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, 2014, 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>
53 #include "osc_internal.h"
62 * State maintained by osc layer for each IO context.
66 struct cl_io_slice oi_cl;
67 /** true if this io is lockless. */
69 /** how many LRU pages are reserved for this IO */
70 unsigned long oi_lru_reserved;
72 /** active extents, we know how many bytes is going to be written,
73 * so having an active extent will prevent it from being fragmented */
74 struct osc_extent *oi_active;
75 /** partially truncated extent, we need to hold this extent to prevent
76 * page writeback from happening. */
77 struct osc_extent *oi_trunc;
78 /** write osc_lock for this IO, used by osc_extent_find(). */
79 struct osc_lock *oi_write_osclock;
81 struct obd_info oi_info;
83 struct osc_async_cbargs {
86 struct completion opc_sync;
91 * State of transfer for osc.
94 struct cl_req_slice or_cl;
98 * State maintained by osc layer for the duration of a system call.
104 #define OTI_PVEC_SIZE 256
105 struct osc_thread_info {
106 struct ldlm_res_id oti_resname;
107 ldlm_policy_data_t oti_policy;
108 struct cl_lock_descr oti_descr;
109 struct cl_attr oti_attr;
110 struct lustre_handle oti_handle;
111 struct cl_page_list oti_plist;
113 void *oti_pvec[OTI_PVEC_SIZE];
115 * Fields used by cl_lock_discard_pages().
117 pgoff_t oti_next_index;
118 pgoff_t oti_fn_index; /* first non-overlapped index */
119 struct cl_sync_io oti_anchor;
123 struct cl_object oo_cl;
124 struct lov_oinfo *oo_oinfo;
126 * True if locking against this stripe got -EUSERS.
129 cfs_time_t oo_contention_time;
130 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
132 * IO context used for invariant checks in osc_lock_has_pages().
134 struct cl_io oo_debug_io;
135 /** Serialization object for osc_object::oo_debug_io. */
136 struct mutex oo_debug_mutex;
139 * List of pages in transfer.
141 struct list_head oo_inflight[CRT_NR];
143 * Lock, protecting osc_page::ops_inflight, because a seat-belt is
144 * locked during take-off and landing.
146 spinlock_t oo_seatbelt;
149 * used by the osc to keep track of what objects to build into rpcs.
150 * Protected by client_obd->cli_loi_list_lock.
152 struct list_head oo_ready_item;
153 struct list_head oo_hp_ready_item;
154 struct list_head oo_write_item;
155 struct list_head oo_read_item;
158 * extent is a red black tree to manage (async) dirty pages.
160 struct rb_root oo_root;
162 * Manage write(dirty) extents.
164 struct list_head oo_hp_exts; /* list of hp extents */
165 struct list_head oo_urgent_exts; /* list of writeback extents */
166 struct list_head oo_rpc_exts;
168 struct list_head oo_reading_exts;
170 atomic_t oo_nr_reads;
171 atomic_t oo_nr_writes;
173 /** Protect extent tree. Will be used to protect
174 * oo_{read|write}_pages soon. */
178 * Radix tree for caching pages
180 struct radix_tree_root oo_tree;
181 spinlock_t oo_tree_lock;
182 unsigned long oo_npages;
184 /* Protect osc_lock this osc_object has */
185 spinlock_t oo_ol_spin;
186 struct list_head oo_ol_list;
189 static inline void osc_object_lock(struct osc_object *obj)
191 spin_lock(&obj->oo_lock);
194 static inline int osc_object_trylock(struct osc_object *obj)
196 return spin_trylock(&obj->oo_lock);
199 static inline void osc_object_unlock(struct osc_object *obj)
201 spin_unlock(&obj->oo_lock);
204 static inline int osc_object_is_locked(struct osc_object *obj)
206 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
207 return spin_is_locked(&obj->oo_lock);
210 * It is not perfect to return true all the time.
211 * But since this function is only used for assertion
212 * and checking, it seems OK.
219 * Lock "micro-states" for osc layer.
221 enum osc_lock_state {
230 * osc-private state of cl_lock.
232 * Interaction with DLM.
234 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
235 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_dlmlock.
237 * This pointer is protected through a reference, acquired by
238 * osc_lock_upcall0(). Also, an additional reference is acquired by
239 * ldlm_lock_addref() call protecting the lock from cancellation, until
240 * osc_lock_unuse() releases it.
242 * Below is a description of how lock references are acquired and released
245 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
246 * - ldlm_lock_create()
247 * - ldlm_lock_new(): initializes a lock with 2 references. One for
248 * the caller (released when reply from the server is received, or on
249 * error), and another for the hash table.
250 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
252 * - When reply is received from the server (osc_enqueue_interpret())
253 * - ldlm_cli_enqueue_fini()
254 * - LDLM_LOCK_PUT(): releases caller reference acquired by
257 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
258 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
260 * - When lock is being cancelled (ldlm_lock_cancel())
261 * - ldlm_lock_destroy()
262 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
265 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
266 * either when lock is cancelled (osc_lock_blocking()), or when locks is
267 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
268 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
272 struct cl_lock_slice ols_cl;
273 /** Internal lock to protect states, etc. */
275 /** Owner sleeps on this channel for state change */
276 struct cl_sync_io *ols_owner;
277 /** waiting list for this lock to be cancelled */
278 struct list_head ols_waiting_list;
279 /** wait entry of ols_waiting_list */
280 struct list_head ols_wait_entry;
281 /** list entry for osc_object::oo_ol_list */
282 struct list_head ols_nextlock_oscobj;
284 /** underlying DLM lock */
285 struct ldlm_lock *ols_dlmlock;
286 /** DLM flags with which osc_lock::ols_lock was enqueued */
288 /** osc_lock::ols_lock handle */
289 struct lustre_handle ols_handle;
290 struct ldlm_enqueue_info ols_einfo;
291 enum osc_lock_state ols_state;
292 /** lock value block */
293 struct ost_lvb ols_lvb;
296 * true, if ldlm_lock_addref() was called against
297 * osc_lock::ols_lock. This is used for sanity checking.
299 * \see osc_lock::ols_has_ref
301 unsigned ols_hold :1,
303 * this is much like osc_lock::ols_hold, except that this bit is
304 * cleared _after_ reference in released in osc_lock_unuse(). This
305 * fine distinction is needed because:
307 * - if ldlm lock still has a reference, osc_ast_data_get() needs
308 * to return associated cl_lock (so that a flag is needed that is
309 * cleared after ldlm_lock_decref() returned), and
311 * - ldlm_lock_decref() can invoke blocking ast (for a
312 * LDLM_FL_CBPENDING lock), and osc_lock functions like
313 * osc_lock_cancel() called from there need to know whether to
314 * release lock reference (so that a flag is needed that is
315 * cleared before ldlm_lock_decref() is called).
319 * inherit the lockless attribute from top level cl_io.
320 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
324 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
325 * the EVAVAIL error as torerable, this will make upper logic happy
326 * to wait all glimpse locks to each OSTs to be completed.
327 * Glimpse lock converts to normal lock if the server lock is
329 * Glimpse lock should be destroyed immediately after use.
333 * For async glimpse lock.
340 * Page state private for osc layer.
343 struct cl_page_slice ops_cl;
345 * Page queues used by osc to detect when RPC can be formed.
347 struct osc_async_page ops_oap;
349 * An offset within page from which next transfer starts. This is used
350 * by cl_page_clip() to submit partial page transfers.
354 * An offset within page at which next transfer ends.
356 * \see osc_page::ops_from.
360 * Boolean, true iff page is under transfer. Used for sanity checking.
362 unsigned ops_transfer_pinned:1,
364 * True for a `temporary page' created by read-ahead code, probably
365 * outside of any DLM lock.
373 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
377 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
379 struct list_head ops_lru;
381 * Linkage into a per-osc_object list of pages in flight. For
384 struct list_head ops_inflight;
386 * Thread that submitted this page for transfer. For debugging.
388 struct task_struct *ops_submitter;
390 * Submit time - the time when the page is starting RPC. For debugging.
392 cfs_time_t ops_submit_time;
395 extern struct kmem_cache *osc_lock_kmem;
396 extern struct kmem_cache *osc_object_kmem;
397 extern struct kmem_cache *osc_thread_kmem;
398 extern struct kmem_cache *osc_session_kmem;
399 extern struct kmem_cache *osc_req_kmem;
400 extern struct kmem_cache *osc_extent_kmem;
402 extern struct lu_device_type osc_device_type;
403 extern struct lu_context_key osc_key;
404 extern struct lu_context_key osc_session_key;
406 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
408 int osc_lock_init(const struct lu_env *env,
409 struct cl_object *obj, struct cl_lock *lock,
410 const struct cl_io *io);
411 int osc_io_init (const struct lu_env *env,
412 struct cl_object *obj, struct cl_io *io);
413 int osc_req_init (const struct lu_env *env, struct cl_device *dev,
415 struct lu_object *osc_object_alloc(const struct lu_env *env,
416 const struct lu_object_header *hdr,
417 struct lu_device *dev);
418 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
419 struct cl_page *page, pgoff_t ind);
421 void osc_index2policy (ldlm_policy_data_t *policy, const struct cl_object *obj,
422 pgoff_t start, pgoff_t end);
423 int osc_lvb_print (const struct lu_env *env, void *cookie,
424 lu_printer_t p, const struct ost_lvb *lvb);
426 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
427 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
428 enum cl_req_type crt, int brw_flags);
429 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
430 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
432 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
433 struct page *page, loff_t offset);
434 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
435 struct osc_page *ops);
436 int osc_page_cache_add(const struct lu_env *env,
437 const struct cl_page_slice *slice, struct cl_io *io);
438 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
439 struct osc_page *ops);
440 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
441 struct osc_page *ops);
442 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
443 struct list_head *list, int cmd, int brw_flags);
444 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
445 struct osc_object *obj, __u64 size);
446 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
447 struct osc_object *obj);
448 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
449 pgoff_t start, pgoff_t end, int hp, int discard);
450 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
451 pgoff_t start, pgoff_t end);
452 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
453 struct osc_object *osc, pdl_policy_t pol);
454 int lru_queue_work(const struct lu_env *env, void *data);
456 void osc_object_set_contended (struct osc_object *obj);
457 void osc_object_clear_contended(struct osc_object *obj);
458 int osc_object_is_contended (struct osc_object *obj);
460 int osc_lock_is_lockless (const struct osc_lock *olck);
462 /*****************************************************************************
468 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
470 struct osc_thread_info *info;
472 info = lu_context_key_get(&env->le_ctx, &osc_key);
473 LASSERT(info != NULL);
477 static inline struct osc_session *osc_env_session(const struct lu_env *env)
479 struct osc_session *ses;
481 ses = lu_context_key_get(env->le_ses, &osc_session_key);
482 LASSERT(ses != NULL);
486 static inline struct osc_io *osc_env_io(const struct lu_env *env)
488 return &osc_env_session(env)->os_io;
491 static inline int osc_is_object(const struct lu_object *obj)
493 return obj->lo_dev->ld_type == &osc_device_type;
496 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
498 LINVRNT(d->ld_type == &osc_device_type);
499 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
502 static inline struct obd_export *osc_export(const struct osc_object *obj)
504 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
507 static inline struct client_obd *osc_cli(const struct osc_object *obj)
509 return &osc_export(obj)->exp_obd->u.cli;
512 static inline struct osc_object *cl2osc(const struct cl_object *obj)
514 LINVRNT(osc_is_object(&obj->co_lu));
515 return container_of0(obj, struct osc_object, oo_cl);
518 static inline struct cl_object *osc2cl(const struct osc_object *obj)
520 return (struct cl_object *)&obj->oo_cl;
523 static inline ldlm_mode_t osc_cl_lock2ldlm(enum cl_lock_mode mode)
525 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
526 if (mode == CLM_READ)
528 else if (mode == CLM_WRITE)
534 static inline enum cl_lock_mode osc_ldlm2cl_lock(ldlm_mode_t mode)
536 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
539 else if (mode == LCK_PW)
545 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
547 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
548 return container_of0(slice, struct osc_page, ops_cl);
551 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
553 return container_of0(oap, struct osc_page, ops_oap);
556 static inline pgoff_t osc_index(struct osc_page *opg)
558 return opg->ops_cl.cpl_index;
561 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
563 return oap2osc(oap)->ops_cl.cpl_page;
566 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
568 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
571 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
573 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
574 return container_of0(slice, struct osc_lock, ols_cl);
577 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
579 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
582 static inline int osc_io_srvlock(struct osc_io *oio)
584 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
587 enum osc_extent_state {
588 OES_INV = 0, /** extent is just initialized or destroyed */
589 OES_ACTIVE = 1, /** process is using this extent */
590 OES_CACHE = 2, /** extent is ready for IO */
591 OES_LOCKING = 3, /** locking page to prepare IO */
592 OES_LOCK_DONE = 4, /** locking finished, ready to send */
593 OES_RPC = 5, /** in RPC */
594 OES_TRUNC = 6, /** being truncated */
599 * osc_extent data to manage dirty pages.
600 * osc_extent has the following attributes:
601 * 1. all pages in the same must be in one RPC in write back;
602 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
603 * 3. must be covered by only 1 osc_lock;
604 * 4. exclusive. It's impossible to have overlapped osc_extent.
606 * The lifetime of an extent is from when the 1st page is dirtied to when
607 * all pages inside it are written out.
611 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
614 /** red-black tree node */
615 struct rb_node oe_node;
616 /** osc_object of this extent */
617 struct osc_object *oe_obj;
618 /** refcount, removed from red-black tree if reaches zero. */
620 /** busy if non-zero */
622 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
623 struct list_head oe_link;
624 /** state of this extent */
625 enum osc_extent_state oe_state;
626 /** flags for this extent. */
627 unsigned int oe_intree:1,
628 /** 0 is write, 1 is read */
630 /** sync extent, queued by osc_queue_sync_pages() */
634 /** an ACTIVE extent is going to be truncated, so when this extent
635 * is released, it will turn into TRUNC state instead of CACHE. */
637 /** this extent should be written asap and someone may wait for the
638 * write to finish. This bit is usually set along with urgent if
639 * the extent was CACHE state.
640 * fsync_wait extent can't be merged because new extent region may
641 * exceed fsync range. */
643 /** covering lock is being canceled */
645 /** this extent should be written back asap. set if one of pages is
646 * called by page WB daemon, or sync write or reading requests. */
648 /** how many grants allocated for this extent.
649 * Grant allocated for this extent. There is no grant allocated
650 * for reading extents and sync write extents. */
651 unsigned int oe_grants;
652 /** # of dirty pages in this extent */
653 unsigned int oe_nr_pages;
654 /** list of pending oap pages. Pages in this list are NOT sorted. */
655 struct list_head oe_pages;
656 /** Since an extent has to be written out in atomic, this is used to
657 * remember the next page need to be locked to write this extent out.
658 * Not used right now.
660 struct osc_page *oe_next_page;
661 /** start and end index of this extent, include start and end
662 * themselves. Page offset here is the page index of osc_pages.
663 * oe_start is used as keyword for red-black tree. */
666 /** maximum ending index of this extent, this is limited by
667 * max_pages_per_rpc, lock extent and chunk size. */
669 /** waitqueue - for those who want to be notified if this extent's
670 * state has changed. */
671 wait_queue_head_t oe_waitq;
672 /** lock covering this extent */
673 struct ldlm_lock *oe_dlmlock;
674 /** terminator of this extent. Must be true if this extent is in IO. */
675 struct task_struct *oe_owner;
676 /** return value of writeback. If somebody is waiting for this extent,
677 * this value can be known by outside world. */
679 /** max pages per rpc when this extent was created */
680 unsigned int oe_mppr;
683 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
685 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
687 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
688 pgoff_t start, pgoff_t end, enum cl_lock_mode mode);
690 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
691 struct osc_page *, void *);
692 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
693 struct osc_object *osc, pgoff_t start, pgoff_t end,
694 osc_page_gang_cbt cb, void *cbdata);
697 #endif /* OSC_CL_INTERNAL_H */