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 osc_async_cbargs {
84 struct completion opc_sync;
89 * State maintained by osc layer for the duration of a system call.
95 #define OTI_PVEC_SIZE 256
96 struct osc_thread_info {
97 struct ldlm_res_id oti_resname;
98 union ldlm_policy_data oti_policy;
99 struct cl_lock_descr oti_descr;
100 struct cl_attr oti_attr;
101 struct lustre_handle oti_handle;
102 struct cl_page_list oti_plist;
104 void *oti_pvec[OTI_PVEC_SIZE];
106 * Fields used by cl_lock_discard_pages().
108 pgoff_t oti_next_index;
109 pgoff_t oti_fn_index; /* first non-overlapped index */
110 struct cl_sync_io oti_anchor;
111 struct cl_req_attr oti_req_attr;
115 struct cl_object oo_cl;
116 struct lov_oinfo *oo_oinfo;
118 * True if locking against this stripe got -EUSERS.
121 cfs_time_t oo_contention_time;
122 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
124 * IO context used for invariant checks in osc_lock_has_pages().
126 struct cl_io oo_debug_io;
127 /** Serialization object for osc_object::oo_debug_io. */
128 struct mutex oo_debug_mutex;
131 * used by the osc to keep track of what objects to build into rpcs.
132 * Protected by client_obd->cli_loi_list_lock.
134 struct list_head oo_ready_item;
135 struct list_head oo_hp_ready_item;
136 struct list_head oo_write_item;
137 struct list_head oo_read_item;
140 * extent is a red black tree to manage (async) dirty pages.
142 struct rb_root oo_root;
144 * Manage write(dirty) extents.
146 struct list_head oo_hp_exts; /* list of hp extents */
147 struct list_head oo_urgent_exts; /* list of writeback extents */
148 struct list_head oo_rpc_exts;
150 struct list_head oo_reading_exts;
152 atomic_t oo_nr_reads;
153 atomic_t oo_nr_writes;
155 /** Protect extent tree. Will be used to protect
156 * oo_{read|write}_pages soon. */
160 * Radix tree for caching pages
162 struct radix_tree_root oo_tree;
163 spinlock_t oo_tree_lock;
164 unsigned long oo_npages;
166 /* Protect osc_lock this osc_object has */
167 spinlock_t oo_ol_spin;
168 struct list_head oo_ol_list;
170 /** number of active IOs of this object */
172 wait_queue_head_t oo_io_waitq;
175 static inline void osc_object_lock(struct osc_object *obj)
177 spin_lock(&obj->oo_lock);
180 static inline int osc_object_trylock(struct osc_object *obj)
182 return spin_trylock(&obj->oo_lock);
185 static inline void osc_object_unlock(struct osc_object *obj)
187 spin_unlock(&obj->oo_lock);
190 static inline int osc_object_is_locked(struct osc_object *obj)
192 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
193 return spin_is_locked(&obj->oo_lock);
196 * It is not perfect to return true all the time.
197 * But since this function is only used for assertion
198 * and checking, it seems OK.
205 * Lock "micro-states" for osc layer.
207 enum osc_lock_state {
216 * osc-private state of cl_lock.
218 * Interaction with DLM.
220 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
221 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_dlmlock.
223 * This pointer is protected through a reference, acquired by
224 * osc_lock_upcall0(). Also, an additional reference is acquired by
225 * ldlm_lock_addref() call protecting the lock from cancellation, until
226 * osc_lock_unuse() releases it.
228 * Below is a description of how lock references are acquired and released
231 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
232 * - ldlm_lock_create()
233 * - ldlm_lock_new(): initializes a lock with 2 references. One for
234 * the caller (released when reply from the server is received, or on
235 * error), and another for the hash table.
236 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
238 * - When reply is received from the server (osc_enqueue_interpret())
239 * - ldlm_cli_enqueue_fini()
240 * - LDLM_LOCK_PUT(): releases caller reference acquired by
243 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
244 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
246 * - When lock is being cancelled (ldlm_lock_cancel())
247 * - ldlm_lock_destroy()
248 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
251 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
252 * either when lock is cancelled (osc_lock_blocking()), or when locks is
253 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
254 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
258 struct cl_lock_slice ols_cl;
259 /** Internal lock to protect states, etc. */
261 /** Owner sleeps on this channel for state change */
262 struct cl_sync_io *ols_owner;
263 /** waiting list for this lock to be cancelled */
264 struct list_head ols_waiting_list;
265 /** wait entry of ols_waiting_list */
266 struct list_head ols_wait_entry;
267 /** list entry for osc_object::oo_ol_list */
268 struct list_head ols_nextlock_oscobj;
270 /** underlying DLM lock */
271 struct ldlm_lock *ols_dlmlock;
272 /** DLM flags with which osc_lock::ols_lock was enqueued */
274 /** osc_lock::ols_lock handle */
275 struct lustre_handle ols_handle;
276 struct ldlm_enqueue_info ols_einfo;
277 enum osc_lock_state ols_state;
278 /** lock value block */
279 struct ost_lvb ols_lvb;
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 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
311 * the EVAVAIL error as torerable, this will make upper logic happy
312 * to wait all glimpse locks to each OSTs to be completed.
313 * Glimpse lock converts to normal lock if the server lock is
315 * Glimpse lock should be destroyed immediately after use.
319 * For async glimpse lock.
326 * Page state private for osc layer.
329 struct cl_page_slice ops_cl;
331 * Page queues used by osc to detect when RPC can be formed.
333 struct osc_async_page ops_oap;
335 * An offset within page from which next transfer starts. This is used
336 * by cl_page_clip() to submit partial page transfers.
340 * An offset within page at which next transfer ends.
342 * \see osc_page::ops_from.
346 * Boolean, true iff page is under transfer. Used for sanity checking.
348 unsigned ops_transfer_pinned:1,
350 * True for a `temporary page' created by read-ahead code, probably
351 * outside of any DLM lock.
359 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
363 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
365 struct list_head ops_lru;
367 * Submit time - the time when the page is starting RPC. For debugging.
369 cfs_time_t ops_submit_time;
372 extern struct kmem_cache *osc_lock_kmem;
373 extern struct kmem_cache *osc_object_kmem;
374 extern struct kmem_cache *osc_thread_kmem;
375 extern struct kmem_cache *osc_session_kmem;
376 extern struct kmem_cache *osc_extent_kmem;
378 extern struct lu_device_type osc_device_type;
379 extern struct lu_context_key osc_key;
380 extern struct lu_context_key osc_session_key;
382 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
384 int osc_lock_init(const struct lu_env *env,
385 struct cl_object *obj, struct cl_lock *lock,
386 const struct cl_io *io);
387 int osc_io_init (const struct lu_env *env,
388 struct cl_object *obj, struct cl_io *io);
389 struct lu_object *osc_object_alloc(const struct lu_env *env,
390 const struct lu_object_header *hdr,
391 struct lu_device *dev);
392 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
393 struct cl_page *page, pgoff_t ind);
395 void osc_index2policy(union ldlm_policy_data *policy,
396 const struct cl_object *obj, pgoff_t start, pgoff_t end);
397 int osc_lvb_print(const struct lu_env *env, void *cookie,
398 lu_printer_t p, const struct ost_lvb *lvb);
400 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
401 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
402 enum cl_req_type crt, int brw_flags);
403 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
404 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
406 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
407 struct page *page, loff_t offset);
408 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
409 struct osc_page *ops);
410 int osc_page_cache_add(const struct lu_env *env,
411 const struct cl_page_slice *slice, struct cl_io *io);
412 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
413 struct osc_page *ops);
414 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
415 struct osc_page *ops);
416 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
417 struct list_head *list, int cmd, int brw_flags);
418 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
419 __u64 size, struct osc_extent **extp);
420 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext);
421 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
422 pgoff_t start, pgoff_t end, int hp, int discard);
423 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
424 pgoff_t start, pgoff_t end);
425 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
426 struct osc_object *osc);
427 int lru_queue_work(const struct lu_env *env, void *data);
429 void osc_object_set_contended (struct osc_object *obj);
430 void osc_object_clear_contended(struct osc_object *obj);
431 int osc_object_is_contended (struct osc_object *obj);
433 int osc_lock_is_lockless (const struct osc_lock *olck);
435 /*****************************************************************************
441 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
443 struct osc_thread_info *info;
445 info = lu_context_key_get(&env->le_ctx, &osc_key);
446 LASSERT(info != NULL);
450 static inline struct osc_session *osc_env_session(const struct lu_env *env)
452 struct osc_session *ses;
454 ses = lu_context_key_get(env->le_ses, &osc_session_key);
455 LASSERT(ses != NULL);
459 static inline struct osc_io *osc_env_io(const struct lu_env *env)
461 return &osc_env_session(env)->os_io;
464 static inline int osc_is_object(const struct lu_object *obj)
466 return obj->lo_dev->ld_type == &osc_device_type;
469 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
471 LINVRNT(d->ld_type == &osc_device_type);
472 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
475 static inline struct obd_export *osc_export(const struct osc_object *obj)
477 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
480 static inline struct client_obd *osc_cli(const struct osc_object *obj)
482 return &osc_export(obj)->exp_obd->u.cli;
485 static inline struct osc_object *cl2osc(const struct cl_object *obj)
487 LINVRNT(osc_is_object(&obj->co_lu));
488 return container_of0(obj, struct osc_object, oo_cl);
491 static inline struct cl_object *osc2cl(const struct osc_object *obj)
493 return (struct cl_object *)&obj->oo_cl;
496 static inline enum ldlm_mode osc_cl_lock2ldlm(enum cl_lock_mode mode)
498 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
499 if (mode == CLM_READ)
501 if (mode == CLM_WRITE)
506 static inline enum cl_lock_mode osc_ldlm2cl_lock(enum ldlm_mode mode)
508 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
516 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
518 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
519 return container_of0(slice, struct osc_page, ops_cl);
522 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
524 return container_of0(oap, struct osc_page, ops_oap);
527 static inline pgoff_t osc_index(struct osc_page *opg)
529 return opg->ops_cl.cpl_index;
532 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
534 return oap2osc(oap)->ops_cl.cpl_page;
537 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
539 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
542 static inline struct osc_page *
543 osc_cl_page_osc(struct cl_page *page, struct osc_object *osc)
545 const struct cl_page_slice *slice;
547 LASSERT(osc != NULL);
548 slice = cl_object_page_slice(&osc->oo_cl, page);
549 return cl2osc_page(slice);
552 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
554 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
555 return container_of0(slice, struct osc_lock, ols_cl);
558 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
560 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
563 static inline int osc_io_srvlock(struct osc_io *oio)
565 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
568 enum osc_extent_state {
569 OES_INV = 0, /** extent is just initialized or destroyed */
570 OES_ACTIVE = 1, /** process is using this extent */
571 OES_CACHE = 2, /** extent is ready for IO */
572 OES_LOCKING = 3, /** locking page to prepare IO */
573 OES_LOCK_DONE = 4, /** locking finished, ready to send */
574 OES_RPC = 5, /** in RPC */
575 OES_TRUNC = 6, /** being truncated */
580 * osc_extent data to manage dirty pages.
581 * osc_extent has the following attributes:
582 * 1. all pages in the same must be in one RPC in write back;
583 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
584 * 3. must be covered by only 1 osc_lock;
585 * 4. exclusive. It's impossible to have overlapped osc_extent.
587 * The lifetime of an extent is from when the 1st page is dirtied to when
588 * all pages inside it are written out.
592 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
595 /** red-black tree node */
596 struct rb_node oe_node;
597 /** osc_object of this extent */
598 struct osc_object *oe_obj;
599 /** refcount, removed from red-black tree if reaches zero. */
601 /** busy if non-zero */
603 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
604 struct list_head oe_link;
605 /** state of this extent */
606 enum osc_extent_state oe_state;
607 /** flags for this extent. */
608 unsigned int oe_intree:1,
609 /** 0 is write, 1 is read */
611 /** sync extent, queued by osc_queue_sync_pages() */
615 /** an ACTIVE extent is going to be truncated, so when this extent
616 * is released, it will turn into TRUNC state instead of CACHE. */
618 /** this extent should be written asap and someone may wait for the
619 * write to finish. This bit is usually set along with urgent if
620 * the extent was CACHE state.
621 * fsync_wait extent can't be merged because new extent region may
622 * exceed fsync range. */
624 /** covering lock is being canceled */
626 /** this extent should be written back asap. set if one of pages is
627 * called by page WB daemon, or sync write or reading requests. */
629 /** how many grants allocated for this extent.
630 * Grant allocated for this extent. There is no grant allocated
631 * for reading extents and sync write extents. */
632 unsigned int oe_grants;
633 /** # of dirty pages in this extent */
634 unsigned int oe_nr_pages;
635 /** list of pending oap pages. Pages in this list are NOT sorted. */
636 struct list_head oe_pages;
637 /** Since an extent has to be written out in atomic, this is used to
638 * remember the next page need to be locked to write this extent out.
639 * Not used right now.
641 struct osc_page *oe_next_page;
642 /** start and end index of this extent, include start and end
643 * themselves. Page offset here is the page index of osc_pages.
644 * oe_start is used as keyword for red-black tree. */
647 /** maximum ending index of this extent, this is limited by
648 * max_pages_per_rpc, lock extent and chunk size. */
650 /** waitqueue - for those who want to be notified if this extent's
651 * state has changed. */
652 wait_queue_head_t oe_waitq;
653 /** lock covering this extent */
654 struct ldlm_lock *oe_dlmlock;
655 /** terminator of this extent. Must be true if this extent is in IO. */
656 struct task_struct *oe_owner;
657 /** return value of writeback. If somebody is waiting for this extent,
658 * this value can be known by outside world. */
660 /** max pages per rpc when this extent was created */
661 unsigned int oe_mppr;
664 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
666 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
668 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
669 pgoff_t start, pgoff_t end, enum cl_lock_mode mode);
671 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
672 struct osc_page *, void *);
673 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
674 struct osc_object *osc, pgoff_t start, pgoff_t end,
675 osc_page_gang_cbt cb, void *cbdata);
678 #endif /* OSC_CL_INTERNAL_H */