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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file is part of Lustre, http://www.lustre.org/
38 * Lustre is a trademark of Sun Microsystems, Inc.
40 * Internal interfaces of OSC layer.
42 * Author: Nikita Danilov <nikita.danilov@sun.com>
43 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
46 #ifndef OSC_CL_INTERNAL_H
47 #define OSC_CL_INTERNAL_H
50 # include <libcfs/libcfs.h>
52 # include <liblustre.h>
56 /* osc_build_res_name() */
58 #include <cl_object.h>
60 #include "osc_internal.h"
69 * State maintained by osc layer for each IO context.
73 struct cl_io_slice oi_cl;
74 /** true if this io is lockless. */
76 /** active extents, we know how many bytes is going to be written,
77 * so having an active extent will prevent it from being fragmented */
78 struct osc_extent *oi_active;
79 /** partially truncated extent, we need to hold this extent to prevent
80 * page writeback from happening. */
81 struct osc_extent *oi_trunc;
85 struct obd_info oi_info;
87 struct osc_async_cbargs {
90 struct completion opc_sync;
95 * State of transfer for osc.
98 struct cl_req_slice or_cl;
102 * State maintained by osc layer for the duration of a system call.
108 #define OTI_PVEC_SIZE 256
109 struct osc_thread_info {
110 struct ldlm_res_id oti_resname;
111 ldlm_policy_data_t oti_policy;
112 struct cl_lock_descr oti_descr;
113 struct cl_attr oti_attr;
114 struct lustre_handle oti_handle;
115 struct cl_page_list oti_plist;
117 void *oti_pvec[OTI_PVEC_SIZE];
119 * Fields used by cl_lock_discard_pages().
121 pgoff_t oti_next_index;
122 pgoff_t oti_fn_index; /* first non-overlapped index */
126 struct cl_object oo_cl;
127 struct lov_oinfo *oo_oinfo;
129 * True if locking against this stripe got -EUSERS.
132 cfs_time_t oo_contention_time;
133 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
135 * IO context used for invariant checks in osc_lock_has_pages().
137 struct cl_io oo_debug_io;
138 /** Serialization object for osc_object::oo_debug_io. */
139 struct mutex oo_debug_mutex;
142 * List of pages in transfer.
144 cfs_list_t oo_inflight[CRT_NR];
146 * Lock, protecting ccc_object::cob_inflight, because a seat-belt is
147 * locked during take-off and landing.
149 spinlock_t oo_seatbelt;
152 * used by the osc to keep track of what objects to build into rpcs.
153 * Protected by client_obd->cli_loi_list_lock.
155 cfs_list_t oo_ready_item;
156 cfs_list_t oo_hp_ready_item;
157 cfs_list_t oo_write_item;
158 cfs_list_t oo_read_item;
161 * extent is a red black tree to manage (async) dirty pages.
163 struct rb_root oo_root;
165 * Manage write(dirty) extents.
167 cfs_list_t oo_hp_exts; /* list of hp extents */
168 cfs_list_t oo_urgent_exts; /* list of writeback extents */
169 cfs_list_t oo_rpc_exts;
171 cfs_list_t oo_reading_exts;
173 cfs_atomic_t oo_nr_reads;
174 cfs_atomic_t oo_nr_writes;
176 /** Protect extent tree. Will be used to protect
177 * oo_{read|write}_pages soon. */
181 * Radix tree for caching pages
183 struct radix_tree_root oo_tree;
184 spinlock_t oo_tree_lock;
185 unsigned long oo_npages;
188 static inline void osc_object_lock(struct osc_object *obj)
190 spin_lock(&obj->oo_lock);
193 static inline int osc_object_trylock(struct osc_object *obj)
195 return spin_trylock(&obj->oo_lock);
198 static inline void osc_object_unlock(struct osc_object *obj)
200 spin_unlock(&obj->oo_lock);
203 static inline int osc_object_is_locked(struct osc_object *obj)
205 return spin_is_locked(&obj->oo_lock);
209 * Lock "micro-states" for osc layer.
211 enum osc_lock_state {
222 * osc-private state of cl_lock.
224 * Interaction with DLM.
226 * CLIO enqueues all DLM locks through ptlrpcd (that is, in "async" mode).
228 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
229 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_lock.
231 * This pointer is protected through a reference, acquired by
232 * osc_lock_upcall0(). Also, an additional reference is acquired by
233 * ldlm_lock_addref() call protecting the lock from cancellation, until
234 * osc_lock_unuse() releases it.
236 * Below is a description of how lock references are acquired and released
239 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
240 * - ldlm_lock_create()
241 * - ldlm_lock_new(): initializes a lock with 2 references. One for
242 * the caller (released when reply from the server is received, or on
243 * error), and another for the hash table.
244 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
246 * - When reply is received from the server (osc_enqueue_interpret())
247 * - ldlm_cli_enqueue_fini()
248 * - LDLM_LOCK_PUT(): releases caller reference acquired by
251 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
252 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
254 * - When lock is being cancelled (ldlm_lock_cancel())
255 * - ldlm_lock_destroy()
256 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
259 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
260 * either when lock is cancelled (osc_lock_blocking()), or when locks is
261 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
262 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
266 struct cl_lock_slice ols_cl;
267 /** underlying DLM lock */
268 struct ldlm_lock *ols_lock;
269 /** lock value block */
270 struct ost_lvb ols_lvb;
271 /** DLM flags with which osc_lock::ols_lock was enqueued */
273 /** osc_lock::ols_lock handle */
274 struct lustre_handle ols_handle;
275 struct ldlm_enqueue_info ols_einfo;
276 enum osc_lock_state ols_state;
279 * How many pages are using this lock for io, currently only used by
280 * read-ahead. If non-zero, the underlying dlm lock won't be cancelled
281 * during recovery to avoid deadlock. see bz16774.
283 * \see osc_page::ops_lock
284 * \see osc_page_addref_lock(), osc_page_putref_lock()
286 cfs_atomic_t ols_pageref;
289 * true, if ldlm_lock_addref() was called against
290 * osc_lock::ols_lock. This is used for sanity checking.
292 * \see osc_lock::ols_has_ref
294 unsigned ols_hold :1,
296 * this is much like osc_lock::ols_hold, except that this bit is
297 * cleared _after_ reference in released in osc_lock_unuse(). This
298 * fine distinction is needed because:
300 * - if ldlm lock still has a reference, osc_ast_data_get() needs
301 * to return associated cl_lock (so that a flag is needed that is
302 * cleared after ldlm_lock_decref() returned), and
304 * - ldlm_lock_decref() can invoke blocking ast (for a
305 * LDLM_FL_CBPENDING lock), and osc_lock functions like
306 * osc_lock_cancel() called from there need to know whether to
307 * release lock reference (so that a flag is needed that is
308 * cleared before ldlm_lock_decref() is called).
312 * inherit the lockless attribute from top level cl_io.
313 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
317 * set by osc_lock_use() to wait until blocking AST enters into
318 * osc_ldlm_blocking_ast0(), so that cl_lock mutex can be used for
319 * further synchronization.
323 * If the data of this lock has been flushed to server side.
327 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
328 * the EVAVAIL error as torerable, this will make upper logic happy
329 * to wait all glimpse locks to each OSTs to be completed.
330 * Glimpse lock converts to normal lock if the server lock is
332 * Glimpse lock should be destroyed immediately after use.
336 * For async glimpse lock.
340 * IO that owns this lock. This field is used for a dead-lock
341 * avoidance by osc_lock_enqueue_wait().
343 * XXX: unfortunately, the owner of a osc_lock is not unique,
344 * the lock may have multiple users, if the lock is granted and
347 struct osc_io *ols_owner;
352 * Page state private for osc layer.
355 struct cl_page_slice ops_cl;
357 * Page queues used by osc to detect when RPC can be formed.
359 struct osc_async_page ops_oap;
361 * An offset within page from which next transfer starts. This is used
362 * by cl_page_clip() to submit partial page transfers.
366 * An offset within page at which next transfer ends.
368 * \see osc_page::ops_from.
372 * Boolean, true iff page is under transfer. Used for sanity checking.
374 unsigned ops_transfer_pinned:1,
376 * True for a `temporary page' created by read-ahead code, probably
377 * outside of any DLM lock.
385 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
389 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
393 * Linkage into a per-osc_object list of pages in flight. For
396 cfs_list_t ops_inflight;
398 * Thread that submitted this page for transfer. For debugging.
400 struct task_struct *ops_submitter;
402 * Submit time - the time when the page is starting RPC. For debugging.
404 cfs_time_t ops_submit_time;
407 * A lock of which we hold a reference covers this page. Only used by
408 * read-ahead: for a readahead page, we hold it's covering lock to
409 * prevent it from being canceled during recovery.
411 * \see osc_lock::ols_pageref
412 * \see osc_page_addref_lock(), osc_page_putref_lock().
414 struct cl_lock *ops_lock;
417 extern struct kmem_cache *osc_lock_kmem;
418 extern struct kmem_cache *osc_object_kmem;
419 extern struct kmem_cache *osc_thread_kmem;
420 extern struct kmem_cache *osc_session_kmem;
421 extern struct kmem_cache *osc_req_kmem;
422 extern struct kmem_cache *osc_extent_kmem;
424 extern struct lu_device_type osc_device_type;
425 extern struct lu_context_key osc_key;
426 extern struct lu_context_key osc_session_key;
428 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
430 int osc_lock_init(const struct lu_env *env,
431 struct cl_object *obj, struct cl_lock *lock,
432 const struct cl_io *io);
433 int osc_io_init (const struct lu_env *env,
434 struct cl_object *obj, struct cl_io *io);
435 int osc_req_init (const struct lu_env *env, struct cl_device *dev,
437 struct lu_object *osc_object_alloc(const struct lu_env *env,
438 const struct lu_object_header *hdr,
439 struct lu_device *dev);
440 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
441 struct cl_page *page, struct page *vmpage);
443 void osc_index2policy (ldlm_policy_data_t *policy, const struct cl_object *obj,
444 pgoff_t start, pgoff_t end);
445 int osc_lvb_print (const struct lu_env *env, void *cookie,
446 lu_printer_t p, const struct ost_lvb *lvb);
448 void osc_lru_add_batch(struct client_obd *cli, cfs_list_t *list);
449 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
450 enum cl_req_type crt, int brw_flags);
451 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
452 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
453 obd_flag async_flags);
454 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
455 struct page *page, loff_t offset);
456 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
457 struct osc_page *ops);
458 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
459 struct osc_page *ops);
460 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
461 struct osc_page *ops);
462 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
463 cfs_list_t *list, int cmd, int brw_flags);
464 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
465 struct osc_object *obj, __u64 size);
466 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
467 struct osc_object *obj);
468 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
469 pgoff_t start, pgoff_t end, int hp, int discard);
470 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
471 pgoff_t start, pgoff_t end);
472 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
473 struct osc_object *osc, pdl_policy_t pol);
474 int lru_queue_work(const struct lu_env *env, void *data);
476 void osc_object_set_contended (struct osc_object *obj);
477 void osc_object_clear_contended(struct osc_object *obj);
478 int osc_object_is_contended (struct osc_object *obj);
480 int osc_lock_is_lockless (const struct osc_lock *olck);
482 /*****************************************************************************
488 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
490 struct osc_thread_info *info;
492 info = lu_context_key_get(&env->le_ctx, &osc_key);
493 LASSERT(info != NULL);
497 static inline struct osc_session *osc_env_session(const struct lu_env *env)
499 struct osc_session *ses;
501 ses = lu_context_key_get(env->le_ses, &osc_session_key);
502 LASSERT(ses != NULL);
506 static inline struct osc_io *osc_env_io(const struct lu_env *env)
508 return &osc_env_session(env)->os_io;
511 static inline int osc_is_object(const struct lu_object *obj)
513 return obj->lo_dev->ld_type == &osc_device_type;
516 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
518 LINVRNT(d->ld_type == &osc_device_type);
519 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
522 static inline struct obd_export *osc_export(const struct osc_object *obj)
524 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
527 static inline struct client_obd *osc_cli(const struct osc_object *obj)
529 return &osc_export(obj)->exp_obd->u.cli;
532 static inline struct osc_object *cl2osc(const struct cl_object *obj)
534 LINVRNT(osc_is_object(&obj->co_lu));
535 return container_of0(obj, struct osc_object, oo_cl);
538 static inline struct cl_object *osc2cl(const struct osc_object *obj)
540 return (struct cl_object *)&obj->oo_cl;
543 static inline ldlm_mode_t osc_cl_lock2ldlm(enum cl_lock_mode mode)
545 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
546 if (mode == CLM_READ)
548 else if (mode == CLM_WRITE)
554 static inline enum cl_lock_mode osc_ldlm2cl_lock(ldlm_mode_t mode)
556 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
559 else if (mode == LCK_PW)
565 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
567 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
568 return container_of0(slice, struct osc_page, ops_cl);
571 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
573 return container_of0(oap, struct osc_page, ops_oap);
576 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
578 return oap2osc(oap)->ops_cl.cpl_page;
581 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
583 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
586 static inline pgoff_t osc_index(struct osc_page *opg)
588 return opg->ops_cl.cpl_page->cp_index;
591 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
593 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
594 return container_of0(slice, struct osc_lock, ols_cl);
597 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
599 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
602 static inline int osc_io_srvlock(struct osc_io *oio)
604 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
607 enum osc_extent_state {
608 OES_INV = 0, /** extent is just initialized or destroyed */
609 OES_ACTIVE = 1, /** process is using this extent */
610 OES_CACHE = 2, /** extent is ready for IO */
611 OES_LOCKING = 3, /** locking page to prepare IO */
612 OES_LOCK_DONE = 4, /** locking finished, ready to send */
613 OES_RPC = 5, /** in RPC */
614 OES_TRUNC = 6, /** being truncated */
619 * osc_extent data to manage dirty pages.
620 * osc_extent has the following attributes:
621 * 1. all pages in the same must be in one RPC in write back;
622 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
623 * 3. must be covered by only 1 osc_lock;
624 * 4. exclusive. It's impossible to have overlapped osc_extent.
626 * The lifetime of an extent is from when the 1st page is dirtied to when
627 * all pages inside it are written out.
631 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
634 /** red-black tree node */
635 struct rb_node oe_node;
636 /** osc_object of this extent */
637 struct osc_object *oe_obj;
638 /** refcount, removed from red-black tree if reaches zero. */
639 cfs_atomic_t oe_refc;
640 /** busy if non-zero */
641 cfs_atomic_t oe_users;
642 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
644 /** state of this extent */
645 unsigned int oe_state;
646 /** flags for this extent. */
647 unsigned int oe_intree:1,
648 /** 0 is write, 1 is read */
652 /** an ACTIVE extent is going to be truncated, so when this extent
653 * is released, it will turn into TRUNC state instead of CACHE. */
655 /** this extent should be written asap and someone may wait for the
656 * write to finish. This bit is usually set along with urgent if
657 * the extent was CACHE state.
658 * fsync_wait extent can't be merged because new extent region may
659 * exceed fsync range. */
661 /** covering lock is being canceled */
663 /** this extent should be written back asap. set if one of pages is
664 * called by page WB daemon, or sync write or reading requests. */
666 /** how many grants allocated for this extent.
667 * Grant allocated for this extent. There is no grant allocated
668 * for reading extents and sync write extents. */
669 unsigned int oe_grants;
670 /** # of dirty pages in this extent */
671 unsigned int oe_nr_pages;
672 /** list of pending oap pages. Pages in this list are NOT sorted. */
674 /** Since an extent has to be written out in atomic, this is used to
675 * remember the next page need to be locked to write this extent out.
676 * Not used right now.
678 struct osc_page *oe_next_page;
679 /** start and end index of this extent, include start and end
680 * themselves. Page offset here is the page index of osc_pages.
681 * oe_start is used as keyword for red-black tree. */
684 /** maximum ending index of this extent, this is limited by
685 * max_pages_per_rpc, lock extent and chunk size. */
687 /** waitqueue - for those who want to be notified if this extent's
688 * state has changed. */
689 wait_queue_head_t oe_waitq;
690 /** lock covering this extent */
691 struct cl_lock *oe_osclock;
692 /** terminator of this extent. Must be true if this extent is in IO. */
693 struct task_struct *oe_owner;
694 /** return value of writeback. If somebody is waiting for this extent,
695 * this value can be known by outside world. */
697 /** max pages per rpc when this extent was created */
698 unsigned int oe_mppr;
701 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
703 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
705 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *lock);
707 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
708 struct osc_page *, void *);
709 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
710 struct osc_object *osc, pgoff_t start, pgoff_t end,
711 osc_page_gang_cbt cb, void *cbdata);
715 #endif /* OSC_CL_INTERNAL_H */