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);
250 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
253 * Definition of request set structure.
254 * Request set is a list of requests (not necessary to the same target) that
255 * once populated with RPCs could be sent in parallel.
256 * There are two kinds of request sets. General purpose and with dedicated
257 * serving thread. Example of the latter is ptlrpcd set.
258 * For general purpose sets once request set started sending it is impossible
259 * to add new requests to such set.
260 * Provides a way to call "completion callbacks" when all requests in the set
263 struct ptlrpc_request_set {
264 cfs_atomic_t set_refcount;
265 /** number of in queue requests */
266 cfs_atomic_t set_new_count;
267 /** number of uncompleted requests */
268 cfs_atomic_t set_remaining;
269 /** wait queue to wait on for request events */
270 cfs_waitq_t set_waitq;
271 cfs_waitq_t *set_wakeup_ptr;
272 /** List of requests in the set */
273 cfs_list_t set_requests;
275 * List of completion callbacks to be called when the set is completed
276 * This is only used if \a set_interpret is NULL.
277 * Links struct ptlrpc_set_cbdata.
279 cfs_list_t set_cblist;
280 /** Completion callback, if only one. */
281 set_interpreter_func set_interpret;
282 /** opaq argument passed to completion \a set_interpret callback. */
284 /** rq_status of requests that have been freed already */
287 * Lock for \a set_new_requests manipulations
288 * locked so that any old caller can communicate requests to
289 * the set holder who can then fold them into the lock-free set
291 cfs_spinlock_t set_new_req_lock;
292 /** List of new yet unsent requests. Only used with ptlrpcd now. */
293 cfs_list_t set_new_requests;
295 /** Additional fields used by the flow control extension */
296 /** Maximum number of RPCs in flight */
297 int set_max_inflight;
298 /** Callback function used to generate RPCs */
299 set_producer_func set_producer;
300 /** opaq argument passed to the producer callback */
301 void *set_producer_arg;
305 * Description of a single ptrlrpc_set callback
307 struct ptlrpc_set_cbdata {
308 /** List linkage item */
310 /** Pointer to interpreting function */
311 set_interpreter_func psc_interpret;
312 /** Opaq argument to pass to the callback */
316 struct ptlrpc_bulk_desc;
319 * ptlrpc callback & work item stuff
321 struct ptlrpc_cb_id {
322 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
323 void *cbid_arg; /* additional arg */
326 /** Maximum number of locks to fit into reply state */
327 #define RS_MAX_LOCKS 8
331 * Structure to define reply state on the server
332 * Reply state holds various reply message information. Also for "difficult"
333 * replies (rep-ack case) we store the state after sending reply and wait
334 * for the client to acknowledge the reception. In these cases locks could be
335 * added to the state for replay/failover consistency guarantees.
337 struct ptlrpc_reply_state {
338 /** Callback description */
339 struct ptlrpc_cb_id rs_cb_id;
340 /** Linkage for list of all reply states in a system */
342 /** Linkage for list of all reply states on same export */
343 cfs_list_t rs_exp_list;
344 /** Linkage for list of all reply states for same obd */
345 cfs_list_t rs_obd_list;
347 cfs_list_t rs_debug_list;
349 /** A spinlock to protect the reply state flags */
350 cfs_spinlock_t rs_lock;
351 /** Reply state flags */
352 unsigned long rs_difficult:1; /* ACK/commit stuff */
353 unsigned long rs_no_ack:1; /* no ACK, even for
354 difficult requests */
355 unsigned long rs_scheduled:1; /* being handled? */
356 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
357 unsigned long rs_handled:1; /* been handled yet? */
358 unsigned long rs_on_net:1; /* reply_out_callback pending? */
359 unsigned long rs_prealloc:1; /* rs from prealloc list */
360 unsigned long rs_committed:1;/* the transaction was committed
361 and the rs was dispatched
362 by ptlrpc_commit_replies */
363 /** Size of the state */
367 /** Transaction number */
371 struct obd_export *rs_export;
372 struct ptlrpc_service *rs_service;
373 /** Lnet metadata handle for the reply */
374 lnet_handle_md_t rs_md_h;
375 cfs_atomic_t rs_refcount;
377 /** Context for the sevice thread */
378 struct ptlrpc_svc_ctx *rs_svc_ctx;
379 /** Reply buffer (actually sent to the client), encoded if needed */
380 struct lustre_msg *rs_repbuf; /* wrapper */
381 /** Size of the reply buffer */
382 int rs_repbuf_len; /* wrapper buf length */
383 /** Size of the reply message */
384 int rs_repdata_len; /* wrapper msg length */
386 * Actual reply message. Its content is encrupted (if needed) to
387 * produce reply buffer for actual sending. In simple case
388 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
390 struct lustre_msg *rs_msg; /* reply message */
392 /** Number of locks awaiting client ACK */
394 /** Handles of locks awaiting client reply ACK */
395 struct lustre_handle rs_locks[RS_MAX_LOCKS];
396 /** Lock modes of locks in \a rs_locks */
397 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
400 struct ptlrpc_thread;
404 RQ_PHASE_NEW = 0xebc0de00,
405 RQ_PHASE_RPC = 0xebc0de01,
406 RQ_PHASE_BULK = 0xebc0de02,
407 RQ_PHASE_INTERPRET = 0xebc0de03,
408 RQ_PHASE_COMPLETE = 0xebc0de04,
409 RQ_PHASE_UNREGISTERING = 0xebc0de05,
410 RQ_PHASE_UNDEFINED = 0xebc0de06
413 /** Type of request interpreter call-back */
414 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
415 struct ptlrpc_request *req,
419 * Definition of request pool structure.
420 * The pool is used to store empty preallocated requests for the case
421 * when we would actually need to send something without performing
422 * any allocations (to avoid e.g. OOM).
424 struct ptlrpc_request_pool {
425 /** Locks the list */
426 cfs_spinlock_t prp_lock;
427 /** list of ptlrpc_request structs */
428 cfs_list_t prp_req_list;
429 /** Maximum message size that would fit into a rquest from this pool */
431 /** Function to allocate more requests for this pool */
432 void (*prp_populate)(struct ptlrpc_request_pool *, int);
441 * Basic request prioritization operations structure.
442 * The whole idea is centered around locks and RPCs that might affect locks.
443 * When a lock is contended we try to give priority to RPCs that might lead
444 * to fastest release of that lock.
445 * Currently only implemented for OSTs only in a way that makes all
446 * IO and truncate RPCs that are coming from a locked region where a lock is
447 * contended a priority over other requests.
449 struct ptlrpc_hpreq_ops {
451 * Check if the lock handle of the given lock is the same as
452 * taken from the request.
454 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
456 * Check if the request is a high priority one.
458 int (*hpreq_check)(struct ptlrpc_request *);
460 * Called after the request has been handled.
462 void (*hpreq_fini)(struct ptlrpc_request *);
466 * Represents remote procedure call.
468 * This is a staple structure used by everybody wanting to send a request
471 struct ptlrpc_request {
472 /* Request type: one of PTL_RPC_MSG_* */
475 * Linkage item through which this request is included into
476 * sending/delayed lists on client and into rqbd list on server
480 * Server side list of incoming unserved requests sorted by arrival
481 * time. Traversed from time to time to notice about to expire
482 * requests and sent back "early replies" to clients to let them
483 * know server is alive and well, just very busy to service their
486 cfs_list_t rq_timed_list;
487 /** server-side history, used for debuging purposes. */
488 cfs_list_t rq_history_list;
489 /** server-side per-export list */
490 cfs_list_t rq_exp_list;
491 /** server-side hp handlers */
492 struct ptlrpc_hpreq_ops *rq_ops;
493 /** history sequence # */
494 __u64 rq_history_seq;
495 /** the index of service's srv_at_array into which request is linked */
497 /** Result of request processing */
499 /** Lock to protect request flags and some other important bits, like
502 cfs_spinlock_t rq_lock;
503 /** client-side flags are serialized by rq_lock */
504 unsigned long rq_intr:1, rq_replied:1, rq_err:1,
505 rq_timedout:1, rq_resend:1, rq_restart:1,
507 * when ->rq_replay is set, request is kept by the client even
508 * after server commits corresponding transaction. This is
509 * used for operations that require sequence of multiple
510 * requests to be replayed. The only example currently is file
511 * open/close. When last request in such a sequence is
512 * committed, ->rq_replay is cleared on all requests in the
516 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
517 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
518 rq_early:1, rq_must_unlink:1,
519 rq_fake:1, /* this fake req */
520 rq_memalloc:1, /* req originated from "kswapd" */
521 /* server-side flags */
522 rq_packed_final:1, /* packed final reply */
523 rq_hp:1, /* high priority RPC */
524 rq_at_linked:1, /* link into service's srv_at_array */
527 /* whether the "rq_set" is a valid one */
531 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
532 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
533 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
534 server-side refcounf for multiple replies */
536 /** initial thread servicing this request */
537 struct ptlrpc_thread *rq_svc_thread;
539 /** Portal to which this request would be sent */
540 int rq_request_portal; /* XXX FIXME bug 249 */
541 /** Portal where to wait for reply and where reply would be sent */
542 int rq_reply_portal; /* XXX FIXME bug 249 */
546 * !rq_truncate : # reply bytes actually received,
547 * rq_truncate : required repbuf_len for resend
550 /** Request length */
552 /** Request message - what client sent */
553 struct lustre_msg *rq_reqmsg;
557 /** Reply message - server response */
558 struct lustre_msg *rq_repmsg;
559 /** Transaction number */
564 * List item to for replay list. Not yet commited requests get linked
566 * Also see \a rq_replay comment above.
568 cfs_list_t rq_replay_list;
571 * security and encryption data
573 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
574 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
575 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
577 struct sptlrpc_flavor rq_flvr; /**< for client & server */
578 enum lustre_sec_part rq_sp_from;
580 unsigned long /* client/server security flags */
581 rq_ctx_init:1, /* context initiation */
582 rq_ctx_fini:1, /* context destroy */
583 rq_bulk_read:1, /* request bulk read */
584 rq_bulk_write:1, /* request bulk write */
585 /* server authentication flags */
586 rq_auth_gss:1, /* authenticated by gss */
587 rq_auth_remote:1, /* authed as remote user */
588 rq_auth_usr_root:1, /* authed as root */
589 rq_auth_usr_mdt:1, /* authed as mdt */
590 rq_auth_usr_ost:1, /* authed as ost */
591 /* security tfm flags */
594 /* doesn't expect reply FIXME */
596 rq_pill_init:1; /* pill initialized */
598 uid_t rq_auth_uid; /* authed uid */
599 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
601 /* (server side), pointed directly into req buffer */
602 struct ptlrpc_user_desc *rq_user_desc;
604 /** early replies go to offset 0, regular replies go after that */
605 unsigned int rq_reply_off;
607 /* various buffer pointers */
608 struct lustre_msg *rq_reqbuf; /* req wrapper */
609 int rq_reqbuf_len; /* req wrapper buf len */
610 int rq_reqdata_len; /* req wrapper msg len */
611 char *rq_repbuf; /* rep buffer */
612 int rq_repbuf_len; /* rep buffer len */
613 struct lustre_msg *rq_repdata; /* rep wrapper msg */
614 int rq_repdata_len; /* rep wrapper msg len */
615 struct lustre_msg *rq_clrbuf; /* only in priv mode */
616 int rq_clrbuf_len; /* only in priv mode */
617 int rq_clrdata_len; /* only in priv mode */
621 /** Fields that help to see if request and reply were swabbed or not */
622 __u32 rq_req_swab_mask;
623 __u32 rq_rep_swab_mask;
625 /** What was import generation when this request was sent */
626 int rq_import_generation;
627 enum lustre_imp_state rq_send_state;
629 /** how many early replies (for stats) */
632 /** client+server request */
633 lnet_handle_md_t rq_req_md_h;
634 struct ptlrpc_cb_id rq_req_cbid;
635 /** optional time limit for send attempts */
636 cfs_duration_t rq_delay_limit;
637 /** time request was first queued */
638 cfs_time_t rq_queued_time;
641 /** request arrival time */
642 struct timeval rq_arrival_time;
643 /** separated reply state */
644 struct ptlrpc_reply_state *rq_reply_state;
645 /** incoming request buffer */
646 struct ptlrpc_request_buffer_desc *rq_rqbd;
648 __u32 rq_uid; /* peer uid, used in MDS only */
651 /** client-only incoming reply */
652 lnet_handle_md_t rq_reply_md_h;
653 cfs_waitq_t rq_reply_waitq;
654 struct ptlrpc_cb_id rq_reply_cbid;
658 /** Peer description (the other side) */
659 lnet_process_id_t rq_peer;
660 /** Server-side, export on which request was received */
661 struct obd_export *rq_export;
662 /** Client side, import where request is being sent */
663 struct obd_import *rq_import;
665 /** Replay callback, called after request is replayed at recovery */
666 void (*rq_replay_cb)(struct ptlrpc_request *);
668 * Commit callback, called when request is committed and about to be
671 void (*rq_commit_cb)(struct ptlrpc_request *);
672 /** Opaq data for replay and commit callbacks. */
675 /** For bulk requests on client only: bulk descriptor */
676 struct ptlrpc_bulk_desc *rq_bulk;
678 /** client outgoing req */
680 * when request/reply sent (secs), or time when request should be sent
683 /** time for request really sent out */
686 /** when request must finish. volatile
687 * so that servers' early reply updates to the deadline aren't
688 * kept in per-cpu cache */
689 volatile time_t rq_deadline;
690 /** when req reply unlink must finish. */
691 time_t rq_reply_deadline;
692 /** when req bulk unlink must finish. */
693 time_t rq_bulk_deadline;
695 * service time estimate (secs)
696 * If the requestsis not served by this time, it is marked as timed out.
700 /** Multi-rpc bits */
701 /** Link item for request set lists */
702 cfs_list_t rq_set_chain;
703 /** Per-request waitq introduced by bug 21938 for recovery waiting */
704 cfs_waitq_t rq_set_waitq;
705 /** Link back to the request set */
706 struct ptlrpc_request_set *rq_set;
707 /** Async completion handler, called when reply is received */
708 ptlrpc_interpterer_t rq_interpret_reply;
709 /** Async completion context */
710 union ptlrpc_async_args rq_async_args;
712 /** Pool if request is from preallocated list */
713 struct ptlrpc_request_pool *rq_pool;
715 struct lu_context rq_session;
716 struct lu_context rq_recov_session;
718 /** request format description */
719 struct req_capsule rq_pill;
723 * Call completion handler for rpc if any, return it's status or original
724 * rc if there was no handler defined for this request.
726 static inline int ptlrpc_req_interpret(const struct lu_env *env,
727 struct ptlrpc_request *req, int rc)
729 if (req->rq_interpret_reply != NULL) {
730 req->rq_status = req->rq_interpret_reply(env, req,
733 return req->rq_status;
739 * Returns 1 if request buffer at offset \a index was already swabbed
741 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
743 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
744 return req->rq_req_swab_mask & (1 << index);
748 * Returns 1 if request reply buffer at offset \a index was already swabbed
750 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
752 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
753 return req->rq_rep_swab_mask & (1 << index);
757 * Returns 1 if request needs to be swabbed into local cpu byteorder
759 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
761 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
765 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
767 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
769 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
773 * Mark request buffer at offset \a index that it was already swabbed
775 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
777 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
778 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
779 req->rq_req_swab_mask |= 1 << index;
783 * Mark request reply buffer at offset \a index that it was already swabbed
785 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
787 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
788 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
789 req->rq_rep_swab_mask |= 1 << index;
793 * Convert numerical request phase value \a phase into text string description
795 static inline const char *
796 ptlrpc_phase2str(enum rq_phase phase)
805 case RQ_PHASE_INTERPRET:
807 case RQ_PHASE_COMPLETE:
809 case RQ_PHASE_UNREGISTERING:
810 return "Unregistering";
817 * Convert numerical request phase of the request \a req into text stringi
820 static inline const char *
821 ptlrpc_rqphase2str(struct ptlrpc_request *req)
823 return ptlrpc_phase2str(req->rq_phase);
827 * Debugging functions and helpers to print request structure into debug log
830 /* Spare the preprocessor, spoil the bugs. */
831 #define FLAG(field, str) (field ? str : "")
833 /** Convert bit flags into a string */
834 #define DEBUG_REQ_FLAGS(req) \
835 ptlrpc_rqphase2str(req), \
836 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
837 FLAG(req->rq_err, "E"), \
838 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
839 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
840 FLAG(req->rq_no_resend, "N"), \
841 FLAG(req->rq_waiting, "W"), \
842 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
843 FLAG(req->rq_committed, "M")
845 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
847 void _debug_req(struct ptlrpc_request *req,
848 struct libcfs_debug_msg_data *data, const char *fmt, ...)
849 __attribute__ ((format (printf, 3, 4)));
852 * Helper that decides if we need to print request accordig to current debug
855 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
857 CFS_CHECK_STACK(msgdata, mask, cdls); \
859 if (((mask) & D_CANTMASK) != 0 || \
860 ((libcfs_debug & (mask)) != 0 && \
861 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
862 _debug_req((req), msgdata, fmt, ##a); \
866 * This is the debug print function you need to use to print request sturucture
867 * content into lustre debug log.
868 * for most callers (level is a constant) this is resolved at compile time */
869 #define DEBUG_REQ(level, req, fmt, args...) \
871 if ((level) & (D_ERROR | D_WARNING)) { \
872 static cfs_debug_limit_state_t cdls; \
873 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
874 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
876 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
877 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
883 * Structure that defines a single page of a bulk transfer
885 struct ptlrpc_bulk_page {
886 /** Linkage to list of pages in a bulk */
889 * Number of bytes in a page to transfer starting from \a bp_pageoffset
892 /** offset within a page */
894 /** The page itself */
895 struct page *bp_page;
898 #define BULK_GET_SOURCE 0
899 #define BULK_PUT_SINK 1
900 #define BULK_GET_SINK 2
901 #define BULK_PUT_SOURCE 3
904 * Definition of buk descriptor.
905 * Bulks are special "Two phase" RPCs where initial request message
906 * is sent first and it is followed bt a transfer (o receiving) of a large
907 * amount of data to be settled into pages referenced from the bulk descriptors.
908 * Bulks transfers (the actual data following the small requests) are done
909 * on separate LNet portals.
910 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
911 * Another user is readpage for MDT.
913 struct ptlrpc_bulk_desc {
914 /** completed successfully */
915 unsigned long bd_success:1;
916 /** accessible to the network (network io potentially in progress) */
917 unsigned long bd_network_rw:1;
918 /** {put,get}{source,sink} */
919 unsigned long bd_type:2;
921 unsigned long bd_registered:1;
922 /** For serialization with callback */
923 cfs_spinlock_t bd_lock;
924 /** Import generation when request for this bulk was sent */
925 int bd_import_generation;
926 /** Server side - export this bulk created for */
927 struct obd_export *bd_export;
928 /** Client side - import this bulk was sent on */
929 struct obd_import *bd_import;
930 /** LNet portal for this bulk */
932 /** Back pointer to the request */
933 struct ptlrpc_request *bd_req;
934 cfs_waitq_t bd_waitq; /* server side only WQ */
935 int bd_iov_count; /* # entries in bd_iov */
936 int bd_max_iov; /* allocated size of bd_iov */
937 int bd_nob; /* # bytes covered */
938 int bd_nob_transferred; /* # bytes GOT/PUT */
942 struct ptlrpc_cb_id bd_cbid; /* network callback info */
943 lnet_handle_md_t bd_md_h; /* associated MD */
944 lnet_nid_t bd_sender; /* stash event::sender */
946 #if defined(__KERNEL__)
948 * encrypt iov, size is either 0 or bd_iov_count.
950 lnet_kiov_t *bd_enc_iov;
952 lnet_kiov_t bd_iov[0];
954 lnet_md_iovec_t bd_iov[0];
959 SVC_STOPPED = 1 << 0,
960 SVC_STOPPING = 1 << 1,
961 SVC_STARTING = 1 << 2,
962 SVC_RUNNING = 1 << 3,
968 * Definition of server service thread structure
970 struct ptlrpc_thread {
972 * List of active threads in svc->srv_threads
976 * thread-private data (preallocated memory)
981 * service thread index, from ptlrpc_start_threads
989 * put watchdog in the structure per thread b=14840
991 struct lc_watchdog *t_watchdog;
993 * the svc this thread belonged to b=18582
995 struct ptlrpc_service *t_svc;
996 cfs_waitq_t t_ctl_waitq;
997 struct lu_env *t_env;
1000 static inline int thread_is_init(struct ptlrpc_thread *thread)
1002 return thread->t_flags == 0;
1005 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1007 return !!(thread->t_flags & SVC_STOPPED);
1010 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1012 return !!(thread->t_flags & SVC_STOPPING);
1015 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1017 return !!(thread->t_flags & SVC_STARTING);
1020 static inline int thread_is_running(struct ptlrpc_thread *thread)
1022 return !!(thread->t_flags & SVC_RUNNING);
1025 static inline int thread_is_event(struct ptlrpc_thread *thread)
1027 return !!(thread->t_flags & SVC_EVENT);
1030 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1032 return !!(thread->t_flags & SVC_SIGNAL);
1035 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1037 thread->t_flags &= ~flags;
1040 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1042 thread->t_flags = flags;
1045 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1047 thread->t_flags |= flags;
1050 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1053 if (thread->t_flags & flags) {
1054 thread->t_flags &= ~flags;
1061 * Request buffer descriptor structure.
1062 * This is a structure that contains one posted request buffer for service.
1063 * Once data land into a buffer, event callback creates actual request and
1064 * notifies wakes one of the service threads to process new incoming request.
1065 * More than one request can fit into the buffer.
1067 struct ptlrpc_request_buffer_desc {
1068 /** Link item for rqbds on a service */
1069 cfs_list_t rqbd_list;
1070 /** History of requests for this buffer */
1071 cfs_list_t rqbd_reqs;
1072 /** Back pointer to service for which this buffer is registered */
1073 struct ptlrpc_service *rqbd_service;
1074 /** LNet descriptor */
1075 lnet_handle_md_t rqbd_md_h;
1077 /** The buffer itself */
1079 struct ptlrpc_cb_id rqbd_cbid;
1081 * This "embedded" request structure is only used for the
1082 * last request to fit into the buffer
1084 struct ptlrpc_request rqbd_req;
1087 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1089 struct ptlrpc_service_ops {
1091 * if non-NULL called during thread creation (ptlrpc_start_thread())
1092 * to initialize service specific per-thread state.
1094 int (*so_thr_init)(struct ptlrpc_thread *thr);
1096 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1097 * destruct state created by ->srv_init().
1099 void (*so_thr_done)(struct ptlrpc_thread *thr);
1101 * Handler function for incoming requests for this service
1103 int (*so_req_handler)(struct ptlrpc_request *req);
1105 * function to determine priority of the request, it's called
1106 * on every new request
1108 int (*so_hpreq_handler)(struct ptlrpc_request *);
1110 * service-specific print fn
1112 void (*so_req_printer)(void *, struct ptlrpc_request *);
1115 #ifndef __cfs_cacheline_aligned
1116 /* NB: put it here for reducing patche dependence */
1117 # define __cfs_cacheline_aligned
1121 * How many high priority requests to serve before serving one normal
1124 #define PTLRPC_SVC_HP_RATIO 10
1127 * Definition of PortalRPC service.
1128 * The service is listening on a particular portal (like tcp port)
1129 * and perform actions for a specific server like IO service for OST
1130 * or general metadata service for MDS.
1132 * ptlrpc service has four locks:
1134 * serialize operations on rqbd and requests waiting for preprocess
1136 * serialize operations active requests sent to this portal
1138 * serialize adaptive timeout stuff
1140 * serialize operations on RS list (reply states)
1142 * We don't have any use-case to take two or more locks at the same time
1143 * for now, so there is no lock order issue.
1145 struct ptlrpc_service {
1146 /** most often accessed fields */
1147 /** chain thru all services */
1148 cfs_list_t srv_list;
1149 /** service operations table */
1150 struct ptlrpc_service_ops srv_ops;
1151 /** only statically allocated strings here; we don't clean them */
1153 /** only statically allocated strings here; we don't clean them */
1154 char *srv_thread_name;
1155 /** service thread list */
1156 cfs_list_t srv_threads;
1157 /** threads to start at beginning of service */
1158 int srv_threads_min;
1159 /** thread upper limit */
1160 int srv_threads_max;
1161 /** always increasing number */
1162 unsigned srv_threads_next_id;
1163 /** # of starting threads */
1164 int srv_threads_starting;
1165 /** # running threads */
1166 int srv_threads_running;
1168 /** Root of /proc dir tree for this service */
1169 cfs_proc_dir_entry_t *srv_procroot;
1170 /** Pointer to statistic data for this service */
1171 struct lprocfs_stats *srv_stats;
1172 /** # hp per lp reqs to handle */
1173 int srv_hpreq_ratio;
1174 /** biggest request to receive */
1175 int srv_max_req_size;
1176 /** biggest reply to send */
1177 int srv_max_reply_size;
1178 /** size of individual buffers */
1180 /** # buffers to allocate in 1 group */
1181 int srv_nbuf_per_group;
1182 /** Local portal on which to receive requests */
1183 __u32 srv_req_portal;
1184 /** Portal on the client to send replies to */
1185 __u32 srv_rep_portal;
1187 * Tags for lu_context associated with this thread, see struct
1191 /** soft watchdog timeout multiplier */
1192 int srv_watchdog_factor;
1193 /** bind threads to CPUs */
1194 unsigned srv_cpu_affinity:1;
1195 /** under unregister_service */
1196 unsigned srv_is_stopping:1;
1199 * serialize the following fields, used for protecting
1200 * rqbd list and incoming requests waiting for preprocess
1202 cfs_spinlock_t srv_lock __cfs_cacheline_aligned;
1203 /** incoming reqs */
1204 cfs_list_t srv_req_in_queue;
1205 /** total # req buffer descs allocated */
1207 /** # posted request buffers */
1208 int srv_nrqbd_receiving;
1209 /** timeout before re-posting reqs, in tick */
1210 cfs_duration_t srv_rqbd_timeout;
1211 /** request buffers to be reposted */
1212 cfs_list_t srv_idle_rqbds;
1213 /** req buffers receiving */
1214 cfs_list_t srv_active_rqbds;
1215 /** request buffer history */
1216 cfs_list_t srv_history_rqbds;
1217 /** # request buffers in history */
1218 int srv_n_history_rqbds;
1219 /** max # request buffers in history */
1220 int srv_max_history_rqbds;
1221 /** request history */
1222 cfs_list_t srv_request_history;
1223 /** next request sequence # */
1224 __u64 srv_request_seq;
1225 /** highest seq culled from history */
1226 __u64 srv_request_max_cull_seq;
1228 * all threads sleep on this. This wait-queue is signalled when new
1229 * incoming request arrives and when difficult reply has to be handled.
1231 cfs_waitq_t srv_waitq;
1234 * serialize the following fields, used for processing requests
1235 * sent to this portal
1237 cfs_spinlock_t srv_rq_lock __cfs_cacheline_aligned;
1238 /** # reqs in either of the queues below */
1239 /** reqs waiting for service */
1240 cfs_list_t srv_request_queue;
1241 /** high priority queue */
1242 cfs_list_t srv_request_hpq;
1243 /** # incoming reqs */
1244 int srv_n_queued_reqs;
1245 /** # reqs being served */
1246 int srv_n_active_reqs;
1247 /** # HPreqs being served */
1248 int srv_n_active_hpreq;
1249 /** # hp requests handled */
1250 int srv_hpreq_count;
1255 * serialize the following fields, used for changes on
1258 cfs_spinlock_t srv_at_lock __cfs_cacheline_aligned;
1259 /** estimated rpc service time */
1260 struct adaptive_timeout srv_at_estimate;
1261 /** reqs waiting for replies */
1262 struct ptlrpc_at_array srv_at_array;
1263 /** early reply timer */
1264 cfs_timer_t srv_at_timer;
1265 /** check early replies */
1266 unsigned srv_at_check;
1268 cfs_time_t srv_at_checktime;
1272 * serialize the following fields, used for processing
1273 * replies for this portal
1275 cfs_spinlock_t srv_rs_lock __cfs_cacheline_aligned;
1276 /** all the active replies */
1277 cfs_list_t srv_active_replies;
1279 /** replies waiting for service */
1280 cfs_list_t srv_reply_queue;
1282 /** List of free reply_states */
1283 cfs_list_t srv_free_rs_list;
1284 /** waitq to run, when adding stuff to srv_free_rs_list */
1285 cfs_waitq_t srv_free_rs_waitq;
1286 /** # 'difficult' replies */
1287 cfs_atomic_t srv_n_difficult_replies;
1288 //struct ptlrpc_srv_ni srv_interfaces[0];
1292 * Declaration of ptlrpcd control structure
1294 struct ptlrpcd_ctl {
1296 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1298 unsigned long pc_flags;
1300 * Thread lock protecting structure fields.
1302 cfs_spinlock_t pc_lock;
1306 cfs_completion_t pc_starting;
1310 cfs_completion_t pc_finishing;
1312 * Thread requests set.
1314 struct ptlrpc_request_set *pc_set;
1316 * Thread name used in cfs_daemonize()
1320 * Environment for request interpreters to run in.
1322 struct lu_env pc_env;
1324 * Index of ptlrpcd thread in the array.
1328 * Number of the ptlrpcd's partners.
1332 * Pointer to the array of partners' ptlrpcd_ctl structure.
1334 struct ptlrpcd_ctl **pc_partners;
1336 * Record the partner index to be processed next.
1341 * Async rpcs flag to make sure that ptlrpcd_check() is called only
1346 * Currently not used.
1350 * User-space async rpcs callback.
1352 void *pc_wait_callback;
1354 * User-space check idle rpcs callback.
1356 void *pc_idle_callback;
1360 /* Bits for pc_flags */
1361 enum ptlrpcd_ctl_flags {
1363 * Ptlrpc thread start flag.
1365 LIOD_START = 1 << 0,
1367 * Ptlrpc thread stop flag.
1371 * Ptlrpc thread force flag (only stop force so far).
1372 * This will cause aborting any inflight rpcs handled
1373 * by thread if LIOD_STOP is specified.
1375 LIOD_FORCE = 1 << 2,
1377 * This is a recovery ptlrpc thread.
1379 LIOD_RECOVERY = 1 << 3,
1381 * The ptlrpcd is bound to some CPU core.
1386 /* ptlrpc/events.c */
1387 extern lnet_handle_eq_t ptlrpc_eq_h;
1388 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1389 lnet_process_id_t *peer, lnet_nid_t *self);
1391 * These callbacks are invoked by LNet when something happened to
1395 extern void request_out_callback(lnet_event_t *ev);
1396 extern void reply_in_callback(lnet_event_t *ev);
1397 extern void client_bulk_callback(lnet_event_t *ev);
1398 extern void request_in_callback(lnet_event_t *ev);
1399 extern void reply_out_callback(lnet_event_t *ev);
1400 #ifdef HAVE_SERVER_SUPPORT
1401 extern void server_bulk_callback(lnet_event_t *ev);
1405 /* ptlrpc/connection.c */
1406 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
1408 struct obd_uuid *uuid);
1409 int ptlrpc_connection_put(struct ptlrpc_connection *c);
1410 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1411 int ptlrpc_connection_init(void);
1412 void ptlrpc_connection_fini(void);
1413 extern lnet_pid_t ptl_get_pid(void);
1415 /* ptlrpc/niobuf.c */
1417 * Actual interfacing with LNet to put/get/register/unregister stuff
1420 #ifdef HAVE_SERVER_SUPPORT
1421 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
1422 int npages, int type, int portal);
1423 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
1424 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
1426 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
1430 LASSERT(desc != NULL);
1432 cfs_spin_lock(&desc->bd_lock);
1433 rc = desc->bd_network_rw;
1434 cfs_spin_unlock(&desc->bd_lock);
1439 int ptlrpc_register_bulk(struct ptlrpc_request *req);
1440 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
1442 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
1444 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1447 LASSERT(req != NULL);
1449 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
1450 req->rq_bulk_deadline > cfs_time_current_sec())
1456 cfs_spin_lock(&desc->bd_lock);
1457 rc = desc->bd_network_rw;
1458 cfs_spin_unlock(&desc->bd_lock);
1462 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
1463 #define PTLRPC_REPLY_EARLY 0x02
1464 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
1465 int ptlrpc_reply(struct ptlrpc_request *req);
1466 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
1467 int ptlrpc_error(struct ptlrpc_request *req);
1468 void ptlrpc_resend_req(struct ptlrpc_request *request);
1469 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
1470 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
1471 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
1474 /* ptlrpc/client.c */
1476 * Client-side portals API. Everything to send requests, receive replies,
1477 * request queues, request management, etc.
1480 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
1481 struct ptlrpc_client *);
1482 void ptlrpc_cleanup_client(struct obd_import *imp);
1483 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
1485 int ptlrpc_queue_wait(struct ptlrpc_request *req);
1486 int ptlrpc_replay_req(struct ptlrpc_request *req);
1487 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
1488 void ptlrpc_restart_req(struct ptlrpc_request *req);
1489 void ptlrpc_abort_inflight(struct obd_import *imp);
1490 void ptlrpc_cleanup_imp(struct obd_import *imp);
1491 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
1493 struct ptlrpc_request_set *ptlrpc_prep_set(void);
1494 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1496 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1497 set_interpreter_func fn, void *data);
1498 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
1499 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
1500 int ptlrpc_set_wait(struct ptlrpc_request_set *);
1501 int ptlrpc_expired_set(void *data);
1502 void ptlrpc_interrupted_set(void *data);
1503 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
1504 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
1505 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
1506 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1507 struct ptlrpc_request *req);
1509 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
1510 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
1512 struct ptlrpc_request_pool *
1513 ptlrpc_init_rq_pool(int, int,
1514 void (*populate_pool)(struct ptlrpc_request_pool *, int));
1516 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
1517 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1518 const struct req_format *format);
1519 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
1520 struct ptlrpc_request_pool *,
1521 const struct req_format *format);
1522 void ptlrpc_request_free(struct ptlrpc_request *request);
1523 int ptlrpc_request_pack(struct ptlrpc_request *request,
1524 __u32 version, int opcode);
1525 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1526 const struct req_format *format,
1527 __u32 version, int opcode);
1528 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
1529 __u32 version, int opcode, char **bufs,
1530 struct ptlrpc_cli_ctx *ctx);
1531 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
1532 unsigned int timeout,
1533 ptlrpc_interpterer_t interpreter);
1534 void ptlrpc_fakereq_finished(struct ptlrpc_request *req);
1536 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
1537 int opcode, int count, __u32 *lengths,
1539 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
1540 __u32 version, int opcode,
1541 int count, __u32 *lengths, char **bufs,
1542 struct ptlrpc_request_pool *pool);
1543 void ptlrpc_req_finished(struct ptlrpc_request *request);
1544 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
1545 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
1546 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
1547 int npages, int type, int portal);
1548 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
1549 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
1550 cfs_page_t *page, int pageoffset, int len);
1551 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1552 struct obd_import *imp);
1553 __u64 ptlrpc_next_xid(void);
1554 __u64 ptlrpc_sample_next_xid(void);
1555 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
1557 /* Set of routines to run a function in ptlrpcd context */
1558 void *ptlrpcd_alloc_work(struct obd_import *imp,
1559 int (*cb)(const struct lu_env *, void *), void *data);
1560 void ptlrpcd_destroy_work(void *handler);
1561 int ptlrpcd_queue_work(void *handler);
1564 struct ptlrpc_service_buf_conf {
1565 /* nbufs is how many buffers to post */
1566 unsigned int bc_nbufs;
1567 /* buffer size to post */
1568 unsigned int bc_buf_size;
1569 /* portal to listed for requests on */
1570 unsigned int bc_req_portal;
1571 /* portal of where to send replies to */
1572 unsigned int bc_rep_portal;
1573 /* maximum request size to be accepted for this service */
1574 unsigned int bc_req_max_size;
1575 /* maximum reply size this service can ever send */
1576 unsigned int bc_rep_max_size;
1579 struct ptlrpc_service_thr_conf {
1580 /* threadname should be 8 characters or less - 6 will be added on */
1582 /* min number of service threads to start */
1583 unsigned int tc_nthrs_min;
1584 /* max number of service threads to start */
1585 unsigned int tc_nthrs_max;
1586 /* set NUMA node affinity for service threads */
1587 unsigned int tc_cpu_affinity;
1588 /* Tags for lu_context associated with service thread */
1592 struct ptlrpc_service_conf {
1595 /* soft watchdog timeout multiplifier to print stuck service traces */
1596 unsigned int psc_watchdog_factor;
1597 /* buffer information */
1598 struct ptlrpc_service_buf_conf psc_buf;
1599 /* thread information */
1600 struct ptlrpc_service_thr_conf psc_thr;
1601 /* function table */
1602 struct ptlrpc_service_ops psc_ops;
1605 /* ptlrpc/service.c */
1607 * Server-side services API. Register/unregister service, request state
1608 * management, service thread management
1612 void ptlrpc_save_lock(struct ptlrpc_request *req,
1613 struct lustre_handle *lock, int mode, int no_ack);
1614 void ptlrpc_commit_replies(struct obd_export *exp);
1615 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
1616 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
1617 struct ptlrpc_service *ptlrpc_register_service(
1618 struct ptlrpc_service_conf *conf,
1619 struct proc_dir_entry *proc_entry);
1620 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
1622 int ptlrpc_start_threads(struct ptlrpc_service *svc);
1623 int ptlrpc_start_thread(struct ptlrpc_service *svc);
1624 int ptlrpc_unregister_service(struct ptlrpc_service *service);
1625 int liblustre_check_services(void *arg);
1626 void ptlrpc_daemonize(char *name);
1627 int ptlrpc_service_health_check(struct ptlrpc_service *);
1628 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req);
1629 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
1632 int ptlrpc_hr_init(void);
1633 void ptlrpc_hr_fini(void);
1635 # define ptlrpc_hr_init() (0)
1636 # define ptlrpc_hr_fini() do {} while(0)
1639 struct ptlrpc_svc_data {
1641 struct ptlrpc_service *svc;
1642 struct ptlrpc_thread *thread;
1646 /* ptlrpc/import.c */
1651 int ptlrpc_connect_import(struct obd_import *imp);
1652 int ptlrpc_init_import(struct obd_import *imp);
1653 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
1654 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
1655 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
1658 /* ptlrpc/pack_generic.c */
1659 int ptlrpc_reconnect_import(struct obd_import *imp);
1663 * ptlrpc msg buffer and swab interface
1667 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
1669 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
1671 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
1672 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
1674 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
1675 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
1677 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
1678 __u32 *lens, char **bufs);
1679 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
1681 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
1682 __u32 *lens, char **bufs, int flags);
1683 #define LPRFL_EARLY_REPLY 1
1684 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
1685 char **bufs, int flags);
1686 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
1687 unsigned int newlen, int move_data);
1688 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
1689 int __lustre_unpack_msg(struct lustre_msg *m, int len);
1690 int lustre_msg_hdr_size(__u32 magic, int count);
1691 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
1692 int lustre_msg_size_v2(int count, __u32 *lengths);
1693 int lustre_packed_msg_size(struct lustre_msg *msg);
1694 int lustre_msg_early_size(void);
1695 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
1696 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
1697 int lustre_msg_buflen(struct lustre_msg *m, int n);
1698 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
1699 int lustre_msg_bufcount(struct lustre_msg *m);
1700 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
1701 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
1702 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
1703 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
1704 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
1705 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
1706 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
1707 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
1708 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
1709 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
1710 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
1711 __u32 lustre_msg_get_type(struct lustre_msg *msg);
1712 __u32 lustre_msg_get_version(struct lustre_msg *msg);
1713 void lustre_msg_add_version(struct lustre_msg *msg, int version);
1714 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
1715 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
1716 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
1717 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
1718 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
1719 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
1720 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
1721 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
1722 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
1723 int lustre_msg_get_status(struct lustre_msg *msg);
1724 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
1725 int lustre_msg_is_v1(struct lustre_msg *msg);
1726 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
1727 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
1728 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
1729 char *lustre_msg_get_jobid(struct lustre_msg *msg);
1730 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
1731 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 0, 0)
1732 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
1734 # warning "remove checksum compatibility support for b1_8"
1735 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
1737 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
1738 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
1739 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
1740 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
1741 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
1742 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
1743 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
1744 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
1745 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
1746 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
1747 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
1748 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
1749 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
1750 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
1751 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
1754 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
1755 unsigned int newlen, int move_data)
1757 LASSERT(req->rq_reply_state);
1758 LASSERT(req->rq_repmsg);
1759 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
1764 /** Change request phase of \a req to \a new_phase */
1766 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
1768 if (req->rq_phase == new_phase)
1771 if (new_phase == RQ_PHASE_UNREGISTERING) {
1772 req->rq_next_phase = req->rq_phase;
1774 cfs_atomic_inc(&req->rq_import->imp_unregistering);
1777 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1779 cfs_atomic_dec(&req->rq_import->imp_unregistering);
1782 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
1783 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
1785 req->rq_phase = new_phase;
1789 * Returns true if request \a req got early reply and hard deadline is not met
1792 ptlrpc_client_early(struct ptlrpc_request *req)
1794 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1795 req->rq_reply_deadline > cfs_time_current_sec())
1797 return req->rq_early;
1801 * Returns true if we got real reply from server for this request
1804 ptlrpc_client_replied(struct ptlrpc_request *req)
1806 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1807 req->rq_reply_deadline > cfs_time_current_sec())
1809 return req->rq_replied;
1812 /** Returns true if request \a req is in process of receiving server reply */
1814 ptlrpc_client_recv(struct ptlrpc_request *req)
1816 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1817 req->rq_reply_deadline > cfs_time_current_sec())
1819 return req->rq_receiving_reply;
1823 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
1827 cfs_spin_lock(&req->rq_lock);
1828 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1829 req->rq_reply_deadline > cfs_time_current_sec()) {
1830 cfs_spin_unlock(&req->rq_lock);
1833 rc = req->rq_receiving_reply || req->rq_must_unlink;
1834 cfs_spin_unlock(&req->rq_lock);
1839 ptlrpc_client_wake_req(struct ptlrpc_request *req)
1841 if (req->rq_set == NULL)
1842 cfs_waitq_signal(&req->rq_reply_waitq);
1844 cfs_waitq_signal(&req->rq_set->set_waitq);
1848 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
1850 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
1851 cfs_atomic_inc(&rs->rs_refcount);
1855 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
1857 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
1858 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
1859 lustre_free_reply_state(rs);
1862 /* Should only be called once per req */
1863 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
1865 if (req->rq_reply_state == NULL)
1866 return; /* shouldn't occur */
1867 ptlrpc_rs_decref(req->rq_reply_state);
1868 req->rq_reply_state = NULL;
1869 req->rq_repmsg = NULL;
1872 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
1874 return lustre_msg_get_magic(req->rq_reqmsg);
1877 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
1879 switch (req->rq_reqmsg->lm_magic) {
1880 case LUSTRE_MSG_MAGIC_V2:
1881 return req->rq_reqmsg->lm_repsize;
1883 LASSERTF(0, "incorrect message magic: %08x\n",
1884 req->rq_reqmsg->lm_magic);
1889 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
1891 if (req->rq_delay_limit != 0 &&
1892 cfs_time_before(cfs_time_add(req->rq_queued_time,
1893 cfs_time_seconds(req->rq_delay_limit)),
1894 cfs_time_current())) {
1900 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
1902 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
1903 cfs_spin_lock(&req->rq_lock);
1904 req->rq_no_resend = 1;
1905 cfs_spin_unlock(&req->rq_lock);
1907 return req->rq_no_resend;
1910 /* ldlm/ldlm_lib.c */
1912 * Target client logic
1915 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
1916 int client_obd_cleanup(struct obd_device *obddev);
1917 int client_connect_import(const struct lu_env *env,
1918 struct obd_export **exp, struct obd_device *obd,
1919 struct obd_uuid *cluuid, struct obd_connect_data *,
1921 int client_disconnect_export(struct obd_export *exp);
1922 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
1924 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
1925 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
1926 struct obd_uuid *uuid);
1927 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
1928 void client_destroy_import(struct obd_import *imp);
1931 #ifdef HAVE_SERVER_SUPPORT
1932 int server_disconnect_export(struct obd_export *exp);
1935 /* ptlrpc/pinger.c */
1937 * Pinger API (client side only)
1940 enum timeout_event {
1943 struct timeout_item;
1944 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
1945 int ptlrpc_pinger_add_import(struct obd_import *imp);
1946 int ptlrpc_pinger_del_import(struct obd_import *imp);
1947 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
1948 timeout_cb_t cb, void *data,
1949 cfs_list_t *obd_list);
1950 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
1951 enum timeout_event event);
1952 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
1953 int ptlrpc_obd_ping(struct obd_device *obd);
1954 cfs_time_t ptlrpc_suspend_wakeup_time(void);
1956 void ping_evictor_start(void);
1957 void ping_evictor_stop(void);
1959 #define ping_evictor_start() do {} while (0)
1960 #define ping_evictor_stop() do {} while (0)
1962 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
1965 /* ptlrpc daemon bind policy */
1967 /* all ptlrpcd threads are free mode */
1968 PDB_POLICY_NONE = 1,
1969 /* all ptlrpcd threads are bound mode */
1970 PDB_POLICY_FULL = 2,
1971 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
1972 PDB_POLICY_PAIR = 3,
1973 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
1974 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
1975 * If kernel supports NUMA, pthrpcd threads are binded and
1976 * grouped by NUMA node */
1977 PDB_POLICY_NEIGHBOR = 4,
1980 /* ptlrpc daemon load policy
1981 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
1982 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
1983 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
1984 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
1985 * depends on which is scheduled firstly, to accelerate the RPC processing. */
1987 /* on the same CPU core as the caller */
1988 PDL_POLICY_SAME = 1,
1989 /* within the same CPU partition, but not the same core as the caller */
1990 PDL_POLICY_LOCAL = 2,
1991 /* round-robin on all CPU cores, but not the same core as the caller */
1992 PDL_POLICY_ROUND = 3,
1993 /* the specified CPU core is preferred, but not enforced */
1994 PDL_POLICY_PREFERRED = 4,
1997 /* ptlrpc/ptlrpcd.c */
1998 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
1999 void ptlrpcd_wake(struct ptlrpc_request *req);
2000 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
2001 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2002 int ptlrpcd_addref(void);
2003 void ptlrpcd_decref(void);
2005 /* ptlrpc/lproc_ptlrpc.c */
2007 * procfs output related functions
2010 const char* ll_opcode2str(__u32 opcode);
2012 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2013 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2014 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2016 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2017 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2018 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2022 /* ptlrpc/llog_server.c */
2023 int llog_origin_handle_create(struct ptlrpc_request *req);
2024 int llog_origin_handle_destroy(struct ptlrpc_request *req);
2025 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2026 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2027 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2028 int llog_origin_handle_close(struct ptlrpc_request *req);
2029 int llog_origin_handle_cancel(struct ptlrpc_request *req);
2030 int llog_catinfo(struct ptlrpc_request *req);
2032 /* ptlrpc/llog_client.c */
2033 extern struct llog_operations llog_client_ops;