4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 #ifndef _LUSTRE_CL_OBJECT_H
32 #define _LUSTRE_CL_OBJECT_H
34 /** \defgroup clio clio
36 * Client objects implement io operations and cache pages.
38 * Examples: lov and osc are implementations of cl interface.
40 * Big Theory Statement.
44 * Client implementation is based on the following data-types:
50 * - cl_lock represents an extent lock on an object.
52 * - cl_io represents high-level i/o activity such as whole read/write
53 * system call, or write-out of pages from under the lock being
54 * canceled. cl_io has sub-ios that can be stopped and resumed
55 * independently, thus achieving high degree of transfer
56 * parallelism. Single cl_io can be advanced forward by
57 * the multiple threads (although in the most usual case of
58 * read/write system call it is associated with the single user
59 * thread, that issued the system call).
63 * - to avoid confusion high-level I/O operation like read or write system
64 * call is referred to as "an io", whereas low-level I/O operation, like
65 * RPC, is referred to as "a transfer"
67 * - "generic code" means generic (not file system specific) code in the
68 * hosting environment. "cl-code" means code (mostly in cl_*.c files) that
69 * is not layer specific.
75 * - cl_object_header::coh_page_guard
78 * See the top comment in cl_object.c for the description of overall locking and
79 * reference-counting design.
81 * See comments below for the description of i/o, page, and dlm-locking
88 * super-class definitions.
90 #include <linux/aio.h>
93 #include <libcfs/libcfs.h>
94 #include <lu_object.h>
95 #include <linux/atomic.h>
96 #include <linux/mutex.h>
97 #include <linux/radix-tree.h>
98 #include <linux/spinlock.h>
99 #include <linux/wait.h>
100 #include <linux/pagevec.h>
101 #include <libcfs/linux/linux-misc.h>
102 #include <lustre_dlm.h>
112 struct cl_page_slice;
114 struct cl_lock_slice;
116 struct cl_lock_operations;
117 struct cl_page_operations;
125 * Device in the client stack.
127 * \see vvp_device, lov_device, lovsub_device, osc_device
131 struct lu_device cd_lu_dev;
134 /** \addtogroup cl_object cl_object
137 * "Data attributes" of cl_object. Data attributes can be updated
138 * independently for a sub-object, and top-object's attributes are calculated
139 * from sub-objects' ones.
142 /** Object size, in bytes */
145 unsigned int cat_kms_valid:1;
147 * Known minimal size, in bytes.
149 * This is only valid when at least one DLM lock is held.
152 /** Modification time. Measured in seconds since epoch. */
154 /** Access time. Measured in seconds since epoch. */
156 /** Change time. Measured in seconds since epoch. */
159 * Blocks allocated to this cl_object on the server file system.
161 * \todo XXX An interface for block size is needed.
165 * User identifier for quota purposes.
169 * Group identifier for quota purposes.
173 /* nlink of the directory */
176 /* Project identifier for quota purpose. */
181 * Fields in cl_attr that are being set.
196 * Sub-class of lu_object with methods common for objects on the client
199 * cl_object: represents a regular file system object, both a file and a
200 * stripe. cl_object is based on lu_object: it is identified by a fid,
201 * layered, cached, hashed, and lrued. Important distinction with the server
202 * side, where md_object and dt_object are used, is that cl_object "fans out"
203 * at the lov/sns level: depending on the file layout, single file is
204 * represented as a set of "sub-objects" (stripes). At the implementation
205 * level, struct lov_object contains an array of cl_objects. Each sub-object
206 * is a full-fledged cl_object, having its fid, living in the lru and hash
209 * This leads to the next important difference with the server side: on the
210 * client, it's quite usual to have objects with the different sequence of
211 * layers. For example, typical top-object is composed of the following
217 * whereas its sub-objects are composed of
222 * layers. Here "lovsub" is a mostly dummy layer, whose purpose is to keep
223 * track of the object-subobject relationship.
225 * Sub-objects are not cached independently: when top-object is about to
226 * be discarded from the memory, all its sub-objects are torn-down and
229 * \see vvp_object, lov_object, lovsub_object, osc_object
233 struct lu_object co_lu;
234 /** per-object-layer operations */
235 const struct cl_object_operations *co_ops;
236 /** offset of page slice in cl_page buffer */
241 * Description of the client object configuration. This is used for the
242 * creation of a new client object that is identified by a more state than
245 struct cl_object_conf {
247 struct lu_object_conf coc_lu;
250 * Object layout. This is consumed by lov.
252 struct lu_buf coc_layout;
254 * Description of particular stripe location in the
255 * cluster. This is consumed by osc.
257 struct lov_oinfo *coc_oinfo;
260 * VFS inode. This is consumed by vvp.
262 struct inode *coc_inode;
264 * Layout lock handle.
266 struct ldlm_lock *coc_lock;
269 * Operation to handle layout, OBJECT_CONF_XYZ.
275 /** configure layout, set up a new stripe, must be called while
276 * holding layout lock. */
278 /** invalidate the current stripe configuration due to losing
280 OBJECT_CONF_INVALIDATE = 1,
281 /** wait for old layout to go away so that new layout can be
287 CL_LAYOUT_GEN_NONE = (u32)-2, /* layout lock was cancelled */
288 CL_LAYOUT_GEN_EMPTY = (u32)-1, /* for empty layout */
292 /** the buffer to return the layout in lov_mds_md format. */
293 struct lu_buf cl_buf;
294 /** size of layout in lov_mds_md format. */
296 /** Layout generation. */
298 /** whether layout is a composite one */
299 bool cl_is_composite;
300 /** Whether layout is a HSM released one */
312 * Operations implemented for each cl object layer.
314 * \see vvp_ops, lov_ops, lovsub_ops, osc_ops
316 struct cl_object_operations {
318 * Initialize page slice for this layer. Called top-to-bottom through
319 * every object layer when a new cl_page is instantiated. Layer
320 * keeping private per-page data, or requiring its own page operations
321 * vector should allocate these data here, and attach then to the page
322 * by calling cl_page_slice_add(). \a vmpage is locked (in the VM
325 * \retval NULL success.
327 * \retval ERR_PTR(errno) failure code.
329 * \retval valid-pointer pointer to already existing referenced page
330 * to be used instead of newly created.
332 int (*coo_page_init)(const struct lu_env *env, struct cl_object *obj,
333 struct cl_page *page, pgoff_t index);
335 * Initialize lock slice for this layer. Called top-to-bottom through
336 * every object layer when a new cl_lock is instantiated. Layer
337 * keeping private per-lock data, or requiring its own lock operations
338 * vector should allocate these data here, and attach then to the lock
339 * by calling cl_lock_slice_add(). Mandatory.
341 int (*coo_lock_init)(const struct lu_env *env,
342 struct cl_object *obj, struct cl_lock *lock,
343 const struct cl_io *io);
345 * Initialize io state for a given layer.
347 * called top-to-bottom once per io existence to initialize io
348 * state. If layer wants to keep some state for this type of io, it
349 * has to embed struct cl_io_slice in lu_env::le_ses, and register
350 * slice with cl_io_slice_add(). It is guaranteed that all threads
351 * participating in this io share the same session.
353 int (*coo_io_init)(const struct lu_env *env,
354 struct cl_object *obj, struct cl_io *io);
356 * Fill portion of \a attr that this layer controls. This method is
357 * called top-to-bottom through all object layers.
359 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
361 * \return 0: to continue
362 * \return +ve: to stop iterating through layers (but 0 is returned
363 * from enclosing cl_object_attr_get())
364 * \return -ve: to signal error
366 int (*coo_attr_get)(const struct lu_env *env, struct cl_object *obj,
367 struct cl_attr *attr);
371 * \a valid is a bitmask composed from enum #cl_attr_valid, and
372 * indicating what attributes are to be set.
374 * \pre cl_object_header::coh_attr_guard of the top-object is locked.
376 * \return the same convention as for
377 * cl_object_operations::coo_attr_get() is used.
379 int (*coo_attr_update)(const struct lu_env *env, struct cl_object *obj,
380 const struct cl_attr *attr, unsigned valid);
382 * Mark the inode dirty. By this way, the inode will add into the
383 * writeback list of the corresponding @bdi_writeback, and then it will
384 * defer to write out the dirty pages to OSTs via the kernel writeback
387 void (*coo_dirty_for_sync)(const struct lu_env *env,
388 struct cl_object *obj);
390 * Update object configuration. Called top-to-bottom to modify object
393 * XXX error conditions and handling.
395 int (*coo_conf_set)(const struct lu_env *env, struct cl_object *obj,
396 const struct cl_object_conf *conf);
398 * Glimpse ast. Executed when glimpse ast arrives for a lock on this
399 * object. Layers are supposed to fill parts of \a lvb that will be
400 * shipped to the glimpse originator as a glimpse result.
402 * \see vvp_object_glimpse(), lovsub_object_glimpse(),
403 * \see osc_object_glimpse()
405 int (*coo_glimpse)(const struct lu_env *env,
406 const struct cl_object *obj, struct ost_lvb *lvb);
408 * Object prune method. Called when the layout is going to change on
409 * this object, therefore each layer has to clean up their cache,
410 * mainly pages and locks.
412 int (*coo_prune)(const struct lu_env *env, struct cl_object *obj);
414 * Object getstripe method.
416 int (*coo_getstripe)(const struct lu_env *env, struct cl_object *obj,
417 struct lov_user_md __user *lum, size_t size);
419 * Get FIEMAP mapping from the object.
421 int (*coo_fiemap)(const struct lu_env *env, struct cl_object *obj,
422 struct ll_fiemap_info_key *fmkey,
423 struct fiemap *fiemap, size_t *buflen);
425 * Get layout and generation of the object.
427 int (*coo_layout_get)(const struct lu_env *env, struct cl_object *obj,
428 struct cl_layout *layout);
430 * Get maximum size of the object.
432 loff_t (*coo_maxbytes)(struct cl_object *obj);
434 * Set request attributes.
436 void (*coo_req_attr_set)(const struct lu_env *env,
437 struct cl_object *obj,
438 struct cl_req_attr *attr);
440 * Flush \a obj data corresponding to \a lock. Used for DoM
441 * locks in llite's cancelling blocking ast callback.
443 int (*coo_object_flush)(const struct lu_env *env,
444 struct cl_object *obj,
445 struct ldlm_lock *lock);
447 * operate upon inode. Used in LOV to lock/unlock inode from vvp layer.
449 int (*coo_inode_ops)(const struct lu_env *env, struct cl_object *obj,
450 enum coo_inode_opc opc, void *data);
454 * Extended header for client object.
456 struct cl_object_header {
457 /** Standard lu_object_header. cl_object::co_lu::lo_header points
459 struct lu_object_header coh_lu;
462 * Parent object. It is assumed that an object has a well-defined
463 * parent, but not a well-defined child (there may be multiple
464 * sub-objects, for the same top-object). cl_object_header::coh_parent
465 * field allows certain code to be written generically, without
466 * limiting possible cl_object layouts unduly.
468 struct cl_object_header *coh_parent;
470 * Protects consistency between cl_attr of parent object and
471 * attributes of sub-objects, that the former is calculated ("merged")
474 * \todo XXX this can be read/write lock if needed.
476 spinlock_t coh_attr_guard;
478 * Size of cl_page + page slices
480 unsigned short coh_page_bufsize;
482 * Number of objects above this one: 0 for a top-object, 1 for its
485 unsigned char coh_nesting;
489 * Helper macro: iterate over all layers of the object \a obj, assigning every
490 * layer top-to-bottom to \a slice.
492 #define cl_object_for_each(slice, obj) \
493 list_for_each_entry((slice), \
494 &(obj)->co_lu.lo_header->loh_layers,\
498 * Helper macro: iterate over all layers of the object \a obj, assigning every
499 * layer bottom-to-top to \a slice.
501 #define cl_object_for_each_reverse(slice, obj) \
502 list_for_each_entry_reverse((slice), \
503 &(obj)->co_lu.lo_header->loh_layers,\
508 #define CL_PAGE_EOF ((pgoff_t)~0ull)
510 /** \addtogroup cl_page cl_page
514 * Layered client page.
516 * cl_page: represents a portion of a file, cached in the memory. All pages
517 * of the given file are of the same size, and are kept in the radix tree
518 * hanging off the cl_object. cl_page doesn't fan out, but as sub-objects
519 * of the top-level file object are first class cl_objects, they have their
520 * own radix trees of pages and hence page is implemented as a sequence of
521 * struct cl_pages's, linked into double-linked list through
522 * cl_page::cp_parent and cl_page::cp_child pointers, each residing in the
523 * corresponding radix tree at the corresponding logical offset.
525 * cl_page is associated with VM page of the hosting environment (struct
526 * page in Linux kernel, for example), struct page. It is assumed, that this
527 * association is implemented by one of cl_page layers (top layer in the
528 * current design) that
530 * - intercepts per-VM-page call-backs made by the environment (e.g.,
533 * - translates state (page flag bits) and locking between lustre and
536 * The association between cl_page and struct page is immutable and
537 * established when cl_page is created.
539 * cl_page can be "owned" by a particular cl_io (see below), guaranteeing
540 * this io an exclusive access to this page w.r.t. other io attempts and
541 * various events changing page state (such as transfer completion, or
542 * eviction of the page from the memory). Note, that in general cl_io
543 * cannot be identified with a particular thread, and page ownership is not
544 * exactly equal to the current thread holding a lock on the page. Layer
545 * implementing association between cl_page and struct page has to implement
546 * ownership on top of available synchronization mechanisms.
548 * While lustre client maintains the notion of an page ownership by io,
549 * hosting MM/VM usually has its own page concurrency control
550 * mechanisms. For example, in Linux, page access is synchronized by the
551 * per-page PG_locked bit-lock, and generic kernel code (generic_file_*())
552 * takes care to acquire and release such locks as necessary around the
553 * calls to the file system methods (->readpage(), ->prepare_write(),
554 * ->commit_write(), etc.). This leads to the situation when there are two
555 * different ways to own a page in the client:
557 * - client code explicitly and voluntary owns the page (cl_page_own());
559 * - VM locks a page and then calls the client, that has "to assume"
560 * the ownership from the VM (cl_page_assume()).
562 * Dual methods to release ownership are cl_page_disown() and
563 * cl_page_unassume().
565 * cl_page is reference counted (cl_page::cp_ref). When reference counter
566 * drops to 0, the page is returned to the cache, unless it is in
567 * cl_page_state::CPS_FREEING state, in which case it is immediately
570 * The general logic guaranteeing the absence of "existential races" for
571 * pages is the following:
573 * - there are fixed known ways for a thread to obtain a new reference
576 * - by doing a lookup in the cl_object radix tree, protected by the
579 * - by starting from VM-locked struct page and following some
580 * hosting environment method (e.g., following ->private pointer in
581 * the case of Linux kernel), see cl_vmpage_page();
583 * - when the page enters cl_page_state::CPS_FREEING state, all these
584 * ways are severed with the proper synchronization
585 * (cl_page_delete());
587 * - entry into cl_page_state::CPS_FREEING is serialized by the VM page
590 * - no new references to the page in cl_page_state::CPS_FREEING state
591 * are allowed (checked in cl_page_get()).
593 * Together this guarantees that when last reference to a
594 * cl_page_state::CPS_FREEING page is released, it is safe to destroy the
595 * page, as neither references to it can be acquired at that point, nor
598 * cl_page is a state machine. States are enumerated in enum
599 * cl_page_state. Possible state transitions are enumerated in
600 * cl_page_state_set(). State transition process (i.e., actual changing of
601 * cl_page::cp_state field) is protected by the lock on the underlying VM
604 * Linux Kernel implementation.
606 * Binding between cl_page and struct page (which is a typedef for
607 * struct page) is implemented in the vvp layer. cl_page is attached to the
608 * ->private pointer of the struct page, together with the setting of
609 * PG_private bit in page->flags, and acquiring additional reference on the
610 * struct page (much like struct buffer_head, or any similar file system
611 * private data structures).
613 * PG_locked lock is used to implement both ownership and transfer
614 * synchronization, that is, page is VM-locked in CPS_{OWNED,PAGE{IN,OUT}}
615 * states. No additional references are acquired for the duration of the
618 * \warning *THIS IS NOT* the behavior expected by the Linux kernel, where
619 * write-out is "protected" by the special PG_writeback bit.
623 * States of cl_page. cl_page.c assumes particular order here.
625 * The page state machine is rather crude, as it doesn't recognize finer page
626 * states like "dirty" or "up to date". This is because such states are not
627 * always well defined for the whole stack (see, for example, the
628 * implementation of the read-ahead, that hides page up-to-dateness to track
629 * cache hits accurately). Such sub-states are maintained by the layers that
630 * are interested in them.
634 * Page is in the cache, un-owned. Page leaves cached state in the
637 * - [cl_page_state::CPS_OWNED] io comes across the page and
640 * - [cl_page_state::CPS_PAGEOUT] page is dirty, the
641 * req-formation engine decides that it wants to include this page
642 * into an RPC being constructed, and yanks it from the cache;
644 * - [cl_page_state::CPS_FREEING] VM callback is executed to
645 * evict the page form the memory;
647 * \invariant cl_page::cp_owner == NULL && cl_page::cp_req == NULL
651 * Page is exclusively owned by some cl_io. Page may end up in this
652 * state as a result of
654 * - io creating new page and immediately owning it;
656 * - [cl_page_state::CPS_CACHED] io finding existing cached page
659 * - [cl_page_state::CPS_OWNED] io finding existing owned page
660 * and waiting for owner to release the page;
662 * Page leaves owned state in the following cases:
664 * - [cl_page_state::CPS_CACHED] io decides to leave the page in
665 * the cache, doing nothing;
667 * - [cl_page_state::CPS_PAGEIN] io starts read transfer for
670 * - [cl_page_state::CPS_PAGEOUT] io starts immediate write
671 * transfer for this page;
673 * - [cl_page_state::CPS_FREEING] io decides to destroy this
674 * page (e.g., as part of truncate or extent lock cancellation).
676 * \invariant cl_page::cp_owner != NULL && cl_page::cp_req == NULL
680 * Page is being written out, as a part of a transfer. This state is
681 * entered when req-formation logic decided that it wants this page to
682 * be sent through the wire _now_. Specifically, it means that once
683 * this state is achieved, transfer completion handler (with either
684 * success or failure indication) is guaranteed to be executed against
685 * this page independently of any locks and any scheduling decisions
686 * made by the hosting environment (that effectively means that the
687 * page is never put into cl_page_state::CPS_PAGEOUT state "in
688 * advance". This property is mentioned, because it is important when
689 * reasoning about possible dead-locks in the system). The page can
690 * enter this state as a result of
692 * - [cl_page_state::CPS_OWNED] an io requesting an immediate
693 * write-out of this page, or
695 * - [cl_page_state::CPS_CACHED] req-forming engine deciding
696 * that it has enough dirty pages cached to issue a "good"
699 * The page leaves cl_page_state::CPS_PAGEOUT state when the transfer
700 * is completed---it is moved into cl_page_state::CPS_CACHED state.
702 * Underlying VM page is locked for the duration of transfer.
704 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
708 * Page is being read in, as a part of a transfer. This is quite
709 * similar to the cl_page_state::CPS_PAGEOUT state, except that
710 * read-in is always "immediate"---there is no such thing a sudden
711 * construction of read request from cached, presumably not up to date,
714 * Underlying VM page is locked for the duration of transfer.
716 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req != NULL
720 * Page is being destroyed. This state is entered when client decides
721 * that page has to be deleted from its host object, as, e.g., a part
724 * Once this state is reached, there is no way to escape it.
726 * \invariant: cl_page::cp_owner == NULL && cl_page::cp_req == NULL
733 /** Host page, the page is from the host inode which the cl_page
737 /** Transient page, the transient cl_page is used to bind a cl_page
738 * to vmpage which is not belonging to the same object of cl_page.
739 * it is used in DirectIO and lockless IO. */
744 #define CP_STATE_BITS 4
745 #define CP_TYPE_BITS 2
746 #define CP_MAX_LAYER 2
749 * Fields are protected by the lock on struct page, except for atomics and
752 * \invariant Data type invariants are in cl_page_invariant(). Basically:
753 * cl_page::cp_parent and cl_page::cp_child are a well-formed double-linked
754 * list, consistent with the parent/child pointers in the cl_page::cp_obj and
755 * cl_page::cp_owner (when set).
758 /** Reference counter. */
760 /** layout_entry + stripe index, composed using lov_comp_index() */
761 unsigned int cp_lov_index;
762 /** page->index of the page within the whole file */
763 pgoff_t cp_page_index;
764 /** An object this page is a part of. Immutable after creation. */
765 struct cl_object *cp_obj;
767 struct page *cp_vmpage;
769 * Assigned if doing direct IO, because in this case cp_vmpage is not
770 * a valid page cache page, hence the inode cannot be inferred from
771 * cp_vmpage->mapping->host.
773 struct inode *cp_inode;
774 /** Linkage of pages within group. Pages must be owned */
775 struct list_head cp_batch;
776 /** array of slices offset. Immutable after creation. */
777 unsigned char cp_layer_offset[CP_MAX_LAYER];
778 /** current slice index */
779 unsigned char cp_layer_count:2;
781 * Page state. This field is const to avoid accidental update, it is
782 * modified only internally within cl_page.c. Protected by a VM lock.
784 enum cl_page_state cp_state:CP_STATE_BITS;
786 * Page type. Only CPT_TRANSIENT is used so far. Immutable after
789 enum cl_page_type cp_type:CP_TYPE_BITS;
790 unsigned cp_defer_uptodate:1,
793 /* which slab kmem index this memory allocated from */
794 short int cp_kmem_index;
797 * Owning IO in cl_page_state::CPS_OWNED state. Sub-page can be owned
798 * by sub-io. Protected by a VM lock.
800 struct cl_io *cp_owner;
801 /** List of references to this page, for debugging. */
802 struct lu_ref cp_reference;
803 /** Link to an object, for debugging. */
804 struct lu_ref_link cp_obj_ref;
805 /** Link to a queue, for debugging. */
806 struct lu_ref_link cp_queue_ref;
807 /** Assigned if doing a sync_io */
808 struct cl_sync_io *cp_sync_io;
812 * Per-layer part of cl_page.
814 * \see vvp_page, lov_page, osc_page
816 struct cl_page_slice {
817 struct cl_page *cpl_page;
818 const struct cl_page_operations *cpl_ops;
822 * Lock mode. For the client extent locks.
834 * Requested transfer type.
843 * Per-layer page operations.
845 * Methods taking an \a io argument are for the activity happening in the
846 * context of given \a io. Page is assumed to be owned by that io, except for
849 * \see vvp_page_ops, lov_page_ops, osc_page_ops
851 struct cl_page_operations {
853 * cl_page<->struct page methods. Only one layer in the stack has to
854 * implement these. Current code assumes that this functionality is
855 * provided by the topmost layer, see __cl_page_disown() as an example.
859 * Update file attributes when all we have is this page. Used for tiny
860 * writes to update attributes when we don't have a full cl_io.
862 void (*cpo_page_touch)(const struct lu_env *env,
863 const struct cl_page_slice *slice, size_t to);
869 * Called when page is truncated from the object. Optional.
871 * \see cl_page_discard()
872 * \see vvp_page_discard(), osc_page_discard()
874 void (*cpo_discard)(const struct lu_env *env,
875 const struct cl_page_slice *slice,
878 * Called when page is removed from the cache, and is about to being
879 * destroyed. Optional.
881 * \see cl_page_delete()
882 * \see vvp_page_delete(), osc_page_delete()
884 void (*cpo_delete)(const struct lu_env *env,
885 const struct cl_page_slice *slice);
887 * Optional debugging helper. Prints given page slice.
889 * \see cl_page_print()
891 int (*cpo_print)(const struct lu_env *env,
892 const struct cl_page_slice *slice,
893 void *cookie, lu_printer_t p);
902 * Request type dependent vector of operations.
904 * Transfer operations depend on transfer mode (cl_req_type). To avoid
905 * passing transfer mode to each and every of these methods, and to
906 * avoid branching on request type inside of the methods, separate
907 * methods for cl_req_type:CRT_READ and cl_req_type:CRT_WRITE are
908 * provided. That is, method invocation usually looks like
910 * slice->cp_ops.io[req->crq_type].cpo_method(env, slice, ...);
914 * Completion handler. This is guaranteed to be eventually
915 * fired after cl_page_prep() or cl_page_make_ready() call.
917 * This method can be called in a non-blocking context. It is
918 * guaranteed however, that the page involved and its object
919 * are pinned in memory (and, hence, calling cl_page_put() is
922 * \see cl_page_completion()
924 void (*cpo_completion)(const struct lu_env *env,
925 const struct cl_page_slice *slice,
929 * Tell transfer engine that only [to, from] part of a page should be
932 * This is used for immediate transfers.
934 * \todo XXX this is not very good interface. It would be much better
935 * if all transfer parameters were supplied as arguments to
936 * cl_io_operations::cio_submit() call, but it is not clear how to do
937 * this for page queues.
939 * \see cl_page_clip()
941 void (*cpo_clip)(const struct lu_env *env,
942 const struct cl_page_slice *slice,
945 * Write out a page by kernel. This is only called by ll_writepage
948 * \see cl_page_flush()
950 int (*cpo_flush)(const struct lu_env *env,
951 const struct cl_page_slice *slice,
957 * Helper macro, dumping detailed information about \a page into a log.
959 #define CL_PAGE_DEBUG(mask, env, page, format, ...) \
961 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
962 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
963 cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
964 CDEBUG(mask, format , ## __VA_ARGS__); \
969 * Helper macro, dumping shorter information about \a page into a log.
971 #define CL_PAGE_HEADER(mask, env, page, format, ...) \
973 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
974 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
975 cl_page_header_print(env, &msgdata, lu_cdebug_printer, page); \
976 CDEBUG(mask, format , ## __VA_ARGS__); \
980 static inline struct page *cl_page_vmpage(const struct cl_page *page)
982 LASSERT(page->cp_vmpage != NULL);
983 return page->cp_vmpage;
986 static inline pgoff_t cl_page_index(const struct cl_page *cp)
988 return cl_page_vmpage(cp)->index;
992 * Check if a cl_page is in use.
994 * Client cache holds a refcount, this refcount will be dropped when
995 * the page is taken out of cache, see vvp_page_delete().
997 static inline bool __page_in_use(const struct cl_page *page, int refc)
999 return (refcount_read(&page->cp_ref) > refc + 1);
1003 * Caller itself holds a refcount of cl_page.
1005 #define cl_page_in_use(pg) __page_in_use(pg, 1)
1007 * Caller doesn't hold a refcount.
1009 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
1013 /** \addtogroup cl_lock cl_lock
1017 * Extent locking on the client.
1021 * The locking model of the new client code is built around
1025 * data-type representing an extent lock on a regular file. cl_lock is a
1026 * layered object (much like cl_object and cl_page), it consists of a header
1027 * (struct cl_lock) and a list of layers (struct cl_lock_slice), linked to
1028 * cl_lock::cll_layers list through cl_lock_slice::cls_linkage.
1030 * Typical cl_lock consists of one layer:
1032 * - lov_lock (lov specific data).
1034 * lov_lock contains an array of sub-locks. Each of these sub-locks is a
1035 * normal cl_lock: it has a header (struct cl_lock) and a list of layers:
1039 * Each sub-lock is associated with a cl_object (representing stripe
1040 * sub-object or the file to which top-level cl_lock is associated to), and is
1041 * linked into that cl_object::coh_locks. In this respect cl_lock is similar to
1042 * cl_object (that at lov layer also fans out into multiple sub-objects), and
1043 * is different from cl_page, that doesn't fan out (there is usually exactly
1044 * one osc_page for every vvp_page). We shall call vvp-lov portion of the lock
1045 * a "top-lock" and its lovsub-osc portion a "sub-lock".
1049 * cl_lock is a cacheless data container for the requirements of locks to
1050 * complete the IO. cl_lock is created before I/O starts and destroyed when the
1053 * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
1054 * to cl_lock at OSC layer. LDLM lock is still cacheable.
1056 * INTERFACE AND USAGE
1058 * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
1059 * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
1060 * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
1061 * consists of multiple sub cl_locks, each sub locks will be enqueued
1062 * correspondingly. At OSC layer, the lock enqueue request will tend to reuse
1063 * cached LDLM lock; otherwise a new LDLM lock will have to be requested from
1066 * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
1067 * method will be called for each layer to release the resource held by this
1068 * lock. At OSC layer, the reference count of LDLM lock, which is held at
1069 * clo_enqueue time, is released.
1071 * LDLM lock can only be canceled if there is no cl_lock using it.
1073 * Overall process of the locking during IO operation is as following:
1075 * - once parameters for IO are setup in cl_io, cl_io_operations::cio_lock()
1076 * is called on each layer. Responsibility of this method is to add locks,
1077 * needed by a given layer into cl_io.ci_lockset.
1079 * - once locks for all layers were collected, they are sorted to avoid
1080 * dead-locks (cl_io_locks_sort()), and enqueued.
1082 * - when all locks are acquired, IO is performed;
1084 * - locks are released after IO is complete.
1086 * Striping introduces major additional complexity into locking. The
1087 * fundamental problem is that it is generally unsafe to actively use (hold)
1088 * two locks on the different OST servers at the same time, as this introduces
1089 * inter-server dependency and can lead to cascading evictions.
1091 * Basic solution is to sub-divide large read/write IOs into smaller pieces so
1092 * that no multi-stripe locks are taken (note that this design abandons POSIX
1093 * read/write semantics). Such pieces ideally can be executed concurrently. At
1094 * the same time, certain types of IO cannot be sub-divived, without
1095 * sacrificing correctness. This includes:
1097 * - O_APPEND write, where [0, EOF] lock has to be taken, to guarantee
1100 * - ftruncate(fd, offset), where [offset, EOF] lock has to be taken.
1102 * Also, in the case of read(fd, buf, count) or write(fd, buf, count), where
1103 * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
1104 * has to be held together with the usual lock on [offset, offset + count].
1106 * Interaction with DLM
1108 * In the expected setup, cl_lock is ultimately backed up by a collection of
1109 * DLM locks (struct ldlm_lock). Association between cl_lock and DLM lock is
1110 * implemented in osc layer, that also matches DLM events (ASTs, cancellation,
1111 * etc.) into cl_lock_operation calls. See struct osc_lock for a more detailed
1112 * description of interaction with DLM.
1118 struct cl_lock_descr {
1119 /** Object this lock is granted for. */
1120 struct cl_object *cld_obj;
1121 /** Index of the first page protected by this lock. */
1123 /** Index of the last page (inclusive) protected by this lock. */
1125 /** Group ID, for group lock */
1128 enum cl_lock_mode cld_mode;
1130 * flags to enqueue lock. A combination of bit-flags from
1131 * enum cl_enq_flags.
1133 __u32 cld_enq_flags;
1136 #define DDESCR "%s(%d):[%lu, %lu]:%x"
1137 #define PDESCR(descr) \
1138 cl_lock_mode_name((descr)->cld_mode), (descr)->cld_mode, \
1139 (descr)->cld_start, (descr)->cld_end, (descr)->cld_enq_flags
1141 const char *cl_lock_mode_name(const enum cl_lock_mode mode);
1144 * Layered client lock.
1147 /** List of slices. Immutable after creation. */
1148 struct list_head cll_layers;
1149 /** lock attribute, extent, cl_object, etc. */
1150 struct cl_lock_descr cll_descr;
1154 * Per-layer part of cl_lock
1156 * \see lov_lock, osc_lock
1158 struct cl_lock_slice {
1159 struct cl_lock *cls_lock;
1160 /** Object slice corresponding to this lock slice. Immutable after
1162 struct cl_object *cls_obj;
1163 const struct cl_lock_operations *cls_ops;
1164 /** Linkage into cl_lock::cll_layers. Immutable after creation. */
1165 struct list_head cls_linkage;
1170 * \see lov_lock_ops, osc_lock_ops
1172 struct cl_lock_operations {
1175 * Attempts to enqueue the lock. Called top-to-bottom.
1177 * \retval 0 this layer has enqueued the lock successfully
1178 * \retval >0 this layer has enqueued the lock, but need to wait on
1179 * @anchor for resources
1180 * \retval -ve failure
1182 * \see lov_lock_enqueue(), osc_lock_enqueue()
1184 int (*clo_enqueue)(const struct lu_env *env,
1185 const struct cl_lock_slice *slice,
1186 struct cl_io *io, struct cl_sync_io *anchor);
1188 * Cancel a lock, release its DLM lock ref, while does not cancel the
1191 void (*clo_cancel)(const struct lu_env *env,
1192 const struct cl_lock_slice *slice);
1195 * Destructor. Frees resources and the slice.
1197 * \see lov_lock_fini(), osc_lock_fini()
1199 void (*clo_fini)(const struct lu_env *env, struct cl_lock_slice *slice);
1201 * Optional debugging helper. Prints given lock slice.
1203 int (*clo_print)(const struct lu_env *env,
1204 void *cookie, lu_printer_t p,
1205 const struct cl_lock_slice *slice);
1208 #define CL_LOCK_DEBUG(mask, env, lock, format, ...) \
1210 if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
1211 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
1212 cl_lock_print(env, &msgdata, lu_cdebug_printer, lock); \
1213 CDEBUG(mask, format , ## __VA_ARGS__); \
1217 #define CL_LOCK_ASSERT(expr, env, lock) do { \
1221 CL_LOCK_DEBUG(D_ERROR, env, lock, "failed at %s.\n", #expr); \
1227 /** \addtogroup cl_page_list cl_page_list
1228 * Page list used to perform collective operations on a group of pages.
1230 * Pages are added to the list one by one. cl_page_list acquires a reference
1231 * for every page in it. Page list is used to perform collective operations on
1234 * - submit pages for an immediate transfer,
1236 * - own pages on behalf of certain io (waiting for each page in turn),
1240 * When list is finalized, it releases references on all pages it still has.
1242 * \todo XXX concurrency control.
1246 struct cl_page_list {
1248 struct list_head pl_pages;
1252 * A 2-queue of pages. A convenience data-type for common use case, 2-queue
1253 * contains an incoming page list and an outgoing page list.
1256 struct cl_page_list c2_qin;
1257 struct cl_page_list c2_qout;
1260 /** @} cl_page_list */
1262 /** \addtogroup cl_io cl_io
1267 * cl_io represents a high level I/O activity like
1268 * read(2)/write(2)/truncate(2) system call, or cancellation of an extent
1271 * cl_io is a layered object, much like cl_{object,page,lock} but with one
1272 * important distinction. We want to minimize number of calls to the allocator
1273 * in the fast path, e.g., in the case of read(2) when everything is cached:
1274 * client already owns the lock over region being read, and data are cached
1275 * due to read-ahead. To avoid allocation of cl_io layers in such situations,
1276 * per-layer io state is stored in the session, associated with the io, see
1277 * struct {vvp,lov,osc}_io for example. Sessions allocation is amortized
1278 * by using free-lists, see cl_env_get().
1280 * There is a small predefined number of possible io types, enumerated in enum
1283 * cl_io is a state machine, that can be advanced concurrently by the multiple
1284 * threads. It is up to these threads to control the concurrency and,
1285 * specifically, to detect when io is done, and its state can be safely
1288 * For read/write io overall execution plan is as following:
1290 * (0) initialize io state through all layers;
1292 * (1) loop: prepare chunk of work to do
1294 * (2) call all layers to collect locks they need to process current chunk
1296 * (3) sort all locks to avoid dead-locks, and acquire them
1298 * (4) process the chunk: call per-page methods
1299 * cl_io_operations::cio_prepare_write(),
1300 * cl_io_operations::cio_commit_write() for write)
1306 * To implement the "parallel IO mode", lov layer creates sub-io's (lazily to
1307 * address allocation efficiency issues mentioned above), and returns with the
1308 * special error condition from per-page method when current sub-io has to
1309 * block. This causes io loop to be repeated, and lov switches to the next
1310 * sub-io in its cl_io_operations::cio_iter_init() implementation.
1315 /** read system call */
1317 /** write system call */
1319 /** truncate, utime system calls */
1321 /** get data version */
1324 * page fault handling
1328 * fsync system call handling
1329 * To write out a range of file
1333 * glimpse. An io context to acquire glimpse lock.
1337 * Miscellaneous io. This is used for occasional io activity that
1338 * doesn't fit into other types. Currently this is used for:
1340 * - cancellation of an extent lock. This io exists as a context
1341 * to write dirty pages from under the lock being canceled back
1344 * - VM induced page write-out. An io context for writing page out
1345 * for memory cleansing;
1347 * - grouplock. An io context to acquire group lock.
1349 * CIT_MISC io is used simply as a context in which locks and pages
1350 * are manipulated. Such io has no internal "process", that is,
1351 * cl_io_loop() is never called for it.
1356 * To give advice about access of a file
1360 * SEEK_HOLE/SEEK_DATA handling to search holes or data
1361 * across all file objects
1368 * States of cl_io state machine
1371 /** Not initialized. */
1375 /** IO iteration started. */
1379 /** Actual IO is in progress. */
1381 /** IO for the current iteration finished. */
1383 /** Locks released. */
1385 /** Iteration completed. */
1387 /** cl_io finalized. */
1392 * IO state private for a layer.
1394 * This is usually embedded into layer session data, rather than allocated
1397 * \see vvp_io, lov_io, osc_io
1399 struct cl_io_slice {
1400 struct cl_io *cis_io;
1401 /** corresponding object slice. Immutable after creation. */
1402 struct cl_object *cis_obj;
1403 /** io operations. Immutable after creation. */
1404 const struct cl_io_operations *cis_iop;
1406 * linkage into a list of all slices for a given cl_io, hanging off
1407 * cl_io::ci_layers. Immutable after creation.
1409 struct list_head cis_linkage;
1412 typedef void (*cl_commit_cbt)(const struct lu_env *, struct cl_io *,
1415 struct cl_read_ahead {
1416 /* Maximum page index the readahead window will end.
1417 * This is determined DLM lock coverage, RPC and stripe boundary.
1418 * cra_end is included. */
1419 pgoff_t cra_end_idx;
1420 /* optimal RPC size for this read, by pages */
1421 unsigned long cra_rpc_pages;
1422 /* Release callback. If readahead holds resources underneath, this
1423 * function should be called to release it. */
1424 void (*cra_release)(const struct lu_env *env,
1425 struct cl_read_ahead *ra);
1427 /* Callback data for cra_release routine */
1431 /* whether lock is in contention */
1432 bool cra_contention;
1435 static inline void cl_read_ahead_release(const struct lu_env *env,
1436 struct cl_read_ahead *ra)
1438 if (ra->cra_release != NULL)
1439 ra->cra_release(env, ra);
1440 memset(ra, 0, sizeof(*ra));
1445 * Per-layer io operations.
1446 * \see vvp_io_ops, lov_io_ops, lovsub_io_ops, osc_io_ops
1448 struct cl_io_operations {
1450 * Vector of io state transition methods for every io type.
1452 * \see cl_page_operations::io
1456 * Prepare io iteration at a given layer.
1458 * Called top-to-bottom at the beginning of each iteration of
1459 * "io loop" (if it makes sense for this type of io). Here
1460 * layer selects what work it will do during this iteration.
1462 * \see cl_io_operations::cio_iter_fini()
1464 int (*cio_iter_init) (const struct lu_env *env,
1465 const struct cl_io_slice *slice);
1467 * Finalize io iteration.
1469 * Called bottom-to-top at the end of each iteration of "io
1470 * loop". Here layers can decide whether IO has to be
1473 * \see cl_io_operations::cio_iter_init()
1475 void (*cio_iter_fini) (const struct lu_env *env,
1476 const struct cl_io_slice *slice);
1478 * Collect locks for the current iteration of io.
1480 * Called top-to-bottom to collect all locks necessary for
1481 * this iteration. This methods shouldn't actually enqueue
1482 * anything, instead it should post a lock through
1483 * cl_io_lock_add(). Once all locks are collected, they are
1484 * sorted and enqueued in the proper order.
1486 int (*cio_lock) (const struct lu_env *env,
1487 const struct cl_io_slice *slice);
1489 * Finalize unlocking.
1491 * Called bottom-to-top to finish layer specific unlocking
1492 * functionality, after generic code released all locks
1493 * acquired by cl_io_operations::cio_lock().
1495 void (*cio_unlock)(const struct lu_env *env,
1496 const struct cl_io_slice *slice);
1498 * Start io iteration.
1500 * Once all locks are acquired, called top-to-bottom to
1501 * commence actual IO. In the current implementation,
1502 * top-level vvp_io_{read,write}_start() does all the work
1503 * synchronously by calling generic_file_*(), so other layers
1504 * are called when everything is done.
1506 int (*cio_start)(const struct lu_env *env,
1507 const struct cl_io_slice *slice);
1509 * Called top-to-bottom at the end of io loop. Here layer
1510 * might wait for an unfinished asynchronous io.
1512 void (*cio_end) (const struct lu_env *env,
1513 const struct cl_io_slice *slice);
1515 * Called bottom-to-top to notify layers that read/write IO
1516 * iteration finished, with \a nob bytes transferred.
1518 void (*cio_advance)(const struct lu_env *env,
1519 const struct cl_io_slice *slice,
1522 * Called once per io, bottom-to-top to release io resources.
1524 void (*cio_fini) (const struct lu_env *env,
1525 const struct cl_io_slice *slice);
1529 * Submit pages from \a queue->c2_qin for IO, and move
1530 * successfully submitted pages into \a queue->c2_qout. Return
1531 * non-zero if failed to submit even the single page. If
1532 * submission failed after some pages were moved into \a
1533 * queue->c2_qout, completion callback with non-zero ioret is
1536 int (*cio_submit)(const struct lu_env *env,
1537 const struct cl_io_slice *slice,
1538 enum cl_req_type crt,
1539 struct cl_2queue *queue);
1541 * Queue async page for write.
1542 * The difference between cio_submit and cio_queue is that
1543 * cio_submit is for urgent request.
1545 int (*cio_commit_async)(const struct lu_env *env,
1546 const struct cl_io_slice *slice,
1547 struct cl_page_list *queue, int from, int to,
1550 * Release active extent.
1552 void (*cio_extent_release)(const struct lu_env *env,
1553 const struct cl_io_slice *slice);
1555 * Decide maximum read ahead extent
1557 * \pre io->ci_type == CIT_READ
1559 int (*cio_read_ahead)(const struct lu_env *env,
1560 const struct cl_io_slice *slice,
1561 pgoff_t start, struct cl_read_ahead *ra);
1564 * Reserve LRU slots before IO.
1566 int (*cio_lru_reserve) (const struct lu_env *env,
1567 const struct cl_io_slice *slice,
1568 loff_t pos, size_t bytes);
1570 * Optional debugging helper. Print given io slice.
1572 int (*cio_print)(const struct lu_env *env, void *cookie,
1573 lu_printer_t p, const struct cl_io_slice *slice);
1577 * Flags to lock enqueue procedure.
1582 * instruct server to not block, if conflicting lock is found. Instead
1583 * -EAGAIN is returned immediately.
1585 CEF_NONBLOCK = 0x00000001,
1587 * Tell lower layers this is a glimpse request, translated to
1588 * LDLM_FL_HAS_INTENT at LDLM layer.
1590 * Also, because glimpse locks never block other locks, we count this
1591 * as automatically compatible with other osc locks.
1592 * (see osc_lock_compatible)
1594 CEF_GLIMPSE = 0x00000002,
1596 * tell the server to instruct (though a flag in the blocking ast) an
1597 * owner of the conflicting lock, that it can drop dirty pages
1598 * protected by this lock, without sending them to the server.
1600 CEF_DISCARD_DATA = 0x00000004,
1602 * tell the sub layers that it must be a `real' lock. This is used for
1603 * mmapped-buffer locks, glimpse locks, manually requested locks
1604 * (LU_LADVISE_LOCKAHEAD) that must never be converted into lockless
1607 * \see vvp_mmap_locks(), cl_glimpse_lock, cl_request_lock().
1609 CEF_MUST = 0x00000008,
1611 * tell the sub layers that never request a `real' lock. This flag is
1612 * not used currently.
1614 * cl_io::ci_lockreq and CEF_{MUST,NEVER} flags specify lockless
1615 * conversion policy: ci_lockreq describes generic information of lock
1616 * requirement for this IO, especially for locks which belong to the
1617 * object doing IO; however, lock itself may have precise requirements
1618 * that are described by the enqueue flags.
1620 CEF_NEVER = 0x00000010,
1622 * tell the dlm layer this is a speculative lock request
1623 * speculative lock requests are locks which are not requested as part
1624 * of an I/O operation. Instead, they are requested because we expect
1625 * to use them in the future. They are requested asynchronously at the
1628 * Currently used for asynchronous glimpse locks and manually requested
1629 * locks (LU_LADVISE_LOCKAHEAD).
1631 CEF_SPECULATIVE = 0x00000020,
1633 * enqueue a lock to test DLM lock existence.
1635 CEF_PEEK = 0x00000040,
1637 * Lock match only. Used by group lock in I/O as group lock
1638 * is known to exist.
1640 CEF_LOCK_MATCH = 0x00000080,
1642 * tell the DLM layer to lock only the requested range
1644 CEF_LOCK_NO_EXPAND = 0x00000100,
1646 * mask of enq_flags.
1648 CEF_MASK = 0x000001ff,
1652 * Link between lock and io. Intermediate structure is needed, because the
1653 * same lock can be part of multiple io's simultaneously.
1655 struct cl_io_lock_link {
1656 /** linkage into one of cl_lockset lists. */
1657 struct list_head cill_linkage;
1658 struct cl_lock cill_lock;
1659 /** optional destructor */
1660 void (*cill_fini)(const struct lu_env *env,
1661 struct cl_io_lock_link *link);
1663 #define cill_descr cill_lock.cll_descr
1666 * Lock-set represents a collection of locks, that io needs at a
1667 * time. Generally speaking, client tries to avoid holding multiple locks when
1670 * - holding extent locks over multiple ost's introduces the danger of
1671 * "cascading timeouts";
1673 * - holding multiple locks over the same ost is still dead-lock prone,
1674 * see comment in osc_lock_enqueue(),
1676 * but there are certain situations where this is unavoidable:
1678 * - O_APPEND writes have to take [0, EOF] lock for correctness;
1680 * - truncate has to take [new-size, EOF] lock for correctness;
1682 * - SNS has to take locks across full stripe for correctness;
1684 * - in the case when user level buffer, supplied to {read,write}(file0),
1685 * is a part of a memory mapped lustre file, client has to take a dlm
1686 * locks on file0, and all files that back up the buffer (or a part of
1687 * the buffer, that is being processed in the current chunk, in any
1688 * case, there are situations where at least 2 locks are necessary).
1690 * In such cases we at least try to take locks in the same consistent
1691 * order. To this end, all locks are first collected, then sorted, and then
1695 /** locks to be acquired. */
1696 struct list_head cls_todo;
1697 /** locks acquired. */
1698 struct list_head cls_done;
1702 * Lock requirements(demand) for IO. It should be cl_io_lock_req,
1703 * but 'req' is always to be thought as 'request' :-)
1705 enum cl_io_lock_dmd {
1706 /** Always lock data (e.g., O_APPEND). */
1708 /** Layers are free to decide between local and global locking. */
1710 /** Never lock: there is no cache (e.g., liblustre). */
1714 enum cl_fsync_mode {
1715 /** start writeback, do not wait for them to finish */
1717 /** start writeback and wait for them to finish */
1719 /** discard all of dirty pages in a specific file range */
1720 CL_FSYNC_DISCARD = 2,
1721 /** start writeback and make sure they have reached storage before
1722 * return. OST_SYNC RPC must be issued and finished */
1724 /** start writeback, thus the kernel can reclaim some memory */
1725 CL_FSYNC_RECLAIM = 4,
1728 struct cl_io_rw_common {
1733 enum cl_setattr_subtype {
1734 /** regular setattr **/
1738 /** fallocate(2) - mode preallocate **/
1739 CL_SETATTR_FALLOCATE
1742 struct cl_io_range {
1748 struct cl_io_pt *cip_next;
1749 struct kiocb cip_iocb;
1750 struct iov_iter cip_iter;
1751 struct file *cip_file;
1752 enum cl_io_type cip_iot;
1753 unsigned int cip_need_restart:1;
1762 * cl_io is shared by all threads participating in this IO (in current
1763 * implementation only one thread advances IO, but parallel IO design and
1764 * concurrent copy_*_user() require multiple threads acting on the same IO. It
1765 * is up to these threads to serialize their activities, including updates to
1766 * mutable cl_io fields.
1769 /** type of this IO. Immutable after creation. */
1770 enum cl_io_type ci_type;
1771 /** current state of cl_io state machine. */
1772 enum cl_io_state ci_state;
1773 /** main object this io is against. Immutable after creation. */
1774 struct cl_object *ci_obj;
1775 /** top level dio_aio */
1776 struct cl_dio_aio *ci_dio_aio;
1778 * Upper layer io, of which this io is a part of. Immutable after
1781 struct cl_io *ci_parent;
1782 /** List of slices. Immutable after creation. */
1783 struct list_head ci_layers;
1784 /** list of locks (to be) acquired by this io. */
1785 struct cl_lockset ci_lockset;
1786 /** lock requirements, this is just a help info for sublayers. */
1787 enum cl_io_lock_dmd ci_lockreq;
1788 /** layout version when this IO occurs */
1789 __u32 ci_layout_version;
1792 struct cl_io_rw_common rd;
1795 struct cl_io_rw_common wr;
1799 struct cl_io_rw_common ci_rw;
1800 struct cl_setattr_io {
1801 struct ost_lvb sa_attr;
1802 unsigned int sa_attr_flags;
1803 unsigned int sa_avalid; /* ATTR_* */
1804 unsigned int sa_xvalid; /* OP_XVALID */
1805 int sa_stripe_index;
1806 struct ost_layout sa_layout;
1807 const struct lu_fid *sa_parent_fid;
1808 /* SETATTR interface is used for regular setattr, */
1809 /* truncate(2) and fallocate(2) subtypes */
1810 enum cl_setattr_subtype sa_subtype;
1811 /* The following are used for fallocate(2) */
1813 loff_t sa_falloc_offset;
1814 loff_t sa_falloc_end;
1815 uid_t sa_falloc_uid;
1816 gid_t sa_falloc_gid;
1817 __u32 sa_falloc_projid;
1819 struct cl_data_version_io {
1820 u64 dv_data_version;
1821 u32 dv_layout_version;
1824 struct cl_fault_io {
1825 /** page index within file. */
1827 /** bytes valid byte on a faulted page. */
1829 /** writable page? for nopage() only */
1831 /** page of an executable? */
1833 /** page_mkwrite() */
1835 /** resulting page */
1836 struct cl_page *ft_page;
1838 struct cl_fsync_io {
1841 /** file system level fid */
1842 struct lu_fid *fi_fid;
1843 enum cl_fsync_mode fi_mode;
1844 /* how many pages were written/discarded */
1845 unsigned int fi_nr_written;
1847 struct cl_ladvise_io {
1850 /** file system level fid */
1851 struct lu_fid *li_fid;
1852 enum lu_ladvise_type li_advice;
1855 struct cl_lseek_io {
1861 time64_t lm_next_rpc_time;
1864 struct cl_2queue ci_queue;
1867 unsigned int ci_continue:1,
1869 * This io has held grouplock, to inform sublayers that
1870 * don't do lockless i/o.
1874 * The whole IO need to be restarted because layout has been changed
1878 * to not refresh layout - the IO issuer knows that the layout won't
1879 * change(page operations, layout change causes all page to be
1880 * discarded), or it doesn't matter if it changes(sync).
1884 * Need MDS intervention to complete a write.
1885 * Write intent is required for the following cases:
1886 * 1. component being written is not initialized, or
1887 * 2. the mirrored files are NOT in WRITE_PENDING state.
1889 ci_need_write_intent:1,
1891 * Check if layout changed after the IO finishes. Mainly for HSM
1892 * requirement. If IO occurs to openning files, it doesn't need to
1893 * verify layout because HSM won't release openning files.
1894 * Right now, only two opertaions need to verify layout: glimpse
1899 * file is released, restore has to to be triggered by vvp layer
1901 ci_restore_needed:1,
1906 /* Tell sublayers not to expand LDLM locks requested for this IO */
1907 ci_lock_no_expand:1,
1909 * Set if non-delay RPC should be used for this IO.
1911 * If this file has multiple mirrors, and if the OSTs of the current
1912 * mirror is inaccessible, non-delay RPC would error out quickly so
1913 * that the upper layer can try to access the next mirror.
1917 * Set if IO is triggered by async workqueue readahead.
1919 ci_async_readahead:1,
1921 * Ignore lockless and do normal locking for this io.
1925 * Set if we've tried all mirrors for this read IO, if it's not set,
1926 * the read IO will check to-be-read OSCs' status, and make fast-switch
1927 * another mirror if some of the OSTs are not healthy.
1929 ci_tried_all_mirrors:1,
1931 * Random read hints, readahead will be disabled.
1935 * Sequential read hints.
1939 * Do parallel (async) submission of DIO RPCs. Note DIO is still sync
1940 * to userspace, only the RPCs are submitted async, then waited for at
1941 * the llite layer before returning.
1945 * this DIO is at least partly unaligned, and so the unaligned DIO
1946 * path is being used for this entire IO
1950 * there is a compat issue with unupgraded ZFS targets which means we
1951 * must refuse to do unaligned DIO to these targets, so this is used
1952 * to annotate that in the IO (since we learn if there is a problematic
1953 * OST/MDT target as we build the IO)
1955 ci_allow_unaligned_dio:1,
1957 * Bypass quota check
1961 * io_uring direct IO with flags IOCB_NOWAIT.
1965 * The filesystem must exclusively acquire invalidate_lock before
1966 * invalidating page cache in truncate / hole punch / DLM extent
1967 * lock blocking AST path (and thus calling into ->invalidatepage)
1968 * to block races between page cache invalidation and page cache
1969 * filling functions (fault, read, ...)
1971 ci_invalidate_page_cache:1;
1974 * How many times the read has retried before this one.
1975 * Set by the top level and consumed by the LOV.
1977 unsigned ci_ndelay_tried;
1979 * Designated mirror index for this I/O.
1981 unsigned ci_designated_mirror;
1983 * Number of pages owned by this IO. For invariant checking.
1985 unsigned ci_owned_nr;
1987 * Range of write intent. Valid if ci_need_write_intent is set.
1989 struct lu_extent ci_write_intent;
1995 * Per-transfer attributes.
1997 struct cl_req_attr {
1998 enum cl_req_type cra_type;
2000 struct cl_page *cra_page;
2001 /** Generic attributes for the server consumption. */
2002 struct obdo *cra_oa;
2004 char cra_jobid[LUSTRE_JOBID_SIZE];
2005 /** uid/gid of the process doing an io */
2010 enum cache_stats_item {
2011 /** how many cache lookups were performed */
2013 /** how many times cache lookup resulted in a hit */
2015 /** how many entities are in the cache right now */
2017 /** how many entities in the cache are actively used (and cannot be
2018 * evicted) right now */
2020 /** how many entities were created at all */
2025 #define CS_NAMES { "lookup", "hit", "total", "busy", "create" }
2028 * Stats for a generic cache (similar to inode, lu_object, etc. caches).
2030 struct cache_stats {
2031 const char *cs_name;
2032 atomic_t cs_stats[CS_NR];
2035 /** These are not exported so far */
2036 void cache_stats_init (struct cache_stats *cs, const char *name);
2039 * Client-side site. This represents particular client stack. "Global"
2040 * variables should (directly or indirectly) be added here to allow multiple
2041 * clients to co-exist in the single address space.
2044 struct lu_site cs_lu;
2046 * Statistical counters. Atomics do not scale, something better like
2047 * per-cpu counters is needed.
2049 * These are exported as /proc/fs/lustre/llite/.../site
2051 * When interpreting keep in mind that both sub-locks (and sub-pages)
2052 * and top-locks (and top-pages) are accounted here.
2054 struct cache_stats cs_pages;
2055 atomic_t cs_pages_state[CPS_NR];
2058 int cl_site_init(struct cl_site *s, struct cl_device *top);
2059 void cl_site_fini(struct cl_site *s);
2060 void cl_stack_fini(const struct lu_env *env, struct cl_device *cl);
2063 * Output client site statistical counters into a buffer. Suitable for
2064 * ll_rd_*()-style functions.
2066 int cl_site_stats_print(const struct cl_site *site, struct seq_file *m);
2071 * Type conversion and accessory functions.
2075 static inline struct cl_site *lu2cl_site(const struct lu_site *site)
2077 return container_of(site, struct cl_site, cs_lu);
2080 static inline struct cl_device *lu2cl_dev(const struct lu_device *d)
2082 LASSERT(d == NULL || IS_ERR(d) || lu_device_is_cl(d));
2083 return container_of_safe(d, struct cl_device, cd_lu_dev);
2086 static inline struct lu_device *cl2lu_dev(struct cl_device *d)
2088 return &d->cd_lu_dev;
2091 static inline struct cl_object *lu2cl(const struct lu_object *o)
2093 LASSERT(o == NULL || IS_ERR(o) || lu_device_is_cl(o->lo_dev));
2094 return container_of_safe(o, struct cl_object, co_lu);
2097 static inline const struct cl_object_conf *
2098 lu2cl_conf(const struct lu_object_conf *conf)
2100 return container_of_safe(conf, struct cl_object_conf, coc_lu);
2103 static inline struct cl_object *cl_object_next(const struct cl_object *obj)
2105 return obj ? lu2cl(lu_object_next(&obj->co_lu)) : NULL;
2108 static inline struct cl_object_header *luh2coh(const struct lu_object_header *h)
2110 return container_of_safe(h, struct cl_object_header, coh_lu);
2113 static inline struct cl_site *cl_object_site(const struct cl_object *obj)
2115 return lu2cl_site(obj->co_lu.lo_dev->ld_site);
2119 struct cl_object_header *cl_object_header(const struct cl_object *obj)
2121 return luh2coh(obj->co_lu.lo_header);
2124 static inline int cl_device_init(struct cl_device *d, struct lu_device_type *t)
2126 return lu_device_init(&d->cd_lu_dev, t);
2129 static inline void cl_device_fini(struct cl_device *d)
2131 lu_device_fini(&d->cd_lu_dev);
2134 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
2135 struct cl_object *obj,
2136 const struct cl_page_operations *ops);
2137 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
2138 struct cl_object *obj,
2139 const struct cl_lock_operations *ops);
2140 void cl_io_slice_add(struct cl_io *io, struct cl_io_slice *slice,
2141 struct cl_object *obj, const struct cl_io_operations *ops);
2144 /** \defgroup cl_object cl_object
2146 struct cl_object *cl_object_top (struct cl_object *o);
2147 struct cl_object *cl_object_find(const struct lu_env *env, struct cl_device *cd,
2148 const struct lu_fid *fid,
2149 const struct cl_object_conf *c);
2151 int cl_object_header_init(struct cl_object_header *h);
2152 void cl_object_header_fini(struct cl_object_header *h);
2153 void cl_object_put (const struct lu_env *env, struct cl_object *o);
2154 void cl_object_get (struct cl_object *o);
2155 void cl_object_attr_lock (struct cl_object *o);
2156 void cl_object_attr_unlock(struct cl_object *o);
2157 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
2158 struct cl_attr *attr);
2159 int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
2160 const struct cl_attr *attr, unsigned valid);
2161 void cl_object_dirty_for_sync(const struct lu_env *env, struct cl_object *obj);
2162 int cl_object_glimpse (const struct lu_env *env, struct cl_object *obj,
2163 struct ost_lvb *lvb);
2164 int cl_conf_set (const struct lu_env *env, struct cl_object *obj,
2165 const struct cl_object_conf *conf);
2166 int cl_object_prune (const struct lu_env *env, struct cl_object *obj);
2167 void cl_object_kill (const struct lu_env *env, struct cl_object *obj);
2168 int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2169 struct lov_user_md __user *lum, size_t size);
2170 int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
2171 struct ll_fiemap_info_key *fmkey, struct fiemap *fiemap,
2173 int cl_object_layout_get(const struct lu_env *env, struct cl_object *obj,
2174 struct cl_layout *cl);
2175 loff_t cl_object_maxbytes(struct cl_object *obj);
2176 int cl_object_flush(const struct lu_env *env, struct cl_object *obj,
2177 struct ldlm_lock *lock);
2178 int cl_object_inode_ops(const struct lu_env *env, struct cl_object *obj,
2179 enum coo_inode_opc opc, void *data);
2183 * Returns true, iff \a o0 and \a o1 are slices of the same object.
2185 static inline int cl_object_same(struct cl_object *o0, struct cl_object *o1)
2187 return cl_object_header(o0) == cl_object_header(o1);
2190 static inline void cl_object_page_init(struct cl_object *clob, int size)
2192 clob->co_slice_off = cl_object_header(clob)->coh_page_bufsize;
2193 cl_object_header(clob)->coh_page_bufsize += round_up(size, 8);
2194 WARN_ON(cl_object_header(clob)->coh_page_bufsize > 512);
2197 static inline void *cl_object_page_slice(struct cl_object *clob,
2198 struct cl_page *page)
2200 return (void *)((char *)page + clob->co_slice_off);
2204 * Return refcount of cl_object.
2206 static inline int cl_object_refc(struct cl_object *clob)
2208 struct lu_object_header *header = clob->co_lu.lo_header;
2209 return atomic_read(&header->loh_ref);
2214 /** \defgroup cl_page cl_page
2216 struct cl_page *cl_page_find (const struct lu_env *env,
2217 struct cl_object *obj,
2218 pgoff_t idx, struct page *vmpage,
2219 enum cl_page_type type);
2220 struct cl_page *cl_page_alloc (const struct lu_env *env,
2221 struct cl_object *o, pgoff_t ind,
2222 struct page *vmpage,
2223 enum cl_page_type type);
2224 void cl_page_get (struct cl_page *page);
2225 void cl_page_put (const struct lu_env *env,
2226 struct cl_page *page);
2227 void cl_pagevec_put (const struct lu_env *env,
2228 struct cl_page *page,
2229 struct pagevec *pvec);
2230 void cl_page_print (const struct lu_env *env, void *cookie,
2231 lu_printer_t printer,
2232 const struct cl_page *pg);
2233 void cl_page_header_print(const struct lu_env *env, void *cookie,
2234 lu_printer_t printer,
2235 const struct cl_page *pg);
2236 struct cl_page *cl_vmpage_page (struct page *vmpage, struct cl_object *obj);
2241 * Functions dealing with the ownership of page by io.
2245 int cl_page_own (const struct lu_env *env,
2246 struct cl_io *io, struct cl_page *page);
2247 int cl_page_own_try (const struct lu_env *env,
2248 struct cl_io *io, struct cl_page *page);
2249 void cl_page_assume (const struct lu_env *env,
2250 struct cl_io *io, struct cl_page *page);
2251 void cl_page_unassume (const struct lu_env *env,
2252 struct cl_io *io, struct cl_page *pg);
2253 void cl_page_disown (const struct lu_env *env,
2254 struct cl_io *io, struct cl_page *page);
2255 int cl_page_is_owned (const struct cl_page *pg, const struct cl_io *io);
2262 * Functions dealing with the preparation of a page for a transfer, and
2263 * tracking transfer state.
2266 int cl_page_prep (const struct lu_env *env, struct cl_io *io,
2267 struct cl_page *pg, enum cl_req_type crt);
2268 void cl_page_completion (const struct lu_env *env,
2269 struct cl_page *pg, enum cl_req_type crt, int ioret);
2270 int cl_page_make_ready (const struct lu_env *env, struct cl_page *pg,
2271 enum cl_req_type crt);
2272 int cl_page_cache_add (const struct lu_env *env, struct cl_io *io,
2273 struct cl_page *pg, enum cl_req_type crt);
2274 void cl_page_clip (const struct lu_env *env, struct cl_page *pg,
2276 int cl_page_flush (const struct lu_env *env, struct cl_io *io,
2277 struct cl_page *pg);
2283 * \name helper routines
2284 * Functions to discard, delete and export a cl_page.
2287 void cl_page_discard(const struct lu_env *env, struct cl_io *io,
2288 struct cl_page *pg);
2289 void cl_page_delete(const struct lu_env *env, struct cl_page *pg);
2290 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
2293 void cl_lock_print(const struct lu_env *env, void *cookie,
2294 lu_printer_t printer, const struct cl_lock *lock);
2295 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2296 lu_printer_t printer,
2297 const struct cl_lock_descr *descr);
2301 * Data structure managing a client's cached pages. A count of
2302 * "unstable" pages is maintained, and an LRU of clean pages is
2303 * maintained. "unstable" pages are pages pinned by the ptlrpc
2304 * layer for recovery purposes.
2306 struct cl_client_cache {
2308 * # of client cache refcount
2309 * # of users (OSCs) + 2 (held by llite and lov)
2311 refcount_t ccc_users;
2313 * # of threads are doing shrinking
2315 unsigned int ccc_lru_shrinkers;
2317 * # of LRU entries available
2319 atomic_long_t ccc_lru_left;
2321 * List of entities(OSCs) for this LRU cache
2323 struct list_head ccc_lru;
2325 * Max # of LRU entries
2327 unsigned long ccc_lru_max;
2329 * Lock to protect ccc_lru list
2331 spinlock_t ccc_lru_lock;
2333 * Set if unstable check is enabled
2335 unsigned int ccc_unstable_check:1;
2337 * # of unstable pages for this mount point
2339 atomic_long_t ccc_unstable_nr;
2341 * Serialize max_cache_mb write operation
2343 struct mutex ccc_max_cache_mb_lock;
2346 * cl_cache functions
2348 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max);
2349 void cl_cache_incref(struct cl_client_cache *cache);
2350 void cl_cache_decref(struct cl_client_cache *cache);
2354 /** \defgroup cl_lock cl_lock
2356 int cl_lock_request(const struct lu_env *env, struct cl_io *io,
2357 struct cl_lock *lock);
2358 int cl_lock_init(const struct lu_env *env, struct cl_lock *lock,
2359 const struct cl_io *io);
2360 void cl_lock_fini(const struct lu_env *env, struct cl_lock *lock);
2361 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
2362 const struct lu_device_type *dtype);
2363 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock);
2365 int cl_lock_enqueue(const struct lu_env *env, struct cl_io *io,
2366 struct cl_lock *lock, struct cl_sync_io *anchor);
2367 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock);
2371 /** \defgroup cl_io cl_io
2374 int cl_io_init(const struct lu_env *env, struct cl_io *io,
2375 enum cl_io_type iot, struct cl_object *obj);
2376 int cl_io_sub_init(const struct lu_env *env, struct cl_io *io,
2377 enum cl_io_type iot, struct cl_object *obj);
2378 int cl_io_rw_init(const struct lu_env *env, struct cl_io *io,
2379 enum cl_io_type iot, loff_t pos, size_t bytes);
2380 int cl_io_loop(const struct lu_env *env, struct cl_io *io);
2382 void cl_io_fini(const struct lu_env *env, struct cl_io *io);
2383 int cl_io_iter_init(const struct lu_env *env, struct cl_io *io);
2384 void cl_io_iter_fini(const struct lu_env *env, struct cl_io *io);
2385 int cl_io_lock(const struct lu_env *env, struct cl_io *io);
2386 void cl_io_unlock(const struct lu_env *env, struct cl_io *io);
2387 int cl_io_start(const struct lu_env *env, struct cl_io *io);
2388 void cl_io_end(const struct lu_env *env, struct cl_io *io);
2389 int cl_io_lock_add(const struct lu_env *env, struct cl_io *io,
2390 struct cl_io_lock_link *link);
2391 int cl_io_lock_alloc_add(const struct lu_env *env, struct cl_io *io,
2392 struct cl_lock_descr *descr);
2393 int cl_io_submit_rw (const struct lu_env *env, struct cl_io *io,
2394 enum cl_req_type iot, struct cl_2queue *queue);
2395 int cl_io_submit_sync (const struct lu_env *env, struct cl_io *io,
2396 enum cl_req_type iot, struct cl_2queue *queue,
2398 int cl_io_commit_async (const struct lu_env *env, struct cl_io *io,
2399 struct cl_page_list *queue, int from, int to,
2401 void cl_io_extent_release (const struct lu_env *env, struct cl_io *io);
2402 int cl_io_lru_reserve(const struct lu_env *env, struct cl_io *io,
2403 loff_t pos, size_t bytes);
2404 int cl_io_read_ahead (const struct lu_env *env, struct cl_io *io,
2405 pgoff_t start, struct cl_read_ahead *ra);
2406 void cl_io_rw_advance(const struct lu_env *env, struct cl_io *io,
2410 * True, iff \a io is an O_APPEND write(2).
2412 static inline int cl_io_is_append(const struct cl_io *io)
2414 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_append;
2417 static inline int cl_io_is_sync_write(const struct cl_io *io)
2419 return io->ci_type == CIT_WRITE && io->u.ci_wr.wr_sync;
2422 static inline int cl_io_is_mkwrite(const struct cl_io *io)
2424 return io->ci_type == CIT_FAULT && io->u.ci_fault.ft_mkwrite;
2428 * True, iff \a io is a truncate(2).
2430 static inline int cl_io_is_trunc(const struct cl_io *io)
2432 return io->ci_type == CIT_SETATTR &&
2433 (io->u.ci_setattr.sa_avalid & ATTR_SIZE) &&
2434 (io->u.ci_setattr.sa_subtype != CL_SETATTR_FALLOCATE);
2437 static inline int cl_io_is_fallocate(const struct cl_io *io)
2439 return (io->ci_type == CIT_SETATTR) &&
2440 (io->u.ci_setattr.sa_subtype == CL_SETATTR_FALLOCATE);
2443 struct cl_io *cl_io_top(struct cl_io *io);
2445 #define CL_IO_SLICE_CLEAN(obj, base) memset_startat(obj, 0, base)
2449 /** \defgroup cl_page_list cl_page_list
2453 * Last page in the page list.
2455 static inline struct cl_page *cl_page_list_last(struct cl_page_list *plist)
2457 LASSERT(plist->pl_nr > 0);
2458 return list_entry(plist->pl_pages.prev, struct cl_page, cp_batch);
2461 static inline struct cl_page *cl_page_list_first(struct cl_page_list *plist)
2463 LASSERT(plist->pl_nr > 0);
2464 return list_first_entry(&plist->pl_pages, struct cl_page, cp_batch);
2468 * Iterate over pages in a page list.
2470 #define cl_page_list_for_each(page, list) \
2471 list_for_each_entry((page), &(list)->pl_pages, cp_batch)
2474 * Iterate over pages in a page list, taking possible removals into account.
2476 #define cl_page_list_for_each_safe(page, temp, list) \
2477 list_for_each_entry_safe((page), (temp), &(list)->pl_pages, cp_batch)
2479 void cl_page_list_init(struct cl_page_list *plist);
2480 void cl_page_list_add(struct cl_page_list *plist, struct cl_page *page,
2482 void cl_page_list_move(struct cl_page_list *dst, struct cl_page_list *src,
2483 struct cl_page *page);
2484 void cl_page_list_move_head(struct cl_page_list *dst, struct cl_page_list *src,
2485 struct cl_page *page);
2486 void cl_page_list_splice(struct cl_page_list *list,
2487 struct cl_page_list *head);
2488 void cl_page_list_del(const struct lu_env *env,
2489 struct cl_page_list *plist, struct cl_page *page);
2490 void cl_page_list_disown(const struct lu_env *env,
2491 struct cl_page_list *plist);
2492 void cl_page_list_assume(const struct lu_env *env,
2493 struct cl_io *io, struct cl_page_list *plist);
2494 void cl_page_list_discard(const struct lu_env *env,
2495 struct cl_io *io, struct cl_page_list *plist);
2496 void cl_page_list_fini(const struct lu_env *env, struct cl_page_list *plist);
2498 void cl_2queue_init(struct cl_2queue *queue);
2499 void cl_2queue_disown(const struct lu_env *env, struct cl_2queue *queue);
2500 void cl_2queue_assume(const struct lu_env *env, struct cl_io *io,
2501 struct cl_2queue *queue);
2502 void cl_2queue_discard(const struct lu_env *env, struct cl_io *io,
2503 struct cl_2queue *queue);
2504 void cl_2queue_fini(const struct lu_env *env, struct cl_2queue *queue);
2505 void cl_2queue_init_page(struct cl_2queue *queue, struct cl_page *page);
2507 /** @} cl_page_list */
2509 void cl_req_attr_set(const struct lu_env *env, struct cl_object *obj,
2510 struct cl_req_attr *attr);
2512 /** \defgroup cl_sync_io cl_sync_io
2519 typedef void (cl_sync_io_end_t)(const struct lu_env *, struct cl_sync_io *);
2521 void cl_sync_io_init_notify(struct cl_sync_io *anchor, int nr, void *dio_aio,
2522 cl_sync_io_end_t *end);
2524 int cl_sync_io_wait(const struct lu_env *env, struct cl_sync_io *anchor,
2526 void cl_sync_io_note(const struct lu_env *env, struct cl_sync_io *anchor,
2528 int cl_sync_io_wait_recycle(const struct lu_env *env, struct cl_sync_io *anchor,
2529 long timeout, int ioret);
2530 struct cl_dio_aio *cl_dio_aio_alloc(struct kiocb *iocb, struct cl_object *obj,
2532 struct cl_sub_dio *cl_sub_dio_alloc(struct cl_dio_aio *ll_aio,
2533 struct iov_iter *iter, bool write,
2534 bool unaligned, bool sync);
2535 void cl_dio_aio_free(const struct lu_env *env, struct cl_dio_aio *aio);
2536 void cl_sub_dio_free(struct cl_sub_dio *sdio);
2537 static inline void cl_sync_io_init(struct cl_sync_io *anchor, int nr)
2539 cl_sync_io_init_notify(anchor, nr, NULL, NULL);
2543 * Anchor for synchronous transfer. This is allocated on a stack by thread
2544 * doing synchronous transfer, and a pointer to this structure is set up in
2545 * every page submitted for transfer. Transfer completion routine updates
2546 * anchor and wakes up waiting thread when transfer is complete.
2549 /** number of pages yet to be transferred. */
2550 atomic_t csi_sync_nr;
2551 /** has this i/o completed? */
2552 atomic_t csi_complete;
2555 /** completion to be signaled when transfer is complete. */
2556 wait_queue_head_t csi_waitq;
2557 /** callback to invoke when this IO is finished */
2558 cl_sync_io_end_t *csi_end_io;
2559 /* private pointer for an associated DIO/AIO */
2563 /** direct IO pages */
2564 struct ll_dio_pages {
2566 * page array for RDMA - for aligned i/o, this is the user provided
2567 * pages, but for unaligned i/o, this is the internal buffer
2569 struct page **ldp_pages;
2570 /** # of pages in the array. */
2572 /* the file offset of the first page. */
2573 loff_t ldp_file_offset;
2576 /* Top level struct used for AIO and DIO */
2578 struct cl_sync_io cda_sync;
2579 struct cl_object *cda_obj;
2580 struct kiocb *cda_iocb;
2582 struct mm_struct *cda_mm;
2583 unsigned cda_no_aio_complete:1,
2587 /* Sub-dio used for splitting DIO (and AIO, because AIO is DIO) according to
2588 * the layout/striping, so we can do parallel submit of DIO RPCs
2591 struct cl_sync_io csd_sync;
2592 struct cl_page_list csd_pages;
2594 struct cl_dio_aio *csd_ll_aio;
2595 struct ll_dio_pages csd_dio_pages;
2596 struct iov_iter csd_iter;
2597 unsigned csd_creator_free:1,
2601 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW) || \
2602 defined(HAVE_DIRECTIO_2ARGS)
2603 #define HAVE_DIO_ITER 1
2606 void ll_release_user_pages(struct page **pages, int npages);
2607 int ll_allocate_dio_buffer(struct ll_dio_pages *pvec, size_t io_size);
2608 void ll_free_dio_buffer(struct ll_dio_pages *pvec);
2609 ssize_t ll_dio_user_copy(struct cl_sub_dio *sdio, struct iov_iter *write_iov);
2611 #ifndef HAVE_KTHREAD_USE_MM
2612 #define kthread_use_mm(mm) use_mm(mm)
2613 #define kthread_unuse_mm(mm) unuse_mm(mm)
2616 /** @} cl_sync_io */
2618 /** \defgroup cl_env cl_env
2620 * lu_env handling for a client.
2622 * lu_env is an environment within which lustre code executes. Its major part
2623 * is lu_context---a fast memory allocation mechanism that is used to conserve
2624 * precious kernel stack space. Originally lu_env was designed for a server,
2627 * - there is a (mostly) fixed number of threads, and
2629 * - call chains have no non-lustre portions inserted between lustre code.
2631 * On a client both these assumtpion fails, because every user thread can
2632 * potentially execute lustre code as part of a system call, and lustre calls
2633 * into VFS or MM that call back into lustre.
2635 * To deal with that, cl_env wrapper functions implement the following
2638 * - allocation and destruction of environment is amortized by caching no
2639 * longer used environments instead of destroying them;
2641 * \see lu_env, lu_context, lu_context_key
2644 struct lu_env *cl_env_get(__u16 *refcheck);
2645 struct lu_env *cl_env_alloc(__u16 *refcheck, __u32 tags);
2646 void cl_env_put(struct lu_env *env, __u16 *refcheck);
2647 unsigned cl_env_cache_purge(unsigned nr);
2648 struct lu_env *cl_env_percpu_get(void);
2649 void cl_env_percpu_put(struct lu_env *env);
2656 void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr);
2657 void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb);
2659 struct cl_device *cl_type_setup(const struct lu_env *env, struct lu_site *site,
2660 struct lu_device_type *ldt,
2661 struct lu_device *next);
2664 int cl_global_init(void);
2665 void cl_global_fini(void);
2667 #endif /* _LINUX_CL_OBJECT_H */