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, 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 "../../include/linux/libcfs/libcfs.h"
51 #include "../include/obd.h"
52 /* osc_build_res_name() */
53 #include "../include/cl_object.h"
54 #include "../include/lclient.h"
55 #include "osc_internal.h"
64 * State maintained by osc layer for each IO context.
68 struct cl_io_slice oi_cl;
69 /** true if this io is lockless. */
71 /** active extents, we know how many bytes is going to be written,
72 * so having an active extent will prevent it from being fragmented */
73 struct osc_extent *oi_active;
74 /** partially truncated extent, we need to hold this extent to prevent
75 * page writeback from happening. */
76 struct osc_extent *oi_trunc;
78 struct obd_info oi_info;
80 struct osc_async_cbargs {
83 struct completion opc_sync;
88 * State of transfer for osc.
91 struct cl_req_slice or_cl;
95 * State maintained by osc layer for the duration of a system call.
101 #define OTI_PVEC_SIZE 64
102 struct osc_thread_info {
103 struct ldlm_res_id oti_resname;
104 ldlm_policy_data_t oti_policy;
105 struct cl_lock_descr oti_descr;
106 struct cl_attr oti_attr;
107 struct lustre_handle oti_handle;
108 struct cl_page_list oti_plist;
110 struct cl_page *oti_pvec[OTI_PVEC_SIZE];
114 struct cl_object oo_cl;
115 struct lov_oinfo *oo_oinfo;
117 * True if locking against this stripe got -EUSERS.
120 unsigned long oo_contention_time;
122 * List of pages in transfer.
124 struct list_head oo_inflight[CRT_NR];
126 * Lock, protecting ccc_object::cob_inflight, because a seat-belt is
127 * locked during take-off and landing.
129 spinlock_t oo_seatbelt;
132 * used by the osc to keep track of what objects to build into rpcs.
133 * Protected by client_obd->cli_loi_list_lock.
135 struct list_head oo_ready_item;
136 struct list_head oo_hp_ready_item;
137 struct list_head oo_write_item;
138 struct list_head oo_read_item;
141 * extent is a red black tree to manage (async) dirty pages.
143 struct rb_root oo_root;
145 * Manage write(dirty) extents.
147 struct list_head oo_hp_exts; /* list of hp extents */
148 struct list_head oo_urgent_exts; /* list of writeback extents */
149 struct list_head oo_rpc_exts;
151 struct list_head oo_reading_exts;
153 atomic_t oo_nr_reads;
154 atomic_t oo_nr_writes;
156 /** Protect extent tree. Will be used to protect
157 * oo_{read|write}_pages soon. */
161 static inline void osc_object_lock(struct osc_object *obj)
163 spin_lock(&obj->oo_lock);
166 static inline int osc_object_trylock(struct osc_object *obj)
168 return spin_trylock(&obj->oo_lock);
171 static inline void osc_object_unlock(struct osc_object *obj)
173 spin_unlock(&obj->oo_lock);
176 static inline int osc_object_is_locked(struct osc_object *obj)
178 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
179 return spin_is_locked(&obj->oo_lock);
182 * It is not perfect to return true all the time.
183 * But since this function is only used for assertion
184 * and checking, it seems OK.
191 * Lock "micro-states" for osc layer.
193 enum osc_lock_state {
204 * osc-private state of cl_lock.
206 * Interaction with DLM.
208 * CLIO enqueues all DLM locks through ptlrpcd (that is, in "async" mode).
210 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
211 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_lock.
213 * This pointer is protected through a reference, acquired by
214 * osc_lock_upcall0(). Also, an additional reference is acquired by
215 * ldlm_lock_addref() call protecting the lock from cancellation, until
216 * osc_lock_unuse() releases it.
218 * Below is a description of how lock references are acquired and released
221 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
222 * - ldlm_lock_create()
223 * - ldlm_lock_new(): initializes a lock with 2 references. One for
224 * the caller (released when reply from the server is received, or on
225 * error), and another for the hash table.
226 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
228 * - When reply is received from the server (osc_enqueue_interpret())
229 * - ldlm_cli_enqueue_fini()
230 * - LDLM_LOCK_PUT(): releases caller reference acquired by
233 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
234 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
236 * - When lock is being cancelled (ldlm_lock_cancel())
237 * - ldlm_lock_destroy()
238 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
241 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
242 * either when lock is cancelled (osc_lock_blocking()), or when locks is
243 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
244 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
248 struct cl_lock_slice ols_cl;
249 /** underlying DLM lock */
250 struct ldlm_lock *ols_lock;
251 /** lock value block */
252 struct ost_lvb ols_lvb;
253 /** DLM flags with which osc_lock::ols_lock was enqueued */
255 /** osc_lock::ols_lock handle */
256 struct lustre_handle ols_handle;
257 struct ldlm_enqueue_info ols_einfo;
258 enum osc_lock_state ols_state;
261 * How many pages are using this lock for io, currently only used by
262 * read-ahead. If non-zero, the underlying dlm lock won't be cancelled
263 * during recovery to avoid deadlock. see bz16774.
265 * \see osc_page::ops_lock
266 * \see osc_page_addref_lock(), osc_page_putref_lock()
268 atomic_t ols_pageref;
271 * true, if ldlm_lock_addref() was called against
272 * osc_lock::ols_lock. This is used for sanity checking.
274 * \see osc_lock::ols_has_ref
276 unsigned ols_hold :1,
278 * this is much like osc_lock::ols_hold, except that this bit is
279 * cleared _after_ reference in released in osc_lock_unuse(). This
280 * fine distinction is needed because:
282 * - if ldlm lock still has a reference, osc_ast_data_get() needs
283 * to return associated cl_lock (so that a flag is needed that is
284 * cleared after ldlm_lock_decref() returned), and
286 * - ldlm_lock_decref() can invoke blocking ast (for a
287 * LDLM_FL_CBPENDING lock), and osc_lock functions like
288 * osc_lock_cancel() called from there need to know whether to
289 * release lock reference (so that a flag is needed that is
290 * cleared before ldlm_lock_decref() is called).
294 * inherit the lockless attribute from top level cl_io.
295 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
299 * set by osc_lock_use() to wait until blocking AST enters into
300 * osc_ldlm_blocking_ast0(), so that cl_lock mutex can be used for
301 * further synchronization.
305 * If the data of this lock has been flushed to server side.
309 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
310 * the EVAVAIL error as tolerable, this will make upper logic happy
311 * to wait all glimpse locks to each OSTs to be completed.
312 * Glimpse lock converts to normal lock if the server lock is
314 * Glimpse lock should be destroyed immediately after use.
318 * For async glimpse lock.
322 * IO that owns this lock. This field is used for a dead-lock
323 * avoidance by osc_lock_enqueue_wait().
325 * XXX: unfortunately, the owner of a osc_lock is not unique,
326 * the lock may have multiple users, if the lock is granted and
329 struct osc_io *ols_owner;
334 * Page state private for osc layer.
337 struct cl_page_slice ops_cl;
339 * Page queues used by osc to detect when RPC can be formed.
341 struct osc_async_page ops_oap;
343 * An offset within page from which next transfer starts. This is used
344 * by cl_page_clip() to submit partial page transfers.
348 * An offset within page at which next transfer ends.
350 * \see osc_page::ops_from.
354 * Boolean, true iff page is under transfer. Used for sanity checking.
356 unsigned ops_transfer_pinned:1,
358 * True for a `temporary page' created by read-ahead code, probably
359 * outside of any DLM lock.
367 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
372 * lru page list. ops_inflight and ops_lru are exclusive so
373 * that they can share the same data.
375 struct list_head ops_lru;
377 * Linkage into a per-osc_object list of pages in flight. For
380 struct list_head ops_inflight;
383 * Thread that submitted this page for transfer. For debugging.
385 struct task_struct *ops_submitter;
387 * Submit time - the time when the page is starting RPC. For debugging.
389 unsigned long ops_submit_time;
392 * A lock of which we hold a reference covers this page. Only used by
393 * read-ahead: for a readahead page, we hold it's covering lock to
394 * prevent it from being canceled during recovery.
396 * \see osc_lock::ols_pageref
397 * \see osc_page_addref_lock(), osc_page_putref_lock().
399 struct cl_lock *ops_lock;
402 extern struct kmem_cache *osc_lock_kmem;
403 extern struct kmem_cache *osc_object_kmem;
404 extern struct kmem_cache *osc_thread_kmem;
405 extern struct kmem_cache *osc_session_kmem;
406 extern struct kmem_cache *osc_req_kmem;
407 extern struct kmem_cache *osc_extent_kmem;
409 extern struct lu_device_type osc_device_type;
410 extern struct lu_context_key osc_key;
411 extern struct lu_context_key osc_session_key;
413 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
415 int osc_lock_init(const struct lu_env *env,
416 struct cl_object *obj, struct cl_lock *lock,
417 const struct cl_io *io);
418 int osc_io_init (const struct lu_env *env,
419 struct cl_object *obj, struct cl_io *io);
420 int osc_req_init (const struct lu_env *env, struct cl_device *dev,
422 struct lu_object *osc_object_alloc(const struct lu_env *env,
423 const struct lu_object_header *hdr,
424 struct lu_device *dev);
425 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
426 struct cl_page *page, struct page *vmpage);
428 void osc_index2policy (ldlm_policy_data_t *policy, const struct cl_object *obj,
429 pgoff_t start, pgoff_t end);
430 int osc_lvb_print (const struct lu_env *env, void *cookie,
431 lu_printer_t p, const struct ost_lvb *lvb);
433 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
434 enum cl_req_type crt, int brw_flags);
435 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
436 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
438 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
439 struct page *page, loff_t offset);
440 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
441 struct osc_page *ops);
442 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
443 struct osc_page *ops);
444 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
445 struct osc_page *ops);
446 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
447 struct list_head *list, int cmd, int brw_flags);
448 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
449 struct osc_object *obj, __u64 size);
450 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
451 struct osc_object *obj);
452 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
453 pgoff_t start, pgoff_t end, int hp, int discard);
454 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
455 pgoff_t start, pgoff_t end);
456 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
457 struct osc_object *osc, pdl_policy_t pol);
459 void osc_object_set_contended (struct osc_object *obj);
460 void osc_object_clear_contended(struct osc_object *obj);
461 int osc_object_is_contended (struct osc_object *obj);
463 int osc_lock_is_lockless (const struct osc_lock *olck);
465 /*****************************************************************************
471 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
473 struct osc_thread_info *info;
475 info = lu_context_key_get(&env->le_ctx, &osc_key);
476 LASSERT(info != NULL);
480 static inline struct osc_session *osc_env_session(const struct lu_env *env)
482 struct osc_session *ses;
484 ses = lu_context_key_get(env->le_ses, &osc_session_key);
485 LASSERT(ses != NULL);
489 static inline struct osc_io *osc_env_io(const struct lu_env *env)
491 return &osc_env_session(env)->os_io;
494 static inline int osc_is_object(const struct lu_object *obj)
496 return obj->lo_dev->ld_type == &osc_device_type;
499 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
501 LINVRNT(d->ld_type == &osc_device_type);
502 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
505 static inline struct obd_export *osc_export(const struct osc_object *obj)
507 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
510 static inline struct client_obd *osc_cli(const struct osc_object *obj)
512 return &osc_export(obj)->exp_obd->u.cli;
515 static inline struct osc_object *cl2osc(const struct cl_object *obj)
517 LINVRNT(osc_is_object(&obj->co_lu));
518 return container_of0(obj, struct osc_object, oo_cl);
521 static inline struct cl_object *osc2cl(const struct osc_object *obj)
523 return (struct cl_object *)&obj->oo_cl;
526 static inline ldlm_mode_t osc_cl_lock2ldlm(enum cl_lock_mode mode)
528 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
529 if (mode == CLM_READ)
531 else if (mode == CLM_WRITE)
537 static inline enum cl_lock_mode osc_ldlm2cl_lock(ldlm_mode_t mode)
539 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
542 else if (mode == LCK_PW)
548 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
550 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
551 return container_of0(slice, struct osc_page, ops_cl);
554 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
556 return container_of0(oap, struct osc_page, ops_oap);
559 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
561 return oap2osc(oap)->ops_cl.cpl_page;
564 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
566 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
569 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
571 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
572 return container_of0(slice, struct osc_lock, ols_cl);
575 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
577 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
580 static inline int osc_io_srvlock(struct osc_io *oio)
582 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
585 enum osc_extent_state {
586 OES_INV = 0, /** extent is just initialized or destroyed */
587 OES_ACTIVE = 1, /** process is using this extent */
588 OES_CACHE = 2, /** extent is ready for IO */
589 OES_LOCKING = 3, /** locking page to prepare IO */
590 OES_LOCK_DONE = 4, /** locking finished, ready to send */
591 OES_RPC = 5, /** in RPC */
592 OES_TRUNC = 6, /** being truncated */
597 * osc_extent data to manage dirty pages.
598 * osc_extent has the following attributes:
599 * 1. all pages in the same must be in one RPC in write back;
600 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
601 * 3. must be covered by only 1 osc_lock;
602 * 4. exclusive. It's impossible to have overlapped osc_extent.
604 * The lifetime of an extent is from when the 1st page is dirtied to when
605 * all pages inside it are written out.
609 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
612 /** red-black tree node */
613 struct rb_node oe_node;
614 /** osc_object of this extent */
615 struct osc_object *oe_obj;
616 /** refcount, removed from red-black tree if reaches zero. */
618 /** busy if non-zero */
620 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
621 struct list_head oe_link;
622 /** state of this extent */
623 unsigned int oe_state;
624 /** flags for this extent. */
625 unsigned int oe_intree:1,
626 /** 0 is write, 1 is read */
630 /** an ACTIVE extent is going to be truncated, so when this extent
631 * is released, it will turn into TRUNC state instead of CACHE. */
633 /** this extent should be written asap and someone may wait for the
634 * write to finish. This bit is usually set along with urgent if
635 * the extent was CACHE state.
636 * fsync_wait extent can't be merged because new extent region may
637 * exceed fsync range. */
639 /** covering lock is being canceled */
641 /** this extent should be written back asap. set if one of pages is
642 * called by page WB daemon, or sync write or reading requests. */
644 /** how many grants allocated for this extent.
645 * Grant allocated for this extent. There is no grant allocated
646 * for reading extents and sync write extents. */
647 unsigned int oe_grants;
648 /** # of dirty pages in this extent */
649 unsigned int oe_nr_pages;
650 /** list of pending oap pages. Pages in this list are NOT sorted. */
651 struct list_head oe_pages;
652 /** Since an extent has to be written out in atomic, this is used to
653 * remember the next page need to be locked to write this extent out.
654 * Not used right now.
656 struct osc_page *oe_next_page;
657 /** start and end index of this extent, include start and end
658 * themselves. Page offset here is the page index of osc_pages.
659 * oe_start is used as keyword for red-black tree. */
662 /** maximum ending index of this extent, this is limited by
663 * max_pages_per_rpc, lock extent and chunk size. */
665 /** waitqueue - for those who want to be notified if this extent's
666 * state has changed. */
667 wait_queue_head_t oe_waitq;
668 /** lock covering this extent */
669 struct cl_lock *oe_osclock;
670 /** terminator of this extent. Must be true if this extent is in IO. */
671 struct task_struct *oe_owner;
672 /** return value of writeback. If somebody is waiting for this extent,
673 * this value can be known by outside world. */
675 /** max pages per rpc when this extent was created */
676 unsigned int oe_mppr;
679 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
681 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
685 #endif /* OSC_CL_INTERNAL_H */