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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Internal interfaces of OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #ifndef OSC_CL_INTERNAL_H
43 #define OSC_CL_INTERNAL_H
45 #include "../../include/linux/libcfs/libcfs.h"
47 #include "../include/obd.h"
48 /* osc_build_res_name() */
49 #include "../include/cl_object.h"
50 #include "osc_internal.h"
59 * State maintained by osc layer for each IO context.
63 struct cl_io_slice oi_cl;
64 /** true if this io is lockless. */
65 unsigned int oi_lockless;
66 /** how many LRU pages are reserved for this IO */
67 unsigned long oi_lru_reserved;
69 /** active extents, we know how many bytes is going to be written,
70 * so having an active extent will prevent it from being fragmented
72 struct osc_extent *oi_active;
73 /** partially truncated extent, we need to hold this extent to prevent
74 * page writeback from happening.
76 struct osc_extent *oi_trunc;
78 /** write osc_lock for this IO, used by osc_extent_find(). */
79 struct osc_lock *oi_write_osclock;
80 struct obd_info oi_info;
82 struct osc_async_cbargs {
85 struct completion opc_sync;
90 * State of transfer for osc.
93 struct cl_req_slice or_cl;
97 * State maintained by osc layer for the duration of a system call.
103 #define OTI_PVEC_SIZE 256
104 struct osc_thread_info {
105 struct ldlm_res_id oti_resname;
106 ldlm_policy_data_t oti_policy;
107 struct cl_lock_descr oti_descr;
108 struct cl_attr oti_attr;
109 struct lustre_handle oti_handle;
110 struct cl_page_list oti_plist;
112 void *oti_pvec[OTI_PVEC_SIZE];
114 * Fields used by cl_lock_discard_pages().
116 pgoff_t oti_next_index;
117 pgoff_t oti_fn_index; /* first non-overlapped index */
118 struct cl_sync_io oti_anchor;
122 struct cl_object oo_cl;
123 struct lov_oinfo *oo_oinfo;
125 * True if locking against this stripe got -EUSERS.
128 unsigned long oo_contention_time;
130 * List of pages in transfer.
132 struct list_head oo_inflight[CRT_NR];
134 * Lock, protecting osc_page::ops_inflight, because a seat-belt is
135 * locked during take-off and landing.
137 spinlock_t oo_seatbelt;
140 * used by the osc to keep track of what objects to build into rpcs.
141 * Protected by client_obd->cli_loi_list_lock.
143 struct list_head oo_ready_item;
144 struct list_head oo_hp_ready_item;
145 struct list_head oo_write_item;
146 struct list_head oo_read_item;
149 * extent is a red black tree to manage (async) dirty pages.
151 struct rb_root oo_root;
153 * Manage write(dirty) extents.
155 struct list_head oo_hp_exts; /* list of hp extents */
156 struct list_head oo_urgent_exts; /* list of writeback extents */
157 struct list_head oo_rpc_exts;
159 struct list_head oo_reading_exts;
161 atomic_t oo_nr_reads;
162 atomic_t oo_nr_writes;
164 /** Protect extent tree. Will be used to protect
165 * oo_{read|write}_pages soon.
170 * Radix tree for caching pages
172 struct radix_tree_root oo_tree;
173 spinlock_t oo_tree_lock;
174 unsigned long oo_npages;
176 /* Protect osc_lock this osc_object has */
177 spinlock_t oo_ol_spin;
178 struct list_head oo_ol_list;
181 static inline void osc_object_lock(struct osc_object *obj)
183 spin_lock(&obj->oo_lock);
186 static inline int osc_object_trylock(struct osc_object *obj)
188 return spin_trylock(&obj->oo_lock);
191 static inline void osc_object_unlock(struct osc_object *obj)
193 spin_unlock(&obj->oo_lock);
196 static inline int osc_object_is_locked(struct osc_object *obj)
198 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
199 return spin_is_locked(&obj->oo_lock);
202 * It is not perfect to return true all the time.
203 * But since this function is only used for assertion
204 * and checking, it seems OK.
211 * Lock "micro-states" for osc layer.
213 enum osc_lock_state {
222 * osc-private state of cl_lock.
224 * Interaction with DLM.
226 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
227 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_dlmlock.
229 * This pointer is protected through a reference, acquired by
230 * osc_lock_upcall0(). Also, an additional reference is acquired by
231 * ldlm_lock_addref() call protecting the lock from cancellation, until
232 * osc_lock_unuse() releases it.
234 * Below is a description of how lock references are acquired and released
237 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
238 * - ldlm_lock_create()
239 * - ldlm_lock_new(): initializes a lock with 2 references. One for
240 * the caller (released when reply from the server is received, or on
241 * error), and another for the hash table.
242 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
244 * - When reply is received from the server (osc_enqueue_interpret())
245 * - ldlm_cli_enqueue_fini()
246 * - LDLM_LOCK_PUT(): releases caller reference acquired by
249 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
250 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
252 * - When lock is being cancelled (ldlm_lock_cancel())
253 * - ldlm_lock_destroy()
254 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
257 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
258 * either when lock is cancelled (osc_lock_blocking()), or when locks is
259 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
260 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
264 struct cl_lock_slice ols_cl;
265 /** Internal lock to protect states, etc. */
267 /** Owner sleeps on this channel for state change */
268 struct cl_sync_io *ols_owner;
269 /** waiting list for this lock to be cancelled */
270 struct list_head ols_waiting_list;
271 /** wait entry of ols_waiting_list */
272 struct list_head ols_wait_entry;
273 /** list entry for osc_object::oo_ol_list */
274 struct list_head ols_nextlock_oscobj;
276 /** underlying DLM lock */
277 struct ldlm_lock *ols_dlmlock;
278 /** DLM flags with which osc_lock::ols_lock was enqueued */
280 /** osc_lock::ols_lock handle */
281 struct lustre_handle ols_handle;
282 struct ldlm_enqueue_info ols_einfo;
283 enum osc_lock_state ols_state;
284 /** lock value block */
285 struct ost_lvb ols_lvb;
288 * true, if ldlm_lock_addref() was called against
289 * osc_lock::ols_lock. This is used for sanity checking.
291 * \see osc_lock::ols_has_ref
293 unsigned ols_hold :1,
295 * this is much like osc_lock::ols_hold, except that this bit is
296 * cleared _after_ reference in released in osc_lock_unuse(). This
297 * fine distinction is needed because:
299 * - if ldlm lock still has a reference, osc_ast_data_get() needs
300 * to return associated cl_lock (so that a flag is needed that is
301 * cleared after ldlm_lock_decref() returned), and
303 * - ldlm_lock_decref() can invoke blocking ast (for a
304 * LDLM_FL_CBPENDING lock), and osc_lock functions like
305 * osc_lock_cancel() called from there need to know whether to
306 * release lock reference (so that a flag is needed that is
307 * cleared before ldlm_lock_decref() is called).
311 * inherit the lockless attribute from top level cl_io.
312 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
316 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
317 * the EVAVAIL error as tolerable, 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.
331 * Page state private for osc layer.
334 struct cl_page_slice ops_cl;
336 * Page queues used by osc to detect when RPC can be formed.
338 struct osc_async_page ops_oap;
340 * An offset within page from which next transfer starts. This is used
341 * by cl_page_clip() to submit partial page transfers.
345 * An offset within page at which next transfer ends.
347 * \see osc_page::ops_from.
351 * Boolean, true iff page is under transfer. Used for sanity checking.
353 unsigned ops_transfer_pinned:1,
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 * Linkage into a per-osc_object list of pages in flight. For
370 struct list_head ops_inflight;
372 * Thread that submitted this page for transfer. For debugging.
374 struct task_struct *ops_submitter;
376 * Submit time - the time when the page is starting RPC. For debugging.
378 unsigned long ops_submit_time;
381 extern struct kmem_cache *osc_lock_kmem;
382 extern struct kmem_cache *osc_object_kmem;
383 extern struct kmem_cache *osc_thread_kmem;
384 extern struct kmem_cache *osc_session_kmem;
385 extern struct kmem_cache *osc_req_kmem;
386 extern struct kmem_cache *osc_extent_kmem;
388 extern struct lu_device_type osc_device_type;
389 extern struct lu_context_key osc_key;
390 extern struct lu_context_key osc_session_key;
392 #define OSC_FLAGS (ASYNC_URGENT | ASYNC_READY)
394 int osc_lock_init(const struct lu_env *env,
395 struct cl_object *obj, struct cl_lock *lock,
396 const struct cl_io *io);
397 int osc_io_init(const struct lu_env *env,
398 struct cl_object *obj, struct cl_io *io);
399 int osc_req_init(const struct lu_env *env, struct cl_device *dev,
401 struct lu_object *osc_object_alloc(const struct lu_env *env,
402 const struct lu_object_header *hdr,
403 struct lu_device *dev);
404 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
405 struct cl_page *page, pgoff_t ind);
407 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
408 pgoff_t start, pgoff_t end);
409 int osc_lvb_print(const struct lu_env *env, void *cookie,
410 lu_printer_t p, const struct ost_lvb *lvb);
412 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
413 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
414 enum cl_req_type crt, int brw_flags);
415 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
416 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
418 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
419 struct page *page, loff_t offset);
420 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
421 struct osc_page *ops);
422 int osc_page_cache_add(const struct lu_env *env,
423 const struct cl_page_slice *slice, struct cl_io *io);
424 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
425 struct osc_page *ops);
426 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
427 struct osc_page *ops);
428 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
429 struct list_head *list, int cmd, int brw_flags);
430 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
431 struct osc_object *obj, __u64 size);
432 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
433 struct osc_object *obj);
434 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
435 pgoff_t start, pgoff_t end, int hp, int discard);
436 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
437 pgoff_t start, pgoff_t end);
438 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
439 struct osc_object *osc);
440 int lru_queue_work(const struct lu_env *env, void *data);
442 void osc_object_set_contended(struct osc_object *obj);
443 void osc_object_clear_contended(struct osc_object *obj);
444 int osc_object_is_contended(struct osc_object *obj);
446 int osc_lock_is_lockless(const struct osc_lock *olck);
448 /*****************************************************************************
454 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
456 struct osc_thread_info *info;
458 info = lu_context_key_get(&env->le_ctx, &osc_key);
463 static inline struct osc_session *osc_env_session(const struct lu_env *env)
465 struct osc_session *ses;
467 ses = lu_context_key_get(env->le_ses, &osc_session_key);
472 static inline struct osc_io *osc_env_io(const struct lu_env *env)
474 return &osc_env_session(env)->os_io;
477 static inline int osc_is_object(const struct lu_object *obj)
479 return obj->lo_dev->ld_type == &osc_device_type;
482 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
484 LINVRNT(d->ld_type == &osc_device_type);
485 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
488 static inline struct obd_export *osc_export(const struct osc_object *obj)
490 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
493 static inline struct client_obd *osc_cli(const struct osc_object *obj)
495 return &osc_export(obj)->exp_obd->u.cli;
498 static inline struct osc_object *cl2osc(const struct cl_object *obj)
500 LINVRNT(osc_is_object(&obj->co_lu));
501 return container_of0(obj, struct osc_object, oo_cl);
504 static inline struct cl_object *osc2cl(const struct osc_object *obj)
506 return (struct cl_object *)&obj->oo_cl;
509 static inline enum ldlm_mode osc_cl_lock2ldlm(enum cl_lock_mode mode)
511 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
512 if (mode == CLM_READ)
514 else if (mode == CLM_WRITE)
520 static inline enum cl_lock_mode osc_ldlm2cl_lock(enum ldlm_mode mode)
522 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
525 else if (mode == LCK_PW)
531 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
533 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
534 return container_of0(slice, struct osc_page, ops_cl);
537 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
539 return container_of0(oap, struct osc_page, ops_oap);
542 static inline pgoff_t osc_index(struct osc_page *opg)
544 return opg->ops_cl.cpl_index;
547 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
549 return oap2osc(oap)->ops_cl.cpl_page;
552 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
554 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
557 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
559 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
560 return container_of0(slice, struct osc_lock, ols_cl);
563 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
565 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
568 static inline int osc_io_srvlock(struct osc_io *oio)
570 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
573 enum osc_extent_state {
574 OES_INV = 0, /** extent is just initialized or destroyed */
575 OES_ACTIVE = 1, /** process is using this extent */
576 OES_CACHE = 2, /** extent is ready for IO */
577 OES_LOCKING = 3, /** locking page to prepare IO */
578 OES_LOCK_DONE = 4, /** locking finished, ready to send */
579 OES_RPC = 5, /** in RPC */
580 OES_TRUNC = 6, /** being truncated */
585 * osc_extent data to manage dirty pages.
586 * osc_extent has the following attributes:
587 * 1. all pages in the same must be in one RPC in write back;
588 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
589 * 3. must be covered by only 1 osc_lock;
590 * 4. exclusive. It's impossible to have overlapped osc_extent.
592 * The lifetime of an extent is from when the 1st page is dirtied to when
593 * all pages inside it are written out.
597 * page lock -> cl_loi_list_lock -> object lock(osc_object::oo_lock)
600 /** red-black tree node */
601 struct rb_node oe_node;
602 /** osc_object of this extent */
603 struct osc_object *oe_obj;
604 /** refcount, removed from red-black tree if reaches zero. */
606 /** busy if non-zero */
608 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
609 struct list_head oe_link;
610 /** state of this extent */
611 enum osc_extent_state oe_state;
612 /** flags for this extent. */
613 unsigned int oe_intree:1,
614 /** 0 is write, 1 is read */
616 /** sync extent, queued by osc_queue_sync_pages() */
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.
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.
631 /** covering lock is being canceled */
633 /** this extent should be written back asap. set if one of pages is
634 * called by page WB daemon, or sync write or reading requests.
637 /** how many grants allocated for this extent.
638 * Grant allocated for this extent. There is no grant allocated
639 * for reading extents and sync write extents.
641 unsigned int oe_grants;
642 /** # of dirty pages in this extent */
643 unsigned int oe_nr_pages;
644 /** list of pending oap pages. Pages in this list are NOT sorted. */
645 struct list_head oe_pages;
646 /** Since an extent has to be written out in atomic, this is used to
647 * remember the next page need to be locked to write this extent out.
648 * Not used right now.
650 struct osc_page *oe_next_page;
651 /** start and end index of this extent, include start and end
652 * themselves. Page offset here is the page index of osc_pages.
653 * oe_start is used as keyword for red-black tree.
657 /** maximum ending index of this extent, this is limited by
658 * max_pages_per_rpc, lock extent and chunk size.
661 /** waitqueue - for those who want to be notified if this extent's
664 wait_queue_head_t oe_waitq;
665 /** lock covering this extent */
666 struct ldlm_lock *oe_dlmlock;
667 /** terminator of this extent. Must be true if this extent is in IO. */
668 struct task_struct *oe_owner;
669 /** return value of writeback. If somebody is waiting for this extent,
670 * this value can be known by outside world.
673 /** max pages per rpc when this extent was created */
674 unsigned int oe_mppr;
677 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
679 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
681 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
682 pgoff_t start, pgoff_t end, enum cl_lock_mode mode);
684 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
685 struct osc_page *, void *);
686 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
687 struct osc_object *osc, pgoff_t start, pgoff_t end,
688 osc_page_gang_cbt cb, void *cbdata);
691 #endif /* OSC_CL_INTERNAL_H */