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, 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 <linux/libcfs/libcfs.h>
52 /* osc_build_res_name() */
54 #include <cl_object.h>
56 #include "osc_internal.h"
65 * State maintained by osc layer for each IO context.
69 struct cl_io_slice oi_cl
;
70 /** true if this io is lockless. */
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
;
79 struct obd_info oi_info
;
81 struct osc_async_cbargs
{
84 struct completion opc_sync
;
89 * State of transfer for osc.
92 struct cl_req_slice or_cl
;
96 * State maintained by osc layer for the duration of a system call.
102 #define OTI_PVEC_SIZE 64
103 struct osc_thread_info
{
104 struct ldlm_res_id oti_resname
;
105 ldlm_policy_data_t oti_policy
;
106 struct cl_lock_descr oti_descr
;
107 struct cl_attr oti_attr
;
108 struct lustre_handle oti_handle
;
109 struct cl_page_list oti_plist
;
111 struct cl_page
*oti_pvec
[OTI_PVEC_SIZE
];
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
;
123 * List of pages in transfer.
125 struct list_head oo_inflight
[CRT_NR
];
127 * Lock, protecting ccc_object::cob_inflight, because a seat-belt is
128 * locked during take-off and landing.
130 spinlock_t oo_seatbelt
;
133 * used by the osc to keep track of what objects to build into rpcs.
134 * Protected by client_obd->cli_loi_list_lock.
136 struct list_head oo_ready_item
;
137 struct list_head oo_hp_ready_item
;
138 struct list_head oo_write_item
;
139 struct list_head oo_read_item
;
142 * extent is a red black tree to manage (async) dirty pages.
144 struct rb_root oo_root
;
146 * Manage write(dirty) extents.
148 struct list_head oo_hp_exts
; /* list of hp extents */
149 struct list_head oo_urgent_exts
; /* list of writeback extents */
150 struct list_head oo_rpc_exts
;
152 struct list_head oo_reading_exts
;
154 atomic_t oo_nr_reads
;
155 atomic_t oo_nr_writes
;
157 /** Protect extent tree. Will be used to protect
158 * oo_{read|write}_pages soon. */
162 static inline void osc_object_lock(struct osc_object
*obj
)
164 spin_lock(&obj
->oo_lock
);
167 static inline int osc_object_trylock(struct osc_object
*obj
)
169 return spin_trylock(&obj
->oo_lock
);
172 static inline void osc_object_unlock(struct osc_object
*obj
)
174 spin_unlock(&obj
->oo_lock
);
177 static inline int osc_object_is_locked(struct osc_object
*obj
)
179 return spin_is_locked(&obj
->oo_lock
);
183 * Lock "micro-states" for osc layer.
185 enum osc_lock_state
{
196 * osc-private state of cl_lock.
198 * Interaction with DLM.
200 * CLIO enqueues all DLM locks through ptlrpcd (that is, in "async" mode).
202 * Once receive upcall is invoked, osc_lock remembers a handle of DLM lock in
203 * osc_lock::ols_handle and a pointer to that lock in osc_lock::ols_lock.
205 * This pointer is protected through a reference, acquired by
206 * osc_lock_upcall0(). Also, an additional reference is acquired by
207 * ldlm_lock_addref() call protecting the lock from cancellation, until
208 * osc_lock_unuse() releases it.
210 * Below is a description of how lock references are acquired and released
213 * - When new lock is created and enqueued to the server (ldlm_cli_enqueue())
214 * - ldlm_lock_create()
215 * - ldlm_lock_new(): initializes a lock with 2 references. One for
216 * the caller (released when reply from the server is received, or on
217 * error), and another for the hash table.
218 * - ldlm_lock_addref_internal(): protects the lock from cancellation.
220 * - When reply is received from the server (osc_enqueue_interpret())
221 * - ldlm_cli_enqueue_fini()
222 * - LDLM_LOCK_PUT(): releases caller reference acquired by
225 * ldlm_lock_decref(): error case: matches ldlm_cli_enqueue().
226 * - ldlm_lock_decref(): for async locks, matches ldlm_cli_enqueue().
228 * - When lock is being cancelled (ldlm_lock_cancel())
229 * - ldlm_lock_destroy()
230 * - LDLM_LOCK_PUT(): releases hash-table reference acquired by
233 * osc_lock is detached from ldlm_lock by osc_lock_detach() that is called
234 * either when lock is cancelled (osc_lock_blocking()), or when locks is
235 * deleted without cancellation (e.g., from cl_locks_prune()). In the latter
236 * case ldlm lock remains in memory, and can be re-attached to osc_lock in the
240 struct cl_lock_slice ols_cl
;
241 /** underlying DLM lock */
242 struct ldlm_lock
*ols_lock
;
243 /** lock value block */
244 struct ost_lvb ols_lvb
;
245 /** DLM flags with which osc_lock::ols_lock was enqueued */
247 /** osc_lock::ols_lock handle */
248 struct lustre_handle ols_handle
;
249 struct ldlm_enqueue_info ols_einfo
;
250 enum osc_lock_state ols_state
;
253 * How many pages are using this lock for io, currently only used by
254 * read-ahead. If non-zero, the underlying dlm lock won't be cancelled
255 * during recovery to avoid deadlock. see bz16774.
257 * \see osc_page::ops_lock
258 * \see osc_page_addref_lock(), osc_page_putref_lock()
260 atomic_t ols_pageref
;
263 * true, if ldlm_lock_addref() was called against
264 * osc_lock::ols_lock. This is used for sanity checking.
266 * \see osc_lock::ols_has_ref
268 unsigned ols_hold
:1,
270 * this is much like osc_lock::ols_hold, except that this bit is
271 * cleared _after_ reference in released in osc_lock_unuse(). This
272 * fine distinction is needed because:
274 * - if ldlm lock still has a reference, osc_ast_data_get() needs
275 * to return associated cl_lock (so that a flag is needed that is
276 * cleared after ldlm_lock_decref() returned), and
278 * - ldlm_lock_decref() can invoke blocking ast (for a
279 * LDLM_FL_CBPENDING lock), and osc_lock functions like
280 * osc_lock_cancel() called from there need to know whether to
281 * release lock reference (so that a flag is needed that is
282 * cleared before ldlm_lock_decref() is called).
286 * inherit the lockless attribute from top level cl_io.
287 * If true, osc_lock_enqueue is able to tolerate the -EUSERS error.
291 * set by osc_lock_use() to wait until blocking AST enters into
292 * osc_ldlm_blocking_ast0(), so that cl_lock mutex can be used for
293 * further synchronization.
297 * If the data of this lock has been flushed to server side.
301 * if set, the osc_lock is a glimpse lock. For glimpse locks, we treat
302 * the EVAVAIL error as tolerable, this will make upper logic happy
303 * to wait all glimpse locks to each OSTs to be completed.
304 * Glimpse lock converts to normal lock if the server lock is
306 * Glimpse lock should be destroyed immediately after use.
310 * For async glimpse lock.
314 * IO that owns this lock. This field is used for a dead-lock
315 * avoidance by osc_lock_enqueue_wait().
317 * XXX: unfortunately, the owner of a osc_lock is not unique,
318 * the lock may have multiple users, if the lock is granted and
321 struct osc_io
*ols_owner
;
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.
364 * lru page list. ops_inflight and ops_lru are exclusive so
365 * that they can share the same data.
367 struct list_head ops_lru
;
369 * Linkage into a per-osc_object list of pages in flight. For
372 struct list_head ops_inflight
;
375 * Thread that submitted this page for transfer. For debugging.
377 struct task_struct
*ops_submitter
;
379 * Submit time - the time when the page is starting RPC. For debugging.
381 cfs_time_t ops_submit_time
;
384 * A lock of which we hold a reference covers this page. Only used by
385 * read-ahead: for a readahead page, we hold it's covering lock to
386 * prevent it from being canceled during recovery.
388 * \see osc_lock::ols_pageref
389 * \see osc_page_addref_lock(), osc_page_putref_lock().
391 struct cl_lock
*ops_lock
;
394 extern struct kmem_cache
*osc_lock_kmem
;
395 extern struct kmem_cache
*osc_object_kmem
;
396 extern struct kmem_cache
*osc_thread_kmem
;
397 extern struct kmem_cache
*osc_session_kmem
;
398 extern struct kmem_cache
*osc_req_kmem
;
399 extern struct kmem_cache
*osc_extent_kmem
;
401 extern struct lu_device_type osc_device_type
;
402 extern struct lu_context_key osc_key
;
403 extern struct lu_context_key osc_session_key
;
405 #define OSC_FLAGS (ASYNC_URGENT|ASYNC_READY)
407 int osc_lock_init(const struct lu_env
*env
,
408 struct cl_object
*obj
, struct cl_lock
*lock
,
409 const struct cl_io
*io
);
410 int osc_io_init (const struct lu_env
*env
,
411 struct cl_object
*obj
, struct cl_io
*io
);
412 int osc_req_init (const struct lu_env
*env
, struct cl_device
*dev
,
414 struct lu_object
*osc_object_alloc(const struct lu_env
*env
,
415 const struct lu_object_header
*hdr
,
416 struct lu_device
*dev
);
417 int osc_page_init(const struct lu_env
*env
, struct cl_object
*obj
,
418 struct cl_page
*page
, struct page
*vmpage
);
420 void osc_index2policy (ldlm_policy_data_t
*policy
, const struct cl_object
*obj
,
421 pgoff_t start
, pgoff_t end
);
422 int osc_lvb_print (const struct lu_env
*env
, void *cookie
,
423 lu_printer_t p
, const struct ost_lvb
*lvb
);
425 void osc_page_submit(const struct lu_env
*env
, struct osc_page
*opg
,
426 enum cl_req_type crt
, int brw_flags
);
427 int osc_cancel_async_page(const struct lu_env
*env
, struct osc_page
*ops
);
428 int osc_set_async_flags(struct osc_object
*obj
, struct osc_page
*opg
,
429 obd_flag async_flags
);
430 int osc_prep_async_page(struct osc_object
*osc
, struct osc_page
*ops
,
431 struct page
*page
, loff_t offset
);
432 int osc_queue_async_io(const struct lu_env
*env
, struct cl_io
*io
,
433 struct osc_page
*ops
);
434 int osc_teardown_async_page(const struct lu_env
*env
, struct osc_object
*obj
,
435 struct osc_page
*ops
);
436 int osc_flush_async_page(const struct lu_env
*env
, struct cl_io
*io
,
437 struct osc_page
*ops
);
438 int osc_queue_sync_pages(const struct lu_env
*env
, struct osc_object
*obj
,
439 struct list_head
*list
, int cmd
, int brw_flags
);
440 int osc_cache_truncate_start(const struct lu_env
*env
, struct osc_io
*oio
,
441 struct osc_object
*obj
, __u64 size
);
442 void osc_cache_truncate_end(const struct lu_env
*env
, struct osc_io
*oio
,
443 struct osc_object
*obj
);
444 int osc_cache_writeback_range(const struct lu_env
*env
, struct osc_object
*obj
,
445 pgoff_t start
, pgoff_t end
, int hp
, int discard
);
446 int osc_cache_wait_range(const struct lu_env
*env
, struct osc_object
*obj
,
447 pgoff_t start
, pgoff_t end
);
448 void osc_io_unplug(const struct lu_env
*env
, struct client_obd
*cli
,
449 struct osc_object
*osc
, pdl_policy_t pol
);
451 void osc_object_set_contended (struct osc_object
*obj
);
452 void osc_object_clear_contended(struct osc_object
*obj
);
453 int osc_object_is_contended (struct osc_object
*obj
);
455 int osc_lock_is_lockless (const struct osc_lock
*olck
);
457 /*****************************************************************************
463 static inline struct osc_thread_info
*osc_env_info(const struct lu_env
*env
)
465 struct osc_thread_info
*info
;
467 info
= lu_context_key_get(&env
->le_ctx
, &osc_key
);
468 LASSERT(info
!= NULL
);
472 static inline struct osc_session
*osc_env_session(const struct lu_env
*env
)
474 struct osc_session
*ses
;
476 ses
= lu_context_key_get(env
->le_ses
, &osc_session_key
);
477 LASSERT(ses
!= NULL
);
481 static inline struct osc_io
*osc_env_io(const struct lu_env
*env
)
483 return &osc_env_session(env
)->os_io
;
486 static inline int osc_is_object(const struct lu_object
*obj
)
488 return obj
->lo_dev
->ld_type
== &osc_device_type
;
491 static inline struct osc_device
*lu2osc_dev(const struct lu_device
*d
)
493 LINVRNT(d
->ld_type
== &osc_device_type
);
494 return container_of0(d
, struct osc_device
, od_cl
.cd_lu_dev
);
497 static inline struct obd_export
*osc_export(const struct osc_object
*obj
)
499 return lu2osc_dev(obj
->oo_cl
.co_lu
.lo_dev
)->od_exp
;
502 static inline struct client_obd
*osc_cli(const struct osc_object
*obj
)
504 return &osc_export(obj
)->exp_obd
->u
.cli
;
507 static inline struct osc_object
*cl2osc(const struct cl_object
*obj
)
509 LINVRNT(osc_is_object(&obj
->co_lu
));
510 return container_of0(obj
, struct osc_object
, oo_cl
);
513 static inline struct cl_object
*osc2cl(const struct osc_object
*obj
)
515 return (struct cl_object
*)&obj
->oo_cl
;
518 static inline ldlm_mode_t
osc_cl_lock2ldlm(enum cl_lock_mode mode
)
520 LASSERT(mode
== CLM_READ
|| mode
== CLM_WRITE
|| mode
== CLM_GROUP
);
521 if (mode
== CLM_READ
)
523 else if (mode
== CLM_WRITE
)
529 static inline enum cl_lock_mode
osc_ldlm2cl_lock(ldlm_mode_t mode
)
531 LASSERT(mode
== LCK_PR
|| mode
== LCK_PW
|| mode
== LCK_GROUP
);
534 else if (mode
== LCK_PW
)
540 static inline struct osc_page
*cl2osc_page(const struct cl_page_slice
*slice
)
542 LINVRNT(osc_is_object(&slice
->cpl_obj
->co_lu
));
543 return container_of0(slice
, struct osc_page
, ops_cl
);
546 static inline struct osc_page
*oap2osc(struct osc_async_page
*oap
)
548 return container_of0(oap
, struct osc_page
, ops_oap
);
551 static inline struct cl_page
*oap2cl_page(struct osc_async_page
*oap
)
553 return oap2osc(oap
)->ops_cl
.cpl_page
;
556 static inline struct osc_page
*oap2osc_page(struct osc_async_page
*oap
)
558 return (struct osc_page
*)container_of(oap
, struct osc_page
, ops_oap
);
561 static inline struct osc_lock
*cl2osc_lock(const struct cl_lock_slice
*slice
)
563 LINVRNT(osc_is_object(&slice
->cls_obj
->co_lu
));
564 return container_of0(slice
, struct osc_lock
, ols_cl
);
567 static inline struct osc_lock
*osc_lock_at(const struct cl_lock
*lock
)
569 return cl2osc_lock(cl_lock_at(lock
, &osc_device_type
));
572 static inline int osc_io_srvlock(struct osc_io
*oio
)
574 return (oio
->oi_lockless
&& !oio
->oi_cl
.cis_io
->ci_no_srvlock
);
577 enum osc_extent_state
{
578 OES_INV
= 0, /** extent is just initialized or destroyed */
579 OES_ACTIVE
= 1, /** process is using this extent */
580 OES_CACHE
= 2, /** extent is ready for IO */
581 OES_LOCKING
= 3, /** locking page to prepare IO */
582 OES_LOCK_DONE
= 4, /** locking finished, ready to send */
583 OES_RPC
= 5, /** in RPC */
584 OES_TRUNC
= 6, /** being truncated */
589 * osc_extent data to manage dirty pages.
590 * osc_extent has the following attributes:
591 * 1. all pages in the same must be in one RPC in write back;
592 * 2. # of pages must be less than max_pages_per_rpc - implied by 1;
593 * 3. must be covered by only 1 osc_lock;
594 * 4. exclusive. It's impossible to have overlapped osc_extent.
596 * The lifetime of an extent is from when the 1st page is dirtied to when
597 * all pages inside it are written out.
601 * page lock -> client_obd_list_lock -> object lock(osc_object::oo_lock)
604 /** red-black tree node */
605 struct rb_node oe_node
;
606 /** osc_object of this extent */
607 struct osc_object
*oe_obj
;
608 /** refcount, removed from red-black tree if reaches zero. */
610 /** busy if non-zero */
612 /** link list of osc_object's oo_{hp|urgent|locking}_exts. */
613 struct list_head oe_link
;
614 /** state of this extent */
615 unsigned int oe_state
;
616 /** flags for this extent. */
617 unsigned int oe_intree
:1,
618 /** 0 is write, 1 is read */
622 /** an ACTIVE extent is going to be truncated, so when this extent
623 * is released, it will turn into TRUNC state instead of CACHE. */
625 /** this extent should be written asap and someone may wait for the
626 * write to finish. This bit is usually set along with urgent if
627 * the extent was CACHE state.
628 * fsync_wait extent can't be merged because new extent region may
629 * 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. */
636 /** how many grants allocated for this extent.
637 * Grant allocated for this extent. There is no grant allocated
638 * for reading extents and sync write extents. */
639 unsigned int oe_grants
;
640 /** # of dirty pages in this extent */
641 unsigned int oe_nr_pages
;
642 /** list of pending oap pages. Pages in this list are NOT sorted. */
643 struct list_head oe_pages
;
644 /** Since an extent has to be written out in atomic, this is used to
645 * remember the next page need to be locked to write this extent out.
646 * Not used right now.
648 struct osc_page
*oe_next_page
;
649 /** start and end index of this extent, include start and end
650 * themselves. Page offset here is the page index of osc_pages.
651 * oe_start is used as keyword for red-black tree. */
654 /** maximum ending index of this extent, this is limited by
655 * max_pages_per_rpc, lock extent and chunk size. */
657 /** waitqueue - for those who want to be notified if this extent's
658 * state has changed. */
659 wait_queue_head_t oe_waitq
;
660 /** lock covering this extent */
661 struct cl_lock
*oe_osclock
;
662 /** terminator of this extent. Must be true if this extent is in IO. */
663 struct task_struct
*oe_owner
;
664 /** return value of writeback. If somebody is waiting for this extent,
665 * this value can be known by outside world. */
667 /** max pages per rpc when this extent was created */
668 unsigned int oe_mppr
;
671 int osc_extent_finish(const struct lu_env
*env
, struct osc_extent
*ext
,
673 int osc_extent_release(const struct lu_env
*env
, struct osc_extent
*ext
);
677 #endif /* OSC_CL_INTERNAL_H */