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
49 #include <libcfs/libcfs.h>
51 /* osc_build_res_name() */
52 #include <cl_object.h>
54 #include "osc_internal.h"
63 * State maintained by osc layer for each IO context.
67 struct cl_io_slice oi_cl;
68 /** true if this io is lockless. */
70 /** active extents, we know how many bytes is going to be written,
71 * so having an active extent will prevent it from being fragmented */
72 struct osc_extent *oi_active;
73 /** partially truncated extent, we need to hold this extent to prevent
74 * page writeback from happening. */
75 struct osc_extent *oi_trunc;
79 struct obd_info oi_info;
81 struct osc_async_cbargs {
84 struct completion opc_sync;
89 * State of transfer for osc.
92 struct cl_req_slice or_cl;
96 * State maintained by osc layer for the duration of a system call.
102 #define OTI_PVEC_SIZE 256
103 struct osc_thread_info {
104 struct ldlm_res_id oti_resname;
105 ldlm_policy_data_t oti_policy;
106 struct cl_lock_descr oti_descr;
107 struct cl_attr oti_attr;
108 struct lustre_handle oti_handle;
109 struct cl_page_list oti_plist;
111 void *oti_pvec[OTI_PVEC_SIZE];
113 * Fields used by cl_lock_discard_pages().
115 pgoff_t oti_next_index;
116 pgoff_t oti_fn_index; /* first non-overlapped index */
120 struct cl_object oo_cl;
121 struct lov_oinfo *oo_oinfo;
123 * True if locking against this stripe got -EUSERS.
126 cfs_time_t oo_contention_time;
127 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
129 * IO context used for invariant checks in osc_lock_has_pages().
131 struct cl_io oo_debug_io;
132 /** Serialization object for osc_object::oo_debug_io. */
133 struct mutex oo_debug_mutex;
136 * List of pages in transfer.
138 struct list_head oo_inflight[CRT_NR];
140 * Lock, protecting ccc_object::cob_inflight, because a seat-belt is
141 * locked during take-off and landing.
143 spinlock_t oo_seatbelt;
146 * used by the osc to keep track of what objects to build into rpcs.
147 * Protected by client_obd->cli_loi_list_lock.
149 struct list_head oo_ready_item;
150 struct list_head oo_hp_ready_item;
151 struct list_head oo_write_item;
152 struct list_head oo_read_item;
155 * extent is a red black tree to manage (async) dirty pages.
157 struct rb_root oo_root;
159 * Manage write(dirty) extents.
161 struct list_head oo_hp_exts; /* list of hp extents */
162 struct list_head oo_urgent_exts; /* list of writeback extents */
163 struct list_head oo_rpc_exts;
165 struct list_head oo_reading_exts;
167 atomic_t oo_nr_reads;
168 atomic_t oo_nr_writes;
170 /** Protect extent tree. Will be used to protect
171 * oo_{read|write}_pages soon. */
175 * Radix tree for caching pages
177 struct radix_tree_root oo_tree;
178 spinlock_t oo_tree_lock;
179 unsigned long oo_npages;
182 static inline void osc_object_lock(struct osc_object *obj)
184 spin_lock(&obj->oo_lock);
187 static inline int osc_object_trylock(struct osc_object *obj)
189 return spin_trylock(&obj->oo_lock);
192 static inline void osc_object_unlock(struct osc_object *obj)
194 spin_unlock(&obj->oo_lock);
197 static inline int osc_object_is_locked(struct osc_object *obj)
199 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
200 return spin_is_locked(&obj->oo_lock);
203 * It is not perfect to return true all the time.
204 * But since this function is only used for assertion
205 * and checking, it seems OK.
212 * Lock "micro-states" for osc layer.
214 enum osc_lock_state {
225 * osc-private state of cl_lock.
227 * Interaction with DLM.
229 * CLIO enqueues all DLM locks through ptlrpcd (that is, in "async" mode).
231 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
232 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_lock.
234 * This pointer is protected through a reference, acquired by
235 * osc_lock_upcall0(). Also, an additional reference is acquired by
236 * ldlm_lock_addref() call protecting the lock from cancellation, until
237 * osc_lock_unuse() releases it.
239 * Below is a description of how lock references are acquired and released
242 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
243 * - ldlm_lock_create()
244 * - ldlm_lock_new(): initializes a lock with 2 references. One for
245 * the caller (released when reply from the server is received, or on
246 * error), and another for the hash table.
247 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
249 * - When reply is received from the server (osc_enqueue_interpret())
250 * - ldlm_cli_enqueue_fini()
251 * - LDLM_LOCK_PUT(): releases caller reference acquired by
254 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
255 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
257 * - When lock is being cancelled (ldlm_lock_cancel())
258 * - ldlm_lock_destroy()
259 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
262 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
263 * either when lock is cancelled (osc_lock_blocking()), or when locks is
264 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
265 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
269 struct cl_lock_slice ols_cl;
270 /** underlying DLM lock */
271 struct ldlm_lock *ols_lock;
272 /** lock value block */
273 struct ost_lvb ols_lvb;
274 /** DLM flags with which osc_lock::ols_lock was enqueued */
276 /** osc_lock::ols_lock handle */
277 struct lustre_handle ols_handle;
278 struct ldlm_enqueue_info ols_einfo;
279 enum osc_lock_state ols_state;
282 * true, if ldlm_lock_addref() was called against
283 * osc_lock::ols_lock. This is used for sanity checking.
285 * \see osc_lock::ols_has_ref
287 unsigned ols_hold :1,
289 * this is much like osc_lock::ols_hold, except that this bit is
290 * cleared _after_ reference in released in osc_lock_unuse(). This
291 * fine distinction is needed because:
293 * - if ldlm lock still has a reference, osc_ast_data_get() needs
294 * to return associated cl_lock (so that a flag is needed that is
295 * cleared after ldlm_lock_decref() returned), and
297 * - ldlm_lock_decref() can invoke blocking ast (for a
298 * LDLM_FL_CBPENDING lock), and osc_lock functions like
299 * osc_lock_cancel() called from there need to know whether to
300 * release lock reference (so that a flag is needed that is
301 * cleared before ldlm_lock_decref() is called).
305 * inherit the lockless attribute from top level cl_io.
306 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
310 * set by osc_lock_use() to wait until blocking AST enters into
311 * osc_ldlm_blocking_ast0(), so that cl_lock mutex can be used for
312 * further synchronization.
316 * If the data of this lock has been flushed to server side.
320 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
321 * the EVAVAIL error as torerable, this will make upper logic happy
322 * to wait all glimpse locks to each OSTs to be completed.
323 * Glimpse lock converts to normal lock if the server lock is
325 * Glimpse lock should be destroyed immediately after use.
329 * For async glimpse lock.
333 * IO that owns this lock. This field is used for a dead-lock
334 * avoidance by osc_lock_enqueue_wait().
336 * XXX: unfortunately, the owner of a osc_lock is not unique,
337 * the lock may have multiple users, if the lock is granted and
340 struct osc_io *ols_owner;
345 * Page state private for osc layer.
348 struct cl_page_slice ops_cl;
350 * Page queues used by osc to detect when RPC can be formed.
352 struct osc_async_page ops_oap;
354 * An offset within page from which next transfer starts. This is used
355 * by cl_page_clip() to submit partial page transfers.
359 * An offset within page at which next transfer ends.
361 * \see osc_page::ops_from.
365 * Boolean, true iff page is under transfer. Used for sanity checking.
367 unsigned ops_transfer_pinned:1,
369 * True for a `temporary page' created by read-ahead code, probably
370 * outside of any DLM lock.
378 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
382 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
384 struct list_head ops_lru;
386 * Linkage into a per-osc_object list of pages in flight. For
389 struct list_head ops_inflight;
391 * Thread that submitted this page for transfer. For debugging.
393 struct task_struct *ops_submitter;
395 * Submit time - the time when the page is starting RPC. For debugging.
397 cfs_time_t ops_submit_time;
400 extern struct kmem_cache *osc_lock_kmem;
401 extern struct kmem_cache *osc_object_kmem;
402 extern struct kmem_cache *osc_thread_kmem;
403 extern struct kmem_cache *osc_session_kmem;
404 extern struct kmem_cache *osc_req_kmem;
405 extern struct kmem_cache *osc_extent_kmem;
407 extern struct lu_device_type osc_device_type;
408 extern struct lu_context_key osc_key;
409 extern struct lu_context_key osc_session_key;
411 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
413 int osc_lock_init(const struct lu_env *env,
414 struct cl_object *obj, struct cl_lock *lock,
415 const struct cl_io *io);
416 int osc_io_init (const struct lu_env *env,
417 struct cl_object *obj, struct cl_io *io);
418 int osc_req_init (const struct lu_env *env, struct cl_device *dev,
420 struct lu_object *osc_object_alloc(const struct lu_env *env,
421 const struct lu_object_header *hdr,
422 struct lu_device *dev);
423 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
424 struct cl_page *page, pgoff_t ind);
426 void osc_index2policy (ldlm_policy_data_t *policy, const struct cl_object *obj,
427 pgoff_t start, pgoff_t end);
428 int osc_lvb_print (const struct lu_env *env, void *cookie,
429 lu_printer_t p, const struct ost_lvb *lvb);
431 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
432 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
433 enum cl_req_type crt, int brw_flags);
434 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
435 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
436 obd_flag async_flags);
437 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
438 struct page *page, loff_t offset);
439 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
440 struct osc_page *ops);
441 int osc_page_cache_add(const struct lu_env *env,
442 const struct cl_page_slice *slice, struct cl_io *io);
443 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
444 struct osc_page *ops);
445 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
446 struct osc_page *ops);
447 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
448 struct list_head *list, int cmd, int brw_flags);
449 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
450 struct osc_object *obj, __u64 size);
451 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
452 struct osc_object *obj);
453 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
454 pgoff_t start, pgoff_t end, int hp, int discard);
455 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
456 pgoff_t start, pgoff_t end);
457 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
458 struct osc_object *osc, pdl_policy_t pol);
459 int lru_queue_work(const struct lu_env *env, void *data);
461 void osc_object_set_contended (struct osc_object *obj);
462 void osc_object_clear_contended(struct osc_object *obj);
463 int osc_object_is_contended (struct osc_object *obj);
465 int osc_lock_is_lockless (const struct osc_lock *olck);
467 /*****************************************************************************
473 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
475 struct osc_thread_info *info;
477 info = lu_context_key_get(&env->le_ctx, &osc_key);
478 LASSERT(info != NULL);
482 static inline struct osc_session *osc_env_session(const struct lu_env *env)
484 struct osc_session *ses;
486 ses = lu_context_key_get(env->le_ses, &osc_session_key);
487 LASSERT(ses != NULL);
491 static inline struct osc_io *osc_env_io(const struct lu_env *env)
493 return &osc_env_session(env)->os_io;
496 static inline int osc_is_object(const struct lu_object *obj)
498 return obj->lo_dev->ld_type == &osc_device_type;
501 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
503 LINVRNT(d->ld_type == &osc_device_type);
504 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
507 static inline struct obd_export *osc_export(const struct osc_object *obj)
509 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
512 static inline struct client_obd *osc_cli(const struct osc_object *obj)
514 return &osc_export(obj)->exp_obd->u.cli;
517 static inline struct osc_object *cl2osc(const struct cl_object *obj)
519 LINVRNT(osc_is_object(&obj->co_lu));
520 return container_of0(obj, struct osc_object, oo_cl);
523 static inline struct cl_object *osc2cl(const struct osc_object *obj)
525 return (struct cl_object *)&obj->oo_cl;
528 static inline ldlm_mode_t osc_cl_lock2ldlm(enum cl_lock_mode mode)
530 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
531 if (mode == CLM_READ)
533 else if (mode == CLM_WRITE)
539 static inline enum cl_lock_mode osc_ldlm2cl_lock(ldlm_mode_t mode)
541 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
544 else if (mode == LCK_PW)
550 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
552 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
553 return container_of0(slice, struct osc_page, ops_cl);
556 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
558 return container_of0(oap, struct osc_page, ops_oap);
561 static inline pgoff_t osc_index(struct osc_page *opg)
563 return opg->ops_cl.cpl_index;
566 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
568 return oap2osc(oap)->ops_cl.cpl_page;
571 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
573 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
576 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
578 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
579 return container_of0(slice, struct osc_lock, ols_cl);
582 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
584 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
587 static inline int osc_io_srvlock(struct osc_io *oio)
589 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
592 enum osc_extent_state {
593 OES_INV = 0, /** extent is just initialized or destroyed */
594 OES_ACTIVE = 1, /** process is using this extent */
595 OES_CACHE = 2, /** extent is ready for IO */
596 OES_LOCKING = 3, /** locking page to prepare IO */
597 OES_LOCK_DONE = 4, /** locking finished, ready to send */
598 OES_RPC = 5, /** in RPC */
599 OES_TRUNC = 6, /** being truncated */
604 * osc_extent data to manage dirty pages.
605 * osc_extent has the following attributes:
606 * 1. all pages in the same must be in one RPC in write back;
607 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
608 * 3. must be covered by only 1 osc_lock;
609 * 4. exclusive. It's impossible to have overlapped osc_extent.
611 * The lifetime of an extent is from when the 1st page is dirtied to when
612 * all pages inside it are written out.
616 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
619 /** red-black tree node */
620 struct rb_node oe_node;
621 /** osc_object of this extent */
622 struct osc_object *oe_obj;
623 /** refcount, removed from red-black tree if reaches zero. */
625 /** busy if non-zero */
627 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
628 struct list_head oe_link;
629 /** state of this extent */
630 unsigned int oe_state;
631 /** flags for this extent. */
632 unsigned int oe_intree:1,
633 /** 0 is write, 1 is read */
637 /** an ACTIVE extent is going to be truncated, so when this extent
638 * is released, it will turn into TRUNC state instead of CACHE. */
640 /** this extent should be written asap and someone may wait for the
641 * write to finish. This bit is usually set along with urgent if
642 * the extent was CACHE state.
643 * fsync_wait extent can't be merged because new extent region may
644 * exceed fsync range. */
646 /** covering lock is being canceled */
648 /** this extent should be written back asap. set if one of pages is
649 * called by page WB daemon, or sync write or reading requests. */
651 /** how many grants allocated for this extent.
652 * Grant allocated for this extent. There is no grant allocated
653 * for reading extents and sync write extents. */
654 unsigned int oe_grants;
655 /** # of dirty pages in this extent */
656 unsigned int oe_nr_pages;
657 /** list of pending oap pages. Pages in this list are NOT sorted. */
658 struct list_head oe_pages;
659 /** Since an extent has to be written out in atomic, this is used to
660 * remember the next page need to be locked to write this extent out.
661 * Not used right now.
663 struct osc_page *oe_next_page;
664 /** start and end index of this extent, include start and end
665 * themselves. Page offset here is the page index of osc_pages.
666 * oe_start is used as keyword for red-black tree. */
669 /** maximum ending index of this extent, this is limited by
670 * max_pages_per_rpc, lock extent and chunk size. */
672 /** waitqueue - for those who want to be notified if this extent's
673 * state has changed. */
674 wait_queue_head_t oe_waitq;
675 /** lock covering this extent */
676 struct cl_lock *oe_osclock;
677 /** terminator of this extent. Must be true if this extent is in IO. */
678 struct task_struct *oe_owner;
679 /** return value of writeback. If somebody is waiting for this extent,
680 * this value can be known by outside world. */
682 /** max pages per rpc when this extent was created */
683 unsigned int oe_mppr;
686 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
688 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
690 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *lock);
692 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
693 struct osc_page *, void *);
694 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
695 struct osc_object *osc, pgoff_t start, pgoff_t end,
696 osc_page_gang_cbt cb, void *cbdata);
700 #endif /* OSC_CL_INTERNAL_H */