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, 2015, 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 /** true if this io is counted as active IO */
71 /** how many LRU pages are reserved for this IO */
72 unsigned long oi_lru_reserved;
74 /** active extents, we know how many bytes is going to be written,
75 * so having an active extent will prevent it from being fragmented */
76 struct osc_extent *oi_active;
77 /** partially truncated extent, we need to hold this extent to prevent
78 * page writeback from happening. */
79 struct osc_extent *oi_trunc;
80 /** write osc_lock for this IO, used by osc_extent_find(). */
81 struct osc_lock *oi_write_osclock;
83 struct osc_async_cbargs {
86 struct completion opc_sync;
91 * State maintained by osc layer for the duration of a system call.
97 #define OTI_PVEC_SIZE 256
98 struct osc_thread_info {
99 struct ldlm_res_id oti_resname;
100 union ldlm_policy_data oti_policy;
101 struct cl_lock_descr oti_descr;
102 struct cl_attr oti_attr;
103 struct lustre_handle oti_handle;
104 struct cl_page_list oti_plist;
106 void *oti_pvec[OTI_PVEC_SIZE];
108 * Fields used by cl_lock_discard_pages().
110 pgoff_t oti_next_index;
111 pgoff_t oti_fn_index; /* first non-overlapped index */
112 struct cl_sync_io oti_anchor;
113 struct cl_req_attr oti_req_attr;
117 struct cl_object oo_cl;
118 struct lov_oinfo *oo_oinfo;
120 * True if locking against this stripe got -EUSERS.
123 cfs_time_t oo_contention_time;
124 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
126 * IO context used for invariant checks in osc_lock_has_pages().
128 struct cl_io oo_debug_io;
129 /** Serialization object for osc_object::oo_debug_io. */
130 struct mutex oo_debug_mutex;
133 * used by the osc to keep track of what objects to build into rpcs.
134 * Protected by client_obd->cli_loi_list_lock.
136 struct list_head oo_ready_item;
137 struct list_head oo_hp_ready_item;
138 struct list_head oo_write_item;
139 struct list_head oo_read_item;
142 * extent is a red black tree to manage (async) dirty pages.
144 struct rb_root oo_root;
146 * Manage write(dirty) extents.
148 struct list_head oo_hp_exts; /* list of hp extents */
149 struct list_head oo_urgent_exts; /* list of writeback extents */
150 struct list_head oo_rpc_exts;
152 struct list_head oo_reading_exts;
154 atomic_t oo_nr_reads;
155 atomic_t oo_nr_writes;
157 /** Protect extent tree. Will be used to protect
158 * oo_{read|write}_pages soon. */
162 * Radix tree for caching pages
164 spinlock_t oo_tree_lock;
165 struct radix_tree_root oo_tree;
166 unsigned long oo_npages;
168 /* Protect osc_lock this osc_object has */
169 struct list_head oo_ol_list;
170 spinlock_t oo_ol_spin;
172 /** number of active IOs of this object */
174 wait_queue_head_t oo_io_waitq;
177 static inline void osc_object_lock(struct osc_object *obj)
179 spin_lock(&obj->oo_lock);
182 static inline int osc_object_trylock(struct osc_object *obj)
184 return spin_trylock(&obj->oo_lock);
187 static inline void osc_object_unlock(struct osc_object *obj)
189 spin_unlock(&obj->oo_lock);
192 static inline int osc_object_is_locked(struct osc_object *obj)
194 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
195 return spin_is_locked(&obj->oo_lock);
198 * It is not perfect to return true all the time.
199 * But since this function is only used for assertion
200 * and checking, it seems OK.
207 * Lock "micro-states" for osc layer.
209 enum osc_lock_state {
218 * osc-private state of cl_lock.
220 * Interaction with DLM.
222 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
223 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_dlmlock.
225 * This pointer is protected through a reference, acquired by
226 * osc_lock_upcall0(). Also, an additional reference is acquired by
227 * ldlm_lock_addref() call protecting the lock from cancellation, until
228 * osc_lock_unuse() releases it.
230 * Below is a description of how lock references are acquired and released
233 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
234 * - ldlm_lock_create()
235 * - ldlm_lock_new(): initializes a lock with 2 references. One for
236 * the caller (released when reply from the server is received, or on
237 * error), and another for the hash table.
238 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
240 * - When reply is received from the server (osc_enqueue_interpret())
241 * - ldlm_cli_enqueue_fini()
242 * - LDLM_LOCK_PUT(): releases caller reference acquired by
245 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
246 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
248 * - When lock is being cancelled (ldlm_lock_cancel())
249 * - ldlm_lock_destroy()
250 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
253 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
254 * either when lock is cancelled (osc_lock_blocking()), or when locks is
255 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
256 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
260 struct cl_lock_slice ols_cl;
261 /** Internal lock to protect states, etc. */
263 /** Owner sleeps on this channel for state change */
264 struct cl_sync_io *ols_owner;
265 /** waiting list for this lock to be cancelled */
266 struct list_head ols_waiting_list;
267 /** wait entry of ols_waiting_list */
268 struct list_head ols_wait_entry;
269 /** list entry for osc_object::oo_ol_list */
270 struct list_head ols_nextlock_oscobj;
272 /** underlying DLM lock */
273 struct ldlm_lock *ols_dlmlock;
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;
280 /** lock value block */
281 struct ost_lvb ols_lvb;
284 * true, if ldlm_lock_addref() was called against
285 * osc_lock::ols_lock. This is used for sanity checking.
287 * \see osc_lock::ols_has_ref
289 unsigned ols_hold :1,
291 * this is much like osc_lock::ols_hold, except that this bit is
292 * cleared _after_ reference in released in osc_lock_unuse(). This
293 * fine distinction is needed because:
295 * - if ldlm lock still has a reference, osc_ast_data_get() needs
296 * to return associated cl_lock (so that a flag is needed that is
297 * cleared after ldlm_lock_decref() returned), and
299 * - ldlm_lock_decref() can invoke blocking ast (for a
300 * LDLM_FL_CBPENDING lock), and osc_lock functions like
301 * osc_lock_cancel() called from there need to know whether to
302 * release lock reference (so that a flag is needed that is
303 * cleared before ldlm_lock_decref() is called).
307 * inherit the lockless attribute from top level cl_io.
308 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
312 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
313 * the EVAVAIL error as torerable, this will make upper logic happy
314 * to wait all glimpse locks to each OSTs to be completed.
315 * Glimpse lock converts to normal lock if the server lock is
317 * Glimpse lock should be destroyed immediately after use.
321 * For async glimpse lock.
328 * Page state private for osc layer.
331 struct cl_page_slice ops_cl;
333 * Page queues used by osc to detect when RPC can be formed.
335 struct osc_async_page ops_oap;
337 * An offset within page from which next transfer starts. This is used
338 * by cl_page_clip() to submit partial page transfers.
342 * An offset within page at which next transfer ends.
344 * \see osc_page::ops_from.
348 * Boolean, true iff page is under transfer. Used for sanity checking.
350 unsigned ops_transfer_pinned:1,
352 * True for a `temporary page' created by read-ahead code, probably
353 * outside of any DLM lock.
361 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
365 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
367 struct list_head ops_lru;
369 * Submit time - the time when the page is starting RPC. For debugging.
371 cfs_time_t ops_submit_time;
374 extern struct kmem_cache *osc_lock_kmem;
375 extern struct kmem_cache *osc_object_kmem;
376 extern struct kmem_cache *osc_thread_kmem;
377 extern struct kmem_cache *osc_session_kmem;
378 extern struct kmem_cache *osc_extent_kmem;
380 extern struct lu_device_type osc_device_type;
381 extern struct lu_context_key osc_key;
382 extern struct lu_context_key osc_session_key;
384 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
386 int osc_lock_init(const struct lu_env *env,
387 struct cl_object *obj, struct cl_lock *lock,
388 const struct cl_io *io);
389 int osc_io_init (const struct lu_env *env,
390 struct cl_object *obj, struct cl_io *io);
391 struct lu_object *osc_object_alloc(const struct lu_env *env,
392 const struct lu_object_header *hdr,
393 struct lu_device *dev);
394 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
395 struct cl_page *page, pgoff_t ind);
397 void osc_index2policy(union ldlm_policy_data *policy,
398 const struct cl_object *obj, pgoff_t start, pgoff_t end);
399 int osc_lvb_print(const struct lu_env *env, void *cookie,
400 lu_printer_t p, const struct ost_lvb *lvb);
402 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
403 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
404 enum cl_req_type crt, int brw_flags);
405 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
406 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
408 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
409 struct page *page, loff_t offset);
410 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
411 struct osc_page *ops);
412 int osc_page_cache_add(const struct lu_env *env,
413 const struct cl_page_slice *slice, struct cl_io *io);
414 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
415 struct osc_page *ops);
416 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
417 struct osc_page *ops);
418 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
419 struct list_head *list, int cmd, int brw_flags);
420 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
421 __u64 size, struct osc_extent **extp);
422 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext);
423 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
424 pgoff_t start, pgoff_t end, int hp, int discard);
425 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
426 pgoff_t start, pgoff_t end);
427 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
428 struct osc_object *osc);
429 int lru_queue_work(const struct lu_env *env, void *data);
431 void osc_object_set_contended (struct osc_object *obj);
432 void osc_object_clear_contended(struct osc_object *obj);
433 int osc_object_is_contended (struct osc_object *obj);
435 int osc_lock_is_lockless (const struct osc_lock *olck);
437 /*****************************************************************************
443 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
445 struct osc_thread_info *info;
447 info = lu_context_key_get(&env->le_ctx, &osc_key);
448 LASSERT(info != NULL);
452 static inline struct osc_session *osc_env_session(const struct lu_env *env)
454 struct osc_session *ses;
456 ses = lu_context_key_get(env->le_ses, &osc_session_key);
457 LASSERT(ses != NULL);
461 static inline struct osc_io *osc_env_io(const struct lu_env *env)
463 return &osc_env_session(env)->os_io;
466 static inline int osc_is_object(const struct lu_object *obj)
468 return obj->lo_dev->ld_type == &osc_device_type;
471 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
473 LINVRNT(d->ld_type == &osc_device_type);
474 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
477 static inline struct obd_export *osc_export(const struct osc_object *obj)
479 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
482 static inline struct client_obd *osc_cli(const struct osc_object *obj)
484 return &osc_export(obj)->exp_obd->u.cli;
487 static inline struct osc_object *cl2osc(const struct cl_object *obj)
489 LINVRNT(osc_is_object(&obj->co_lu));
490 return container_of0(obj, struct osc_object, oo_cl);
493 static inline struct cl_object *osc2cl(const struct osc_object *obj)
495 return (struct cl_object *)&obj->oo_cl;
498 static inline enum ldlm_mode osc_cl_lock2ldlm(enum cl_lock_mode mode)
500 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
501 if (mode == CLM_READ)
503 if (mode == CLM_WRITE)
508 static inline enum cl_lock_mode osc_ldlm2cl_lock(enum ldlm_mode mode)
510 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
518 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
520 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
521 return container_of0(slice, struct osc_page, ops_cl);
524 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
526 return container_of0(oap, struct osc_page, ops_oap);
529 static inline pgoff_t osc_index(struct osc_page *opg)
531 return opg->ops_cl.cpl_index;
534 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
536 return oap2osc(oap)->ops_cl.cpl_page;
539 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
541 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
544 static inline struct osc_page *
545 osc_cl_page_osc(struct cl_page *page, struct osc_object *osc)
547 const struct cl_page_slice *slice;
549 LASSERT(osc != NULL);
550 slice = cl_object_page_slice(&osc->oo_cl, page);
551 return cl2osc_page(slice);
554 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
556 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
557 return container_of0(slice, struct osc_lock, ols_cl);
560 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
562 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
565 static inline int osc_io_srvlock(struct osc_io *oio)
567 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
570 enum osc_extent_state {
571 OES_INV = 0, /** extent is just initialized or destroyed */
572 OES_ACTIVE = 1, /** process is using this extent */
573 OES_CACHE = 2, /** extent is ready for IO */
574 OES_LOCKING = 3, /** locking page to prepare IO */
575 OES_LOCK_DONE = 4, /** locking finished, ready to send */
576 OES_RPC = 5, /** in RPC */
577 OES_TRUNC = 6, /** being truncated */
582 * osc_extent data to manage dirty pages.
583 * osc_extent has the following attributes:
584 * 1. all pages in the same must be in one RPC in write back;
585 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
586 * 3. must be covered by only 1 osc_lock;
587 * 4. exclusive. It's impossible to have overlapped osc_extent.
589 * The lifetime of an extent is from when the 1st page is dirtied to when
590 * all pages inside it are written out.
594 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
597 /** red-black tree node */
598 struct rb_node oe_node;
599 /** osc_object of this extent */
600 struct osc_object *oe_obj;
601 /** refcount, removed from red-black tree if reaches zero. */
603 /** busy if non-zero */
605 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
606 struct list_head oe_link;
607 /** state of this extent */
608 enum osc_extent_state oe_state;
609 /** flags for this extent. */
610 unsigned int oe_intree:1,
611 /** 0 is write, 1 is read */
613 /** sync extent, queued by osc_queue_sync_pages() */
615 /** set if this extent has partial, sync pages.
616 * Extents with partial page(s) can't merge with others in RPC */
620 /** an ACTIVE extent is going to be truncated, so when this extent
621 * is released, it will turn into TRUNC state instead of CACHE. */
623 /** this extent should be written asap and someone may wait for the
624 * write to finish. This bit is usually set along with urgent if
625 * the extent was CACHE state.
626 * fsync_wait extent can't be merged because new extent region may
627 * exceed fsync range. */
629 /** covering lock is being canceled */
631 /** this extent should be written back asap. set if one of pages is
632 * called by page WB daemon, or sync write or reading requests. */
634 /** how many grants allocated for this extent.
635 * Grant allocated for this extent. There is no grant allocated
636 * for reading extents and sync write extents. */
637 unsigned int oe_grants;
638 /** # of dirty pages in this extent */
639 unsigned int oe_nr_pages;
640 /** list of pending oap pages. Pages in this list are NOT sorted. */
641 struct list_head oe_pages;
642 /** Since an extent has to be written out in atomic, this is used to
643 * remember the next page need to be locked to write this extent out.
644 * Not used right now.
646 struct osc_page *oe_next_page;
647 /** start and end index of this extent, include start and end
648 * themselves. Page offset here is the page index of osc_pages.
649 * oe_start is used as keyword for red-black tree. */
652 /** maximum ending index of this extent, this is limited by
653 * max_pages_per_rpc, lock extent and chunk size. */
655 /** waitqueue - for those who want to be notified if this extent's
656 * state has changed. */
657 wait_queue_head_t oe_waitq;
658 /** lock covering this extent */
659 struct ldlm_lock *oe_dlmlock;
660 /** terminator of this extent. Must be true if this extent is in IO. */
661 struct task_struct *oe_owner;
662 /** return value of writeback. If somebody is waiting for this extent,
663 * this value can be known by outside world. */
665 /** max pages per rpc when this extent was created */
666 unsigned int oe_mppr;
669 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
671 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
673 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
674 pgoff_t start, pgoff_t end, enum cl_lock_mode mode);
676 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
677 struct osc_page *, void *);
678 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
679 struct osc_object *osc, pgoff_t start, pgoff_t end,
680 osc_page_gang_cbt cb, void *cbdata);
683 #endif /* OSC_CL_INTERNAL_H */