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. */
68 unsigned int oi_lockless:1,
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;
114 struct lu_buf oti_ladvise_buf;
118 struct cl_object oo_cl;
119 struct lov_oinfo *oo_oinfo;
121 * True if locking against this stripe got -EUSERS.
124 cfs_time_t oo_contention_time;
125 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
127 * IO context used for invariant checks in osc_lock_has_pages().
129 struct cl_io oo_debug_io;
130 /** Serialization object for osc_object::oo_debug_io. */
131 struct mutex oo_debug_mutex;
134 * used by the osc to keep track of what objects to build into rpcs.
135 * Protected by client_obd->cli_loi_list_lock.
137 struct list_head oo_ready_item;
138 struct list_head oo_hp_ready_item;
139 struct list_head oo_write_item;
140 struct list_head oo_read_item;
143 * extent is a red black tree to manage (async) dirty pages.
145 struct rb_root oo_root;
147 * Manage write(dirty) extents.
149 struct list_head oo_hp_exts; /* list of hp extents */
150 struct list_head oo_urgent_exts; /* list of writeback extents */
151 struct list_head oo_rpc_exts;
153 struct list_head oo_reading_exts;
155 atomic_t oo_nr_reads;
156 atomic_t oo_nr_writes;
158 /** Protect extent tree. Will be used to protect
159 * oo_{read|write}_pages soon. */
163 * Radix tree for caching pages
165 spinlock_t oo_tree_lock;
166 struct radix_tree_root oo_tree;
167 unsigned long oo_npages;
169 /* Protect osc_lock this osc_object has */
170 struct list_head oo_ol_list;
171 spinlock_t oo_ol_spin;
173 /** number of active IOs of this object */
175 wait_queue_head_t oo_io_waitq;
178 static inline void osc_object_lock(struct osc_object *obj)
180 spin_lock(&obj->oo_lock);
183 static inline int osc_object_trylock(struct osc_object *obj)
185 return spin_trylock(&obj->oo_lock);
188 static inline void osc_object_unlock(struct osc_object *obj)
190 spin_unlock(&obj->oo_lock);
193 static inline int osc_object_is_locked(struct osc_object *obj)
195 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
196 return spin_is_locked(&obj->oo_lock);
199 * It is not perfect to return true all the time.
200 * But since this function is only used for assertion
201 * and checking, it seems OK.
208 * Lock "micro-states" for osc layer.
210 enum osc_lock_state {
219 * osc-private state of cl_lock.
221 * Interaction with DLM.
223 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
224 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_dlmlock.
226 * This pointer is protected through a reference, acquired by
227 * osc_lock_upcall0(). Also, an additional reference is acquired by
228 * ldlm_lock_addref() call protecting the lock from cancellation, until
229 * osc_lock_unuse() releases it.
231 * Below is a description of how lock references are acquired and released
234 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
235 * - ldlm_lock_create()
236 * - ldlm_lock_new(): initializes a lock with 2 references. One for
237 * the caller (released when reply from the server is received, or on
238 * error), and another for the hash table.
239 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
241 * - When reply is received from the server (osc_enqueue_interpret())
242 * - ldlm_cli_enqueue_fini()
243 * - LDLM_LOCK_PUT(): releases caller reference acquired by
246 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
247 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
249 * - When lock is being cancelled (ldlm_lock_cancel())
250 * - ldlm_lock_destroy()
251 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
254 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
255 * either when lock is cancelled (osc_lock_blocking()), or when locks is
256 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
257 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
261 struct cl_lock_slice ols_cl;
262 /** Internal lock to protect states, etc. */
264 /** Owner sleeps on this channel for state change */
265 struct cl_sync_io *ols_owner;
266 /** waiting list for this lock to be cancelled */
267 struct list_head ols_waiting_list;
268 /** wait entry of ols_waiting_list */
269 struct list_head ols_wait_entry;
270 /** list entry for osc_object::oo_ol_list */
271 struct list_head ols_nextlock_oscobj;
273 /** underlying DLM lock */
274 struct ldlm_lock *ols_dlmlock;
275 /** DLM flags with which osc_lock::ols_lock was enqueued */
277 /** osc_lock::ols_lock handle */
278 struct lustre_handle ols_handle;
279 struct ldlm_enqueue_info ols_einfo;
280 enum osc_lock_state ols_state;
281 /** lock value block */
282 struct ost_lvb ols_lvb;
285 * true, if ldlm_lock_addref() was called against
286 * osc_lock::ols_lock. This is used for sanity checking.
288 * \see osc_lock::ols_has_ref
290 unsigned ols_hold :1,
292 * this is much like osc_lock::ols_hold, except that this bit is
293 * cleared _after_ reference in released in osc_lock_unuse(). This
294 * fine distinction is needed because:
296 * - if ldlm lock still has a reference, osc_ast_data_get() needs
297 * to return associated cl_lock (so that a flag is needed that is
298 * cleared after ldlm_lock_decref() returned), and
300 * - ldlm_lock_decref() can invoke blocking ast (for a
301 * LDLM_FL_CBPENDING lock), and osc_lock functions like
302 * osc_lock_cancel() called from there need to know whether to
303 * release lock reference (so that a flag is needed that is
304 * cleared before ldlm_lock_decref() is called).
308 * inherit the lockless attribute from top level cl_io.
309 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
313 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
314 * the EVAVAIL error as torerable, this will make upper logic happy
315 * to wait all glimpse locks to each OSTs to be completed.
316 * Glimpse lock converts to normal lock if the server lock is
318 * Glimpse lock should be destroyed immediately after use.
322 * For async glimpse lock.
329 * Page state private for osc layer.
332 struct cl_page_slice ops_cl;
334 * Page queues used by osc to detect when RPC can be formed.
336 struct osc_async_page ops_oap;
338 * An offset within page from which next transfer starts. This is used
339 * by cl_page_clip() to submit partial page transfers.
343 * An offset within page at which next transfer ends.
345 * \see osc_page::ops_from.
349 * Boolean, true iff page is under transfer. Used for sanity checking.
351 unsigned ops_transfer_pinned:1,
353 * True for a `temporary page' created by read-ahead code, probably
354 * outside of any DLM lock.
362 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
366 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
368 struct list_head ops_lru;
370 * Submit time - the time when the page is starting RPC. For debugging.
372 cfs_time_t ops_submit_time;
375 extern struct kmem_cache *osc_lock_kmem;
376 extern struct kmem_cache *osc_object_kmem;
377 extern struct kmem_cache *osc_thread_kmem;
378 extern struct kmem_cache *osc_session_kmem;
379 extern struct kmem_cache *osc_extent_kmem;
381 extern struct lu_device_type osc_device_type;
382 extern struct lu_context_key osc_key;
383 extern struct lu_context_key osc_session_key;
385 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
387 int osc_lock_init(const struct lu_env *env,
388 struct cl_object *obj, struct cl_lock *lock,
389 const struct cl_io *io);
390 int osc_io_init (const struct lu_env *env,
391 struct cl_object *obj, struct cl_io *io);
392 struct lu_object *osc_object_alloc(const struct lu_env *env,
393 const struct lu_object_header *hdr,
394 struct lu_device *dev);
395 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
396 struct cl_page *page, pgoff_t ind);
398 void osc_index2policy(union ldlm_policy_data *policy,
399 const struct cl_object *obj, pgoff_t start, pgoff_t end);
400 int osc_lvb_print(const struct lu_env *env, void *cookie,
401 lu_printer_t p, const struct ost_lvb *lvb);
403 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
404 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
405 enum cl_req_type crt, int brw_flags);
406 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
407 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
409 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
410 struct page *page, loff_t offset);
411 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
412 struct osc_page *ops);
413 int osc_page_cache_add(const struct lu_env *env,
414 const struct cl_page_slice *slice, struct cl_io *io);
415 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
416 struct osc_page *ops);
417 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
418 struct osc_page *ops);
419 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
420 struct list_head *list, int cmd, int brw_flags);
421 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
422 __u64 size, struct osc_extent **extp);
423 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext);
424 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
425 pgoff_t start, pgoff_t end, int hp, int discard);
426 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
427 pgoff_t start, pgoff_t end);
428 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
429 struct osc_object *osc);
430 int lru_queue_work(const struct lu_env *env, void *data);
432 void osc_object_set_contended (struct osc_object *obj);
433 void osc_object_clear_contended(struct osc_object *obj);
434 int osc_object_is_contended (struct osc_object *obj);
436 int osc_lock_is_lockless (const struct osc_lock *olck);
438 /*****************************************************************************
444 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
446 struct osc_thread_info *info;
448 info = lu_context_key_get(&env->le_ctx, &osc_key);
449 LASSERT(info != NULL);
453 static inline struct osc_session *osc_env_session(const struct lu_env *env)
455 struct osc_session *ses;
457 ses = lu_context_key_get(env->le_ses, &osc_session_key);
458 LASSERT(ses != NULL);
462 static inline struct osc_io *osc_env_io(const struct lu_env *env)
464 return &osc_env_session(env)->os_io;
467 static inline int osc_is_object(const struct lu_object *obj)
469 return obj->lo_dev->ld_type == &osc_device_type;
472 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
474 LINVRNT(d->ld_type == &osc_device_type);
475 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
478 static inline struct obd_export *osc_export(const struct osc_object *obj)
480 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
483 static inline struct client_obd *osc_cli(const struct osc_object *obj)
485 return &osc_export(obj)->exp_obd->u.cli;
488 static inline struct osc_object *cl2osc(const struct cl_object *obj)
490 LINVRNT(osc_is_object(&obj->co_lu));
491 return container_of0(obj, struct osc_object, oo_cl);
494 static inline struct cl_object *osc2cl(const struct osc_object *obj)
496 return (struct cl_object *)&obj->oo_cl;
499 static inline enum ldlm_mode osc_cl_lock2ldlm(enum cl_lock_mode mode)
501 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
502 if (mode == CLM_READ)
504 if (mode == CLM_WRITE)
509 static inline enum cl_lock_mode osc_ldlm2cl_lock(enum ldlm_mode mode)
511 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
519 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
521 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
522 return container_of0(slice, struct osc_page, ops_cl);
525 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
527 return container_of0(oap, struct osc_page, ops_oap);
530 static inline pgoff_t osc_index(struct osc_page *opg)
532 return opg->ops_cl.cpl_index;
535 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
537 return oap2osc(oap)->ops_cl.cpl_page;
540 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
542 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
545 static inline struct osc_page *
546 osc_cl_page_osc(struct cl_page *page, struct osc_object *osc)
548 const struct cl_page_slice *slice;
550 LASSERT(osc != NULL);
551 slice = cl_object_page_slice(&osc->oo_cl, page);
552 return cl2osc_page(slice);
555 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
557 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
558 return container_of0(slice, struct osc_lock, ols_cl);
561 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
563 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
566 static inline int osc_io_srvlock(struct osc_io *oio)
568 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
571 enum osc_extent_state {
572 OES_INV = 0, /** extent is just initialized or destroyed */
573 OES_ACTIVE = 1, /** process is using this extent */
574 OES_CACHE = 2, /** extent is ready for IO */
575 OES_LOCKING = 3, /** locking page to prepare IO */
576 OES_LOCK_DONE = 4, /** locking finished, ready to send */
577 OES_RPC = 5, /** in RPC */
578 OES_TRUNC = 6, /** being truncated */
583 * osc_extent data to manage dirty pages.
584 * osc_extent has the following attributes:
585 * 1. all pages in the same must be in one RPC in write back;
586 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
587 * 3. must be covered by only 1 osc_lock;
588 * 4. exclusive. It's impossible to have overlapped osc_extent.
590 * The lifetime of an extent is from when the 1st page is dirtied to when
591 * all pages inside it are written out.
595 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
598 /** red-black tree node */
599 struct rb_node oe_node;
600 /** osc_object of this extent */
601 struct osc_object *oe_obj;
602 /** refcount, removed from red-black tree if reaches zero. */
604 /** busy if non-zero */
606 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
607 struct list_head oe_link;
608 /** state of this extent */
609 enum osc_extent_state oe_state;
610 /** flags for this extent. */
611 unsigned int oe_intree:1,
612 /** 0 is write, 1 is read */
614 /** sync extent, queued by osc_queue_sync_pages() */
616 /** set if this extent has partial, sync pages.
617 * Extents with partial page(s) can't merge with others in RPC */
621 /** an ACTIVE extent is going to be truncated, so when this extent
622 * is released, it will turn into TRUNC state instead of CACHE. */
624 /** this extent should be written asap and someone may wait for the
625 * write to finish. This bit is usually set along with urgent if
626 * the extent was CACHE state.
627 * fsync_wait extent can't be merged because new extent region may
628 * exceed fsync range. */
630 /** covering lock is being canceled */
632 /** this extent should be written back asap. set if one of pages is
633 * called by page WB daemon, or sync write or reading requests. */
635 /** how many grants allocated for this extent.
636 * Grant allocated for this extent. There is no grant allocated
637 * for reading extents and sync write extents. */
638 unsigned int oe_grants;
639 /** # of dirty pages in this extent */
640 unsigned int oe_nr_pages;
641 /** list of pending oap pages. Pages in this list are NOT sorted. */
642 struct list_head oe_pages;
643 /** Since an extent has to be written out in atomic, this is used to
644 * remember the next page need to be locked to write this extent out.
645 * Not used right now.
647 struct osc_page *oe_next_page;
648 /** start and end index of this extent, include start and end
649 * themselves. Page offset here is the page index of osc_pages.
650 * oe_start is used as keyword for red-black tree. */
653 /** maximum ending index of this extent, this is limited by
654 * max_pages_per_rpc, lock extent and chunk size. */
656 /** waitqueue - for those who want to be notified if this extent's
657 * state has changed. */
658 wait_queue_head_t oe_waitq;
659 /** lock covering this extent */
660 struct ldlm_lock *oe_dlmlock;
661 /** terminator of this extent. Must be true if this extent is in IO. */
662 struct task_struct *oe_owner;
663 /** return value of writeback. If somebody is waiting for this extent,
664 * this value can be known by outside world. */
666 /** max pages per rpc when this extent was created */
667 unsigned int oe_mppr;
670 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
672 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
674 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
675 pgoff_t start, pgoff_t end, enum cl_lock_mode mode);
677 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
678 struct osc_page *, void *);
679 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
680 struct osc_object *osc, pgoff_t start, pgoff_t end,
681 osc_page_gang_cbt cb, void *cbdata);
684 #endif /* OSC_CL_INTERNAL_H */