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() */
57 #include <cl_object.h>
59 #include "osc_internal.h"
68 * State maintained by osc layer for each IO context.
72 struct cl_io_slice oi_cl;
73 /** true if this io is lockless. */
75 /** active extents, we know how many bytes is going to be written,
76 * so having an active extent will prevent it from being fragmented */
77 struct osc_extent *oi_active;
78 /** partially truncated extent, we need to hold this extent to prevent
79 * page writeback from happening. */
80 struct osc_extent *oi_trunc;
84 struct obd_info oi_info;
86 struct osc_async_cbargs {
89 struct completion opc_sync;
94 * State of transfer for osc.
97 struct cl_req_slice or_cl;
101 * State maintained by osc layer for the duration of a system call.
107 #define OTI_PVEC_SIZE 256
108 struct osc_thread_info {
109 struct ldlm_res_id oti_resname;
110 ldlm_policy_data_t oti_policy;
111 struct cl_lock_descr oti_descr;
112 struct cl_attr oti_attr;
113 struct lustre_handle oti_handle;
114 struct cl_page_list oti_plist;
116 void *oti_pvec[OTI_PVEC_SIZE];
118 * Fields used by cl_lock_discard_pages().
120 pgoff_t oti_next_index;
121 pgoff_t oti_fn_index; /* first non-overlapped index */
125 struct cl_object oo_cl;
126 struct lov_oinfo *oo_oinfo;
128 * True if locking against this stripe got -EUSERS.
131 cfs_time_t oo_contention_time;
132 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
134 * IO context used for invariant checks in osc_lock_has_pages().
136 struct cl_io oo_debug_io;
137 /** Serialization object for osc_object::oo_debug_io. */
138 struct mutex oo_debug_mutex;
141 * List of pages in transfer.
143 cfs_list_t oo_inflight[CRT_NR];
145 * Lock, protecting ccc_object::cob_inflight, because a seat-belt is
146 * locked during take-off and landing.
148 spinlock_t oo_seatbelt;
151 * used by the osc to keep track of what objects to build into rpcs.
152 * Protected by client_obd->cli_loi_list_lock.
154 cfs_list_t oo_ready_item;
155 cfs_list_t oo_hp_ready_item;
156 cfs_list_t oo_write_item;
157 cfs_list_t oo_read_item;
160 * extent is a red black tree to manage (async) dirty pages.
162 struct rb_root oo_root;
164 * Manage write(dirty) extents.
166 cfs_list_t oo_hp_exts; /* list of hp extents */
167 cfs_list_t oo_urgent_exts; /* list of writeback extents */
168 cfs_list_t oo_rpc_exts;
170 cfs_list_t oo_reading_exts;
172 atomic_t oo_nr_reads;
173 atomic_t oo_nr_writes;
175 /** Protect extent tree. Will be used to protect
176 * oo_{read|write}_pages soon. */
180 * Radix tree for caching pages
182 struct radix_tree_root oo_tree;
183 spinlock_t oo_tree_lock;
184 unsigned long oo_npages;
187 static inline void osc_object_lock(struct osc_object *obj)
189 spin_lock(&obj->oo_lock);
192 static inline int osc_object_trylock(struct osc_object *obj)
194 return spin_trylock(&obj->oo_lock);
197 static inline void osc_object_unlock(struct osc_object *obj)
199 spin_unlock(&obj->oo_lock);
202 static inline int osc_object_is_locked(struct osc_object *obj)
204 return spin_is_locked(&obj->oo_lock);
208 * Lock "micro-states" for osc layer.
210 enum osc_lock_state {
221 * osc-private state of cl_lock.
223 * Interaction with DLM.
225 * CLIO enqueues all DLM locks through ptlrpcd (that is, in "async" mode).
227 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
228 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_lock.
230 * This pointer is protected through a reference, acquired by
231 * osc_lock_upcall0(). Also, an additional reference is acquired by
232 * ldlm_lock_addref() call protecting the lock from cancellation, until
233 * osc_lock_unuse() releases it.
235 * Below is a description of how lock references are acquired and released
238 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
239 * - ldlm_lock_create()
240 * - ldlm_lock_new(): initializes a lock with 2 references. One for
241 * the caller (released when reply from the server is received, or on
242 * error), and another for the hash table.
243 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
245 * - When reply is received from the server (osc_enqueue_interpret())
246 * - ldlm_cli_enqueue_fini()
247 * - LDLM_LOCK_PUT(): releases caller reference acquired by
250 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
251 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
253 * - When lock is being cancelled (ldlm_lock_cancel())
254 * - ldlm_lock_destroy()
255 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
258 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
259 * either when lock is cancelled (osc_lock_blocking()), or when locks is
260 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
261 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
265 struct cl_lock_slice ols_cl;
266 /** underlying DLM lock */
267 struct ldlm_lock *ols_lock;
268 /** lock value block */
269 struct ost_lvb ols_lvb;
270 /** DLM flags with which osc_lock::ols_lock was enqueued */
272 /** osc_lock::ols_lock handle */
273 struct lustre_handle ols_handle;
274 struct ldlm_enqueue_info ols_einfo;
275 enum osc_lock_state ols_state;
278 * true, if ldlm_lock_addref() was called against
279 * osc_lock::ols_lock. This is used for sanity checking.
281 * \see osc_lock::ols_has_ref
283 unsigned ols_hold :1,
285 * this is much like osc_lock::ols_hold, except that this bit is
286 * cleared _after_ reference in released in osc_lock_unuse(). This
287 * fine distinction is needed because:
289 * - if ldlm lock still has a reference, osc_ast_data_get() needs
290 * to return associated cl_lock (so that a flag is needed that is
291 * cleared after ldlm_lock_decref() returned), and
293 * - ldlm_lock_decref() can invoke blocking ast (for a
294 * LDLM_FL_CBPENDING lock), and osc_lock functions like
295 * osc_lock_cancel() called from there need to know whether to
296 * release lock reference (so that a flag is needed that is
297 * cleared before ldlm_lock_decref() is called).
301 * inherit the lockless attribute from top level cl_io.
302 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
306 * set by osc_lock_use() to wait until blocking AST enters into
307 * osc_ldlm_blocking_ast0(), so that cl_lock mutex can be used for
308 * further synchronization.
312 * If the data of this lock has been flushed to server side.
316 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
317 * the EVAVAIL error as torerable, this will make upper logic happy
318 * to wait all glimpse locks to each OSTs to be completed.
319 * Glimpse lock converts to normal lock if the server lock is
321 * Glimpse lock should be destroyed immediately after use.
325 * For async glimpse lock.
329 * IO that owns this lock. This field is used for a dead-lock
330 * avoidance by osc_lock_enqueue_wait().
332 * XXX: unfortunately, the owner of a osc_lock is not unique,
333 * the lock may have multiple users, if the lock is granted and
336 struct osc_io *ols_owner;
341 * Page state private for osc layer.
344 struct cl_page_slice ops_cl;
346 * Page queues used by osc to detect when RPC can be formed.
348 struct osc_async_page ops_oap;
350 * An offset within page from which next transfer starts. This is used
351 * by cl_page_clip() to submit partial page transfers.
355 * An offset within page at which next transfer ends.
357 * \see osc_page::ops_from.
361 * Boolean, true iff page is under transfer. Used for sanity checking.
363 unsigned ops_transfer_pinned:1,
365 * True for a `temporary page' created by read-ahead code, probably
366 * outside of any DLM lock.
374 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
378 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
382 * Linkage into a per-osc_object list of pages in flight. For
385 cfs_list_t ops_inflight;
387 * Thread that submitted this page for transfer. For debugging.
389 struct task_struct *ops_submitter;
391 * Submit time - the time when the page is starting RPC. For debugging.
393 cfs_time_t ops_submit_time;
396 extern struct kmem_cache *osc_lock_kmem;
397 extern struct kmem_cache *osc_object_kmem;
398 extern struct kmem_cache *osc_thread_kmem;
399 extern struct kmem_cache *osc_session_kmem;
400 extern struct kmem_cache *osc_req_kmem;
401 extern struct kmem_cache *osc_extent_kmem;
403 extern struct lu_device_type osc_device_type;
404 extern struct lu_context_key osc_key;
405 extern struct lu_context_key osc_session_key;
407 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
409 int osc_lock_init(const struct lu_env *env,
410 struct cl_object *obj, struct cl_lock *lock,
411 const struct cl_io *io);
412 int osc_io_init (const struct lu_env *env,
413 struct cl_object *obj, struct cl_io *io);
414 int osc_req_init (const struct lu_env *env, struct cl_device *dev,
416 struct lu_object *osc_object_alloc(const struct lu_env *env,
417 const struct lu_object_header *hdr,
418 struct lu_device *dev);
419 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
420 struct cl_page *page, pgoff_t ind);
422 void osc_index2policy (ldlm_policy_data_t *policy, const struct cl_object *obj,
423 pgoff_t start, pgoff_t end);
424 int osc_lvb_print (const struct lu_env *env, void *cookie,
425 lu_printer_t p, const struct ost_lvb *lvb);
427 void osc_lru_add_batch(struct client_obd *cli, cfs_list_t *list);
428 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
429 enum cl_req_type crt, int brw_flags);
430 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
431 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
432 obd_flag async_flags);
433 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
434 struct page *page, loff_t offset);
435 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
436 struct osc_page *ops);
437 int osc_page_cache_add(const struct lu_env *env,
438 const struct cl_page_slice *slice, struct cl_io *io);
439 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
440 struct osc_page *ops);
441 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
442 struct osc_page *ops);
443 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
444 cfs_list_t *list, int cmd, int brw_flags);
445 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
446 struct osc_object *obj, __u64 size);
447 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
448 struct osc_object *obj);
449 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
450 pgoff_t start, pgoff_t end, int hp, int discard);
451 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
452 pgoff_t start, pgoff_t end);
453 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
454 struct osc_object *osc, pdl_policy_t pol);
455 int lru_queue_work(const struct lu_env *env, void *data);
457 void osc_object_set_contended (struct osc_object *obj);
458 void osc_object_clear_contended(struct osc_object *obj);
459 int osc_object_is_contended (struct osc_object *obj);
461 int osc_lock_is_lockless (const struct osc_lock *olck);
463 /*****************************************************************************
469 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
471 struct osc_thread_info *info;
473 info = lu_context_key_get(&env->le_ctx, &osc_key);
474 LASSERT(info != NULL);
478 static inline struct osc_session *osc_env_session(const struct lu_env *env)
480 struct osc_session *ses;
482 ses = lu_context_key_get(env->le_ses, &osc_session_key);
483 LASSERT(ses != NULL);
487 static inline struct osc_io *osc_env_io(const struct lu_env *env)
489 return &osc_env_session(env)->os_io;
492 static inline int osc_is_object(const struct lu_object *obj)
494 return obj->lo_dev->ld_type == &osc_device_type;
497 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
499 LINVRNT(d->ld_type == &osc_device_type);
500 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
503 static inline struct obd_export *osc_export(const struct osc_object *obj)
505 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
508 static inline struct client_obd *osc_cli(const struct osc_object *obj)
510 return &osc_export(obj)->exp_obd->u.cli;
513 static inline struct osc_object *cl2osc(const struct cl_object *obj)
515 LINVRNT(osc_is_object(&obj->co_lu));
516 return container_of0(obj, struct osc_object, oo_cl);
519 static inline struct cl_object *osc2cl(const struct osc_object *obj)
521 return (struct cl_object *)&obj->oo_cl;
524 static inline ldlm_mode_t osc_cl_lock2ldlm(enum cl_lock_mode mode)
526 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
527 if (mode == CLM_READ)
529 else if (mode == CLM_WRITE)
535 static inline enum cl_lock_mode osc_ldlm2cl_lock(ldlm_mode_t mode)
537 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
540 else if (mode == LCK_PW)
546 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
548 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
549 return container_of0(slice, struct osc_page, ops_cl);
552 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
554 return container_of0(oap, struct osc_page, ops_oap);
557 static inline pgoff_t osc_index(struct osc_page *opg)
559 return opg->ops_cl.cpl_index;
562 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
564 return oap2osc(oap)->ops_cl.cpl_page;
567 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
569 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
572 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
574 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
575 return container_of0(slice, struct osc_lock, ols_cl);
578 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
580 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
583 static inline int osc_io_srvlock(struct osc_io *oio)
585 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
588 enum osc_extent_state {
589 OES_INV = 0, /** extent is just initialized or destroyed */
590 OES_ACTIVE = 1, /** process is using this extent */
591 OES_CACHE = 2, /** extent is ready for IO */
592 OES_LOCKING = 3, /** locking page to prepare IO */
593 OES_LOCK_DONE = 4, /** locking finished, ready to send */
594 OES_RPC = 5, /** in RPC */
595 OES_TRUNC = 6, /** being truncated */
600 * osc_extent data to manage dirty pages.
601 * osc_extent has the following attributes:
602 * 1. all pages in the same must be in one RPC in write back;
603 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
604 * 3. must be covered by only 1 osc_lock;
605 * 4. exclusive. It's impossible to have overlapped osc_extent.
607 * The lifetime of an extent is from when the 1st page is dirtied to when
608 * all pages inside it are written out.
612 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
615 /** red-black tree node */
616 struct rb_node oe_node;
617 /** osc_object of this extent */
618 struct osc_object *oe_obj;
619 /** refcount, removed from red-black tree if reaches zero. */
621 /** busy if non-zero */
623 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
625 /** state of this extent */
626 unsigned int oe_state;
627 /** flags for this extent. */
628 unsigned int oe_intree:1,
629 /** 0 is write, 1 is read */
633 /** an ACTIVE extent is going to be truncated, so when this extent
634 * is released, it will turn into TRUNC state instead of CACHE. */
636 /** this extent should be written asap and someone may wait for the
637 * write to finish. This bit is usually set along with urgent if
638 * the extent was CACHE state.
639 * fsync_wait extent can't be merged because new extent region may
640 * exceed fsync range. */
642 /** covering lock is being canceled */
644 /** this extent should be written back asap. set if one of pages is
645 * called by page WB daemon, or sync write or reading requests. */
647 /** how many grants allocated for this extent.
648 * Grant allocated for this extent. There is no grant allocated
649 * for reading extents and sync write extents. */
650 unsigned int oe_grants;
651 /** # of dirty pages in this extent */
652 unsigned int oe_nr_pages;
653 /** list of pending oap pages. Pages in this list are NOT sorted. */
655 /** Since an extent has to be written out in atomic, this is used to
656 * remember the next page need to be locked to write this extent out.
657 * Not used right now.
659 struct osc_page *oe_next_page;
660 /** start and end index of this extent, include start and end
661 * themselves. Page offset here is the page index of osc_pages.
662 * oe_start is used as keyword for red-black tree. */
665 /** maximum ending index of this extent, this is limited by
666 * max_pages_per_rpc, lock extent and chunk size. */
668 /** waitqueue - for those who want to be notified if this extent's
669 * state has changed. */
670 wait_queue_head_t oe_waitq;
671 /** lock covering this extent */
672 struct cl_lock *oe_osclock;
673 /** terminator of this extent. Must be true if this extent is in IO. */
674 struct task_struct *oe_owner;
675 /** return value of writeback. If somebody is waiting for this extent,
676 * this value can be known by outside world. */
678 /** max pages per rpc when this extent was created */
679 unsigned int oe_mppr;
682 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
684 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
686 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *lock);
688 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
689 struct osc_page *, void *);
690 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
691 struct osc_object *osc, pgoff_t start, pgoff_t end,
692 osc_page_gang_cbt cb, void *cbdata);
696 #endif /* OSC_CL_INTERNAL_H */