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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 /** \defgroup PtlRPC Portal RPC and networking module.
38 * PortalRPC is the layer used by rest of lustre code to achieve network
39 * communications: establish connections with corresponding export and import
40 * states, listen for a service, send and receive RPCs.
41 * PortalRPC also includes base recovery framework: packet resending and
42 * replaying, reconnections, pinger.
44 * PortalRPC utilizes LNet as its transport layer.
58 #if defined(__linux__)
59 #include <linux/lustre_net.h>
60 #elif defined(__APPLE__)
61 #include <darwin/lustre_net.h>
62 #elif defined(__WINNT__)
63 #include <winnt/lustre_net.h>
65 #error Unsupported operating system.
68 #include <libcfs/libcfs.h>
70 #include <lnet/lnet.h>
71 #include <lustre/lustre_idl.h>
72 #include <lustre_ha.h>
73 #include <lustre_sec.h>
74 #include <lustre_import.h>
75 #include <lprocfs_status.h>
76 #include <lu_object.h>
77 #include <lustre_req_layout.h>
79 #include <obd_support.h>
80 #include <lustre_ver.h>
82 /* MD flags we _always_ use */
83 #define PTLRPC_MD_OPTIONS 0
86 * Define maxima for bulk I/O
87 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
88 * these limits are system wide and not interface-local. */
89 #define PTLRPC_MAX_BRW_BITS LNET_MTU_BITS
90 #define PTLRPC_MAX_BRW_SIZE (1<<LNET_MTU_BITS)
91 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
93 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
95 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
96 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
98 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE))
99 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE"
101 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU)
102 # error "PTLRPC_MAX_BRW_SIZE too big"
104 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV)
105 # error "PTLRPC_MAX_BRW_PAGES too big"
107 #endif /* __KERNEL__ */
110 * The following constants determine how memory is used to buffer incoming
113 * ?_NBUFS # buffers to allocate when growing the pool
114 * ?_BUFSIZE # bytes in a single request buffer
115 * ?_MAXREQSIZE # maximum request service will receive
117 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
118 * of ?_NBUFS is added to the pool.
120 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
121 * considered full when less than ?_MAXREQSIZE is left in them.
123 #define LDLM_THREADS_AUTO_MIN (2)
124 #define LDLM_THREADS_AUTO_MAX min_t(unsigned, cfs_num_online_cpus() * \
125 cfs_num_online_cpus() * 32, 128)
126 #define LDLM_BL_THREADS LDLM_THREADS_AUTO_MIN
127 #define LDLM_NBUFS (64 * cfs_num_online_cpus())
128 #define LDLM_BUFSIZE (8 * 1024)
129 #define LDLM_MAXREQSIZE (5 * 1024)
130 #define LDLM_MAXREPSIZE (1024)
132 /** Absolute limits */
133 #define MDT_MIN_THREADS 2UL
134 #ifndef MDT_MAX_THREADS
135 #define MDT_MAX_THREADS 512UL
137 #define MDS_NBUFS (64 * cfs_num_online_cpus())
139 * Assume file name length = FNAME_MAX = 256 (true for ext3).
140 * path name length = PATH_MAX = 4096
141 * LOV MD size max = EA_MAX = 48000 (2000 stripes)
142 * symlink: FNAME_MAX + PATH_MAX <- largest
143 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
144 * rename: FNAME_MAX + FNAME_MAX
145 * open: FNAME_MAX + EA_MAX
147 * MDS_MAXREQSIZE ~= 4736 bytes =
148 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
149 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
150 * or, for mds_close() and mds_reint_unlink() on a many-OST filesystem:
151 * = 9210 bytes = lustre_msg + mdt_body + 160 * (easize + cookiesize)
153 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
154 * except in the open case where there are a large number of OSTs in a LOV.
156 #define MDS_MAXREPSIZE max(10 * 1024, 362 + LOV_MAX_STRIPE_COUNT * 56)
157 #define MDS_MAXREQSIZE MDS_MAXREPSIZE
159 /** MDS_BUFSIZE = max_reqsize + max sptlrpc payload size */
160 #define MDS_BUFSIZE (MDS_MAXREQSIZE + 1024)
162 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
163 #define FLD_MAXREQSIZE (160)
165 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
166 #define FLD_MAXREPSIZE (152)
169 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
171 #define SEQ_MAXREQSIZE (160)
173 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
174 #define SEQ_MAXREPSIZE (152)
176 /** MGS threads must be >= 3, see bug 22458 comment #28 */
177 #define MGS_THREADS_AUTO_MIN 3
178 #define MGS_THREADS_AUTO_MAX 32
179 #define MGS_NBUFS (64 * cfs_num_online_cpus())
180 #define MGS_BUFSIZE (8 * 1024)
181 #define MGS_MAXREQSIZE (7 * 1024)
182 #define MGS_MAXREPSIZE (9 * 1024)
184 /** Absolute OSS limits */
185 #define OSS_THREADS_MIN 3 /* difficult replies, HPQ, others */
186 #define OSS_THREADS_MAX 512
187 #define OST_NBUFS (64 * cfs_num_online_cpus())
188 #define OST_BUFSIZE (8 * 1024)
191 * OST_MAXREQSIZE ~= 4768 bytes =
192 * lustre_msg + obdo + 16 * obd_ioobj + 256 * niobuf_remote
194 * - single object with 16 pages is 512 bytes
195 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
197 #define OST_MAXREQSIZE (5 * 1024)
198 #define OST_MAXREPSIZE (9 * 1024)
200 /* Macro to hide a typecast. */
201 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
204 * Structure to single define portal connection.
206 struct ptlrpc_connection {
207 /** linkage for connections hash table */
208 cfs_hlist_node_t c_hash;
209 /** Our own lnet nid for this connection */
211 /** Remote side nid for this connection */
212 lnet_process_id_t c_peer;
213 /** UUID of the other side */
214 struct obd_uuid c_remote_uuid;
215 /** reference counter for this connection */
216 cfs_atomic_t c_refcount;
219 /** Client definition for PortalRPC */
220 struct ptlrpc_client {
221 /** What lnet portal does this client send messages to by default */
222 __u32 cli_request_portal;
223 /** What portal do we expect replies on */
224 __u32 cli_reply_portal;
225 /** Name of the client */
229 /** state flags of requests */
230 /* XXX only ones left are those used by the bulk descs as well! */
231 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
232 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
234 #define REQ_MAX_ACK_LOCKS 8
236 union ptlrpc_async_args {
238 * Scratchpad for passing args to completion interpreter. Users
239 * cast to the struct of their choosing, and CLASSERT that this is
240 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
241 * a pointer to it here. The pointer_arg ensures this struct is at
242 * least big enough for that.
244 void *pointer_arg[11];
248 struct ptlrpc_request_set;
249 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
252 * Definition of request set structure.
253 * Request set is a list of requests (not necessary to the same target) that
254 * once populated with RPCs could be sent in parallel.
255 * There are two kinds of request sets. General purpose and with dedicated
256 * serving thread. Example of the latter is ptlrpcd set.
257 * For general purpose sets once request set started sending it is impossible
258 * to add new requests to such set.
259 * Provides a way to call "completion callbacks" when all requests in the set
262 struct ptlrpc_request_set {
263 cfs_atomic_t set_refcount;
264 /** number of in queue requests */
265 cfs_atomic_t set_new_count;
266 /** number of uncompleted requests */
267 cfs_atomic_t set_remaining;
268 /** wait queue to wait on for request events */
269 cfs_waitq_t set_waitq;
270 cfs_waitq_t *set_wakeup_ptr;
271 /** List of requests in the set */
272 cfs_list_t set_requests;
274 * List of completion callbacks to be called when the set is completed
275 * This is only used if \a set_interpret is NULL.
276 * Links struct ptlrpc_set_cbdata.
278 cfs_list_t set_cblist;
279 /** Completion callback, if only one. */
280 set_interpreter_func set_interpret;
281 /** opaq argument passed to completion \a set_interpret callback. */
284 * Lock for \a set_new_requests manipulations
285 * locked so that any old caller can communicate requests to
286 * the set holder who can then fold them into the lock-free set
288 cfs_spinlock_t set_new_req_lock;
289 /** List of new yet unsent requests. Only used with ptlrpcd now. */
290 cfs_list_t set_new_requests;
294 * Description of a single ptrlrpc_set callback
296 struct ptlrpc_set_cbdata {
297 /** List linkage item */
299 /** Pointer to interpreting function */
300 set_interpreter_func psc_interpret;
301 /** Opaq argument to pass to the callback */
305 struct ptlrpc_bulk_desc;
308 * ptlrpc callback & work item stuff
310 struct ptlrpc_cb_id {
311 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
312 void *cbid_arg; /* additional arg */
315 /** Maximum number of locks to fit into reply state */
316 #define RS_MAX_LOCKS 8
320 * Structure to define reply state on the server
321 * Reply state holds various reply message information. Also for "difficult"
322 * replies (rep-ack case) we store the state after sending reply and wait
323 * for the client to acknowledge the reception. In these cases locks could be
324 * added to the state for replay/failover consistency guarantees.
326 struct ptlrpc_reply_state {
327 /** Callback description */
328 struct ptlrpc_cb_id rs_cb_id;
329 /** Linkage for list of all reply states in a system */
331 /** Linkage for list of all reply states on same export */
332 cfs_list_t rs_exp_list;
333 /** Linkage for list of all reply states for same obd */
334 cfs_list_t rs_obd_list;
336 cfs_list_t rs_debug_list;
338 /** A spinlock to protect the reply state flags */
339 cfs_spinlock_t rs_lock;
340 /** Reply state flags */
341 unsigned long rs_difficult:1; /* ACK/commit stuff */
342 unsigned long rs_no_ack:1; /* no ACK, even for
343 difficult requests */
344 unsigned long rs_scheduled:1; /* being handled? */
345 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
346 unsigned long rs_handled:1; /* been handled yet? */
347 unsigned long rs_on_net:1; /* reply_out_callback pending? */
348 unsigned long rs_prealloc:1; /* rs from prealloc list */
349 unsigned long rs_committed:1;/* the transaction was committed
350 and the rs was dispatched
351 by ptlrpc_commit_replies */
352 /** Size of the state */
356 /** Transaction number */
360 struct obd_export *rs_export;
361 struct ptlrpc_service *rs_service;
362 /** Lnet metadata handle for the reply */
363 lnet_handle_md_t rs_md_h;
364 cfs_atomic_t rs_refcount;
366 /** Context for the sevice thread */
367 struct ptlrpc_svc_ctx *rs_svc_ctx;
368 /** Reply buffer (actually sent to the client), encoded if needed */
369 struct lustre_msg *rs_repbuf; /* wrapper */
370 /** Size of the reply buffer */
371 int rs_repbuf_len; /* wrapper buf length */
372 /** Size of the reply message */
373 int rs_repdata_len; /* wrapper msg length */
375 * Actual reply message. Its content is encrupted (if needed) to
376 * produce reply buffer for actual sending. In simple case
377 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
379 struct lustre_msg *rs_msg; /* reply message */
381 /** Number of locks awaiting client ACK */
383 /** Handles of locks awaiting client reply ACK */
384 struct lustre_handle rs_locks[RS_MAX_LOCKS];
385 /** Lock modes of locks in \a rs_locks */
386 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
389 struct ptlrpc_thread;
393 RQ_PHASE_NEW = 0xebc0de00,
394 RQ_PHASE_RPC = 0xebc0de01,
395 RQ_PHASE_BULK = 0xebc0de02,
396 RQ_PHASE_INTERPRET = 0xebc0de03,
397 RQ_PHASE_COMPLETE = 0xebc0de04,
398 RQ_PHASE_UNREGISTERING = 0xebc0de05,
399 RQ_PHASE_UNDEFINED = 0xebc0de06
402 /** Type of request interpreter call-back */
403 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
404 struct ptlrpc_request *req,
408 * Definition of request pool structure.
409 * The pool is used to store empty preallocated requests for the case
410 * when we would actually need to send something without performing
411 * any allocations (to avoid e.g. OOM).
413 struct ptlrpc_request_pool {
414 /** Locks the list */
415 cfs_spinlock_t prp_lock;
416 /** list of ptlrpc_request structs */
417 cfs_list_t prp_req_list;
418 /** Maximum message size that would fit into a rquest from this pool */
420 /** Function to allocate more requests for this pool */
421 void (*prp_populate)(struct ptlrpc_request_pool *, int);
430 * Basic request prioritization operations structure.
431 * The whole idea is centered around locks and RPCs that might affect locks.
432 * When a lock is contended we try to give priority to RPCs that might lead
433 * to fastest release of that lock.
434 * Currently only implemented for OSTs only in a way that makes all
435 * IO and truncate RPCs that are coming from a locked region where a lock is
436 * contended a priority over other requests.
438 struct ptlrpc_hpreq_ops {
440 * Check if the lock handle of the given lock is the same as
441 * taken from the request.
443 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
445 * Check if the request is a high priority one.
447 int (*hpreq_check)(struct ptlrpc_request *);
449 * Called after the request has been handled.
451 void (*hpreq_fini)(struct ptlrpc_request *);
455 * Represents remote procedure call.
457 * This is a staple structure used by everybody wanting to send a request
460 struct ptlrpc_request {
461 /* Request type: one of PTL_RPC_MSG_* */
464 * Linkage item through which this request is included into
465 * sending/delayed lists on client and into rqbd list on server
469 * Server side list of incoming unserved requests sorted by arrival
470 * time. Traversed from time to time to notice about to expire
471 * requests and sent back "early replies" to clients to let them
472 * know server is alive and well, just very busy to service their
475 cfs_list_t rq_timed_list;
476 /** server-side history, used for debuging purposes. */
477 cfs_list_t rq_history_list;
478 /** server-side per-export list */
479 cfs_list_t rq_exp_list;
480 /** server-side hp handlers */
481 struct ptlrpc_hpreq_ops *rq_ops;
482 /** history sequence # */
483 __u64 rq_history_seq;
484 /** the index of service's srv_at_array into which request is linked */
486 /** Result of request processing */
488 /** Lock to protect request flags and some other important bits, like
491 cfs_spinlock_t rq_lock;
492 /** client-side flags are serialized by rq_lock */
493 unsigned long rq_intr:1, rq_replied:1, rq_err:1,
494 rq_timedout:1, rq_resend:1, rq_restart:1,
496 * when ->rq_replay is set, request is kept by the client even
497 * after server commits corresponding transaction. This is
498 * used for operations that require sequence of multiple
499 * requests to be replayed. The only example currently is file
500 * open/close. When last request in such a sequence is
501 * committed, ->rq_replay is cleared on all requests in the
505 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
506 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
507 rq_early:1, rq_must_unlink:1,
508 rq_fake:1, /* this fake req */
509 rq_memalloc:1, /* req originated from "kswapd" */
510 /* server-side flags */
511 rq_packed_final:1, /* packed final reply */
512 rq_hp:1, /* high priority RPC */
513 rq_at_linked:1, /* link into service's srv_at_array */
516 /* whether the "rq_set" is a valid one */
520 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
521 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
522 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
523 server-side refcounf for multiple replies */
525 /** initial thread servicing this request */
526 struct ptlrpc_thread *rq_svc_thread;
528 /** Portal to which this request would be sent */
529 int rq_request_portal; /* XXX FIXME bug 249 */
530 /** Portal where to wait for reply and where reply would be sent */
531 int rq_reply_portal; /* XXX FIXME bug 249 */
535 * !rq_truncate : # reply bytes actually received,
536 * rq_truncate : required repbuf_len for resend
539 /** Request length */
541 /** Request message - what client sent */
542 struct lustre_msg *rq_reqmsg;
546 /** Reply message - server response */
547 struct lustre_msg *rq_repmsg;
548 /** Transaction number */
553 * List item to for replay list. Not yet commited requests get linked
555 * Also see \a rq_replay comment above.
557 cfs_list_t rq_replay_list;
560 * security and encryption data
562 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
563 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
564 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
566 struct sptlrpc_flavor rq_flvr; /**< for client & server */
567 enum lustre_sec_part rq_sp_from;
569 unsigned long /* client/server security flags */
570 rq_ctx_init:1, /* context initiation */
571 rq_ctx_fini:1, /* context destroy */
572 rq_bulk_read:1, /* request bulk read */
573 rq_bulk_write:1, /* request bulk write */
574 /* server authentication flags */
575 rq_auth_gss:1, /* authenticated by gss */
576 rq_auth_remote:1, /* authed as remote user */
577 rq_auth_usr_root:1, /* authed as root */
578 rq_auth_usr_mdt:1, /* authed as mdt */
579 rq_auth_usr_ost:1, /* authed as ost */
580 /* security tfm flags */
583 /* doesn't expect reply FIXME */
585 rq_pill_init:1; /* pill initialized */
587 uid_t rq_auth_uid; /* authed uid */
588 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
590 /* (server side), pointed directly into req buffer */
591 struct ptlrpc_user_desc *rq_user_desc;
593 /** early replies go to offset 0, regular replies go after that */
594 unsigned int rq_reply_off;
596 /* various buffer pointers */
597 struct lustre_msg *rq_reqbuf; /* req wrapper */
598 int rq_reqbuf_len; /* req wrapper buf len */
599 int rq_reqdata_len; /* req wrapper msg len */
600 char *rq_repbuf; /* rep buffer */
601 int rq_repbuf_len; /* rep buffer len */
602 struct lustre_msg *rq_repdata; /* rep wrapper msg */
603 int rq_repdata_len; /* rep wrapper msg len */
604 struct lustre_msg *rq_clrbuf; /* only in priv mode */
605 int rq_clrbuf_len; /* only in priv mode */
606 int rq_clrdata_len; /* only in priv mode */
610 /** Fields that help to see if request and reply were swabbed or not */
611 __u32 rq_req_swab_mask;
612 __u32 rq_rep_swab_mask;
614 /** What was import generation when this request was sent */
615 int rq_import_generation;
616 enum lustre_imp_state rq_send_state;
618 /** how many early replies (for stats) */
621 /** client+server request */
622 lnet_handle_md_t rq_req_md_h;
623 struct ptlrpc_cb_id rq_req_cbid;
624 /** optional time limit for send attempts */
625 cfs_duration_t rq_delay_limit;
626 /** time request was first queued */
627 cfs_time_t rq_queued_time;
630 /** request arrival time */
631 struct timeval rq_arrival_time;
632 /** separated reply state */
633 struct ptlrpc_reply_state *rq_reply_state;
634 /** incoming request buffer */
635 struct ptlrpc_request_buffer_desc *rq_rqbd;
637 __u32 rq_uid; /* peer uid, used in MDS only */
640 /** client-only incoming reply */
641 lnet_handle_md_t rq_reply_md_h;
642 cfs_waitq_t rq_reply_waitq;
643 struct ptlrpc_cb_id rq_reply_cbid;
647 /** Peer description (the other side) */
648 lnet_process_id_t rq_peer;
649 /** Server-side, export on which request was received */
650 struct obd_export *rq_export;
651 /** Client side, import where request is being sent */
652 struct obd_import *rq_import;
654 /** Replay callback, called after request is replayed at recovery */
655 void (*rq_replay_cb)(struct ptlrpc_request *);
657 * Commit callback, called when request is committed and about to be
660 void (*rq_commit_cb)(struct ptlrpc_request *);
661 /** Opaq data for replay and commit callbacks. */
664 /** For bulk requests on client only: bulk descriptor */
665 struct ptlrpc_bulk_desc *rq_bulk;
667 /** client outgoing req */
669 * when request/reply sent (secs), or time when request should be sent
672 /** time for request really sent out */
675 /** when request must finish. volatile
676 * so that servers' early reply updates to the deadline aren't
677 * kept in per-cpu cache */
678 volatile time_t rq_deadline;
679 /** when req reply unlink must finish. */
680 time_t rq_reply_deadline;
681 /** when req bulk unlink must finish. */
682 time_t rq_bulk_deadline;
684 * service time estimate (secs)
685 * If the requestsis not served by this time, it is marked as timed out.
689 /** Multi-rpc bits */
690 /** Link item for request set lists */
691 cfs_list_t rq_set_chain;
692 /** Per-request waitq introduced by bug 21938 for recovery waiting */
693 cfs_waitq_t rq_set_waitq;
694 /** Link back to the request set */
695 struct ptlrpc_request_set *rq_set;
696 /** Async completion handler, called when reply is received */
697 ptlrpc_interpterer_t rq_interpret_reply;
698 /** Async completion context */
699 union ptlrpc_async_args rq_async_args;
701 /** Pool if request is from preallocated list */
702 struct ptlrpc_request_pool *rq_pool;
704 struct lu_context rq_session;
705 struct lu_context rq_recov_session;
707 /** request format description */
708 struct req_capsule rq_pill;
712 * Call completion handler for rpc if any, return it's status or original
713 * rc if there was no handler defined for this request.
715 static inline int ptlrpc_req_interpret(const struct lu_env *env,
716 struct ptlrpc_request *req, int rc)
718 if (req->rq_interpret_reply != NULL) {
719 req->rq_status = req->rq_interpret_reply(env, req,
722 return req->rq_status;
728 * Returns 1 if request buffer at offset \a index was already swabbed
730 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
732 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
733 return req->rq_req_swab_mask & (1 << index);
737 * Returns 1 if request reply buffer at offset \a index was already swabbed
739 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
741 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
742 return req->rq_rep_swab_mask & (1 << index);
746 * Returns 1 if request needs to be swabbed into local cpu byteorder
748 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
750 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
754 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
756 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
758 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
762 * Mark request buffer at offset \a index that it was already swabbed
764 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
766 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
767 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
768 req->rq_req_swab_mask |= 1 << index;
772 * Mark request reply buffer at offset \a index that it was already swabbed
774 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
776 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
777 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
778 req->rq_rep_swab_mask |= 1 << index;
782 * Convert numerical request phase value \a phase into text string description
784 static inline const char *
785 ptlrpc_phase2str(enum rq_phase phase)
794 case RQ_PHASE_INTERPRET:
796 case RQ_PHASE_COMPLETE:
798 case RQ_PHASE_UNREGISTERING:
799 return "Unregistering";
806 * Convert numerical request phase of the request \a req into text stringi
809 static inline const char *
810 ptlrpc_rqphase2str(struct ptlrpc_request *req)
812 return ptlrpc_phase2str(req->rq_phase);
816 * Debugging functions and helpers to print request structure into debug log
819 /* Spare the preprocessor, spoil the bugs. */
820 #define FLAG(field, str) (field ? str : "")
822 /** Convert bit flags into a string */
823 #define DEBUG_REQ_FLAGS(req) \
824 ptlrpc_rqphase2str(req), \
825 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
826 FLAG(req->rq_err, "E"), \
827 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
828 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
829 FLAG(req->rq_no_resend, "N"), \
830 FLAG(req->rq_waiting, "W"), \
831 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
832 FLAG(req->rq_committed, "M")
834 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
836 void _debug_req(struct ptlrpc_request *req,
837 struct libcfs_debug_msg_data *data, const char *fmt, ...)
838 __attribute__ ((format (printf, 3, 4)));
841 * Helper that decides if we need to print request accordig to current debug
844 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
846 CFS_CHECK_STACK(msgdata, mask, cdls); \
848 if (((mask) & D_CANTMASK) != 0 || \
849 ((libcfs_debug & (mask)) != 0 && \
850 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
851 _debug_req((req), msgdata, fmt, ##a); \
855 * This is the debug print function you need to use to print request sturucture
856 * content into lustre debug log.
857 * for most callers (level is a constant) this is resolved at compile time */
858 #define DEBUG_REQ(level, req, fmt, args...) \
860 if ((level) & (D_ERROR | D_WARNING)) { \
861 static cfs_debug_limit_state_t cdls; \
862 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
863 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
865 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
866 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
872 * Structure that defines a single page of a bulk transfer
874 struct ptlrpc_bulk_page {
875 /** Linkage to list of pages in a bulk */
878 * Number of bytes in a page to transfer starting from \a bp_pageoffset
881 /** offset within a page */
883 /** The page itself */
884 struct page *bp_page;
887 #define BULK_GET_SOURCE 0
888 #define BULK_PUT_SINK 1
889 #define BULK_GET_SINK 2
890 #define BULK_PUT_SOURCE 3
893 * Definition of buk descriptor.
894 * Bulks are special "Two phase" RPCs where initial request message
895 * is sent first and it is followed bt a transfer (o receiving) of a large
896 * amount of data to be settled into pages referenced from the bulk descriptors.
897 * Bulks transfers (the actual data following the small requests) are done
898 * on separate LNet portals.
899 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
900 * Another user is readpage for MDT.
902 struct ptlrpc_bulk_desc {
903 /** completed successfully */
904 unsigned long bd_success:1;
905 /** accessible to the network (network io potentially in progress) */
906 unsigned long bd_network_rw:1;
907 /** {put,get}{source,sink} */
908 unsigned long bd_type:2;
910 unsigned long bd_registered:1;
911 /** For serialization with callback */
912 cfs_spinlock_t bd_lock;
913 /** Import generation when request for this bulk was sent */
914 int bd_import_generation;
915 /** Server side - export this bulk created for */
916 struct obd_export *bd_export;
917 /** Client side - import this bulk was sent on */
918 struct obd_import *bd_import;
919 /** LNet portal for this bulk */
921 /** Back pointer to the request */
922 struct ptlrpc_request *bd_req;
923 cfs_waitq_t bd_waitq; /* server side only WQ */
924 int bd_iov_count; /* # entries in bd_iov */
925 int bd_max_iov; /* allocated size of bd_iov */
926 int bd_nob; /* # bytes covered */
927 int bd_nob_transferred; /* # bytes GOT/PUT */
931 struct ptlrpc_cb_id bd_cbid; /* network callback info */
932 lnet_handle_md_t bd_md_h; /* associated MD */
933 lnet_nid_t bd_sender; /* stash event::sender */
935 #if defined(__KERNEL__)
937 * encrypt iov, size is either 0 or bd_iov_count.
939 lnet_kiov_t *bd_enc_iov;
941 lnet_kiov_t bd_iov[0];
943 lnet_md_iovec_t bd_iov[0];
948 SVC_STOPPED = 1 << 0,
949 SVC_STOPPING = 1 << 1,
950 SVC_STARTING = 1 << 2,
951 SVC_RUNNING = 1 << 3,
957 * Definition of server service thread structure
959 struct ptlrpc_thread {
961 * List of active threads in svc->srv_threads
965 * thread-private data (preallocated memory)
970 * service thread index, from ptlrpc_start_threads
978 * put watchdog in the structure per thread b=14840
980 struct lc_watchdog *t_watchdog;
982 * the svc this thread belonged to b=18582
984 struct ptlrpc_service *t_svc;
985 cfs_waitq_t t_ctl_waitq;
986 struct lu_env *t_env;
989 static inline int thread_is_init(struct ptlrpc_thread *thread)
991 return thread->t_flags == 0;
994 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
996 return !!(thread->t_flags & SVC_STOPPED);
999 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1001 return !!(thread->t_flags & SVC_STOPPING);
1004 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1006 return !!(thread->t_flags & SVC_STARTING);
1009 static inline int thread_is_running(struct ptlrpc_thread *thread)
1011 return !!(thread->t_flags & SVC_RUNNING);
1014 static inline int thread_is_event(struct ptlrpc_thread *thread)
1016 return !!(thread->t_flags & SVC_EVENT);
1019 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1021 return !!(thread->t_flags & SVC_SIGNAL);
1024 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1026 thread->t_flags &= ~flags;
1029 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1031 thread->t_flags = flags;
1034 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1036 thread->t_flags |= flags;
1039 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1042 if (thread->t_flags & flags) {
1043 thread->t_flags &= ~flags;
1050 * Request buffer descriptor structure.
1051 * This is a structure that contains one posted request buffer for service.
1052 * Once data land into a buffer, event callback creates actual request and
1053 * notifies wakes one of the service threads to process new incoming request.
1054 * More than one request can fit into the buffer.
1056 struct ptlrpc_request_buffer_desc {
1057 /** Link item for rqbds on a service */
1058 cfs_list_t rqbd_list;
1059 /** History of requests for this buffer */
1060 cfs_list_t rqbd_reqs;
1061 /** Back pointer to service for which this buffer is registered */
1062 struct ptlrpc_service *rqbd_service;
1063 /** LNet descriptor */
1064 lnet_handle_md_t rqbd_md_h;
1066 /** The buffer itself */
1068 struct ptlrpc_cb_id rqbd_cbid;
1070 * This "embedded" request structure is only used for the
1071 * last request to fit into the buffer
1073 struct ptlrpc_request rqbd_req;
1076 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1078 struct ptlrpc_service_ops {
1080 * if non-NULL called during thread creation (ptlrpc_start_thread())
1081 * to initialize service specific per-thread state.
1083 int (*so_thr_init)(struct ptlrpc_thread *thr);
1085 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1086 * destruct state created by ->srv_init().
1088 void (*so_thr_done)(struct ptlrpc_thread *thr);
1090 * Handler function for incoming requests for this service
1092 int (*so_req_handler)(struct ptlrpc_request *req);
1094 * function to determine priority of the request, it's called
1095 * on every new request
1097 int (*so_hpreq_handler)(struct ptlrpc_request *);
1099 * service-specific print fn
1101 void (*so_req_printer)(void *, struct ptlrpc_request *);
1104 #ifndef __cfs_cacheline_aligned
1105 /* NB: put it here for reducing patche dependence */
1106 # define __cfs_cacheline_aligned
1110 * How many high priority requests to serve before serving one normal
1113 #define PTLRPC_SVC_HP_RATIO 10
1116 * Definition of PortalRPC service.
1117 * The service is listening on a particular portal (like tcp port)
1118 * and perform actions for a specific server like IO service for OST
1119 * or general metadata service for MDS.
1121 * ptlrpc service has four locks:
1123 * serialize operations on rqbd and requests waiting for preprocess
1125 * serialize operations active requests sent to this portal
1127 * serialize adaptive timeout stuff
1129 * serialize operations on RS list (reply states)
1131 * We don't have any use-case to take two or more locks at the same time
1132 * for now, so there is no lock order issue.
1134 struct ptlrpc_service {
1135 /** most often accessed fields */
1136 /** chain thru all services */
1137 cfs_list_t srv_list;
1138 /** service operations table */
1139 struct ptlrpc_service_ops srv_ops;
1140 /** only statically allocated strings here; we don't clean them */
1142 /** only statically allocated strings here; we don't clean them */
1143 char *srv_thread_name;
1144 /** service thread list */
1145 cfs_list_t srv_threads;
1146 /** threads to start at beginning of service */
1147 int srv_threads_min;
1148 /** thread upper limit */
1149 int srv_threads_max;
1150 /** always increasing number */
1151 unsigned srv_threads_next_id;
1152 /** # of starting threads */
1153 int srv_threads_starting;
1154 /** # running threads */
1155 int srv_threads_running;
1157 /** Root of /proc dir tree for this service */
1158 cfs_proc_dir_entry_t *srv_procroot;
1159 /** Pointer to statistic data for this service */
1160 struct lprocfs_stats *srv_stats;
1161 /** # hp per lp reqs to handle */
1162 int srv_hpreq_ratio;
1163 /** biggest request to receive */
1164 int srv_max_req_size;
1165 /** biggest reply to send */
1166 int srv_max_reply_size;
1167 /** size of individual buffers */
1169 /** # buffers to allocate in 1 group */
1170 int srv_nbuf_per_group;
1171 /** Local portal on which to receive requests */
1172 __u32 srv_req_portal;
1173 /** Portal on the client to send replies to */
1174 __u32 srv_rep_portal;
1176 * Tags for lu_context associated with this thread, see struct
1180 /** soft watchdog timeout multiplier */
1181 int srv_watchdog_factor;
1182 /** bind threads to CPUs */
1183 unsigned srv_cpu_affinity:1;
1184 /** under unregister_service */
1185 unsigned srv_is_stopping:1;
1188 * serialize the following fields, used for protecting
1189 * rqbd list and incoming requests waiting for preprocess
1191 cfs_spinlock_t srv_lock __cfs_cacheline_aligned;
1192 /** incoming reqs */
1193 cfs_list_t srv_req_in_queue;
1194 /** total # req buffer descs allocated */
1196 /** # posted request buffers */
1197 int srv_nrqbd_receiving;
1198 /** timeout before re-posting reqs, in tick */
1199 cfs_duration_t srv_rqbd_timeout;
1200 /** request buffers to be reposted */
1201 cfs_list_t srv_idle_rqbds;
1202 /** req buffers receiving */
1203 cfs_list_t srv_active_rqbds;
1204 /** request buffer history */
1205 cfs_list_t srv_history_rqbds;
1206 /** # request buffers in history */
1207 int srv_n_history_rqbds;
1208 /** max # request buffers in history */
1209 int srv_max_history_rqbds;
1210 /** request history */
1211 cfs_list_t srv_request_history;
1212 /** next request sequence # */
1213 __u64 srv_request_seq;
1214 /** highest seq culled from history */
1215 __u64 srv_request_max_cull_seq;
1217 * all threads sleep on this. This wait-queue is signalled when new
1218 * incoming request arrives and when difficult reply has to be handled.
1220 cfs_waitq_t srv_waitq;
1223 * serialize the following fields, used for processing requests
1224 * sent to this portal
1226 cfs_spinlock_t srv_rq_lock __cfs_cacheline_aligned;
1227 /** # reqs in either of the queues below */
1228 /** reqs waiting for service */
1229 cfs_list_t srv_request_queue;
1230 /** high priority queue */
1231 cfs_list_t srv_request_hpq;
1232 /** # incoming reqs */
1233 int srv_n_queued_reqs;
1234 /** # reqs being served */
1235 int srv_n_active_reqs;
1236 /** # HPreqs being served */
1237 int srv_n_active_hpreq;
1238 /** # hp requests handled */
1239 int srv_hpreq_count;
1244 * serialize the following fields, used for changes on
1247 cfs_spinlock_t srv_at_lock __cfs_cacheline_aligned;
1248 /** estimated rpc service time */
1249 struct adaptive_timeout srv_at_estimate;
1250 /** reqs waiting for replies */
1251 struct ptlrpc_at_array srv_at_array;
1252 /** early reply timer */
1253 cfs_timer_t srv_at_timer;
1254 /** check early replies */
1255 unsigned srv_at_check;
1257 cfs_time_t srv_at_checktime;
1261 * serialize the following fields, used for processing
1262 * replies for this portal
1264 cfs_spinlock_t srv_rs_lock __cfs_cacheline_aligned;
1265 /** all the active replies */
1266 cfs_list_t srv_active_replies;
1268 /** replies waiting for service */
1269 cfs_list_t srv_reply_queue;
1271 /** List of free reply_states */
1272 cfs_list_t srv_free_rs_list;
1273 /** waitq to run, when adding stuff to srv_free_rs_list */
1274 cfs_waitq_t srv_free_rs_waitq;
1275 /** # 'difficult' replies */
1276 cfs_atomic_t srv_n_difficult_replies;
1277 //struct ptlrpc_srv_ni srv_interfaces[0];
1281 * Declaration of ptlrpcd control structure
1283 struct ptlrpcd_ctl {
1285 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1287 unsigned long pc_flags;
1289 * Thread lock protecting structure fields.
1291 cfs_spinlock_t pc_lock;
1295 cfs_completion_t pc_starting;
1299 cfs_completion_t pc_finishing;
1301 * Thread requests set.
1303 struct ptlrpc_request_set *pc_set;
1305 * Thread name used in cfs_daemonize()
1309 * Environment for request interpreters to run in.
1311 struct lu_env pc_env;
1313 * Index of ptlrpcd thread in the array.
1317 * Number of the ptlrpcd's partners.
1321 * Pointer to the array of partners' ptlrpcd_ctl structure.
1323 struct ptlrpcd_ctl **pc_partners;
1325 * Record the partner index to be processed next.
1330 * Async rpcs flag to make sure that ptlrpcd_check() is called only
1335 * Currently not used.
1339 * User-space async rpcs callback.
1341 void *pc_wait_callback;
1343 * User-space check idle rpcs callback.
1345 void *pc_idle_callback;
1349 /* Bits for pc_flags */
1350 enum ptlrpcd_ctl_flags {
1352 * Ptlrpc thread start flag.
1354 LIOD_START = 1 << 0,
1356 * Ptlrpc thread stop flag.
1360 * Ptlrpc thread force flag (only stop force so far).
1361 * This will cause aborting any inflight rpcs handled
1362 * by thread if LIOD_STOP is specified.
1364 LIOD_FORCE = 1 << 2,
1366 * This is a recovery ptlrpc thread.
1368 LIOD_RECOVERY = 1 << 3,
1370 * The ptlrpcd is bound to some CPU core.
1375 /* ptlrpc/events.c */
1376 extern lnet_handle_eq_t ptlrpc_eq_h;
1377 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1378 lnet_process_id_t *peer, lnet_nid_t *self);
1380 * These callbacks are invoked by LNet when something happened to
1384 extern void request_out_callback(lnet_event_t *ev);
1385 extern void reply_in_callback(lnet_event_t *ev);
1386 extern void client_bulk_callback(lnet_event_t *ev);
1387 extern void request_in_callback(lnet_event_t *ev);
1388 extern void reply_out_callback(lnet_event_t *ev);
1389 #ifdef HAVE_SERVER_SUPPORT
1390 extern void server_bulk_callback(lnet_event_t *ev);
1394 /* ptlrpc/connection.c */
1395 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
1397 struct obd_uuid *uuid);
1398 int ptlrpc_connection_put(struct ptlrpc_connection *c);
1399 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1400 int ptlrpc_connection_init(void);
1401 void ptlrpc_connection_fini(void);
1402 extern lnet_pid_t ptl_get_pid(void);
1404 /* ptlrpc/niobuf.c */
1406 * Actual interfacing with LNet to put/get/register/unregister stuff
1409 #ifdef HAVE_SERVER_SUPPORT
1410 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
1411 int npages, int type, int portal);
1412 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
1413 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
1415 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
1419 LASSERT(desc != NULL);
1421 cfs_spin_lock(&desc->bd_lock);
1422 rc = desc->bd_network_rw;
1423 cfs_spin_unlock(&desc->bd_lock);
1428 int ptlrpc_register_bulk(struct ptlrpc_request *req);
1429 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
1431 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
1433 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1436 LASSERT(req != NULL);
1438 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
1439 req->rq_bulk_deadline > cfs_time_current_sec())
1445 cfs_spin_lock(&desc->bd_lock);
1446 rc = desc->bd_network_rw;
1447 cfs_spin_unlock(&desc->bd_lock);
1451 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
1452 #define PTLRPC_REPLY_EARLY 0x02
1453 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
1454 int ptlrpc_reply(struct ptlrpc_request *req);
1455 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
1456 int ptlrpc_error(struct ptlrpc_request *req);
1457 void ptlrpc_resend_req(struct ptlrpc_request *request);
1458 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
1459 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
1460 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
1463 /* ptlrpc/client.c */
1465 * Client-side portals API. Everything to send requests, receive replies,
1466 * request queues, request management, etc.
1469 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
1470 struct ptlrpc_client *);
1471 void ptlrpc_cleanup_client(struct obd_import *imp);
1472 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
1474 int ptlrpc_queue_wait(struct ptlrpc_request *req);
1475 int ptlrpc_replay_req(struct ptlrpc_request *req);
1476 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
1477 void ptlrpc_restart_req(struct ptlrpc_request *req);
1478 void ptlrpc_abort_inflight(struct obd_import *imp);
1479 void ptlrpc_cleanup_imp(struct obd_import *imp);
1480 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
1482 struct ptlrpc_request_set *ptlrpc_prep_set(void);
1483 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1484 set_interpreter_func fn, void *data);
1485 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
1486 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
1487 int ptlrpc_set_wait(struct ptlrpc_request_set *);
1488 int ptlrpc_expired_set(void *data);
1489 void ptlrpc_interrupted_set(void *data);
1490 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
1491 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
1492 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
1493 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1494 struct ptlrpc_request *req);
1496 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
1497 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
1499 struct ptlrpc_request_pool *
1500 ptlrpc_init_rq_pool(int, int,
1501 void (*populate_pool)(struct ptlrpc_request_pool *, int));
1503 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
1504 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1505 const struct req_format *format);
1506 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
1507 struct ptlrpc_request_pool *,
1508 const struct req_format *format);
1509 void ptlrpc_request_free(struct ptlrpc_request *request);
1510 int ptlrpc_request_pack(struct ptlrpc_request *request,
1511 __u32 version, int opcode);
1512 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1513 const struct req_format *format,
1514 __u32 version, int opcode);
1515 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
1516 __u32 version, int opcode, char **bufs,
1517 struct ptlrpc_cli_ctx *ctx);
1518 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
1519 unsigned int timeout,
1520 ptlrpc_interpterer_t interpreter);
1521 void ptlrpc_fakereq_finished(struct ptlrpc_request *req);
1523 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
1524 int opcode, int count, __u32 *lengths,
1526 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
1527 __u32 version, int opcode,
1528 int count, __u32 *lengths, char **bufs,
1529 struct ptlrpc_request_pool *pool);
1530 void ptlrpc_req_finished(struct ptlrpc_request *request);
1531 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
1532 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
1533 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
1534 int npages, int type, int portal);
1535 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
1536 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
1537 cfs_page_t *page, int pageoffset, int len);
1538 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1539 struct obd_import *imp);
1540 __u64 ptlrpc_next_xid(void);
1541 __u64 ptlrpc_sample_next_xid(void);
1542 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
1544 /* Set of routines to run a function in ptlrpcd context */
1545 void *ptlrpcd_alloc_work(struct obd_import *imp,
1546 int (*cb)(const struct lu_env *, void *), void *data);
1547 void ptlrpcd_destroy_work(void *handler);
1548 int ptlrpcd_queue_work(void *handler);
1551 struct ptlrpc_service_buf_conf {
1552 /* nbufs is how many buffers to post */
1553 unsigned int bc_nbufs;
1554 /* buffer size to post */
1555 unsigned int bc_buf_size;
1556 /* portal to listed for requests on */
1557 unsigned int bc_req_portal;
1558 /* portal of where to send replies to */
1559 unsigned int bc_rep_portal;
1560 /* maximum request size to be accepted for this service */
1561 unsigned int bc_req_max_size;
1562 /* maximum reply size this service can ever send */
1563 unsigned int bc_rep_max_size;
1566 struct ptlrpc_service_thr_conf {
1567 /* threadname should be 8 characters or less - 6 will be added on */
1569 /* min number of service threads to start */
1570 unsigned int tc_nthrs_min;
1571 /* max number of service threads to start */
1572 unsigned int tc_nthrs_max;
1573 /* set NUMA node affinity for service threads */
1574 unsigned int tc_cpu_affinity;
1575 /* Tags for lu_context associated with service thread */
1579 struct ptlrpc_service_conf {
1582 /* soft watchdog timeout multiplifier to print stuck service traces */
1583 unsigned int psc_watchdog_factor;
1584 /* buffer information */
1585 struct ptlrpc_service_buf_conf psc_buf;
1586 /* thread information */
1587 struct ptlrpc_service_thr_conf psc_thr;
1588 /* function table */
1589 struct ptlrpc_service_ops psc_ops;
1592 /* ptlrpc/service.c */
1594 * Server-side services API. Register/unregister service, request state
1595 * management, service thread management
1599 void ptlrpc_save_lock(struct ptlrpc_request *req,
1600 struct lustre_handle *lock, int mode, int no_ack);
1601 void ptlrpc_commit_replies(struct obd_export *exp);
1602 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
1603 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
1604 struct ptlrpc_service *ptlrpc_register_service(
1605 struct ptlrpc_service_conf *conf,
1606 struct proc_dir_entry *proc_entry);
1607 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
1609 int ptlrpc_start_threads(struct ptlrpc_service *svc);
1610 int ptlrpc_start_thread(struct ptlrpc_service *svc);
1611 int ptlrpc_unregister_service(struct ptlrpc_service *service);
1612 int liblustre_check_services(void *arg);
1613 void ptlrpc_daemonize(char *name);
1614 int ptlrpc_service_health_check(struct ptlrpc_service *);
1615 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req);
1616 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
1619 int ptlrpc_hr_init(void);
1620 void ptlrpc_hr_fini(void);
1622 # define ptlrpc_hr_init() (0)
1623 # define ptlrpc_hr_fini() do {} while(0)
1626 struct ptlrpc_svc_data {
1628 struct ptlrpc_service *svc;
1629 struct ptlrpc_thread *thread;
1633 /* ptlrpc/import.c */
1638 int ptlrpc_connect_import(struct obd_import *imp);
1639 int ptlrpc_init_import(struct obd_import *imp);
1640 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
1641 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
1642 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
1645 /* ptlrpc/pack_generic.c */
1646 int ptlrpc_reconnect_import(struct obd_import *imp);
1650 * ptlrpc msg buffer and swab interface
1654 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
1656 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
1658 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
1659 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
1661 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
1662 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
1664 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
1665 __u32 *lens, char **bufs);
1666 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
1668 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
1669 __u32 *lens, char **bufs, int flags);
1670 #define LPRFL_EARLY_REPLY 1
1671 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
1672 char **bufs, int flags);
1673 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
1674 unsigned int newlen, int move_data);
1675 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
1676 int __lustre_unpack_msg(struct lustre_msg *m, int len);
1677 int lustre_msg_hdr_size(__u32 magic, int count);
1678 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
1679 int lustre_msg_size_v2(int count, __u32 *lengths);
1680 int lustre_packed_msg_size(struct lustre_msg *msg);
1681 int lustre_msg_early_size(void);
1682 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
1683 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
1684 int lustre_msg_buflen(struct lustre_msg *m, int n);
1685 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
1686 int lustre_msg_bufcount(struct lustre_msg *m);
1687 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
1688 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
1689 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
1690 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
1691 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
1692 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
1693 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
1694 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
1695 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
1696 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
1697 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
1698 __u32 lustre_msg_get_type(struct lustre_msg *msg);
1699 __u32 lustre_msg_get_version(struct lustre_msg *msg);
1700 void lustre_msg_add_version(struct lustre_msg *msg, int version);
1701 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
1702 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
1703 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
1704 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
1705 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
1706 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
1707 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
1708 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
1709 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
1710 int lustre_msg_get_status(struct lustre_msg *msg);
1711 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
1712 int lustre_msg_is_v1(struct lustre_msg *msg);
1713 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
1714 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
1715 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
1716 char *lustre_msg_get_jobid(struct lustre_msg *msg);
1717 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
1718 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 0, 0)
1719 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
1721 # warning "remove checksum compatibility support for b1_8"
1722 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
1724 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
1725 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
1726 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
1727 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
1728 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
1729 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
1730 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
1731 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
1732 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
1733 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
1734 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
1735 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
1736 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
1737 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
1738 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
1741 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
1742 unsigned int newlen, int move_data)
1744 LASSERT(req->rq_reply_state);
1745 LASSERT(req->rq_repmsg);
1746 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
1751 /** Change request phase of \a req to \a new_phase */
1753 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
1755 if (req->rq_phase == new_phase)
1758 if (new_phase == RQ_PHASE_UNREGISTERING) {
1759 req->rq_next_phase = req->rq_phase;
1761 cfs_atomic_inc(&req->rq_import->imp_unregistering);
1764 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1766 cfs_atomic_dec(&req->rq_import->imp_unregistering);
1769 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
1770 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
1772 req->rq_phase = new_phase;
1776 * Returns true if request \a req got early reply and hard deadline is not met
1779 ptlrpc_client_early(struct ptlrpc_request *req)
1781 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1782 req->rq_reply_deadline > cfs_time_current_sec())
1784 return req->rq_early;
1788 * Returns true if we got real reply from server for this request
1791 ptlrpc_client_replied(struct ptlrpc_request *req)
1793 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1794 req->rq_reply_deadline > cfs_time_current_sec())
1796 return req->rq_replied;
1799 /** Returns true if request \a req is in process of receiving server reply */
1801 ptlrpc_client_recv(struct ptlrpc_request *req)
1803 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1804 req->rq_reply_deadline > cfs_time_current_sec())
1806 return req->rq_receiving_reply;
1810 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
1814 cfs_spin_lock(&req->rq_lock);
1815 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1816 req->rq_reply_deadline > cfs_time_current_sec()) {
1817 cfs_spin_unlock(&req->rq_lock);
1820 rc = req->rq_receiving_reply || req->rq_must_unlink;
1821 cfs_spin_unlock(&req->rq_lock);
1826 ptlrpc_client_wake_req(struct ptlrpc_request *req)
1828 if (req->rq_set == NULL)
1829 cfs_waitq_signal(&req->rq_reply_waitq);
1831 cfs_waitq_signal(&req->rq_set->set_waitq);
1835 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
1837 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
1838 cfs_atomic_inc(&rs->rs_refcount);
1842 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
1844 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
1845 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
1846 lustre_free_reply_state(rs);
1849 /* Should only be called once per req */
1850 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
1852 if (req->rq_reply_state == NULL)
1853 return; /* shouldn't occur */
1854 ptlrpc_rs_decref(req->rq_reply_state);
1855 req->rq_reply_state = NULL;
1856 req->rq_repmsg = NULL;
1859 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
1861 return lustre_msg_get_magic(req->rq_reqmsg);
1864 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
1866 switch (req->rq_reqmsg->lm_magic) {
1867 case LUSTRE_MSG_MAGIC_V2:
1868 return req->rq_reqmsg->lm_repsize;
1870 LASSERTF(0, "incorrect message magic: %08x\n",
1871 req->rq_reqmsg->lm_magic);
1876 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
1878 if (req->rq_delay_limit != 0 &&
1879 cfs_time_before(cfs_time_add(req->rq_queued_time,
1880 cfs_time_seconds(req->rq_delay_limit)),
1881 cfs_time_current())) {
1887 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
1889 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
1890 cfs_spin_lock(&req->rq_lock);
1891 req->rq_no_resend = 1;
1892 cfs_spin_unlock(&req->rq_lock);
1894 return req->rq_no_resend;
1897 /* ldlm/ldlm_lib.c */
1899 * Target client logic
1902 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
1903 int client_obd_cleanup(struct obd_device *obddev);
1904 int client_connect_import(const struct lu_env *env,
1905 struct obd_export **exp, struct obd_device *obd,
1906 struct obd_uuid *cluuid, struct obd_connect_data *,
1908 int client_disconnect_export(struct obd_export *exp);
1909 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
1911 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
1912 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
1913 struct obd_uuid *uuid);
1914 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
1915 void client_destroy_import(struct obd_import *imp);
1918 #ifdef HAVE_SERVER_SUPPORT
1919 int server_disconnect_export(struct obd_export *exp);
1922 /* ptlrpc/pinger.c */
1924 * Pinger API (client side only)
1927 enum timeout_event {
1930 struct timeout_item;
1931 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
1932 int ptlrpc_pinger_add_import(struct obd_import *imp);
1933 int ptlrpc_pinger_del_import(struct obd_import *imp);
1934 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
1935 timeout_cb_t cb, void *data,
1936 cfs_list_t *obd_list);
1937 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
1938 enum timeout_event event);
1939 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
1940 int ptlrpc_obd_ping(struct obd_device *obd);
1941 cfs_time_t ptlrpc_suspend_wakeup_time(void);
1943 void ping_evictor_start(void);
1944 void ping_evictor_stop(void);
1946 #define ping_evictor_start() do {} while (0)
1947 #define ping_evictor_stop() do {} while (0)
1949 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
1952 /* ptlrpc daemon bind policy */
1954 /* all ptlrpcd threads are free mode */
1955 PDB_POLICY_NONE = 1,
1956 /* all ptlrpcd threads are bound mode */
1957 PDB_POLICY_FULL = 2,
1958 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
1959 PDB_POLICY_PAIR = 3,
1960 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
1961 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
1962 * If kernel supports NUMA, pthrpcd threads are binded and
1963 * grouped by NUMA node */
1964 PDB_POLICY_NEIGHBOR = 4,
1967 /* ptlrpc daemon load policy
1968 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
1969 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
1970 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
1971 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
1972 * depends on which is scheduled firstly, to accelerate the RPC processing. */
1974 /* on the same CPU core as the caller */
1975 PDL_POLICY_SAME = 1,
1976 /* within the same CPU partition, but not the same core as the caller */
1977 PDL_POLICY_LOCAL = 2,
1978 /* round-robin on all CPU cores, but not the same core as the caller */
1979 PDL_POLICY_ROUND = 3,
1980 /* the specified CPU core is preferred, but not enforced */
1981 PDL_POLICY_PREFERRED = 4,
1984 /* ptlrpc/ptlrpcd.c */
1985 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
1986 void ptlrpcd_wake(struct ptlrpc_request *req);
1987 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
1988 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
1989 int ptlrpcd_addref(void);
1990 void ptlrpcd_decref(void);
1992 /* ptlrpc/lproc_ptlrpc.c */
1994 * procfs output related functions
1997 const char* ll_opcode2str(__u32 opcode);
1999 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2000 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2001 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2003 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2004 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2005 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2009 /* ptlrpc/llog_server.c */
2010 int llog_origin_handle_create(struct ptlrpc_request *req);
2011 int llog_origin_handle_destroy(struct ptlrpc_request *req);
2012 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2013 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2014 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2015 int llog_origin_handle_close(struct ptlrpc_request *req);
2016 int llog_origin_handle_cancel(struct ptlrpc_request *req);
2017 int llog_catinfo(struct ptlrpc_request *req);
2019 /* ptlrpc/llog_client.c */
2020 extern struct llog_operations llog_client_ops;