_t p, const struct cl_io_slice *slice);
1626 * Flags to lock enqueue procedure.
1631 * instruct server to not block, if conflicting lock is found. Instead
1632 * -EAGAIN is returned immediately.
1634 CEF_NONBLOCK = 0x00000001,
1636 * Tell lower layers this is a glimpse request, translated to
1637 * LDLM_FL_HAS_INTENT at LDLM layer.
1639 * Also, because glimpse locks never block other locks, we count this
1640 * as automatically compatible with other osc locks.
1641 * (see osc_lock_compatible)
1643 CEF_GLIMPSE = 0x00000002,
1645 * tell the server to instruct (though a flag in the blocking ast) an
1646 * owner of the conflicting lock, that it can drop dirty pages
1647 * protected by this lock, without sending them to the server.
1649 CEF_DISCARD_DATA = 0x00000004,
1651 * tell the sub layers that it must be a `real' lock. This is used for
1652 * mmapped-buffer locks, glimpse locks, manually requested locks
1653 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1656 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1658 CEF_MUST = 0x00000008,
1660 * tell the sub layers that never request a `real' lock. This flag is
1661 * not used currently.
1663 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1664 * conversion policy: ci_lockreq describes generic information of lock
1665 * requirement for this IO, especially for locks which belong to the
1666 * object doing IO; however, lock itself may have precise requirements
1667 * that are described by the enqueue flags.
1669 CEF_NEVER = 0x00000010,
1671 * tell the dlm layer this is a speculative lock request
1672 * speculative lock requests are locks which are not requested as part
1673 * of an I/O operation. Instead, they are requested because we expect
1674 * to use them in the future. They are requested asynchronously at the
1677 * Currently used for asynchronous glimpse locks and manually requested
1678 * locks (LU_LADVISE_LOCKAHEAD).
1680 CEF_SPECULATIVE = 0x00000020,
1682 * enqueue a lock to test DLM lock existence.
1684 CEF_PEEK = 0x00000040,
1686 * Lock match only. Used by group lock in I/O as group lock
1687 * is known to exist.
1689 CEF_LOCK_MATCH = 0x00000080,
1691 * tell the DLM layer to lock only the requested range
1693 CEF_LOCK_NO_EXPAND = 0x00000100,
1695 * mask of enq_flags.
1697 CEF_MASK = 0x000001ff,
1701 * Link between lock and io. Intermediate structure is needed, because the
1702 * same lock can be part of multiple io's simultaneously.
1704 struct cl_io_lock_link {
1705 /** linkage into one of cl_lockset lists. */
1706 struct list_head cill_linkage;
1707 struct cl_lock cill_lock;
1708 /** optional destructor */
1709 void (*cill_fini)(const struct lu_env *env,
1710 struct cl_io_lock_link *link);
1712 #define cill_descr cill_lock.cll_descr
1715 * Lock-set represents a collection of locks, that io needs at a
1716 * time. Generally speaking, client tries to avoid holding multiple locks when
1719 * - holding extent locks over multiple ost's introduces the danger of
1720 * "cascading timeouts";
1722 * - holding multiple locks over the same ost is still dead-lock prone,
1723 * see comment in osc_lock_enqueue(),
1725 * but there are certain situations where this is unavoidable:
1727 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1729 * - truncate has to take [new-size, EOF] lock for correctness;
1731 * - SNS has to take locks across full stripe for correctness;
1733 * - in the case when user level buffer, supplied to {read,write}(file0),
1734 * is a part of a memory mapped lustre file, client has to take a dlm
1735 * locks on file0, and all files that back up the buffer (or a part of
1736 * the buffer, that is being processed in the current chunk, in any
1737 * case, there are situations where at least 2 locks are necessary).
1739 * In such cases we at least try to take locks in the same consistent
1740 * order. To this end, all locks are first collected, then sorted, and then
1744 /** locks to be acquired. */
1745 struct list_head cls_todo;
1746 /** locks acquired. */
1747 struct list_head cls_done;
1751 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1752 * but 'req' is always to be thought as 'request' :-)
1754 enum cl_io_lock_dmd {
1755 /** Always lock data (e.g., O_APPEND). */
1757 /** Layers are free to decide between local and global locking. */
1759 /** Never lock: there is no cache (e.g., liblustre). */
1763 enum cl_fsync_mode {
1764 /** start writeback, do not wait for them to finish */
1766 /** start writeback and wait for them to finish */
1768 /** discard all of dirty pages in a specific file range */
1769 CL_FSYNC_DISCARD = 2,
1770 /** start writeback and make sure they have reached storage before
1771 * return. OST_SYNC RPC must be issued and finished */
1775 struct cl_io_rw_common {
1780 enum cl_setattr_subtype {
1781 /** regular setattr **/
1785 /** fallocate(2) - mode preallocate **/
1786 CL_SETATTR_FALLOCATE
1789 struct cl_io_range {
1795 struct cl_io_pt *cip_next;
1796 struct kiocb cip_iocb;
1797 struct iov_iter cip_iter;
1798 struct file *cip_file;
1799 enum cl_io_type cip_iot;
1800 unsigned int cip_need_restart:1;
1809 * cl_io is shared by all threads participating in this IO (in current
1810 * implementation only one thread advances IO, but parallel IO design and
1811 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1812 * is up to these threads to serialize their activities, including updates to
1813 * mutable cl_io fields.
1816 /** type of this IO. Immutable after creation. */
1817 enum cl_io_type ci_type;
1818 /** current state of cl_io state machine. */
1819 enum cl_io_state ci_state;
1820 /** main object this io is against. Immutable after creation. */
1821 struct cl_object *ci_obj;
1822 /** one AIO request might be split in cl_io_loop */
1823 struct cl_dio_aio *ci_aio;
1825 * Upper layer io, of which this io is a part of. Immutable after
1828 struct cl_io *ci_parent;
1829 /** List of slices. Immutable after creation. */
1830 struct list_head ci_layers;
1831 /** list of locks (to be) acquired by this io. */
1832 struct cl_lockset ci_lockset;
1833 /** lock requirements, this is just a help info for sublayers. */
1834 enum cl_io_lock_dmd ci_lockreq;
1835 /** layout version when this IO occurs */
1836 __u32 ci_layout_version;
1839 struct cl_io_rw_common rd;
1842 struct cl_io_rw_common wr;
1846 struct cl_io_rw_common ci_rw;
1847 struct cl_setattr_io {
1848 struct ost_lvb sa_attr;
1849 unsigned int sa_attr_flags;
1850 unsigned int sa_avalid; /* ATTR_* */
1851 unsigned int sa_xvalid; /* OP_XVALID */
1852 int sa_stripe_index;
1853 struct ost_layout sa_layout;
1854 const struct lu_fid *sa_parent_fid;
1855 /* SETATTR interface is used for regular setattr, */
1856 /* truncate(2) and fallocate(2) subtypes */
1857 enum cl_setattr_subtype sa_subtype;
1858 /* The following are used for fallocate(2) */
1860 loff_t sa_falloc_offset;
1861 loff_t sa_falloc_end;
1862 uid_t sa_falloc_uid;
1863 gid_t sa_falloc_gid;
1864 __u32 sa_falloc_projid;
1866 struct cl_data_version_io {
1867 u64 dv_data_version;
1868 u32 dv_layout_version;
1871 struct cl_fault_io {
1872 /** page index within file. */
1874 /** bytes valid byte on a faulted page. */
1876 /** writable page? for nopage() only */
1878 /** page of an executable? */
1880 /** page_mkwrite() */
1882 /** resulting page */
1883 struct cl_page *ft_page;
1885 struct cl_fsync_io {
1888 /** file system level fid */
1889 struct lu_fid *fi_fid;
1890 enum cl_fsync_mode fi_mode;
1891 /* how many pages were written/discarded */
1892 unsigned int fi_nr_written;
1894 struct cl_ladvise_io {
1897 /** file system level fid */
1898 struct lu_fid *li_fid;
1899 enum lu_ladvise_type li_advice;
1902 struct cl_lseek_io {
1908 time64_t lm_next_rpc_time;
1911 struct cl_2queue ci_queue;
1914 unsigned int ci_continue:1,
1916 * This io has held grouplock, to inform sublayers that
1917 * don't do lockless i/o.
1921 * The whole IO need to be restarted because layout has been changed
1925 * to not refresh layout - the IO issuer knows that the layout won't
1926 * change(page operations, layout change causes all page to be
1927 * discarded), or it doesn't matter if it changes(sync).
1931 * Need MDS intervention to complete a write.
1932 * Write intent is required for the following cases:
1933 * 1. component being written is not initialized, or
1934 * 2. the mirrored files are NOT in WRITE_PENDING state.
1936 ci_need_write_intent:1,
1938 * Check if layout changed after the IO finishes. Mainly for HSM
1939 * requirement. If IO occurs to openning files, it doesn't need to
1940 * verify layout because HSM won't release openning files.
1941 * Right now, only two opertaions need to verify layout: glimpse
1946 * file is released, restore has to to be triggered by vvp layer
1948 ci_restore_needed:1,
1953 /* Tell sublayers not to expand LDLM locks requested for this IO */
1954 ci_lock_no_expand:1,
1956 * Set if non-delay RPC should be used for this IO.
1958 * If this file has multiple mirrors, and if the OSTs of the current
1959 * mirror is inaccessible, non-delay RPC would error out quickly so
1960 * that the upper layer can try to access the next mirror.
1964 * Set if IO is triggered by async workqueue readahead.
1966 ci_async_readahead:1,
1968 * Ignore lockless and do normal locking for this io.
1972 * Set if we've tried all mirrors for this read IO, if it's not set,
1973 * the read IO will check to-be-read OSCs' status, and make fast-switch
1974 * another mirror if some of the OSTs are not healthy.
1976 ci_tried_all_mirrors:1,
1978 * Random read hints, readahead will be disabled.
1982 * Sequential read hints.
1986 * Do parallel (async) submission of DIO RPCs. Note DIO is still sync
1987 * to userspace, only the RPCs are submitted async, then waited for at
1988 * the llite layer before returning.
1992 * Bypass quota check
1994 unsigned ci_noquota:1;
1996 * How many times the read has retried before this one.
1997 * Set by the top level and consumed by the LOV.
1999 unsigned ci_ndelay_tried;
2001 * Designated mirror index for this I/O.
2003 unsigned ci_designated_mirror;
2005 * Number of pages owned by this IO. For invariant checking.
2007 unsigned ci_owned_nr;
2009 * Range of write intent. Valid if ci_need_write_intent is set.
2011 struct lu_extent ci_write_intent;
2017 * Per-transfer attributes.
2019 struct cl_req_attr {
2020 enum cl_req_type cra_type;
2022 struct cl_page *cra_page;
2023 /** Generic attributes for the server consumption. */
2024 struct obdo *cra_oa;
2026 char cra_jobid[LUSTRE_JOBID_SIZE];
2029 enum cache_stats_item {
2030 /** how many cache lookups were performed */
2032 /** how many times cache lookup resulted in a hit */
2034 /** how many entities are in the cache right now */
2036 /** how many entities in the cache are actively used (and cannot be
2037 * evicted) right now */
2039 /** how many entities were created at all */
2044 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2047 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2049 struct cache_stats {
2050 const char *cs_name;
2051 atomic_t cs_stats[CS_NR];
2054 /** These are not exported so far */
2055 void cache_stats_init (struct cache_stats *cs, const char *name);
2058 * Client-side site. This represents particular client stack. "Global"
2059 * variables should (directly or indirectly) be added here to allow multiple
2060 * clients to co-exist in the single address space.
2063 struct lu_site cs_lu;
2065 * Statistical counters. Atomics do not scale, something better like
2066 * per-cpu counters is needed.
2068 * These are exported as /proc/fs/lustre/llite/.../site
2070 * When interpreting keep in mind that both sub-locks (and sub-pages)
2071 * and top-locks (and top-pages) are accounted here.
2073 struct cache_stats cs_pages;
2074 atomic_t cs_pages_state[CPS_NR];
2077 int cl_site_init(struct cl_site *s, struct cl_device *top);
2078 void cl_site_fini(struct cl_site *s);
2079 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2082 * Output client site statistical counters into a buffer. Suitable for
2083 * ll_rd_*()-style functions.
2085 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2090 * Type conversion and accessory functions.
2094 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2096 return container_of(site, struct cl_site, cs_lu);
2099 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2101 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2102 return container_of_safe(d, struct cl_device, cd_lu_dev);
2105 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2107 return &d->cd_lu_dev;
2110 static inline struct cl_object *lu2cl(const struct lu_object *o)
2112 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2113 return container_of_safe(o, struct cl_object, co_lu);
2116 static inline const struct cl_object_conf *
2117 lu2cl_conf(const struct lu_object_conf *conf)
2119 return container_of_safe(conf, struct cl_object_conf, coc_lu);
2122 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2124 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2127 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2129 return container_of_safe(h, struct cl_object_header, coh_lu);
2132 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2134 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2138 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2140 return luh2coh(obj->co_lu.lo_header);
2143 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2145 return lu_device_init(&d->cd_lu_dev, t);
2148 static inline void cl_device_fini(struct cl_device *d)
2150 lu_device_fini(&d->cd_lu_dev);
2153 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2154 struct cl_object *obj,
2155 const struct cl_page_operations *ops);
2156 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2157 struct cl_object *obj,
2158 const struct cl_lock_operations *ops);
2159 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2160 struct cl_object *obj, const struct cl_io_operations *ops);
2163 /** \defgroup cl_object cl_object
2165 struct cl_object *cl_object_top (struct cl_object *o);
2166 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2167 const struct lu_fid *fid,
2168 const struct cl_object_conf *c);
2170 int cl_object_header_init(struct cl_object_header *h);
2171 void cl_object_header_fini(struct cl_object_header *h);
2172 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2173 void cl_object_get (struct cl_object *o);
2174 void cl_object_attr_lock (struct cl_object *o);
2175 void cl_object_attr_unlock(struct cl_object *o);
2176 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2177 struct cl_attr *attr);
2178 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2179 const struct cl_attr *attr, unsigned valid);
2180 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2181 struct ost_lvb *lvb);
2182 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2183 const struct cl_object_conf *conf);
2184 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2185 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2186 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2187 struct lov_user_md __user *lum, size_t size);
2188 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2189 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2191 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2192 struct cl_layout *cl);
2193 loff_t cl_object_maxbytes(struct cl_object *obj);
2194 int cl_object_flush(const struct lu_env *env, struct cl_object *obj,
2195 struct ldlm_lock *lock);
2199 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2201 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2203 return cl_object_header(o0) == cl_object_header(o1);
2206 static inline void cl_object_page_init(struct cl_object *clob, int size)
2208 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2209 cl_object_header(clob)->coh_page_bufsize += cfs_size_round(size);
2210 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2213 static inline void *cl_object_page_slice(struct cl_object *clob,
2214 struct cl_page *page)
2216 return (void *)((char *)page + clob->co_slice_off);
2220 * Return refcount of cl_object.
2222 static inline int cl_object_refc(struct cl_object *clob)
2224 struct lu_object_header *header = clob->co_lu.lo_header;
2225 return atomic_read(&header->loh_ref);
2230 /** \defgroup cl_page cl_page
2232 struct cl_page *cl_page_find (const struct lu_env *env,
2233 struct cl_object *obj,
2234 pgoff_t idx, struct page *vmpage,
2235 enum cl_page_type type);
2236 struct cl_page *cl_page_alloc (const struct lu_env *env,
2237 struct cl_object *o, pgoff_t ind,
2238 struct page *vmpage,
2239 enum cl_page_type type);
2240 void cl_page_get (struct cl_page *page);
2241 void cl_page_put (const struct lu_env *env,
2242 struct cl_page *page);
2243 void cl_pagevec_put (const struct lu_env *env,
2244 struct cl_page *page,
2245 struct pagevec *pvec);
2246 void cl_page_print (const struct lu_env *env, void *cookie,
2247 lu_printer_t printer,
2248 const struct cl_page *pg);
2249 void cl_page_header_print(const struct lu_env *env, void *cookie,
2250 lu_printer_t printer,
2251 const struct cl_page *pg);
2252 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2253 struct cl_page *cl_page_top (struct cl_page *page);
2255 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
2256 const struct lu_device_type *dtype);
2261 * Functions dealing with the ownership of page by io.
2265 int cl_page_own (const struct lu_env *env,
2266 struct cl_io *io, struct cl_page *page);
2267 int cl_page_own_try (const struct lu_env *env,
2268 struct cl_io *io, struct cl_page *page);
2269 void cl_page_assume (const struct lu_env *env,
2270 struct cl_io *io, struct cl_page *page);
2271 void cl_page_unassume (const struct lu_env *env,
2272 struct cl_io *io, struct cl_page *pg);
2273 void cl_page_disown (const struct lu_env *env,
2274 struct cl_io *io, struct cl_page *page);
2275 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2282 * Functions dealing with the preparation of a page for a transfer, and
2283 * tracking transfer state.
2286 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2287 struct cl_page *pg, enum cl_req_type crt);
2288 void cl_page_completion (const struct lu_env *env,
2289 struct cl_page *pg, enum cl_req_type crt, int ioret);
2290 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2291 enum cl_req_type crt);
2292 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2293 struct cl_page *pg, enum cl_req_type crt);
2294 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2296 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2297 struct cl_page *pg);
2303 * \name helper routines
2304 * Functions to discard, delete and export a cl_page.
2307 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2308 struct cl_page *pg);
2309 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2310 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
2312 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx);
2313 pgoff_t cl_index(const struct cl_object *obj, loff_t offset);
2314 size_t cl_page_size(const struct cl_object *obj);
2316 void cl_lock_print(const struct lu_env *env, void *cookie,
2317 lu_printer_t printer, const struct cl_lock *lock);
2318 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2319 lu_printer_t printer,
2320 const struct cl_lock_descr *descr);
2324 * Data structure managing a client's cached pages. A count of
2325 * "unstable" pages is maintained, and an LRU of clean pages is
2326 * maintained. "unstable" pages are pages pinned by the ptlrpc
2327 * layer for recovery purposes.
2329 struct cl_client_cache {
2331 * # of client cache refcount
2332 * # of users (OSCs) + 2 (held by llite and lov)
2336 * # of threads are doing shrinking
2338 unsigned int ccc_lru_shrinkers;
2340 * # of LRU entries available
2342 atomic_long_t ccc_lru_left;
2344 * List of entities(OSCs) for this LRU cache
2346 struct list_head ccc_lru;
2348 * Max # of LRU entries
2350 unsigned long ccc_lru_max;
2352 * Lock to protect ccc_lru list
2354 spinlock_t ccc_lru_lock;
2356 * Set if unstable check is enabled
2358 unsigned int ccc_unstable_check:1;
2360 * # of unstable pages for this mount point
2362 atomic_long_t ccc_unstable_nr;
2364 * Waitq for awaiting unstable pages to reach zero.
2365 * Used at umounting time and signaled on BRW commit
2367 wait_queue_head_t ccc_unstable_waitq;
2369 * Serialize max_cache_mb write operation
2371 struct mutex ccc_max_cache_mb_lock;
2374 * cl_cache functions
2376 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2377 void cl_cache_incref(struct cl_client_cache *cache);
2378 void cl_cache_decref(struct cl_client_cache *cache);
2382 /** \defgroup cl_lock cl_lock
2384 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2385 struct cl_lock *lock);
2386 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2387 const struct cl_io *io);
2388 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2389 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2390 const struct lu_device_type *dtype);
2391 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2393 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2394 struct cl_lock *lock, struct cl_sync_io *anchor);
2395 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2399 /** \defgroup cl_io cl_io
2402 int cl_io_init (const struct lu_env *env, struct cl_io *io,
2403 enum cl_io_type iot, struct cl_object *obj);
2404 int cl_io_sub_init (const struct lu_env *env, struct cl_io *io,
2405 enum cl_io_type iot, struct cl_object *obj);
2406 int cl_io_rw_init (const struct lu_env *env, struct cl_io *io,
2407 enum cl_io_type iot, loff_t pos, size_t count);
2408 int cl_io_loop (const struct lu_env *env, struct cl_io *io);
2410 void cl_io_fini (const struct lu_env *env, struct cl_io *io);
2411 int cl_io_iter_init (const struct lu_env *env, struct cl_io *io);
2412 void cl_io_iter_fini (const struct lu_env *env, struct cl_io *io);
2413 int cl_io_lock (const struct lu_env *env, struct cl_io *io);
2414 void cl_io_unlock (const struct lu_env *env, struct cl_io *io);
2415 int cl_io_start (const struct lu_env *env, struct cl_io *io);
2416 void cl_io_end (const struct lu_env *env, struct cl_io *io);
2417 int cl_io_lock_add (const struct lu_env *env, struct cl_io *io,
2418 struct cl_io_lock_link *link);
2419 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2420 struct cl_lock_descr *descr);
2421 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2422 enum cl_req_type iot, struct cl_2queue *queue);
2423 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2424 enum cl_req_type iot, struct cl_2queue *queue,
2426 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2427 struct cl_page_list *queue, int from, int to,
2429 void cl_io_extent_release (const struct lu_env *env, struct cl_io *io);
2430 int cl_io_lru_reserve(const struct lu_env *env, struct cl_io *io,
2431 loff_t pos, size_t bytes);
2432 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2433 pgoff_t start, struct cl_read_ahead *ra);
2434 void cl_io_rw_advance (const struct lu_env *env, struct cl_io *io,
2438 * True, iff \a io is an O_APPEND write(2).
2440 static inline int cl_io_is_append(const struct cl_io *io)
2442 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2445 static inline int cl_io_is_sync_write(const struct cl_io *io)
2447 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2450 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2452 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2456 * True, iff \a io is a truncate(2).
2458 static inline int cl_io_is_trunc(const struct cl_io *io)
2460 return io->ci_type == CIT_SETATTR &&
2461 (io->u.ci_setattr.sa_avalid & ATTR_SIZE) &&
2462 (io->u.ci_setattr.sa_subtype != CL_SETATTR_FALLOCATE);
2465 static inline int cl_io_is_fallocate(const struct cl_io *io)
2467 return (io->ci_type == CIT_SETATTR) &&
2468 (io->u.ci_setattr.sa_subtype == CL_SETATTR_FALLOCATE);
2471 struct cl_io *cl_io_top(struct cl_io *io);
2473 void cl_io_print(const struct lu_env *env, void *cookie,
2474 lu_printer_t printer, const struct cl_io *io);
2476 #define CL_IO_SLICE_CLEAN(foo_io, base) \
2478 typeof(foo_io) __foo_io = (foo_io); \
2480 memset(&__foo_io->base, 0, \
2481 sizeof(*__foo_io) - offsetof(typeof(*__foo_io), base)); \
2486 /** \defgroup cl_page_list cl_page_list
2490 * Last page in the page list.
2492 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2494 LASSERT(plist->pl_nr > 0);
2495 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2498 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2500 LASSERT(plist->pl_nr > 0);
2501 return list_first_entry(&plist->pl_pages, struct cl_page, cp_batch);
2505 * Iterate over pages in a page list.
2507 #define cl_page_list_for_each(page, list) \
2508 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2511 * Iterate over pages in a page list, taking possible removals into account.
2513 #define cl_page_list_for_each_safe(page, temp, list) \
2514 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2516 void cl_page_list_init(struct cl_page_list *plist);
2517 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page,
2519 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2520 struct cl_page *page);
2521 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2522 struct cl_page *page);
2523 void cl_page_list_splice(struct cl_page_list *list,
2524 struct cl_page_list *head);
2525 void cl_page_list_del(const struct lu_env *env,
2526 struct cl_page_list *plist, struct cl_page *page);
2527 void cl_page_list_disown(const struct lu_env *env,
2528 struct cl_io *io, struct cl_page_list *plist);
2529 void cl_page_list_assume(const struct lu_env *env,
2530 struct cl_io *io, struct cl_page_list *plist);
2531 void cl_page_list_discard(const struct lu_env *env,
2532 struct cl_io *io, struct cl_page_list *plist);
2533 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2535 void cl_2queue_init(struct cl_2queue *queue);
2536 void cl_2queue_add(struct cl_2queue *queue, struct cl_page *page,
2538 void cl_2queue_disown(const struct lu_env *env, struct cl_io *io,
2539 struct cl_2queue *queue);
2540 void cl_2queue_assume(const struct lu_env *env, struct cl_io *io,
2541 struct cl_2queue *queue);
2542 void cl_2queue_discard(const struct lu_env *env, struct cl_io *io,
2543 struct cl_2queue *queue);
2544 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2545 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2547 /** @} cl_page_list */
2549 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2550 struct cl_req_attr *attr);
2552 /** \defgroup cl_sync_io cl_sync_io
2558 typedef void (cl_sync_io_end_t)(const struct lu_env *, struct cl_sync_io *);
2560 void cl_sync_io_init_notify(struct cl_sync_io *anchor, int nr,
2561 struct cl_dio_aio *aio, cl_sync_io_end_t *end);
2563 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2565 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2567 int cl_sync_io_wait_recycle(const struct lu_env *env, struct cl_sync_io *anchor,
2568 long timeout, int ioret);
2569 struct cl_dio_aio *cl_aio_alloc(struct kiocb *iocb, struct cl_object *obj,
2570 struct cl_dio_aio *ll_aio);
2571 void cl_aio_free(const struct lu_env *env, struct cl_dio_aio *aio);
2572 static inline void cl_sync_io_init(struct cl_sync_io *anchor, int nr)
2574 cl_sync_io_init_notify(anchor, nr, NULL, NULL);
2578 * Anchor for synchronous transfer. This is allocated on a stack by thread
2579 * doing synchronous transfer, and a pointer to this structure is set up in
2580 * every page submitted for transfer. Transfer completion routine updates
2581 * anchor and wakes up waiting thread when transfer is complete.
2584 /** number of pages yet to be transferred. */
2585 atomic_t csi_sync_nr;
2588 /** completion to be signaled when transfer is complete. */
2589 wait_queue_head_t csi_waitq;
2590 /** callback to invoke when this IO is finished */
2591 cl_sync_io_end_t *csi_end_io;
2592 /** aio private data */
2593 struct cl_dio_aio *csi_aio;
2596 /** direct IO pages */
2597 struct ll_dio_pages {
2599 * page array to be written. we don't support
2600 * partial pages except the last one.
2602 struct page **ldp_pages;
2603 /** # of pages in the array. */
2605 /* the file offset of the first page. */
2606 loff_t ldp_file_offset;
2609 /** To support Direct AIO */
2611 struct cl_sync_io cda_sync;
2612 struct cl_page_list cda_pages;
2613 struct cl_object *cda_obj;
2614 struct kiocb *cda_iocb;
2616 struct cl_dio_aio *cda_ll_aio;
2617 struct ll_dio_pages cda_dio_pages;
2618 unsigned cda_no_aio_complete:1,
2622 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW) || \
2623 defined(HAVE_DIRECTIO_2ARGS)
2624 #define HAVE_DIO_ITER 1
2627 void ll_release_user_pages(struct page **pages, int npages);
2629 /** @} cl_sync_io */
2631 /** \defgroup cl_env cl_env
2633 * lu_env handling for a client.
2635 * lu_env is an environment within which lustre code executes. Its major part
2636 * is lu_context---a fast memory allocation mechanism that is used to conserve
2637 * precious kernel stack space. Originally lu_env was designed for a server,
2640 * - there is a (mostly) fixed number of threads, and
2642 * - call chains have no non-lustre portions inserted between lustre code.
2644 * On a client both these assumtpion fails, because every user thread can
2645 * potentially execute lustre code as part of a system call, and lustre calls
2646 * into VFS or MM that call back into lustre.
2648 * To deal with that, cl_env wrapper functions implement the following
2651 * - allocation and destruction of environment is amortized by caching no
2652 * longer used environments instead of destroying them;
2654 * \see lu_env, lu_context, lu_context_key
2657 struct lu_env *cl_env_get(__u16 *refcheck);
2658 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2659 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2660 unsigned cl_env_cache_purge(unsigned nr);
2661 struct lu_env *cl_env_percpu_get(void);
2662 void cl_env_percpu_put(struct lu_env *env);
2669 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2670 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2672 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2673 struct lu_device_type *ldt,
2674 struct lu_device *next);
2677 int cl_global_init(void);
2678 void cl_global_fini(void);
2680 #endif /* _LINUX_CL_OBJECT_H */