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 <linux/libcfs/libcfs.h>
48 /* osc_build_res_name() */
49 #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:1,
66 /** true if this io is counted as active IO */
68 /** how many LRU pages are reserved for this IO */
69 unsigned long oi_lru_reserved;
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
74 struct osc_extent *oi_active;
75 /** partially truncated extent, we need to hold this extent to prevent
76 * page writeback from happening.
78 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;
117 struct cl_object oo_cl;
118 struct lov_oinfo *oo_oinfo;
120 * True if locking against this stripe got -EUSERS.
123 unsigned long oo_contention_time;
125 * used by the osc to keep track of what objects to build into rpcs.
126 * Protected by client_obd->cli_loi_list_lock.
128 struct list_head oo_ready_item;
129 struct list_head oo_hp_ready_item;
130 struct list_head oo_write_item;
131 struct list_head oo_read_item;
134 * extent is a red black tree to manage (async) dirty pages.
136 struct rb_root oo_root;
138 * Manage write(dirty) extents.
140 struct list_head oo_hp_exts; /* list of hp extents */
141 struct list_head oo_urgent_exts; /* list of writeback extents */
142 struct list_head oo_rpc_exts;
144 struct list_head oo_reading_exts;
146 atomic_t oo_nr_reads;
147 atomic_t oo_nr_writes;
149 /** Protect extent tree. Will be used to protect
150 * oo_{read|write}_pages soon.
155 * Radix tree for caching pages
157 struct radix_tree_root oo_tree;
158 spinlock_t oo_tree_lock;
159 unsigned long oo_npages;
161 /* Protect osc_lock this osc_object has */
162 spinlock_t oo_ol_spin;
163 struct list_head oo_ol_list;
165 /** number of active IOs of this object */
167 wait_queue_head_t oo_io_waitq;
170 static inline void osc_object_lock(struct osc_object *obj)
172 spin_lock(&obj->oo_lock);
175 static inline int osc_object_trylock(struct osc_object *obj)
177 return spin_trylock(&obj->oo_lock);
180 static inline void osc_object_unlock(struct osc_object *obj)
182 spin_unlock(&obj->oo_lock);
185 static inline int osc_object_is_locked(struct osc_object *obj)
187 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
188 return spin_is_locked(&obj->oo_lock);
191 * It is not perfect to return true all the time.
192 * But since this function is only used for assertion
193 * and checking, it seems OK.
200 * Lock "micro-states" for osc layer.
202 enum osc_lock_state {
211 * osc-private state of cl_lock.
213 * Interaction with DLM.
215 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
216 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_dlmlock.
218 * This pointer is protected through a reference, acquired by
219 * osc_lock_upcall0(). Also, an additional reference is acquired by
220 * ldlm_lock_addref() call protecting the lock from cancellation, until
221 * osc_lock_unuse() releases it.
223 * Below is a description of how lock references are acquired and released
226 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
227 * - ldlm_lock_create()
228 * - ldlm_lock_new(): initializes a lock with 2 references. One for
229 * the caller (released when reply from the server is received, or on
230 * error), and another for the hash table.
231 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
233 * - When reply is received from the server (osc_enqueue_interpret())
234 * - ldlm_cli_enqueue_fini()
235 * - LDLM_LOCK_PUT(): releases caller reference acquired by
238 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
239 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
241 * - When lock is being cancelled (ldlm_lock_cancel())
242 * - ldlm_lock_destroy()
243 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
246 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
247 * either when lock is cancelled (osc_lock_blocking()), or when locks is
248 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
249 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
253 struct cl_lock_slice ols_cl;
254 /** Internal lock to protect states, etc. */
256 /** Owner sleeps on this channel for state change */
257 struct cl_sync_io *ols_owner;
258 /** waiting list for this lock to be cancelled */
259 struct list_head ols_waiting_list;
260 /** wait entry of ols_waiting_list */
261 struct list_head ols_wait_entry;
262 /** list entry for osc_object::oo_ol_list */
263 struct list_head ols_nextlock_oscobj;
265 /** underlying DLM lock */
266 struct ldlm_lock *ols_dlmlock;
267 /** DLM flags with which osc_lock::ols_lock was enqueued */
269 /** osc_lock::ols_lock handle */
270 struct lustre_handle ols_handle;
271 struct ldlm_enqueue_info ols_einfo;
272 enum osc_lock_state ols_state;
273 /** lock value block */
274 struct ost_lvb ols_lvb;
277 * true, if ldlm_lock_addref() was called against
278 * osc_lock::ols_lock. This is used for sanity checking.
280 * \see osc_lock::ols_has_ref
282 unsigned ols_hold :1,
284 * this is much like osc_lock::ols_hold, except that this bit is
285 * cleared _after_ reference in released in osc_lock_unuse(). This
286 * fine distinction is needed because:
288 * - if ldlm lock still has a reference, osc_ast_data_get() needs
289 * to return associated cl_lock (so that a flag is needed that is
290 * cleared after ldlm_lock_decref() returned), and
292 * - ldlm_lock_decref() can invoke blocking ast (for a
293 * LDLM_FL_CBPENDING lock), and osc_lock functions like
294 * osc_lock_cancel() called from there need to know whether to
295 * release lock reference (so that a flag is needed that is
296 * cleared before ldlm_lock_decref() is called).
300 * inherit the lockless attribute from top level cl_io.
301 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
305 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
306 * the EVAVAIL error as tolerable, this will make upper logic happy
307 * to wait all glimpse locks to each OSTs to be completed.
308 * Glimpse lock converts to normal lock if the server lock is
310 * Glimpse lock should be destroyed immediately after use.
314 * For async glimpse lock.
320 * Page state private for osc layer.
323 struct cl_page_slice ops_cl;
325 * Page queues used by osc to detect when RPC can be formed.
327 struct osc_async_page ops_oap;
329 * An offset within page from which next transfer starts. This is used
330 * by cl_page_clip() to submit partial page transfers.
334 * An offset within page at which next transfer ends.
336 * \see osc_page::ops_from.
340 * Boolean, true iff page is under transfer. Used for sanity checking.
342 unsigned ops_transfer_pinned:1,
348 * Set if the page must be transferred with OBD_BRW_SRVLOCK.
352 * lru page list. See osc_lru_{del|use}() in osc_page.c for usage.
354 struct list_head ops_lru;
356 * Submit time - the time when the page is starting RPC. For debugging.
358 unsigned long ops_submit_time;
361 extern struct kmem_cache *osc_lock_kmem;
362 extern struct kmem_cache *osc_object_kmem;
363 extern struct kmem_cache *osc_thread_kmem;
364 extern struct kmem_cache *osc_session_kmem;
365 extern struct kmem_cache *osc_extent_kmem;
367 extern struct lu_device_type osc_device_type;
368 extern struct lu_context_key osc_key;
369 extern struct lu_context_key osc_session_key;
371 #define OSC_FLAGS (ASYNC_URGENT | ASYNC_READY)
373 int osc_lock_init(const struct lu_env *env,
374 struct cl_object *obj, struct cl_lock *lock,
375 const struct cl_io *io);
376 int osc_io_init(const struct lu_env *env,
377 struct cl_object *obj, struct cl_io *io);
378 struct lu_object *osc_object_alloc(const struct lu_env *env,
379 const struct lu_object_header *hdr,
380 struct lu_device *dev);
381 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
382 struct cl_page *page, pgoff_t ind);
384 void osc_index2policy(union ldlm_policy_data *policy,
385 const struct cl_object *obj,
386 pgoff_t start, pgoff_t end);
387 int osc_lvb_print(const struct lu_env *env, void *cookie,
388 lu_printer_t p, const struct ost_lvb *lvb);
390 void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
391 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
392 enum cl_req_type crt, int brw_flags);
393 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops);
394 int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
396 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
397 struct page *page, loff_t offset);
398 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
399 struct osc_page *ops);
400 int osc_page_cache_add(const struct lu_env *env,
401 const struct cl_page_slice *slice, struct cl_io *io);
402 int osc_teardown_async_page(const struct lu_env *env, struct osc_object *obj,
403 struct osc_page *ops);
404 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
405 struct osc_page *ops);
406 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
407 struct list_head *list, int cmd, int brw_flags);
408 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
409 u64 size, struct osc_extent **extp);
410 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext);
411 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
412 pgoff_t start, pgoff_t end, int hp, int discard);
413 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
414 pgoff_t start, pgoff_t end);
415 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
416 struct osc_object *osc);
417 int lru_queue_work(const struct lu_env *env, void *data);
419 void osc_object_set_contended(struct osc_object *obj);
420 void osc_object_clear_contended(struct osc_object *obj);
421 int osc_object_is_contended(struct osc_object *obj);
423 int osc_lock_is_lockless(const struct osc_lock *olck);
425 /*****************************************************************************
431 static inline struct osc_thread_info *osc_env_info(const struct lu_env *env)
433 struct osc_thread_info *info;
435 info = lu_context_key_get(&env->le_ctx, &osc_key);
440 static inline struct osc_session *osc_env_session(const struct lu_env *env)
442 struct osc_session *ses;
444 ses = lu_context_key_get(env->le_ses, &osc_session_key);
449 static inline struct osc_io *osc_env_io(const struct lu_env *env)
451 return &osc_env_session(env)->os_io;
454 static inline int osc_is_object(const struct lu_object *obj)
456 return obj->lo_dev->ld_type == &osc_device_type;
459 static inline struct osc_device *lu2osc_dev(const struct lu_device *d)
461 LINVRNT(d->ld_type == &osc_device_type);
462 return container_of0(d, struct osc_device, od_cl.cd_lu_dev);
465 static inline struct obd_export *osc_export(const struct osc_object *obj)
467 return lu2osc_dev(obj->oo_cl.co_lu.lo_dev)->od_exp;
470 static inline struct client_obd *osc_cli(const struct osc_object *obj)
472 return &osc_export(obj)->exp_obd->u.cli;
475 static inline struct osc_object *cl2osc(const struct cl_object *obj)
477 LINVRNT(osc_is_object(&obj->co_lu));
478 return container_of0(obj, struct osc_object, oo_cl);
481 static inline struct cl_object *osc2cl(const struct osc_object *obj)
483 return (struct cl_object *)&obj->oo_cl;
486 static inline enum ldlm_mode osc_cl_lock2ldlm(enum cl_lock_mode mode)
488 LASSERT(mode == CLM_READ || mode == CLM_WRITE || mode == CLM_GROUP);
489 if (mode == CLM_READ)
491 else if (mode == CLM_WRITE)
497 static inline enum cl_lock_mode osc_ldlm2cl_lock(enum ldlm_mode mode)
499 LASSERT(mode == LCK_PR || mode == LCK_PW || mode == LCK_GROUP);
502 else if (mode == LCK_PW)
508 static inline struct osc_page *cl2osc_page(const struct cl_page_slice *slice)
510 LINVRNT(osc_is_object(&slice->cpl_obj->co_lu));
511 return container_of0(slice, struct osc_page, ops_cl);
514 static inline struct osc_page *oap2osc(struct osc_async_page *oap)
516 return container_of0(oap, struct osc_page, ops_oap);
519 static inline pgoff_t osc_index(struct osc_page *opg)
521 return opg->ops_cl.cpl_index;
524 static inline struct cl_page *oap2cl_page(struct osc_async_page *oap)
526 return oap2osc(oap)->ops_cl.cpl_page;
529 static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
531 return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
534 static inline struct osc_page *
535 osc_cl_page_osc(struct cl_page *page, struct osc_object *osc)
537 const struct cl_page_slice *slice;
540 slice = cl_object_page_slice(&osc->oo_cl, page);
541 return cl2osc_page(slice);
544 static inline struct osc_lock *cl2osc_lock(const struct cl_lock_slice *slice)
546 LINVRNT(osc_is_object(&slice->cls_obj->co_lu));
547 return container_of0(slice, struct osc_lock, ols_cl);
550 static inline struct osc_lock *osc_lock_at(const struct cl_lock *lock)
552 return cl2osc_lock(cl_lock_at(lock, &osc_device_type));
555 static inline int osc_io_srvlock(struct osc_io *oio)
557 return (oio->oi_lockless && !oio->oi_cl.cis_io->ci_no_srvlock);
560 enum osc_extent_state {
561 OES_INV = 0, /** extent is just initialized or destroyed */
562 OES_ACTIVE = 1, /** process is using this extent */
563 OES_CACHE = 2, /** extent is ready for IO */
564 OES_LOCKING = 3, /** locking page to prepare IO */
565 OES_LOCK_DONE = 4, /** locking finished, ready to send */
566 OES_RPC = 5, /** in RPC */
567 OES_TRUNC = 6, /** being truncated */
572 * osc_extent data to manage dirty pages.
573 * osc_extent has the following attributes:
574 * 1. all pages in the same must be in one RPC in write back;
575 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
576 * 3. must be covered by only 1 osc_lock;
577 * 4. exclusive. It's impossible to have overlapped osc_extent.
579 * The lifetime of an extent is from when the 1st page is dirtied to when
580 * all pages inside it are written out.
584 * page lock -> cl_loi_list_lock -> object lock(osc_object::oo_lock)
587 /** red-black tree node */
588 struct rb_node oe_node;
589 /** osc_object of this extent */
590 struct osc_object *oe_obj;
591 /** refcount, removed from red-black tree if reaches zero. */
593 /** busy if non-zero */
595 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
596 struct list_head oe_link;
597 /** state of this extent */
598 enum osc_extent_state oe_state;
599 /** flags for this extent. */
600 unsigned int oe_intree:1,
601 /** 0 is write, 1 is read */
603 /** sync extent, queued by osc_queue_sync_pages() */
605 /** set if this extent has partial, sync pages.
606 * Extents with partial page(s) can't merge with others in RPC
611 /** an ACTIVE extent is going to be truncated, so when this extent
612 * is released, it will turn into TRUNC state instead of CACHE.
615 /** this extent should be written asap and someone may wait for the
616 * write to finish. This bit is usually set along with urgent if
617 * the extent was CACHE state.
618 * fsync_wait extent can't be merged because new extent region may
619 * exceed fsync range.
622 /** covering lock is being canceled */
624 /** this extent should be written back asap. set if one of pages is
625 * called by page WB daemon, or sync write or reading requests.
628 /** how many grants allocated for this extent.
629 * Grant allocated for this extent. There is no grant allocated
630 * 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.
648 /** maximum ending index of this extent, this is limited by
649 * max_pages_per_rpc, lock extent and chunk size.
652 /** waitqueue - for those who want to be notified if this extent's
655 wait_queue_head_t oe_waitq;
656 /** lock covering this extent */
657 struct ldlm_lock *oe_dlmlock;
658 /** terminator of this extent. Must be true if this extent is in IO. */
659 struct task_struct *oe_owner;
660 /** return value of writeback. If somebody is waiting for this extent,
661 * this value can be known by outside world.
664 /** max pages per rpc when this extent was created */
665 unsigned int oe_mppr;
668 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
670 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext);
672 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
673 pgoff_t start, pgoff_t end, enum cl_lock_mode mode);
675 typedef int (*osc_page_gang_cbt)(const struct lu_env *, struct cl_io *,
676 struct osc_page *, void *);
677 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
678 struct osc_object *osc, pgoff_t start, pgoff_t end,
679 osc_page_gang_cbt cb, void *cbdata);
682 #endif /* OSC_CL_INTERNAL_H */