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__ */
109 #define PTLRPC_NTHRS_MIN 2
112 * The following constants determine how memory is used to buffer incoming
115 * ?_NBUFS # buffers to allocate when growing the pool
116 * ?_BUFSIZE # bytes in a single request buffer
117 * ?_MAXREQSIZE # maximum request service will receive
119 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
120 * of ?_NBUFS is added to the pool.
122 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
123 * considered full when less than ?_MAXREQSIZE is left in them.
125 #define LDLM_THREADS_AUTO_MIN (2)
126 #define LDLM_THREADS_AUTO_MAX min_t(unsigned, cfs_num_online_cpus() * \
127 cfs_num_online_cpus() * 32, 128)
128 #define LDLM_BL_THREADS LDLM_THREADS_AUTO_MIN
129 #define LDLM_NBUFS (64 * cfs_num_online_cpus())
130 #define LDLM_BUFSIZE (8 * 1024)
131 #define LDLM_MAXREQSIZE (5 * 1024)
132 #define LDLM_MAXREPSIZE (1024)
134 /** Absolute limits */
135 #ifndef MDT_MAX_THREADS
136 #define MDT_MIN_THREADS PTLRPC_NTHRS_MIN
137 #define MDT_MAX_THREADS 512UL
139 #define MDS_NBUFS (64 * cfs_num_online_cpus())
141 * Assume file name length = FNAME_MAX = 256 (true for ext3).
142 * path name length = PATH_MAX = 4096
143 * LOV MD size max = EA_MAX = 48000 (2000 stripes)
144 * symlink: FNAME_MAX + PATH_MAX <- largest
145 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
146 * rename: FNAME_MAX + FNAME_MAX
147 * open: FNAME_MAX + EA_MAX
149 * MDS_MAXREQSIZE ~= 4736 bytes =
150 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
151 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
152 * or, for mds_close() and mds_reint_unlink() on a many-OST filesystem:
153 * = 9210 bytes = lustre_msg + mdt_body + 160 * (easize + cookiesize)
155 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
156 * except in the open case where there are a large number of OSTs in a LOV.
158 #define MDS_MAXREPSIZE max(10 * 1024, 362 + LOV_MAX_STRIPE_COUNT * 56)
159 #define MDS_MAXREQSIZE MDS_MAXREPSIZE
161 /** MDS_BUFSIZE = max_reqsize + max sptlrpc payload size */
162 #define MDS_BUFSIZE (MDS_MAXREQSIZE + 1024)
164 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
165 #define FLD_MAXREQSIZE (160)
167 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
168 #define FLD_MAXREPSIZE (152)
171 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
173 #define SEQ_MAXREQSIZE (160)
175 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
176 #define SEQ_MAXREPSIZE (152)
178 /** MGS threads must be >= 3, see bug 22458 comment #28 */
179 #define MGS_THREADS_AUTO_MIN 3
180 #define MGS_THREADS_AUTO_MAX 32
181 #define MGS_NBUFS (64 * cfs_num_online_cpus())
182 #define MGS_BUFSIZE (8 * 1024)
183 #define MGS_MAXREQSIZE (7 * 1024)
184 #define MGS_MAXREPSIZE (9 * 1024)
186 /** Absolute OSS limits */
187 #define OSS_THREADS_MIN 3 /* difficult replies, HPQ, others */
188 #define OSS_THREADS_MAX 512
189 #define OST_NBUFS (64 * cfs_num_online_cpus())
190 #define OST_BUFSIZE (8 * 1024)
193 * OST_MAXREQSIZE ~= 4768 bytes =
194 * lustre_msg + obdo + 16 * obd_ioobj + 256 * niobuf_remote
196 * - single object with 16 pages is 512 bytes
197 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
199 #define OST_MAXREQSIZE (5 * 1024)
200 #define OST_MAXREPSIZE (9 * 1024)
202 /* Macro to hide a typecast. */
203 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
206 * Structure to single define portal connection.
208 struct ptlrpc_connection {
209 /** linkage for connections hash table */
210 cfs_hlist_node_t c_hash;
211 /** Our own lnet nid for this connection */
213 /** Remote side nid for this connection */
214 lnet_process_id_t c_peer;
215 /** UUID of the other side */
216 struct obd_uuid c_remote_uuid;
217 /** reference counter for this connection */
218 cfs_atomic_t c_refcount;
221 /** Client definition for PortalRPC */
222 struct ptlrpc_client {
223 /** What lnet portal does this client send messages to by default */
224 __u32 cli_request_portal;
225 /** What portal do we expect replies on */
226 __u32 cli_reply_portal;
227 /** Name of the client */
231 /** state flags of requests */
232 /* XXX only ones left are those used by the bulk descs as well! */
233 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
234 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
236 #define REQ_MAX_ACK_LOCKS 8
238 union ptlrpc_async_args {
240 * Scratchpad for passing args to completion interpreter. Users
241 * cast to the struct of their choosing, and CLASSERT that this is
242 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
243 * a pointer to it here. The pointer_arg ensures this struct is at
244 * least big enough for that.
246 void *pointer_arg[11];
250 struct ptlrpc_request_set;
251 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
252 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
255 * Definition of request set structure.
256 * Request set is a list of requests (not necessary to the same target) that
257 * once populated with RPCs could be sent in parallel.
258 * There are two kinds of request sets. General purpose and with dedicated
259 * serving thread. Example of the latter is ptlrpcd set.
260 * For general purpose sets once request set started sending it is impossible
261 * to add new requests to such set.
262 * Provides a way to call "completion callbacks" when all requests in the set
265 struct ptlrpc_request_set {
266 cfs_atomic_t set_refcount;
267 /** number of in queue requests */
268 cfs_atomic_t set_new_count;
269 /** number of uncompleted requests */
270 cfs_atomic_t set_remaining;
271 /** wait queue to wait on for request events */
272 cfs_waitq_t set_waitq;
273 cfs_waitq_t *set_wakeup_ptr;
274 /** List of requests in the set */
275 cfs_list_t set_requests;
277 * List of completion callbacks to be called when the set is completed
278 * This is only used if \a set_interpret is NULL.
279 * Links struct ptlrpc_set_cbdata.
281 cfs_list_t set_cblist;
282 /** Completion callback, if only one. */
283 set_interpreter_func set_interpret;
284 /** opaq argument passed to completion \a set_interpret callback. */
286 /** rq_status of requests that have been freed already */
289 * Lock for \a set_new_requests manipulations
290 * locked so that any old caller can communicate requests to
291 * the set holder who can then fold them into the lock-free set
293 cfs_spinlock_t set_new_req_lock;
294 /** List of new yet unsent requests. Only used with ptlrpcd now. */
295 cfs_list_t set_new_requests;
297 /** Additional fields used by the flow control extension */
298 /** Maximum number of RPCs in flight */
299 int set_max_inflight;
300 /** Callback function used to generate RPCs */
301 set_producer_func set_producer;
302 /** opaq argument passed to the producer callback */
303 void *set_producer_arg;
307 * Description of a single ptrlrpc_set callback
309 struct ptlrpc_set_cbdata {
310 /** List linkage item */
312 /** Pointer to interpreting function */
313 set_interpreter_func psc_interpret;
314 /** Opaq argument to pass to the callback */
318 struct ptlrpc_bulk_desc;
321 * ptlrpc callback & work item stuff
323 struct ptlrpc_cb_id {
324 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
325 void *cbid_arg; /* additional arg */
328 /** Maximum number of locks to fit into reply state */
329 #define RS_MAX_LOCKS 8
333 * Structure to define reply state on the server
334 * Reply state holds various reply message information. Also for "difficult"
335 * replies (rep-ack case) we store the state after sending reply and wait
336 * for the client to acknowledge the reception. In these cases locks could be
337 * added to the state for replay/failover consistency guarantees.
339 struct ptlrpc_reply_state {
340 /** Callback description */
341 struct ptlrpc_cb_id rs_cb_id;
342 /** Linkage for list of all reply states in a system */
344 /** Linkage for list of all reply states on same export */
345 cfs_list_t rs_exp_list;
346 /** Linkage for list of all reply states for same obd */
347 cfs_list_t rs_obd_list;
349 cfs_list_t rs_debug_list;
351 /** A spinlock to protect the reply state flags */
352 cfs_spinlock_t rs_lock;
353 /** Reply state flags */
354 unsigned long rs_difficult:1; /* ACK/commit stuff */
355 unsigned long rs_no_ack:1; /* no ACK, even for
356 difficult requests */
357 unsigned long rs_scheduled:1; /* being handled? */
358 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
359 unsigned long rs_handled:1; /* been handled yet? */
360 unsigned long rs_on_net:1; /* reply_out_callback pending? */
361 unsigned long rs_prealloc:1; /* rs from prealloc list */
362 unsigned long rs_committed:1;/* the transaction was committed
363 and the rs was dispatched
364 by ptlrpc_commit_replies */
365 /** Size of the state */
369 /** Transaction number */
373 struct obd_export *rs_export;
374 struct ptlrpc_service *rs_service;
375 /** Lnet metadata handle for the reply */
376 lnet_handle_md_t rs_md_h;
377 cfs_atomic_t rs_refcount;
379 /** Context for the sevice thread */
380 struct ptlrpc_svc_ctx *rs_svc_ctx;
381 /** Reply buffer (actually sent to the client), encoded if needed */
382 struct lustre_msg *rs_repbuf; /* wrapper */
383 /** Size of the reply buffer */
384 int rs_repbuf_len; /* wrapper buf length */
385 /** Size of the reply message */
386 int rs_repdata_len; /* wrapper msg length */
388 * Actual reply message. Its content is encrupted (if needed) to
389 * produce reply buffer for actual sending. In simple case
390 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
392 struct lustre_msg *rs_msg; /* reply message */
394 /** Number of locks awaiting client ACK */
396 /** Handles of locks awaiting client reply ACK */
397 struct lustre_handle rs_locks[RS_MAX_LOCKS];
398 /** Lock modes of locks in \a rs_locks */
399 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
402 struct ptlrpc_thread;
406 RQ_PHASE_NEW = 0xebc0de00,
407 RQ_PHASE_RPC = 0xebc0de01,
408 RQ_PHASE_BULK = 0xebc0de02,
409 RQ_PHASE_INTERPRET = 0xebc0de03,
410 RQ_PHASE_COMPLETE = 0xebc0de04,
411 RQ_PHASE_UNREGISTERING = 0xebc0de05,
412 RQ_PHASE_UNDEFINED = 0xebc0de06
415 /** Type of request interpreter call-back */
416 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
417 struct ptlrpc_request *req,
421 * Definition of request pool structure.
422 * The pool is used to store empty preallocated requests for the case
423 * when we would actually need to send something without performing
424 * any allocations (to avoid e.g. OOM).
426 struct ptlrpc_request_pool {
427 /** Locks the list */
428 cfs_spinlock_t prp_lock;
429 /** list of ptlrpc_request structs */
430 cfs_list_t prp_req_list;
431 /** Maximum message size that would fit into a rquest from this pool */
433 /** Function to allocate more requests for this pool */
434 void (*prp_populate)(struct ptlrpc_request_pool *, int);
443 * Basic request prioritization operations structure.
444 * The whole idea is centered around locks and RPCs that might affect locks.
445 * When a lock is contended we try to give priority to RPCs that might lead
446 * to fastest release of that lock.
447 * Currently only implemented for OSTs only in a way that makes all
448 * IO and truncate RPCs that are coming from a locked region where a lock is
449 * contended a priority over other requests.
451 struct ptlrpc_hpreq_ops {
453 * Check if the lock handle of the given lock is the same as
454 * taken from the request.
456 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
458 * Check if the request is a high priority one.
460 int (*hpreq_check)(struct ptlrpc_request *);
462 * Called after the request has been handled.
464 void (*hpreq_fini)(struct ptlrpc_request *);
468 * Represents remote procedure call.
470 * This is a staple structure used by everybody wanting to send a request
473 struct ptlrpc_request {
474 /* Request type: one of PTL_RPC_MSG_* */
477 * Linkage item through which this request is included into
478 * sending/delayed lists on client and into rqbd list on server
482 * Server side list of incoming unserved requests sorted by arrival
483 * time. Traversed from time to time to notice about to expire
484 * requests and sent back "early replies" to clients to let them
485 * know server is alive and well, just very busy to service their
488 cfs_list_t rq_timed_list;
489 /** server-side history, used for debuging purposes. */
490 cfs_list_t rq_history_list;
491 /** server-side per-export list */
492 cfs_list_t rq_exp_list;
493 /** server-side hp handlers */
494 struct ptlrpc_hpreq_ops *rq_ops;
495 /** history sequence # */
496 __u64 rq_history_seq;
497 /** the index of service's srv_at_array into which request is linked */
499 /** Result of request processing */
501 /** Lock to protect request flags and some other important bits, like
504 cfs_spinlock_t rq_lock;
505 /** client-side flags are serialized by rq_lock */
506 unsigned long rq_intr:1, rq_replied:1, rq_err:1,
507 rq_timedout:1, rq_resend:1, rq_restart:1,
509 * when ->rq_replay is set, request is kept by the client even
510 * after server commits corresponding transaction. This is
511 * used for operations that require sequence of multiple
512 * requests to be replayed. The only example currently is file
513 * open/close. When last request in such a sequence is
514 * committed, ->rq_replay is cleared on all requests in the
518 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
519 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
520 rq_early:1, rq_must_unlink:1,
521 rq_fake:1, /* this fake req */
522 rq_memalloc:1, /* req originated from "kswapd" */
523 /* server-side flags */
524 rq_packed_final:1, /* packed final reply */
525 rq_hp:1, /* high priority RPC */
526 rq_at_linked:1, /* link into service's srv_at_array */
529 /* whether the "rq_set" is a valid one */
533 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
534 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
535 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
536 server-side refcounf for multiple replies */
538 /** initial thread servicing this request */
539 struct ptlrpc_thread *rq_svc_thread;
541 /** Portal to which this request would be sent */
542 int rq_request_portal; /* XXX FIXME bug 249 */
543 /** Portal where to wait for reply and where reply would be sent */
544 int rq_reply_portal; /* XXX FIXME bug 249 */
548 * !rq_truncate : # reply bytes actually received,
549 * rq_truncate : required repbuf_len for resend
552 /** Request length */
554 /** Request message - what client sent */
555 struct lustre_msg *rq_reqmsg;
559 /** Reply message - server response */
560 struct lustre_msg *rq_repmsg;
561 /** Transaction number */
566 * List item to for replay list. Not yet commited requests get linked
568 * Also see \a rq_replay comment above.
570 cfs_list_t rq_replay_list;
573 * security and encryption data
575 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
576 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
577 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
579 struct sptlrpc_flavor rq_flvr; /**< for client & server */
580 enum lustre_sec_part rq_sp_from;
582 unsigned long /* client/server security flags */
583 rq_ctx_init:1, /* context initiation */
584 rq_ctx_fini:1, /* context destroy */
585 rq_bulk_read:1, /* request bulk read */
586 rq_bulk_write:1, /* request bulk write */
587 /* server authentication flags */
588 rq_auth_gss:1, /* authenticated by gss */
589 rq_auth_remote:1, /* authed as remote user */
590 rq_auth_usr_root:1, /* authed as root */
591 rq_auth_usr_mdt:1, /* authed as mdt */
592 rq_auth_usr_ost:1, /* authed as ost */
593 /* security tfm flags */
596 /* doesn't expect reply FIXME */
598 rq_pill_init:1; /* pill initialized */
600 uid_t rq_auth_uid; /* authed uid */
601 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
603 /* (server side), pointed directly into req buffer */
604 struct ptlrpc_user_desc *rq_user_desc;
606 /** early replies go to offset 0, regular replies go after that */
607 unsigned int rq_reply_off;
609 /* various buffer pointers */
610 struct lustre_msg *rq_reqbuf; /* req wrapper */
611 int rq_reqbuf_len; /* req wrapper buf len */
612 int rq_reqdata_len; /* req wrapper msg len */
613 char *rq_repbuf; /* rep buffer */
614 int rq_repbuf_len; /* rep buffer len */
615 struct lustre_msg *rq_repdata; /* rep wrapper msg */
616 int rq_repdata_len; /* rep wrapper msg len */
617 struct lustre_msg *rq_clrbuf; /* only in priv mode */
618 int rq_clrbuf_len; /* only in priv mode */
619 int rq_clrdata_len; /* only in priv mode */
623 /** Fields that help to see if request and reply were swabbed or not */
624 __u32 rq_req_swab_mask;
625 __u32 rq_rep_swab_mask;
627 /** What was import generation when this request was sent */
628 int rq_import_generation;
629 enum lustre_imp_state rq_send_state;
631 /** how many early replies (for stats) */
634 /** client+server request */
635 lnet_handle_md_t rq_req_md_h;
636 struct ptlrpc_cb_id rq_req_cbid;
637 /** optional time limit for send attempts */
638 cfs_duration_t rq_delay_limit;
639 /** time request was first queued */
640 cfs_time_t rq_queued_time;
643 /** request arrival time */
644 struct timeval rq_arrival_time;
645 /** separated reply state */
646 struct ptlrpc_reply_state *rq_reply_state;
647 /** incoming request buffer */
648 struct ptlrpc_request_buffer_desc *rq_rqbd;
650 __u32 rq_uid; /* peer uid, used in MDS only */
653 /** client-only incoming reply */
654 lnet_handle_md_t rq_reply_md_h;
655 cfs_waitq_t rq_reply_waitq;
656 struct ptlrpc_cb_id rq_reply_cbid;
660 /** Peer description (the other side) */
661 lnet_process_id_t rq_peer;
662 /** Server-side, export on which request was received */
663 struct obd_export *rq_export;
664 /** Client side, import where request is being sent */
665 struct obd_import *rq_import;
667 /** Replay callback, called after request is replayed at recovery */
668 void (*rq_replay_cb)(struct ptlrpc_request *);
670 * Commit callback, called when request is committed and about to be
673 void (*rq_commit_cb)(struct ptlrpc_request *);
674 /** Opaq data for replay and commit callbacks. */
677 /** For bulk requests on client only: bulk descriptor */
678 struct ptlrpc_bulk_desc *rq_bulk;
680 /** client outgoing req */
682 * when request/reply sent (secs), or time when request should be sent
685 /** time for request really sent out */
688 /** when request must finish. volatile
689 * so that servers' early reply updates to the deadline aren't
690 * kept in per-cpu cache */
691 volatile time_t rq_deadline;
692 /** when req reply unlink must finish. */
693 time_t rq_reply_deadline;
694 /** when req bulk unlink must finish. */
695 time_t rq_bulk_deadline;
697 * service time estimate (secs)
698 * If the requestsis not served by this time, it is marked as timed out.
702 /** Multi-rpc bits */
703 /** Link item for request set lists */
704 cfs_list_t rq_set_chain;
705 /** Per-request waitq introduced by bug 21938 for recovery waiting */
706 cfs_waitq_t rq_set_waitq;
707 /** Link back to the request set */
708 struct ptlrpc_request_set *rq_set;
709 /** Async completion handler, called when reply is received */
710 ptlrpc_interpterer_t rq_interpret_reply;
711 /** Async completion context */
712 union ptlrpc_async_args rq_async_args;
714 /** Pool if request is from preallocated list */
715 struct ptlrpc_request_pool *rq_pool;
717 struct lu_context rq_session;
718 struct lu_context rq_recov_session;
720 /** request format description */
721 struct req_capsule rq_pill;
725 * Call completion handler for rpc if any, return it's status or original
726 * rc if there was no handler defined for this request.
728 static inline int ptlrpc_req_interpret(const struct lu_env *env,
729 struct ptlrpc_request *req, int rc)
731 if (req->rq_interpret_reply != NULL) {
732 req->rq_status = req->rq_interpret_reply(env, req,
735 return req->rq_status;
741 * Returns 1 if request buffer at offset \a index was already swabbed
743 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
745 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
746 return req->rq_req_swab_mask & (1 << index);
750 * Returns 1 if request reply buffer at offset \a index was already swabbed
752 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
754 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
755 return req->rq_rep_swab_mask & (1 << index);
759 * Returns 1 if request needs to be swabbed into local cpu byteorder
761 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
763 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
767 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
769 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
771 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
775 * Mark request buffer at offset \a index that it was already swabbed
777 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
779 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
780 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
781 req->rq_req_swab_mask |= 1 << index;
785 * Mark request reply buffer at offset \a index that it was already swabbed
787 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
789 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
790 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
791 req->rq_rep_swab_mask |= 1 << index;
795 * Convert numerical request phase value \a phase into text string description
797 static inline const char *
798 ptlrpc_phase2str(enum rq_phase phase)
807 case RQ_PHASE_INTERPRET:
809 case RQ_PHASE_COMPLETE:
811 case RQ_PHASE_UNREGISTERING:
812 return "Unregistering";
819 * Convert numerical request phase of the request \a req into text stringi
822 static inline const char *
823 ptlrpc_rqphase2str(struct ptlrpc_request *req)
825 return ptlrpc_phase2str(req->rq_phase);
829 * Debugging functions and helpers to print request structure into debug log
832 /* Spare the preprocessor, spoil the bugs. */
833 #define FLAG(field, str) (field ? str : "")
835 /** Convert bit flags into a string */
836 #define DEBUG_REQ_FLAGS(req) \
837 ptlrpc_rqphase2str(req), \
838 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
839 FLAG(req->rq_err, "E"), \
840 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
841 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
842 FLAG(req->rq_no_resend, "N"), \
843 FLAG(req->rq_waiting, "W"), \
844 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
845 FLAG(req->rq_committed, "M")
847 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
849 void _debug_req(struct ptlrpc_request *req,
850 struct libcfs_debug_msg_data *data, const char *fmt, ...)
851 __attribute__ ((format (printf, 3, 4)));
854 * Helper that decides if we need to print request accordig to current debug
857 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
859 CFS_CHECK_STACK(msgdata, mask, cdls); \
861 if (((mask) & D_CANTMASK) != 0 || \
862 ((libcfs_debug & (mask)) != 0 && \
863 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
864 _debug_req((req), msgdata, fmt, ##a); \
868 * This is the debug print function you need to use to print request sturucture
869 * content into lustre debug log.
870 * for most callers (level is a constant) this is resolved at compile time */
871 #define DEBUG_REQ(level, req, fmt, args...) \
873 if ((level) & (D_ERROR | D_WARNING)) { \
874 static cfs_debug_limit_state_t cdls; \
875 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
876 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
878 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
879 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
885 * Structure that defines a single page of a bulk transfer
887 struct ptlrpc_bulk_page {
888 /** Linkage to list of pages in a bulk */
891 * Number of bytes in a page to transfer starting from \a bp_pageoffset
894 /** offset within a page */
896 /** The page itself */
897 struct page *bp_page;
900 #define BULK_GET_SOURCE 0
901 #define BULK_PUT_SINK 1
902 #define BULK_GET_SINK 2
903 #define BULK_PUT_SOURCE 3
906 * Definition of buk descriptor.
907 * Bulks are special "Two phase" RPCs where initial request message
908 * is sent first and it is followed bt a transfer (o receiving) of a large
909 * amount of data to be settled into pages referenced from the bulk descriptors.
910 * Bulks transfers (the actual data following the small requests) are done
911 * on separate LNet portals.
912 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
913 * Another user is readpage for MDT.
915 struct ptlrpc_bulk_desc {
916 /** completed successfully */
917 unsigned long bd_success:1;
918 /** accessible to the network (network io potentially in progress) */
919 unsigned long bd_network_rw:1;
920 /** {put,get}{source,sink} */
921 unsigned long bd_type:2;
923 unsigned long bd_registered:1;
924 /** For serialization with callback */
925 cfs_spinlock_t bd_lock;
926 /** Import generation when request for this bulk was sent */
927 int bd_import_generation;
928 /** Server side - export this bulk created for */
929 struct obd_export *bd_export;
930 /** Client side - import this bulk was sent on */
931 struct obd_import *bd_import;
932 /** LNet portal for this bulk */
934 /** Back pointer to the request */
935 struct ptlrpc_request *bd_req;
936 cfs_waitq_t bd_waitq; /* server side only WQ */
937 int bd_iov_count; /* # entries in bd_iov */
938 int bd_max_iov; /* allocated size of bd_iov */
939 int bd_nob; /* # bytes covered */
940 int bd_nob_transferred; /* # bytes GOT/PUT */
944 struct ptlrpc_cb_id bd_cbid; /* network callback info */
945 lnet_handle_md_t bd_md_h; /* associated MD */
946 lnet_nid_t bd_sender; /* stash event::sender */
948 #if defined(__KERNEL__)
950 * encrypt iov, size is either 0 or bd_iov_count.
952 lnet_kiov_t *bd_enc_iov;
954 lnet_kiov_t bd_iov[0];
956 lnet_md_iovec_t bd_iov[0];
961 SVC_STOPPED = 1 << 0,
962 SVC_STOPPING = 1 << 1,
963 SVC_STARTING = 1 << 2,
964 SVC_RUNNING = 1 << 3,
970 * Definition of server service thread structure
972 struct ptlrpc_thread {
974 * List of active threads in svc->srv_threads
978 * thread-private data (preallocated memory)
983 * service thread index, from ptlrpc_start_threads
991 * put watchdog in the structure per thread b=14840
993 struct lc_watchdog *t_watchdog;
995 * the svc this thread belonged to b=18582
997 struct ptlrpc_service *t_svc;
998 cfs_waitq_t t_ctl_waitq;
999 struct lu_env *t_env;
1002 static inline int thread_is_init(struct ptlrpc_thread *thread)
1004 return thread->t_flags == 0;
1007 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1009 return !!(thread->t_flags & SVC_STOPPED);
1012 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1014 return !!(thread->t_flags & SVC_STOPPING);
1017 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1019 return !!(thread->t_flags & SVC_STARTING);
1022 static inline int thread_is_running(struct ptlrpc_thread *thread)
1024 return !!(thread->t_flags & SVC_RUNNING);
1027 static inline int thread_is_event(struct ptlrpc_thread *thread)
1029 return !!(thread->t_flags & SVC_EVENT);
1032 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1034 return !!(thread->t_flags & SVC_SIGNAL);
1037 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1039 thread->t_flags &= ~flags;
1042 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1044 thread->t_flags = flags;
1047 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1049 thread->t_flags |= flags;
1052 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1055 if (thread->t_flags & flags) {
1056 thread->t_flags &= ~flags;
1063 * Request buffer descriptor structure.
1064 * This is a structure that contains one posted request buffer for service.
1065 * Once data land into a buffer, event callback creates actual request and
1066 * notifies wakes one of the service threads to process new incoming request.
1067 * More than one request can fit into the buffer.
1069 struct ptlrpc_request_buffer_desc {
1070 /** Link item for rqbds on a service */
1071 cfs_list_t rqbd_list;
1072 /** History of requests for this buffer */
1073 cfs_list_t rqbd_reqs;
1074 /** Back pointer to service for which this buffer is registered */
1075 struct ptlrpc_service *rqbd_service;
1076 /** LNet descriptor */
1077 lnet_handle_md_t rqbd_md_h;
1079 /** The buffer itself */
1081 struct ptlrpc_cb_id rqbd_cbid;
1083 * This "embedded" request structure is only used for the
1084 * last request to fit into the buffer
1086 struct ptlrpc_request rqbd_req;
1089 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1091 struct ptlrpc_service_ops {
1093 * if non-NULL called during thread creation (ptlrpc_start_thread())
1094 * to initialize service specific per-thread state.
1096 int (*so_thr_init)(struct ptlrpc_thread *thr);
1098 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1099 * destruct state created by ->srv_init().
1101 void (*so_thr_done)(struct ptlrpc_thread *thr);
1103 * Handler function for incoming requests for this service
1105 int (*so_req_handler)(struct ptlrpc_request *req);
1107 * function to determine priority of the request, it's called
1108 * on every new request
1110 int (*so_hpreq_handler)(struct ptlrpc_request *);
1112 * service-specific print fn
1114 void (*so_req_printer)(void *, struct ptlrpc_request *);
1117 #ifndef __cfs_cacheline_aligned
1118 /* NB: put it here for reducing patche dependence */
1119 # define __cfs_cacheline_aligned
1123 * How many high priority requests to serve before serving one normal
1126 #define PTLRPC_SVC_HP_RATIO 10
1129 * Definition of PortalRPC service.
1130 * The service is listening on a particular portal (like tcp port)
1131 * and perform actions for a specific server like IO service for OST
1132 * or general metadata service for MDS.
1134 * ptlrpc service has four locks:
1136 * serialize operations on rqbd and requests waiting for preprocess
1138 * serialize operations active requests sent to this portal
1140 * serialize adaptive timeout stuff
1142 * serialize operations on RS list (reply states)
1144 * We don't have any use-case to take two or more locks at the same time
1145 * for now, so there is no lock order issue.
1147 struct ptlrpc_service {
1148 /** most often accessed fields */
1149 /** chain thru all services */
1150 cfs_list_t srv_list;
1151 /** service operations table */
1152 struct ptlrpc_service_ops srv_ops;
1153 /** only statically allocated strings here; we don't clean them */
1155 /** only statically allocated strings here; we don't clean them */
1156 char *srv_thread_name;
1157 /** service thread list */
1158 cfs_list_t srv_threads;
1159 /** threads to start at beginning of service */
1160 int srv_threads_min;
1161 /** thread upper limit */
1162 int srv_threads_max;
1163 /** always increasing number */
1164 unsigned srv_threads_next_id;
1165 /** # of starting threads */
1166 int srv_threads_starting;
1167 /** # running threads */
1168 int srv_threads_running;
1170 /** Root of /proc dir tree for this service */
1171 cfs_proc_dir_entry_t *srv_procroot;
1172 /** Pointer to statistic data for this service */
1173 struct lprocfs_stats *srv_stats;
1174 /** # hp per lp reqs to handle */
1175 int srv_hpreq_ratio;
1176 /** biggest request to receive */
1177 int srv_max_req_size;
1178 /** biggest reply to send */
1179 int srv_max_reply_size;
1180 /** size of individual buffers */
1182 /** # buffers to allocate in 1 group */
1183 int srv_nbuf_per_group;
1184 /** Local portal on which to receive requests */
1185 __u32 srv_req_portal;
1186 /** Portal on the client to send replies to */
1187 __u32 srv_rep_portal;
1189 * Tags for lu_context associated with this thread, see struct
1193 /** soft watchdog timeout multiplier */
1194 int srv_watchdog_factor;
1195 /** bind threads to CPUs */
1196 unsigned srv_cpu_affinity:1;
1197 /** under unregister_service */
1198 unsigned srv_is_stopping:1;
1201 * serialize the following fields, used for protecting
1202 * rqbd list and incoming requests waiting for preprocess
1204 cfs_spinlock_t srv_lock __cfs_cacheline_aligned;
1205 /** incoming reqs */
1206 cfs_list_t srv_req_in_queue;
1207 /** total # req buffer descs allocated */
1209 /** # posted request buffers */
1210 int srv_nrqbd_receiving;
1211 /** timeout before re-posting reqs, in tick */
1212 cfs_duration_t srv_rqbd_timeout;
1213 /** request buffers to be reposted */
1214 cfs_list_t srv_idle_rqbds;
1215 /** req buffers receiving */
1216 cfs_list_t srv_active_rqbds;
1217 /** request buffer history */
1218 cfs_list_t srv_history_rqbds;
1219 /** # request buffers in history */
1220 int srv_n_history_rqbds;
1221 /** max # request buffers in history */
1222 int srv_max_history_rqbds;
1223 /** request history */
1224 cfs_list_t srv_request_history;
1225 /** next request sequence # */
1226 __u64 srv_request_seq;
1227 /** highest seq culled from history */
1228 __u64 srv_request_max_cull_seq;
1230 * all threads sleep on this. This wait-queue is signalled when new
1231 * incoming request arrives and when difficult reply has to be handled.
1233 cfs_waitq_t srv_waitq;
1236 * serialize the following fields, used for processing requests
1237 * sent to this portal
1239 cfs_spinlock_t srv_rq_lock __cfs_cacheline_aligned;
1240 /** # reqs in either of the queues below */
1241 /** reqs waiting for service */
1242 cfs_list_t srv_request_queue;
1243 /** high priority queue */
1244 cfs_list_t srv_request_hpq;
1245 /** # incoming reqs */
1246 int srv_n_queued_reqs;
1247 /** # reqs being served */
1248 int srv_n_active_reqs;
1249 /** # HPreqs being served */
1250 int srv_n_active_hpreq;
1251 /** # hp requests handled */
1252 int srv_hpreq_count;
1257 * serialize the following fields, used for changes on
1260 cfs_spinlock_t srv_at_lock __cfs_cacheline_aligned;
1261 /** estimated rpc service time */
1262 struct adaptive_timeout srv_at_estimate;
1263 /** reqs waiting for replies */
1264 struct ptlrpc_at_array srv_at_array;
1265 /** early reply timer */
1266 cfs_timer_t srv_at_timer;
1267 /** check early replies */
1268 unsigned srv_at_check;
1270 cfs_time_t srv_at_checktime;
1274 * serialize the following fields, used for processing
1275 * replies for this portal
1277 cfs_spinlock_t srv_rs_lock __cfs_cacheline_aligned;
1278 /** all the active replies */
1279 cfs_list_t srv_active_replies;
1281 /** replies waiting for service */
1282 cfs_list_t srv_reply_queue;
1284 /** List of free reply_states */
1285 cfs_list_t srv_free_rs_list;
1286 /** waitq to run, when adding stuff to srv_free_rs_list */
1287 cfs_waitq_t srv_free_rs_waitq;
1288 /** # 'difficult' replies */
1289 cfs_atomic_t srv_n_difficult_replies;
1290 //struct ptlrpc_srv_ni srv_interfaces[0];
1294 * Declaration of ptlrpcd control structure
1296 struct ptlrpcd_ctl {
1298 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1300 unsigned long pc_flags;
1302 * Thread lock protecting structure fields.
1304 cfs_spinlock_t pc_lock;
1308 cfs_completion_t pc_starting;
1312 cfs_completion_t pc_finishing;
1314 * Thread requests set.
1316 struct ptlrpc_request_set *pc_set;
1318 * Thread name used in cfs_daemonize()
1322 * Environment for request interpreters to run in.
1324 struct lu_env pc_env;
1326 * Index of ptlrpcd thread in the array.
1330 * Number of the ptlrpcd's partners.
1334 * Pointer to the array of partners' ptlrpcd_ctl structure.
1336 struct ptlrpcd_ctl **pc_partners;
1338 * Record the partner index to be processed next.
1343 * Async rpcs flag to make sure that ptlrpcd_check() is called only
1348 * Currently not used.
1352 * User-space async rpcs callback.
1354 void *pc_wait_callback;
1356 * User-space check idle rpcs callback.
1358 void *pc_idle_callback;
1362 /* Bits for pc_flags */
1363 enum ptlrpcd_ctl_flags {
1365 * Ptlrpc thread start flag.
1367 LIOD_START = 1 << 0,
1369 * Ptlrpc thread stop flag.
1373 * Ptlrpc thread force flag (only stop force so far).
1374 * This will cause aborting any inflight rpcs handled
1375 * by thread if LIOD_STOP is specified.
1377 LIOD_FORCE = 1 << 2,
1379 * This is a recovery ptlrpc thread.
1381 LIOD_RECOVERY = 1 << 3,
1383 * The ptlrpcd is bound to some CPU core.
1388 /* ptlrpc/events.c */
1389 extern lnet_handle_eq_t ptlrpc_eq_h;
1390 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1391 lnet_process_id_t *peer, lnet_nid_t *self);
1393 * These callbacks are invoked by LNet when something happened to
1397 extern void request_out_callback(lnet_event_t *ev);
1398 extern void reply_in_callback(lnet_event_t *ev);
1399 extern void client_bulk_callback(lnet_event_t *ev);
1400 extern void request_in_callback(lnet_event_t *ev);
1401 extern void reply_out_callback(lnet_event_t *ev);
1402 #ifdef HAVE_SERVER_SUPPORT
1403 extern void server_bulk_callback(lnet_event_t *ev);
1407 /* ptlrpc/connection.c */
1408 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
1410 struct obd_uuid *uuid);
1411 int ptlrpc_connection_put(struct ptlrpc_connection *c);
1412 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1413 int ptlrpc_connection_init(void);
1414 void ptlrpc_connection_fini(void);
1415 extern lnet_pid_t ptl_get_pid(void);
1417 /* ptlrpc/niobuf.c */
1419 * Actual interfacing with LNet to put/get/register/unregister stuff
1422 #ifdef HAVE_SERVER_SUPPORT
1423 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
1424 int npages, int type, int portal);
1425 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
1426 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
1428 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
1432 LASSERT(desc != NULL);
1434 cfs_spin_lock(&desc->bd_lock);
1435 rc = desc->bd_network_rw;
1436 cfs_spin_unlock(&desc->bd_lock);
1441 int ptlrpc_register_bulk(struct ptlrpc_request *req);
1442 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
1444 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
1446 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1449 LASSERT(req != NULL);
1451 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
1452 req->rq_bulk_deadline > cfs_time_current_sec())
1458 cfs_spin_lock(&desc->bd_lock);
1459 rc = desc->bd_network_rw;
1460 cfs_spin_unlock(&desc->bd_lock);
1464 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
1465 #define PTLRPC_REPLY_EARLY 0x02
1466 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
1467 int ptlrpc_reply(struct ptlrpc_request *req);
1468 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
1469 int ptlrpc_error(struct ptlrpc_request *req);
1470 void ptlrpc_resend_req(struct ptlrpc_request *request);
1471 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
1472 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
1473 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
1476 /* ptlrpc/client.c */
1478 * Client-side portals API. Everything to send requests, receive replies,
1479 * request queues, request management, etc.
1482 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
1483 struct ptlrpc_client *);
1484 void ptlrpc_cleanup_client(struct obd_import *imp);
1485 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
1487 int ptlrpc_queue_wait(struct ptlrpc_request *req);
1488 int ptlrpc_replay_req(struct ptlrpc_request *req);
1489 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
1490 void ptlrpc_restart_req(struct ptlrpc_request *req);
1491 void ptlrpc_abort_inflight(struct obd_import *imp);
1492 void ptlrpc_cleanup_imp(struct obd_import *imp);
1493 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
1495 struct ptlrpc_request_set *ptlrpc_prep_set(void);
1496 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1498 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1499 set_interpreter_func fn, void *data);
1500 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
1501 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
1502 int ptlrpc_set_wait(struct ptlrpc_request_set *);
1503 int ptlrpc_expired_set(void *data);
1504 void ptlrpc_interrupted_set(void *data);
1505 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
1506 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
1507 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
1508 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1509 struct ptlrpc_request *req);
1511 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
1512 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
1514 struct ptlrpc_request_pool *
1515 ptlrpc_init_rq_pool(int, int,
1516 void (*populate_pool)(struct ptlrpc_request_pool *, int));
1518 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
1519 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1520 const struct req_format *format);
1521 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
1522 struct ptlrpc_request_pool *,
1523 const struct req_format *format);
1524 void ptlrpc_request_free(struct ptlrpc_request *request);
1525 int ptlrpc_request_pack(struct ptlrpc_request *request,
1526 __u32 version, int opcode);
1527 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1528 const struct req_format *format,
1529 __u32 version, int opcode);
1530 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
1531 __u32 version, int opcode, char **bufs,
1532 struct ptlrpc_cli_ctx *ctx);
1533 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
1534 unsigned int timeout,
1535 ptlrpc_interpterer_t interpreter);
1536 void ptlrpc_fakereq_finished(struct ptlrpc_request *req);
1538 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
1539 int opcode, int count, __u32 *lengths,
1541 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
1542 __u32 version, int opcode,
1543 int count, __u32 *lengths, char **bufs,
1544 struct ptlrpc_request_pool *pool);
1545 void ptlrpc_req_finished(struct ptlrpc_request *request);
1546 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
1547 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
1548 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
1549 int npages, int type, int portal);
1550 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
1551 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
1552 cfs_page_t *page, int pageoffset, int len);
1553 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1554 struct obd_import *imp);
1555 __u64 ptlrpc_next_xid(void);
1556 __u64 ptlrpc_sample_next_xid(void);
1557 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
1559 /* Set of routines to run a function in ptlrpcd context */
1560 void *ptlrpcd_alloc_work(struct obd_import *imp,
1561 int (*cb)(const struct lu_env *, void *), void *data);
1562 void ptlrpcd_destroy_work(void *handler);
1563 int ptlrpcd_queue_work(void *handler);
1566 struct ptlrpc_service_buf_conf {
1567 /* nbufs is how many buffers to post */
1568 unsigned int bc_nbufs;
1569 /* buffer size to post */
1570 unsigned int bc_buf_size;
1571 /* portal to listed for requests on */
1572 unsigned int bc_req_portal;
1573 /* portal of where to send replies to */
1574 unsigned int bc_rep_portal;
1575 /* maximum request size to be accepted for this service */
1576 unsigned int bc_req_max_size;
1577 /* maximum reply size this service can ever send */
1578 unsigned int bc_rep_max_size;
1581 struct ptlrpc_service_thr_conf {
1582 /* threadname should be 8 characters or less - 6 will be added on */
1584 /* min number of service threads to start */
1585 unsigned int tc_nthrs_min;
1586 /* max number of service threads to start */
1587 unsigned int tc_nthrs_max;
1588 /* user specified threads number, it will be validated due to
1589 * other members of this structure. */
1590 unsigned int tc_nthrs_user;
1591 /* set NUMA node affinity for service threads */
1592 unsigned int tc_cpu_affinity;
1593 /* Tags for lu_context associated with service thread */
1597 struct ptlrpc_service_conf {
1600 /* soft watchdog timeout multiplifier to print stuck service traces */
1601 unsigned int psc_watchdog_factor;
1602 /* buffer information */
1603 struct ptlrpc_service_buf_conf psc_buf;
1604 /* thread information */
1605 struct ptlrpc_service_thr_conf psc_thr;
1606 /* function table */
1607 struct ptlrpc_service_ops psc_ops;
1610 /* ptlrpc/service.c */
1612 * Server-side services API. Register/unregister service, request state
1613 * management, service thread management
1617 void ptlrpc_save_lock(struct ptlrpc_request *req,
1618 struct lustre_handle *lock, int mode, int no_ack);
1619 void ptlrpc_commit_replies(struct obd_export *exp);
1620 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
1621 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
1622 struct ptlrpc_service *ptlrpc_register_service(
1623 struct ptlrpc_service_conf *conf,
1624 struct proc_dir_entry *proc_entry);
1625 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
1627 int ptlrpc_start_threads(struct ptlrpc_service *svc);
1628 int ptlrpc_start_thread(struct ptlrpc_service *svc);
1629 int ptlrpc_unregister_service(struct ptlrpc_service *service);
1630 int liblustre_check_services(void *arg);
1631 void ptlrpc_daemonize(char *name);
1632 int ptlrpc_service_health_check(struct ptlrpc_service *);
1633 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req);
1634 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
1637 int ptlrpc_hr_init(void);
1638 void ptlrpc_hr_fini(void);
1640 # define ptlrpc_hr_init() (0)
1641 # define ptlrpc_hr_fini() do {} while(0)
1644 struct ptlrpc_svc_data {
1646 struct ptlrpc_service *svc;
1647 struct ptlrpc_thread *thread;
1651 /* ptlrpc/import.c */
1656 int ptlrpc_connect_import(struct obd_import *imp);
1657 int ptlrpc_init_import(struct obd_import *imp);
1658 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
1659 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
1660 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
1663 /* ptlrpc/pack_generic.c */
1664 int ptlrpc_reconnect_import(struct obd_import *imp);
1668 * ptlrpc msg buffer and swab interface
1672 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
1674 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
1676 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
1677 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
1679 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
1680 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
1682 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
1683 __u32 *lens, char **bufs);
1684 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
1686 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
1687 __u32 *lens, char **bufs, int flags);
1688 #define LPRFL_EARLY_REPLY 1
1689 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
1690 char **bufs, int flags);
1691 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
1692 unsigned int newlen, int move_data);
1693 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
1694 int __lustre_unpack_msg(struct lustre_msg *m, int len);
1695 int lustre_msg_hdr_size(__u32 magic, int count);
1696 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
1697 int lustre_msg_size_v2(int count, __u32 *lengths);
1698 int lustre_packed_msg_size(struct lustre_msg *msg);
1699 int lustre_msg_early_size(void);
1700 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
1701 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
1702 int lustre_msg_buflen(struct lustre_msg *m, int n);
1703 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
1704 int lustre_msg_bufcount(struct lustre_msg *m);
1705 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
1706 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
1707 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
1708 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
1709 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
1710 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
1711 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
1712 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
1713 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
1714 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
1715 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
1716 __u32 lustre_msg_get_type(struct lustre_msg *msg);
1717 __u32 lustre_msg_get_version(struct lustre_msg *msg);
1718 void lustre_msg_add_version(struct lustre_msg *msg, int version);
1719 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
1720 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
1721 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
1722 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
1723 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
1724 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
1725 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
1726 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
1727 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
1728 int lustre_msg_get_status(struct lustre_msg *msg);
1729 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
1730 int lustre_msg_is_v1(struct lustre_msg *msg);
1731 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
1732 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
1733 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
1734 char *lustre_msg_get_jobid(struct lustre_msg *msg);
1735 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
1736 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 0, 0)
1737 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
1739 # warning "remove checksum compatibility support for b1_8"
1740 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
1742 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
1743 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
1744 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
1745 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
1746 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
1747 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
1748 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
1749 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
1750 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
1751 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
1752 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
1753 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
1754 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
1755 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
1756 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
1759 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
1760 unsigned int newlen, int move_data)
1762 LASSERT(req->rq_reply_state);
1763 LASSERT(req->rq_repmsg);
1764 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
1769 /** Change request phase of \a req to \a new_phase */
1771 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
1773 if (req->rq_phase == new_phase)
1776 if (new_phase == RQ_PHASE_UNREGISTERING) {
1777 req->rq_next_phase = req->rq_phase;
1779 cfs_atomic_inc(&req->rq_import->imp_unregistering);
1782 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1784 cfs_atomic_dec(&req->rq_import->imp_unregistering);
1787 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
1788 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
1790 req->rq_phase = new_phase;
1794 * Returns true if request \a req got early reply and hard deadline is not met
1797 ptlrpc_client_early(struct ptlrpc_request *req)
1799 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1800 req->rq_reply_deadline > cfs_time_current_sec())
1802 return req->rq_early;
1806 * Returns true if we got real reply from server for this request
1809 ptlrpc_client_replied(struct ptlrpc_request *req)
1811 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1812 req->rq_reply_deadline > cfs_time_current_sec())
1814 return req->rq_replied;
1817 /** Returns true if request \a req is in process of receiving server reply */
1819 ptlrpc_client_recv(struct ptlrpc_request *req)
1821 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1822 req->rq_reply_deadline > cfs_time_current_sec())
1824 return req->rq_receiving_reply;
1828 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
1832 cfs_spin_lock(&req->rq_lock);
1833 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1834 req->rq_reply_deadline > cfs_time_current_sec()) {
1835 cfs_spin_unlock(&req->rq_lock);
1838 rc = req->rq_receiving_reply || req->rq_must_unlink;
1839 cfs_spin_unlock(&req->rq_lock);
1844 ptlrpc_client_wake_req(struct ptlrpc_request *req)
1846 if (req->rq_set == NULL)
1847 cfs_waitq_signal(&req->rq_reply_waitq);
1849 cfs_waitq_signal(&req->rq_set->set_waitq);
1853 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
1855 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
1856 cfs_atomic_inc(&rs->rs_refcount);
1860 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
1862 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
1863 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
1864 lustre_free_reply_state(rs);
1867 /* Should only be called once per req */
1868 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
1870 if (req->rq_reply_state == NULL)
1871 return; /* shouldn't occur */
1872 ptlrpc_rs_decref(req->rq_reply_state);
1873 req->rq_reply_state = NULL;
1874 req->rq_repmsg = NULL;
1877 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
1879 return lustre_msg_get_magic(req->rq_reqmsg);
1882 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
1884 switch (req->rq_reqmsg->lm_magic) {
1885 case LUSTRE_MSG_MAGIC_V2:
1886 return req->rq_reqmsg->lm_repsize;
1888 LASSERTF(0, "incorrect message magic: %08x\n",
1889 req->rq_reqmsg->lm_magic);
1894 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
1896 if (req->rq_delay_limit != 0 &&
1897 cfs_time_before(cfs_time_add(req->rq_queued_time,
1898 cfs_time_seconds(req->rq_delay_limit)),
1899 cfs_time_current())) {
1905 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
1907 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
1908 cfs_spin_lock(&req->rq_lock);
1909 req->rq_no_resend = 1;
1910 cfs_spin_unlock(&req->rq_lock);
1912 return req->rq_no_resend;
1915 /* ldlm/ldlm_lib.c */
1917 * Target client logic
1920 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
1921 int client_obd_cleanup(struct obd_device *obddev);
1922 int client_connect_import(const struct lu_env *env,
1923 struct obd_export **exp, struct obd_device *obd,
1924 struct obd_uuid *cluuid, struct obd_connect_data *,
1926 int client_disconnect_export(struct obd_export *exp);
1927 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
1929 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
1930 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
1931 struct obd_uuid *uuid);
1932 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
1933 void client_destroy_import(struct obd_import *imp);
1936 #ifdef HAVE_SERVER_SUPPORT
1937 int server_disconnect_export(struct obd_export *exp);
1940 /* ptlrpc/pinger.c */
1942 * Pinger API (client side only)
1945 enum timeout_event {
1948 struct timeout_item;
1949 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
1950 int ptlrpc_pinger_add_import(struct obd_import *imp);
1951 int ptlrpc_pinger_del_import(struct obd_import *imp);
1952 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
1953 timeout_cb_t cb, void *data,
1954 cfs_list_t *obd_list);
1955 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
1956 enum timeout_event event);
1957 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
1958 int ptlrpc_obd_ping(struct obd_device *obd);
1959 cfs_time_t ptlrpc_suspend_wakeup_time(void);
1961 void ping_evictor_start(void);
1962 void ping_evictor_stop(void);
1964 #define ping_evictor_start() do {} while (0)
1965 #define ping_evictor_stop() do {} while (0)
1967 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
1970 /* ptlrpc daemon bind policy */
1972 /* all ptlrpcd threads are free mode */
1973 PDB_POLICY_NONE = 1,
1974 /* all ptlrpcd threads are bound mode */
1975 PDB_POLICY_FULL = 2,
1976 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
1977 PDB_POLICY_PAIR = 3,
1978 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
1979 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
1980 * If kernel supports NUMA, pthrpcd threads are binded and
1981 * grouped by NUMA node */
1982 PDB_POLICY_NEIGHBOR = 4,
1985 /* ptlrpc daemon load policy
1986 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
1987 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
1988 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
1989 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
1990 * depends on which is scheduled firstly, to accelerate the RPC processing. */
1992 /* on the same CPU core as the caller */
1993 PDL_POLICY_SAME = 1,
1994 /* within the same CPU partition, but not the same core as the caller */
1995 PDL_POLICY_LOCAL = 2,
1996 /* round-robin on all CPU cores, but not the same core as the caller */
1997 PDL_POLICY_ROUND = 3,
1998 /* the specified CPU core is preferred, but not enforced */
1999 PDL_POLICY_PREFERRED = 4,
2002 /* ptlrpc/ptlrpcd.c */
2003 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2004 void ptlrpcd_wake(struct ptlrpc_request *req);
2005 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
2006 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2007 int ptlrpcd_addref(void);
2008 void ptlrpcd_decref(void);
2010 /* ptlrpc/lproc_ptlrpc.c */
2012 * procfs output related functions
2015 const char* ll_opcode2str(__u32 opcode);
2017 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2018 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2019 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2021 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2022 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2023 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2027 /* ptlrpc/llog_server.c */
2028 int llog_origin_handle_create(struct ptlrpc_request *req);
2029 int llog_origin_handle_destroy(struct ptlrpc_request *req);
2030 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2031 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2032 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2033 int llog_origin_handle_close(struct ptlrpc_request *req);
2034 int llog_origin_handle_cancel(struct ptlrpc_request *req);
2035 int llog_catinfo(struct ptlrpc_request *req);
2037 /* ptlrpc/llog_client.c */
2038 extern struct llog_operations llog_client_ops;