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_INIT 2
114 * Constants determine how memory is used to buffer incoming service requests.
116 * ?_NBUFS # buffers to allocate when growing the pool
117 * ?_BUFSIZE # bytes in a single request buffer
118 * ?_MAXREQSIZE # maximum request service will receive
120 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
121 * of ?_NBUFS is added to the pool.
123 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
124 * considered full when less than ?_MAXREQSIZE is left in them.
129 * Constants determine how threads are created for ptlrpc service.
131 * ?_NTHRS_INIT # threads to create for each service partition on
132 * initializing. If it's non-affinity service and
133 * there is only one partition, it's the overall #
134 * threads for the service while initializing.
135 * ?_NTHRS_BASE # threads should be created at least for each
136 * ptlrpc partition to keep the service healthy.
137 * It's the low-water mark of threads upper-limit
138 * for each partition.
139 * ?_THR_FACTOR # threads can be added on threads upper-limit for
140 * each CPU core. This factor is only for reference,
141 * we might decrease value of factor if number of cores
142 * per CPT is above a limit.
143 * ?_NTHRS_MAX # overall threads can be created for a service,
144 * it's a soft limit because if service is running
145 * on machine with hundreds of cores and tens of
146 * CPU partitions, we need to guarantee each partition
147 * has ?_NTHRS_BASE threads, which means total threads
148 * will be ?_NTHRS_BASE * number_of_cpts which can
149 * exceed ?_NTHRS_MAX.
153 * #define MDT_NTHRS_INIT 2
154 * #define MDT_NTHRS_BASE 64
155 * #define MDT_NTHRS_FACTOR 8
156 * #define MDT_NTHRS_MAX 1024
159 * ---------------------------------------------------------------------
160 * Server(A) has 16 cores, user configured it to 4 partitions so each
161 * partition has 4 cores, then actual number of service threads on each
163 * MDT_NTHRS_BASE(64) + cores(4) * MDT_NTHRS_FACTOR(8) = 96
165 * Total number of threads for the service is:
166 * 96 * partitions(4) = 384
169 * ---------------------------------------------------------------------
170 * Server(B) has 32 cores, user configured it to 4 partitions so each
171 * partition has 8 cores, then actual number of service threads on each
173 * MDT_NTHRS_BASE(64) + cores(8) * MDT_NTHRS_FACTOR(8) = 128
175 * Total number of threads for the service is:
176 * 128 * partitions(4) = 512
179 * ---------------------------------------------------------------------
180 * Server(B) has 96 cores, user configured it to 8 partitions so each
181 * partition has 12 cores, then actual number of service threads on each
183 * MDT_NTHRS_BASE(64) + cores(12) * MDT_NTHRS_FACTOR(8) = 160
185 * Total number of threads for the service is:
186 * 160 * partitions(8) = 1280
188 * However, it's above the soft limit MDT_NTHRS_MAX, so we choose this number
189 * as upper limit of threads number for each partition:
190 * MDT_NTHRS_MAX(1024) / partitions(8) = 128
193 * ---------------------------------------------------------------------
194 * Server(C) have a thousand of cores and user configured it to 32 partitions
195 * MDT_NTHRS_BASE(64) * 32 = 2048
197 * which is already above soft limit MDT_NTHRS_MAX(1024), but we still need
198 * to guarantee that each partition has at least MDT_NTHRS_BASE(64) threads
199 * to keep service healthy, so total number of threads will just be 2048.
201 * NB: we don't suggest to choose server with that many cores because backend
202 * filesystem itself, buffer cache, or underlying network stack might
203 * have some SMP scalability issues at that large scale.
205 * If user already has a fat machine with hundreds or thousands of cores,
206 * there are two choices for configuration:
207 * a) create CPU table from subset of all CPUs and run Lustre on
209 * b) bind service threads on a few partitions, see modparameters of
210 * MDS and OSS for details
212 * NB: these calculations (and examples below) are simplified to help
213 * understanding, the real implementation is a little more complex,
214 * please see ptlrpc_server_nthreads_check() for details.
219 * LDLM threads constants:
221 * Given 8 as factor and 24 as base threads number
224 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
227 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
228 * threads for each partition and total threads number will be 112.
231 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
232 * threads for each partition to keep service healthy, so total threads
233 * number should be 24 * 8 = 192.
235 * So with these constants, threads number wil be at the similar level
236 * of old versions, unless target machine has over a hundred cores
238 #define LDLM_THR_FACTOR 8
239 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
240 #define LDLM_NTHRS_BASE 24
241 #define LDLM_NTHRS_MAX (cfs_num_online_cpus() == 1 ? 64 : 128)
243 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
244 #define LDLM_NBUFS (64 * cfs_num_online_cpus())
245 #define LDLM_BUFSIZE (8 * 1024)
246 #define LDLM_MAXREQSIZE (5 * 1024)
247 #define LDLM_MAXREPSIZE (1024)
250 * MDS threads constants:
252 * Please see examples in "Thread Constants", MDS threads number will be at
253 * the comparable level of old versions, unless the server has many cores.
255 #ifndef MDT_MAX_THREADS
256 #define MDT_MAX_THREADS 1024
257 #define MDT_MAX_OTHR_THREADS 256
259 #else /* MDT_MAX_THREADS */
260 #if MDT_MAX_THREADS < PTLRPC_NTHRS_INIT
261 #undef MDT_MAX_THREADS
262 #define MDT_MAX_THREADS PTLRPC_NTHRS_INIT
264 #define MDT_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDT_MAX_THREADS / 2)
267 /* default service */
268 #define MDT_THR_FACTOR 8
269 #define MDT_NTHRS_INIT PTLRPC_NTHRS_INIT
270 #define MDT_NTHRS_MAX MDT_MAX_THREADS
271 #define MDT_NTHRS_BASE min(64, MDT_NTHRS_MAX)
273 /* read-page service */
274 #define MDT_RDPG_THR_FACTOR 4
275 #define MDT_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
276 #define MDT_RDPG_NTHRS_MAX MDT_MAX_OTHR_THREADS
277 #define MDT_RDPG_NTHRS_BASE min(48, MDT_RDPG_NTHRS_MAX)
279 /* these should be removed when we remove setattr service in the future */
280 #define MDT_SETA_THR_FACTOR 4
281 #define MDT_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
282 #define MDT_SETA_NTHRS_MAX MDT_MAX_OTHR_THREADS
283 #define MDT_SETA_NTHRS_BASE min(48, MDT_SETA_NTHRS_MAX)
285 /* non-affinity threads */
286 #define MDT_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
287 #define MDT_OTHR_NTHRS_MAX MDT_MAX_OTHR_THREADS
289 #define MDS_NBUFS (64 * cfs_num_online_cpus())
291 * Assume file name length = FNAME_MAX = 256 (true for ext3).
292 * path name length = PATH_MAX = 4096
293 * LOV MD size max = EA_MAX = 48000 (2000 stripes)
294 * symlink: FNAME_MAX + PATH_MAX <- largest
295 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
296 * rename: FNAME_MAX + FNAME_MAX
297 * open: FNAME_MAX + EA_MAX
299 * MDS_MAXREQSIZE ~= 4736 bytes =
300 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
301 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
302 * or, for mds_close() and mds_reint_unlink() on a many-OST filesystem:
303 * = 9210 bytes = lustre_msg + mdt_body + 160 * (easize + cookiesize)
305 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
306 * except in the open case where there are a large number of OSTs in a LOV.
308 #define MDS_MAXREPSIZE max(10 * 1024, 362 + LOV_MAX_STRIPE_COUNT * 56)
309 #define MDS_MAXREQSIZE MDS_MAXREPSIZE
311 /** MDS_BUFSIZE = max_reqsize + max sptlrpc payload size */
312 #define MDS_BUFSIZE (MDS_MAXREQSIZE + 1024)
314 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
315 #define FLD_MAXREQSIZE (160)
317 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
318 #define FLD_MAXREPSIZE (152)
321 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
323 #define SEQ_MAXREQSIZE (160)
325 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
326 #define SEQ_MAXREPSIZE (152)
328 /** MGS threads must be >= 3, see bug 22458 comment #28 */
329 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
330 #define MGS_NTHRS_MAX 32
332 #define MGS_NBUFS (64 * cfs_num_online_cpus())
333 #define MGS_BUFSIZE (8 * 1024)
334 #define MGS_MAXREQSIZE (7 * 1024)
335 #define MGS_MAXREPSIZE (9 * 1024)
338 * OSS threads constants:
340 * Given 8 as factor and 64 as base threads number
343 * On 8-core server configured to 2 partitions, we will have
344 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
347 * On 32-core machine configured to 4 partitions, we will have
348 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
349 * will be 112 * 4 = 448.
352 * On 64-core machine configured to 4 partitions, we will have
353 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
354 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
355 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
356 * for each partition.
358 * So we can see that with these constants, threads number wil be at the
359 * similar level of old versions, unless the server has many cores.
361 /* depress threads factor for VM with small memory size */
362 #define OSS_THR_FACTOR min_t(int, 8, \
363 CFS_NUM_CACHEPAGES >> (28 - CFS_PAGE_SHIFT))
364 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
365 #define OSS_NTHRS_BASE 64
366 #define OSS_NTHRS_MAX 512
368 /* threads for handling "create" request */
369 #define OSS_CR_THR_FACTOR 1
370 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
371 #define OSS_CR_NTHRS_BASE 8
372 #define OSS_CR_NTHRS_MAX 64
374 #define OST_NBUFS (64 * cfs_num_online_cpus())
375 #define OST_BUFSIZE (8 * 1024)
378 * OST_MAXREQSIZE ~= 4768 bytes =
379 * lustre_msg + obdo + 16 * obd_ioobj + 256 * niobuf_remote
381 * - single object with 16 pages is 512 bytes
382 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
384 #define OST_MAXREQSIZE (5 * 1024)
385 #define OST_MAXREPSIZE (9 * 1024)
387 /* Macro to hide a typecast. */
388 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
391 * Structure to single define portal connection.
393 struct ptlrpc_connection {
394 /** linkage for connections hash table */
395 cfs_hlist_node_t c_hash;
396 /** Our own lnet nid for this connection */
398 /** Remote side nid for this connection */
399 lnet_process_id_t c_peer;
400 /** UUID of the other side */
401 struct obd_uuid c_remote_uuid;
402 /** reference counter for this connection */
403 cfs_atomic_t c_refcount;
406 /** Client definition for PortalRPC */
407 struct ptlrpc_client {
408 /** What lnet portal does this client send messages to by default */
409 __u32 cli_request_portal;
410 /** What portal do we expect replies on */
411 __u32 cli_reply_portal;
412 /** Name of the client */
416 /** state flags of requests */
417 /* XXX only ones left are those used by the bulk descs as well! */
418 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
419 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
421 #define REQ_MAX_ACK_LOCKS 8
423 union ptlrpc_async_args {
425 * Scratchpad for passing args to completion interpreter. Users
426 * cast to the struct of their choosing, and CLASSERT that this is
427 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
428 * a pointer to it here. The pointer_arg ensures this struct is at
429 * least big enough for that.
431 void *pointer_arg[11];
435 struct ptlrpc_request_set;
436 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
437 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
440 * Definition of request set structure.
441 * Request set is a list of requests (not necessary to the same target) that
442 * once populated with RPCs could be sent in parallel.
443 * There are two kinds of request sets. General purpose and with dedicated
444 * serving thread. Example of the latter is ptlrpcd set.
445 * For general purpose sets once request set started sending it is impossible
446 * to add new requests to such set.
447 * Provides a way to call "completion callbacks" when all requests in the set
450 struct ptlrpc_request_set {
451 cfs_atomic_t set_refcount;
452 /** number of in queue requests */
453 cfs_atomic_t set_new_count;
454 /** number of uncompleted requests */
455 cfs_atomic_t set_remaining;
456 /** wait queue to wait on for request events */
457 cfs_waitq_t set_waitq;
458 cfs_waitq_t *set_wakeup_ptr;
459 /** List of requests in the set */
460 cfs_list_t set_requests;
462 * List of completion callbacks to be called when the set is completed
463 * This is only used if \a set_interpret is NULL.
464 * Links struct ptlrpc_set_cbdata.
466 cfs_list_t set_cblist;
467 /** Completion callback, if only one. */
468 set_interpreter_func set_interpret;
469 /** opaq argument passed to completion \a set_interpret callback. */
472 * Lock for \a set_new_requests manipulations
473 * locked so that any old caller can communicate requests to
474 * the set holder who can then fold them into the lock-free set
476 cfs_spinlock_t set_new_req_lock;
477 /** List of new yet unsent requests. Only used with ptlrpcd now. */
478 cfs_list_t set_new_requests;
480 /** rq_status of requests that have been freed already */
482 /** Additional fields used by the flow control extension */
483 /** Maximum number of RPCs in flight */
484 int set_max_inflight;
485 /** Callback function used to generate RPCs */
486 set_producer_func set_producer;
487 /** opaq argument passed to the producer callback */
488 void *set_producer_arg;
492 * Description of a single ptrlrpc_set callback
494 struct ptlrpc_set_cbdata {
495 /** List linkage item */
497 /** Pointer to interpreting function */
498 set_interpreter_func psc_interpret;
499 /** Opaq argument to pass to the callback */
503 struct ptlrpc_bulk_desc;
504 struct ptlrpc_service_part;
507 * ptlrpc callback & work item stuff
509 struct ptlrpc_cb_id {
510 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
511 void *cbid_arg; /* additional arg */
514 /** Maximum number of locks to fit into reply state */
515 #define RS_MAX_LOCKS 8
519 * Structure to define reply state on the server
520 * Reply state holds various reply message information. Also for "difficult"
521 * replies (rep-ack case) we store the state after sending reply and wait
522 * for the client to acknowledge the reception. In these cases locks could be
523 * added to the state for replay/failover consistency guarantees.
525 struct ptlrpc_reply_state {
526 /** Callback description */
527 struct ptlrpc_cb_id rs_cb_id;
528 /** Linkage for list of all reply states in a system */
530 /** Linkage for list of all reply states on same export */
531 cfs_list_t rs_exp_list;
532 /** Linkage for list of all reply states for same obd */
533 cfs_list_t rs_obd_list;
535 cfs_list_t rs_debug_list;
537 /** A spinlock to protect the reply state flags */
538 cfs_spinlock_t rs_lock;
539 /** Reply state flags */
540 unsigned long rs_difficult:1; /* ACK/commit stuff */
541 unsigned long rs_no_ack:1; /* no ACK, even for
542 difficult requests */
543 unsigned long rs_scheduled:1; /* being handled? */
544 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
545 unsigned long rs_handled:1; /* been handled yet? */
546 unsigned long rs_on_net:1; /* reply_out_callback pending? */
547 unsigned long rs_prealloc:1; /* rs from prealloc list */
548 unsigned long rs_committed:1;/* the transaction was committed
549 and the rs was dispatched
550 by ptlrpc_commit_replies */
551 /** Size of the state */
555 /** Transaction number */
559 struct obd_export *rs_export;
560 struct ptlrpc_service_part *rs_svcpt;
561 /** Lnet metadata handle for the reply */
562 lnet_handle_md_t rs_md_h;
563 cfs_atomic_t rs_refcount;
565 /** Context for the sevice thread */
566 struct ptlrpc_svc_ctx *rs_svc_ctx;
567 /** Reply buffer (actually sent to the client), encoded if needed */
568 struct lustre_msg *rs_repbuf; /* wrapper */
569 /** Size of the reply buffer */
570 int rs_repbuf_len; /* wrapper buf length */
571 /** Size of the reply message */
572 int rs_repdata_len; /* wrapper msg length */
574 * Actual reply message. Its content is encrupted (if needed) to
575 * produce reply buffer for actual sending. In simple case
576 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
578 struct lustre_msg *rs_msg; /* reply message */
580 /** Number of locks awaiting client ACK */
582 /** Handles of locks awaiting client reply ACK */
583 struct lustre_handle rs_locks[RS_MAX_LOCKS];
584 /** Lock modes of locks in \a rs_locks */
585 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
588 struct ptlrpc_thread;
592 RQ_PHASE_NEW = 0xebc0de00,
593 RQ_PHASE_RPC = 0xebc0de01,
594 RQ_PHASE_BULK = 0xebc0de02,
595 RQ_PHASE_INTERPRET = 0xebc0de03,
596 RQ_PHASE_COMPLETE = 0xebc0de04,
597 RQ_PHASE_UNREGISTERING = 0xebc0de05,
598 RQ_PHASE_UNDEFINED = 0xebc0de06
601 /** Type of request interpreter call-back */
602 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
603 struct ptlrpc_request *req,
607 * Definition of request pool structure.
608 * The pool is used to store empty preallocated requests for the case
609 * when we would actually need to send something without performing
610 * any allocations (to avoid e.g. OOM).
612 struct ptlrpc_request_pool {
613 /** Locks the list */
614 cfs_spinlock_t prp_lock;
615 /** list of ptlrpc_request structs */
616 cfs_list_t prp_req_list;
617 /** Maximum message size that would fit into a rquest from this pool */
619 /** Function to allocate more requests for this pool */
620 void (*prp_populate)(struct ptlrpc_request_pool *, int);
629 * Basic request prioritization operations structure.
630 * The whole idea is centered around locks and RPCs that might affect locks.
631 * When a lock is contended we try to give priority to RPCs that might lead
632 * to fastest release of that lock.
633 * Currently only implemented for OSTs only in a way that makes all
634 * IO and truncate RPCs that are coming from a locked region where a lock is
635 * contended a priority over other requests.
637 struct ptlrpc_hpreq_ops {
639 * Check if the lock handle of the given lock is the same as
640 * taken from the request.
642 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
644 * Check if the request is a high priority one.
646 int (*hpreq_check)(struct ptlrpc_request *);
648 * Called after the request has been handled.
650 void (*hpreq_fini)(struct ptlrpc_request *);
654 * Represents remote procedure call.
656 * This is a staple structure used by everybody wanting to send a request
659 struct ptlrpc_request {
660 /* Request type: one of PTL_RPC_MSG_* */
662 /** Result of request processing */
665 * Linkage item through which this request is included into
666 * sending/delayed lists on client and into rqbd list on server
670 * Server side list of incoming unserved requests sorted by arrival
671 * time. Traversed from time to time to notice about to expire
672 * requests and sent back "early replies" to clients to let them
673 * know server is alive and well, just very busy to service their
676 cfs_list_t rq_timed_list;
677 /** server-side history, used for debuging purposes. */
678 cfs_list_t rq_history_list;
679 /** server-side per-export list */
680 cfs_list_t rq_exp_list;
681 /** server-side hp handlers */
682 struct ptlrpc_hpreq_ops *rq_ops;
684 /** initial thread servicing this request */
685 struct ptlrpc_thread *rq_svc_thread;
687 /** history sequence # */
688 __u64 rq_history_seq;
689 /** the index of service's srv_at_array into which request is linked */
691 /** Lock to protect request flags and some other important bits, like
694 cfs_spinlock_t rq_lock;
695 /** client-side flags are serialized by rq_lock */
696 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
697 rq_timedout:1, rq_resend:1, rq_restart:1,
699 * when ->rq_replay is set, request is kept by the client even
700 * after server commits corresponding transaction. This is
701 * used for operations that require sequence of multiple
702 * requests to be replayed. The only example currently is file
703 * open/close. When last request in such a sequence is
704 * committed, ->rq_replay is cleared on all requests in the
708 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
709 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
710 rq_early:1, rq_must_unlink:1,
711 rq_fake:1, /* this fake req */
712 rq_memalloc:1, /* req originated from "kswapd" */
713 /* server-side flags */
714 rq_packed_final:1, /* packed final reply */
715 rq_hp:1, /* high priority RPC */
716 rq_at_linked:1, /* link into service's srv_at_array */
719 /* whether the "rq_set" is a valid one */
722 /* do not resend request on -EINPROGRESS */
723 rq_no_retry_einprogress:1;
725 unsigned int rq_nr_resend;
727 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
728 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
729 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
730 server-side refcounf for multiple replies */
732 /** Portal to which this request would be sent */
733 short rq_request_portal; /* XXX FIXME bug 249 */
734 /** Portal where to wait for reply and where reply would be sent */
735 short rq_reply_portal; /* XXX FIXME bug 249 */
739 * !rq_truncate : # reply bytes actually received,
740 * rq_truncate : required repbuf_len for resend
743 /** Request length */
747 /** Request message - what client sent */
748 struct lustre_msg *rq_reqmsg;
749 /** Reply message - server response */
750 struct lustre_msg *rq_repmsg;
751 /** Transaction number */
756 * List item to for replay list. Not yet commited requests get linked
758 * Also see \a rq_replay comment above.
760 cfs_list_t rq_replay_list;
763 * security and encryption data
765 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
766 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
767 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
769 struct sptlrpc_flavor rq_flvr; /**< for client & server */
770 enum lustre_sec_part rq_sp_from;
772 /* client/server security flags */
774 rq_ctx_init:1, /* context initiation */
775 rq_ctx_fini:1, /* context destroy */
776 rq_bulk_read:1, /* request bulk read */
777 rq_bulk_write:1, /* request bulk write */
778 /* server authentication flags */
779 rq_auth_gss:1, /* authenticated by gss */
780 rq_auth_remote:1, /* authed as remote user */
781 rq_auth_usr_root:1, /* authed as root */
782 rq_auth_usr_mdt:1, /* authed as mdt */
783 rq_auth_usr_ost:1, /* authed as ost */
784 /* security tfm flags */
787 /* doesn't expect reply FIXME */
789 rq_pill_init:1; /* pill initialized */
791 uid_t rq_auth_uid; /* authed uid */
792 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
794 /* (server side), pointed directly into req buffer */
795 struct ptlrpc_user_desc *rq_user_desc;
797 /* various buffer pointers */
798 struct lustre_msg *rq_reqbuf; /* req wrapper */
799 char *rq_repbuf; /* rep buffer */
800 struct lustre_msg *rq_repdata; /* rep wrapper msg */
801 struct lustre_msg *rq_clrbuf; /* only in priv mode */
802 int rq_reqbuf_len; /* req wrapper buf len */
803 int rq_reqdata_len; /* req wrapper msg len */
804 int rq_repbuf_len; /* rep buffer len */
805 int rq_repdata_len; /* rep wrapper msg len */
806 int rq_clrbuf_len; /* only in priv mode */
807 int rq_clrdata_len; /* only in priv mode */
809 /** early replies go to offset 0, regular replies go after that */
810 unsigned int rq_reply_off;
814 /** Fields that help to see if request and reply were swabbed or not */
815 __u32 rq_req_swab_mask;
816 __u32 rq_rep_swab_mask;
818 /** What was import generation when this request was sent */
819 int rq_import_generation;
820 enum lustre_imp_state rq_send_state;
822 /** how many early replies (for stats) */
825 /** client+server request */
826 lnet_handle_md_t rq_req_md_h;
827 struct ptlrpc_cb_id rq_req_cbid;
828 /** optional time limit for send attempts */
829 cfs_duration_t rq_delay_limit;
830 /** time request was first queued */
831 cfs_time_t rq_queued_time;
834 /** request arrival time */
835 struct timeval rq_arrival_time;
836 /** separated reply state */
837 struct ptlrpc_reply_state *rq_reply_state;
838 /** incoming request buffer */
839 struct ptlrpc_request_buffer_desc *rq_rqbd;
841 /** client-only incoming reply */
842 lnet_handle_md_t rq_reply_md_h;
843 cfs_waitq_t rq_reply_waitq;
844 struct ptlrpc_cb_id rq_reply_cbid;
848 /** Peer description (the other side) */
849 lnet_process_id_t rq_peer;
850 /** Server-side, export on which request was received */
851 struct obd_export *rq_export;
852 /** Client side, import where request is being sent */
853 struct obd_import *rq_import;
855 /** Replay callback, called after request is replayed at recovery */
856 void (*rq_replay_cb)(struct ptlrpc_request *);
858 * Commit callback, called when request is committed and about to be
861 void (*rq_commit_cb)(struct ptlrpc_request *);
862 /** Opaq data for replay and commit callbacks. */
865 /** For bulk requests on client only: bulk descriptor */
866 struct ptlrpc_bulk_desc *rq_bulk;
868 /** client outgoing req */
870 * when request/reply sent (secs), or time when request should be sent
873 /** time for request really sent out */
876 /** when request must finish. volatile
877 * so that servers' early reply updates to the deadline aren't
878 * kept in per-cpu cache */
879 volatile time_t rq_deadline;
880 /** when req reply unlink must finish. */
881 time_t rq_reply_deadline;
882 /** when req bulk unlink must finish. */
883 time_t rq_bulk_deadline;
885 * service time estimate (secs)
886 * If the requestsis not served by this time, it is marked as timed out.
890 /** Multi-rpc bits */
891 /** Per-request waitq introduced by bug 21938 for recovery waiting */
892 cfs_waitq_t rq_set_waitq;
893 /** Link item for request set lists */
894 cfs_list_t rq_set_chain;
895 /** Link back to the request set */
896 struct ptlrpc_request_set *rq_set;
897 /** Async completion handler, called when reply is received */
898 ptlrpc_interpterer_t rq_interpret_reply;
899 /** Async completion context */
900 union ptlrpc_async_args rq_async_args;
902 /** Pool if request is from preallocated list */
903 struct ptlrpc_request_pool *rq_pool;
905 struct lu_context rq_session;
906 struct lu_context rq_recov_session;
908 /** request format description */
909 struct req_capsule rq_pill;
913 * Call completion handler for rpc if any, return it's status or original
914 * rc if there was no handler defined for this request.
916 static inline int ptlrpc_req_interpret(const struct lu_env *env,
917 struct ptlrpc_request *req, int rc)
919 if (req->rq_interpret_reply != NULL) {
920 req->rq_status = req->rq_interpret_reply(env, req,
923 return req->rq_status;
929 * Returns 1 if request buffer at offset \a index was already swabbed
931 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
933 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
934 return req->rq_req_swab_mask & (1 << index);
938 * Returns 1 if request reply buffer at offset \a index was already swabbed
940 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
942 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
943 return req->rq_rep_swab_mask & (1 << index);
947 * Returns 1 if request needs to be swabbed into local cpu byteorder
949 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
951 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
955 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
957 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
959 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
963 * Mark request buffer at offset \a index that it was already swabbed
965 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
967 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
968 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
969 req->rq_req_swab_mask |= 1 << index;
973 * Mark request reply buffer at offset \a index that it was already swabbed
975 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
977 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
978 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
979 req->rq_rep_swab_mask |= 1 << index;
983 * Convert numerical request phase value \a phase into text string description
985 static inline const char *
986 ptlrpc_phase2str(enum rq_phase phase)
995 case RQ_PHASE_INTERPRET:
997 case RQ_PHASE_COMPLETE:
999 case RQ_PHASE_UNREGISTERING:
1000 return "Unregistering";
1007 * Convert numerical request phase of the request \a req into text stringi
1010 static inline const char *
1011 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1013 return ptlrpc_phase2str(req->rq_phase);
1017 * Debugging functions and helpers to print request structure into debug log
1020 /* Spare the preprocessor, spoil the bugs. */
1021 #define FLAG(field, str) (field ? str : "")
1023 /** Convert bit flags into a string */
1024 #define DEBUG_REQ_FLAGS(req) \
1025 ptlrpc_rqphase2str(req), \
1026 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1027 FLAG(req->rq_err, "E"), \
1028 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1029 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1030 FLAG(req->rq_no_resend, "N"), \
1031 FLAG(req->rq_waiting, "W"), \
1032 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1033 FLAG(req->rq_committed, "M")
1035 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1037 void _debug_req(struct ptlrpc_request *req,
1038 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1039 __attribute__ ((format (printf, 3, 4)));
1042 * Helper that decides if we need to print request accordig to current debug
1045 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1047 CFS_CHECK_STACK(msgdata, mask, cdls); \
1049 if (((mask) & D_CANTMASK) != 0 || \
1050 ((libcfs_debug & (mask)) != 0 && \
1051 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1052 _debug_req((req), msgdata, fmt, ##a); \
1056 * This is the debug print function you need to use to print request sturucture
1057 * content into lustre debug log.
1058 * for most callers (level is a constant) this is resolved at compile time */
1059 #define DEBUG_REQ(level, req, fmt, args...) \
1061 if ((level) & (D_ERROR | D_WARNING)) { \
1062 static cfs_debug_limit_state_t cdls; \
1063 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1064 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1066 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1067 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1073 * Structure that defines a single page of a bulk transfer
1075 struct ptlrpc_bulk_page {
1076 /** Linkage to list of pages in a bulk */
1079 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1082 /** offset within a page */
1084 /** The page itself */
1085 struct page *bp_page;
1088 #define BULK_GET_SOURCE 0
1089 #define BULK_PUT_SINK 1
1090 #define BULK_GET_SINK 2
1091 #define BULK_PUT_SOURCE 3
1094 * Definition of buk descriptor.
1095 * Bulks are special "Two phase" RPCs where initial request message
1096 * is sent first and it is followed bt a transfer (o receiving) of a large
1097 * amount of data to be settled into pages referenced from the bulk descriptors.
1098 * Bulks transfers (the actual data following the small requests) are done
1099 * on separate LNet portals.
1100 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1101 * Another user is readpage for MDT.
1103 struct ptlrpc_bulk_desc {
1104 /** completed successfully */
1105 unsigned long bd_success:1;
1106 /** accessible to the network (network io potentially in progress) */
1107 unsigned long bd_network_rw:1;
1108 /** {put,get}{source,sink} */
1109 unsigned long bd_type:2;
1111 unsigned long bd_registered:1;
1112 /** For serialization with callback */
1113 cfs_spinlock_t bd_lock;
1114 /** Import generation when request for this bulk was sent */
1115 int bd_import_generation;
1116 /** Server side - export this bulk created for */
1117 struct obd_export *bd_export;
1118 /** Client side - import this bulk was sent on */
1119 struct obd_import *bd_import;
1120 /** LNet portal for this bulk */
1122 /** Back pointer to the request */
1123 struct ptlrpc_request *bd_req;
1124 cfs_waitq_t bd_waitq; /* server side only WQ */
1125 int bd_iov_count; /* # entries in bd_iov */
1126 int bd_max_iov; /* allocated size of bd_iov */
1127 int bd_nob; /* # bytes covered */
1128 int bd_nob_transferred; /* # bytes GOT/PUT */
1132 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1133 lnet_handle_md_t bd_md_h; /* associated MD */
1134 lnet_nid_t bd_sender; /* stash event::sender */
1136 #if defined(__KERNEL__)
1138 * encrypt iov, size is either 0 or bd_iov_count.
1140 lnet_kiov_t *bd_enc_iov;
1142 lnet_kiov_t bd_iov[0];
1144 lnet_md_iovec_t bd_iov[0];
1149 SVC_STOPPED = 1 << 0,
1150 SVC_STOPPING = 1 << 1,
1151 SVC_STARTING = 1 << 2,
1152 SVC_RUNNING = 1 << 3,
1154 SVC_SIGNAL = 1 << 5,
1157 #define PTLRPC_THR_NAME_LEN 32
1159 * Definition of server service thread structure
1161 struct ptlrpc_thread {
1163 * List of active threads in svc->srv_threads
1167 * thread-private data (preallocated memory)
1172 * service thread index, from ptlrpc_start_threads
1176 * service thread pid
1180 * put watchdog in the structure per thread b=14840
1182 struct lc_watchdog *t_watchdog;
1184 * the svc this thread belonged to b=18582
1186 struct ptlrpc_service_part *t_svcpt;
1187 cfs_waitq_t t_ctl_waitq;
1188 struct lu_env *t_env;
1189 char t_name[PTLRPC_THR_NAME_LEN];
1192 static inline int thread_is_init(struct ptlrpc_thread *thread)
1194 return thread->t_flags == 0;
1197 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1199 return !!(thread->t_flags & SVC_STOPPED);
1202 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1204 return !!(thread->t_flags & SVC_STOPPING);
1207 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1209 return !!(thread->t_flags & SVC_STARTING);
1212 static inline int thread_is_running(struct ptlrpc_thread *thread)
1214 return !!(thread->t_flags & SVC_RUNNING);
1217 static inline int thread_is_event(struct ptlrpc_thread *thread)
1219 return !!(thread->t_flags & SVC_EVENT);
1222 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1224 return !!(thread->t_flags & SVC_SIGNAL);
1227 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1229 thread->t_flags &= ~flags;
1232 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1234 thread->t_flags = flags;
1237 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1239 thread->t_flags |= flags;
1242 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1245 if (thread->t_flags & flags) {
1246 thread->t_flags &= ~flags;
1253 * Request buffer descriptor structure.
1254 * This is a structure that contains one posted request buffer for service.
1255 * Once data land into a buffer, event callback creates actual request and
1256 * notifies wakes one of the service threads to process new incoming request.
1257 * More than one request can fit into the buffer.
1259 struct ptlrpc_request_buffer_desc {
1260 /** Link item for rqbds on a service */
1261 cfs_list_t rqbd_list;
1262 /** History of requests for this buffer */
1263 cfs_list_t rqbd_reqs;
1264 /** Back pointer to service for which this buffer is registered */
1265 struct ptlrpc_service_part *rqbd_svcpt;
1266 /** LNet descriptor */
1267 lnet_handle_md_t rqbd_md_h;
1269 /** The buffer itself */
1271 struct ptlrpc_cb_id rqbd_cbid;
1273 * This "embedded" request structure is only used for the
1274 * last request to fit into the buffer
1276 struct ptlrpc_request rqbd_req;
1279 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1281 struct ptlrpc_service_ops {
1283 * if non-NULL called during thread creation (ptlrpc_start_thread())
1284 * to initialize service specific per-thread state.
1286 int (*so_thr_init)(struct ptlrpc_thread *thr);
1288 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1289 * destruct state created by ->srv_init().
1291 void (*so_thr_done)(struct ptlrpc_thread *thr);
1293 * Handler function for incoming requests for this service
1295 int (*so_req_handler)(struct ptlrpc_request *req);
1297 * function to determine priority of the request, it's called
1298 * on every new request
1300 int (*so_hpreq_handler)(struct ptlrpc_request *);
1302 * service-specific print fn
1304 void (*so_req_printer)(void *, struct ptlrpc_request *);
1307 #ifndef __cfs_cacheline_aligned
1308 /* NB: put it here for reducing patche dependence */
1309 # define __cfs_cacheline_aligned
1313 * How many high priority requests to serve before serving one normal
1316 #define PTLRPC_SVC_HP_RATIO 10
1319 * Definition of PortalRPC service.
1320 * The service is listening on a particular portal (like tcp port)
1321 * and perform actions for a specific server like IO service for OST
1322 * or general metadata service for MDS.
1324 struct ptlrpc_service {
1325 /** serialize /proc operations */
1326 cfs_spinlock_t srv_lock;
1327 /** most often accessed fields */
1328 /** chain thru all services */
1329 cfs_list_t srv_list;
1330 /** service operations table */
1331 struct ptlrpc_service_ops srv_ops;
1332 /** only statically allocated strings here; we don't clean them */
1334 /** only statically allocated strings here; we don't clean them */
1335 char *srv_thread_name;
1336 /** service thread list */
1337 cfs_list_t srv_threads;
1338 /** threads # should be created for each partition on initializing */
1339 int srv_nthrs_cpt_init;
1340 /** limit of threads number for each partition */
1341 int srv_nthrs_cpt_limit;
1342 /** Root of /proc dir tree for this service */
1343 cfs_proc_dir_entry_t *srv_procroot;
1344 /** Pointer to statistic data for this service */
1345 struct lprocfs_stats *srv_stats;
1346 /** # hp per lp reqs to handle */
1347 int srv_hpreq_ratio;
1348 /** biggest request to receive */
1349 int srv_max_req_size;
1350 /** biggest reply to send */
1351 int srv_max_reply_size;
1352 /** size of individual buffers */
1354 /** # buffers to allocate in 1 group */
1355 int srv_nbuf_per_group;
1356 /** Local portal on which to receive requests */
1357 __u32 srv_req_portal;
1358 /** Portal on the client to send replies to */
1359 __u32 srv_rep_portal;
1361 * Tags for lu_context associated with this thread, see struct
1365 /** soft watchdog timeout multiplier */
1366 int srv_watchdog_factor;
1367 /** under unregister_service */
1368 unsigned srv_is_stopping:1;
1370 /** max # request buffers in history per partition */
1371 int srv_hist_nrqbds_cpt_max;
1372 /** number of CPTs this service bound on */
1374 /** CPTs array this service bound on */
1376 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1378 /** CPT table this service is running over */
1379 struct cfs_cpt_table *srv_cptable;
1381 * partition data for ptlrpc service
1383 struct ptlrpc_service_part *srv_parts[0];
1387 * Definition of PortalRPC service partition data.
1388 * Although a service only has one instance of it right now, but we
1389 * will have multiple instances very soon (instance per CPT).
1391 * it has four locks:
1393 * serialize operations on rqbd and requests waiting for preprocess
1395 * serialize operations active requests sent to this portal
1397 * serialize adaptive timeout stuff
1399 * serialize operations on RS list (reply states)
1401 * We don't have any use-case to take two or more locks at the same time
1402 * for now, so there is no lock order issue.
1404 struct ptlrpc_service_part {
1405 /** back reference to owner */
1406 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1407 /* CPT id, reserved */
1409 /** always increasing number */
1411 /** # of starting threads */
1412 int scp_nthrs_starting;
1413 /** # of stopping threads, reserved for shrinking threads */
1414 int scp_nthrs_stopping;
1415 /** # running threads */
1416 int scp_nthrs_running;
1417 /** service threads list */
1418 cfs_list_t scp_threads;
1421 * serialize the following fields, used for protecting
1422 * rqbd list and incoming requests waiting for preprocess,
1423 * threads starting & stopping are also protected by this lock.
1425 cfs_spinlock_t scp_lock __cfs_cacheline_aligned;
1426 /** total # req buffer descs allocated */
1427 int scp_nrqbds_total;
1428 /** # posted request buffers for receiving */
1429 int scp_nrqbds_posted;
1430 /** # incoming reqs */
1431 int scp_nreqs_incoming;
1432 /** request buffers to be reposted */
1433 cfs_list_t scp_rqbd_idle;
1434 /** req buffers receiving */
1435 cfs_list_t scp_rqbd_posted;
1436 /** incoming reqs */
1437 cfs_list_t scp_req_incoming;
1438 /** timeout before re-posting reqs, in tick */
1439 cfs_duration_t scp_rqbd_timeout;
1441 * all threads sleep on this. This wait-queue is signalled when new
1442 * incoming request arrives and when difficult reply has to be handled.
1444 cfs_waitq_t scp_waitq;
1446 /** request history */
1447 cfs_list_t scp_hist_reqs;
1448 /** request buffer history */
1449 cfs_list_t scp_hist_rqbds;
1450 /** # request buffers in history */
1451 int scp_hist_nrqbds;
1452 /** sequence number for request */
1454 /** highest seq culled from history */
1455 __u64 scp_hist_seq_culled;
1458 * serialize the following fields, used for processing requests
1459 * sent to this portal
1461 cfs_spinlock_t scp_req_lock __cfs_cacheline_aligned;
1462 /** # reqs in either of the queues below */
1463 /** reqs waiting for service */
1464 cfs_list_t scp_req_pending;
1465 /** high priority queue */
1466 cfs_list_t scp_hreq_pending;
1467 /** # reqs being served */
1468 int scp_nreqs_active;
1469 /** # HPreqs being served */
1470 int scp_nhreqs_active;
1471 /** # hp requests handled */
1477 * serialize the following fields, used for changes on
1480 cfs_spinlock_t scp_at_lock __cfs_cacheline_aligned;
1481 /** estimated rpc service time */
1482 struct adaptive_timeout scp_at_estimate;
1483 /** reqs waiting for replies */
1484 struct ptlrpc_at_array scp_at_array;
1485 /** early reply timer */
1486 cfs_timer_t scp_at_timer;
1488 cfs_time_t scp_at_checktime;
1489 /** check early replies */
1490 unsigned scp_at_check;
1494 * serialize the following fields, used for processing
1495 * replies for this portal
1497 cfs_spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1498 /** all the active replies */
1499 cfs_list_t scp_rep_active;
1501 /** replies waiting for service */
1502 cfs_list_t scp_rep_queue;
1504 /** List of free reply_states */
1505 cfs_list_t scp_rep_idle;
1506 /** waitq to run, when adding stuff to srv_free_rs_list */
1507 cfs_waitq_t scp_rep_waitq;
1508 /** # 'difficult' replies */
1509 cfs_atomic_t scp_nreps_difficult;
1512 #define ptlrpc_service_for_each_part(part, i, svc) \
1514 i < (svc)->srv_ncpts && \
1515 (svc)->srv_parts != NULL && \
1516 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1519 * Declaration of ptlrpcd control structure
1521 struct ptlrpcd_ctl {
1523 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1525 unsigned long pc_flags;
1527 * Thread lock protecting structure fields.
1529 cfs_spinlock_t pc_lock;
1533 cfs_completion_t pc_starting;
1537 cfs_completion_t pc_finishing;
1539 * Thread requests set.
1541 struct ptlrpc_request_set *pc_set;
1543 * Thread name used in cfs_daemonize()
1547 * Environment for request interpreters to run in.
1549 struct lu_env pc_env;
1551 * Index of ptlrpcd thread in the array.
1555 * Number of the ptlrpcd's partners.
1559 * Pointer to the array of partners' ptlrpcd_ctl structure.
1561 struct ptlrpcd_ctl **pc_partners;
1563 * Record the partner index to be processed next.
1568 * Async rpcs flag to make sure that ptlrpcd_check() is called only
1573 * Currently not used.
1577 * User-space async rpcs callback.
1579 void *pc_wait_callback;
1581 * User-space check idle rpcs callback.
1583 void *pc_idle_callback;
1587 /* Bits for pc_flags */
1588 enum ptlrpcd_ctl_flags {
1590 * Ptlrpc thread start flag.
1592 LIOD_START = 1 << 0,
1594 * Ptlrpc thread stop flag.
1598 * Ptlrpc thread force flag (only stop force so far).
1599 * This will cause aborting any inflight rpcs handled
1600 * by thread if LIOD_STOP is specified.
1602 LIOD_FORCE = 1 << 2,
1604 * This is a recovery ptlrpc thread.
1606 LIOD_RECOVERY = 1 << 3,
1608 * The ptlrpcd is bound to some CPU core.
1613 /* ptlrpc/events.c */
1614 extern lnet_handle_eq_t ptlrpc_eq_h;
1615 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1616 lnet_process_id_t *peer, lnet_nid_t *self);
1618 * These callbacks are invoked by LNet when something happened to
1622 extern void request_out_callback(lnet_event_t *ev);
1623 extern void reply_in_callback(lnet_event_t *ev);
1624 extern void client_bulk_callback(lnet_event_t *ev);
1625 extern void request_in_callback(lnet_event_t *ev);
1626 extern void reply_out_callback(lnet_event_t *ev);
1627 #ifdef HAVE_SERVER_SUPPORT
1628 extern void server_bulk_callback(lnet_event_t *ev);
1632 /* ptlrpc/connection.c */
1633 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
1635 struct obd_uuid *uuid);
1636 int ptlrpc_connection_put(struct ptlrpc_connection *c);
1637 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1638 int ptlrpc_connection_init(void);
1639 void ptlrpc_connection_fini(void);
1640 extern lnet_pid_t ptl_get_pid(void);
1642 /* ptlrpc/niobuf.c */
1644 * Actual interfacing with LNet to put/get/register/unregister stuff
1647 #ifdef HAVE_SERVER_SUPPORT
1648 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
1649 int npages, int type, int portal);
1650 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
1651 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
1653 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
1657 LASSERT(desc != NULL);
1659 cfs_spin_lock(&desc->bd_lock);
1660 rc = desc->bd_network_rw;
1661 cfs_spin_unlock(&desc->bd_lock);
1666 int ptlrpc_register_bulk(struct ptlrpc_request *req);
1667 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
1669 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
1671 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1674 LASSERT(req != NULL);
1676 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
1677 req->rq_bulk_deadline > cfs_time_current_sec())
1683 cfs_spin_lock(&desc->bd_lock);
1684 rc = desc->bd_network_rw;
1685 cfs_spin_unlock(&desc->bd_lock);
1689 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
1690 #define PTLRPC_REPLY_EARLY 0x02
1691 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
1692 int ptlrpc_reply(struct ptlrpc_request *req);
1693 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
1694 int ptlrpc_error(struct ptlrpc_request *req);
1695 void ptlrpc_resend_req(struct ptlrpc_request *request);
1696 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
1697 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
1698 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
1701 /* ptlrpc/client.c */
1703 * Client-side portals API. Everything to send requests, receive replies,
1704 * request queues, request management, etc.
1707 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
1708 struct ptlrpc_client *);
1709 void ptlrpc_cleanup_client(struct obd_import *imp);
1710 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
1712 int ptlrpc_queue_wait(struct ptlrpc_request *req);
1713 int ptlrpc_replay_req(struct ptlrpc_request *req);
1714 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
1715 void ptlrpc_restart_req(struct ptlrpc_request *req);
1716 void ptlrpc_abort_inflight(struct obd_import *imp);
1717 void ptlrpc_cleanup_imp(struct obd_import *imp);
1718 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
1720 struct ptlrpc_request_set *ptlrpc_prep_set(void);
1721 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1723 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1724 set_interpreter_func fn, void *data);
1725 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
1726 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
1727 int ptlrpc_set_wait(struct ptlrpc_request_set *);
1728 int ptlrpc_expired_set(void *data);
1729 void ptlrpc_interrupted_set(void *data);
1730 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
1731 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
1732 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
1733 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1734 struct ptlrpc_request *req);
1736 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
1737 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
1739 struct ptlrpc_request_pool *
1740 ptlrpc_init_rq_pool(int, int,
1741 void (*populate_pool)(struct ptlrpc_request_pool *, int));
1743 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
1744 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1745 const struct req_format *format);
1746 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
1747 struct ptlrpc_request_pool *,
1748 const struct req_format *format);
1749 void ptlrpc_request_free(struct ptlrpc_request *request);
1750 int ptlrpc_request_pack(struct ptlrpc_request *request,
1751 __u32 version, int opcode);
1752 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1753 const struct req_format *format,
1754 __u32 version, int opcode);
1755 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
1756 __u32 version, int opcode, char **bufs,
1757 struct ptlrpc_cli_ctx *ctx);
1758 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
1759 unsigned int timeout,
1760 ptlrpc_interpterer_t interpreter);
1761 void ptlrpc_fakereq_finished(struct ptlrpc_request *req);
1763 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
1764 int opcode, int count, __u32 *lengths,
1766 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
1767 __u32 version, int opcode,
1768 int count, __u32 *lengths, char **bufs,
1769 struct ptlrpc_request_pool *pool);
1770 void ptlrpc_req_finished(struct ptlrpc_request *request);
1771 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
1772 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
1773 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
1774 int npages, int type, int portal);
1775 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
1776 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
1777 cfs_page_t *page, int pageoffset, int len);
1778 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1779 struct obd_import *imp);
1780 __u64 ptlrpc_next_xid(void);
1781 __u64 ptlrpc_sample_next_xid(void);
1782 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
1784 /* Set of routines to run a function in ptlrpcd context */
1785 void *ptlrpcd_alloc_work(struct obd_import *imp,
1786 int (*cb)(const struct lu_env *, void *), void *data);
1787 void ptlrpcd_destroy_work(void *handler);
1788 int ptlrpcd_queue_work(void *handler);
1791 struct ptlrpc_service_buf_conf {
1792 /* nbufs is how many buffers to post */
1793 unsigned int bc_nbufs;
1794 /* buffer size to post */
1795 unsigned int bc_buf_size;
1796 /* portal to listed for requests on */
1797 unsigned int bc_req_portal;
1798 /* portal of where to send replies to */
1799 unsigned int bc_rep_portal;
1800 /* maximum request size to be accepted for this service */
1801 unsigned int bc_req_max_size;
1802 /* maximum reply size this service can ever send */
1803 unsigned int bc_rep_max_size;
1806 struct ptlrpc_service_thr_conf {
1807 /* threadname should be 8 characters or less - 6 will be added on */
1809 /* threads increasing factor for each CPU */
1810 unsigned int tc_thr_factor;
1811 /* service threads # to start on each partition while initializing */
1812 unsigned int tc_nthrs_init;
1814 * low water of threads # upper-limit on each partition while running,
1815 * service availability may be impacted if threads number is lower
1816 * than this value. It can be ZERO if the service doesn't require
1817 * CPU affinity or there is only one partition.
1819 unsigned int tc_nthrs_base;
1820 /* "soft" limit for total threads number */
1821 unsigned int tc_nthrs_max;
1822 /* user specified threads number, it will be validated due to
1823 * other members of this structure. */
1824 unsigned int tc_nthrs_user;
1825 /* set NUMA node affinity for service threads */
1826 unsigned int tc_cpu_affinity;
1827 /* Tags for lu_context associated with service thread */
1831 struct ptlrpc_service_cpt_conf {
1832 struct cfs_cpt_table *cc_cptable;
1833 /* string pattern to describe CPTs for a service */
1837 struct ptlrpc_service_conf {
1840 /* soft watchdog timeout multiplifier to print stuck service traces */
1841 unsigned int psc_watchdog_factor;
1842 /* buffer information */
1843 struct ptlrpc_service_buf_conf psc_buf;
1844 /* thread information */
1845 struct ptlrpc_service_thr_conf psc_thr;
1846 /* CPU partition information */
1847 struct ptlrpc_service_cpt_conf psc_cpt;
1848 /* function table */
1849 struct ptlrpc_service_ops psc_ops;
1852 /* ptlrpc/service.c */
1854 * Server-side services API. Register/unregister service, request state
1855 * management, service thread management
1859 void ptlrpc_save_lock(struct ptlrpc_request *req,
1860 struct lustre_handle *lock, int mode, int no_ack);
1861 void ptlrpc_commit_replies(struct obd_export *exp);
1862 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
1863 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
1864 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
1865 struct ptlrpc_service *ptlrpc_register_service(
1866 struct ptlrpc_service_conf *conf,
1867 struct proc_dir_entry *proc_entry);
1868 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
1870 int ptlrpc_start_threads(struct ptlrpc_service *svc);
1871 int ptlrpc_unregister_service(struct ptlrpc_service *service);
1872 int liblustre_check_services(void *arg);
1873 void ptlrpc_daemonize(char *name);
1874 int ptlrpc_service_health_check(struct ptlrpc_service *);
1875 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req);
1876 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
1879 int ptlrpc_hr_init(void);
1880 void ptlrpc_hr_fini(void);
1882 # define ptlrpc_hr_init() (0)
1883 # define ptlrpc_hr_fini() do {} while(0)
1888 /* ptlrpc/import.c */
1893 int ptlrpc_connect_import(struct obd_import *imp);
1894 int ptlrpc_init_import(struct obd_import *imp);
1895 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
1896 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
1897 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
1900 /* ptlrpc/pack_generic.c */
1901 int ptlrpc_reconnect_import(struct obd_import *imp);
1905 * ptlrpc msg buffer and swab interface
1909 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
1911 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
1913 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
1914 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
1916 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
1917 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
1919 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
1920 __u32 *lens, char **bufs);
1921 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
1923 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
1924 __u32 *lens, char **bufs, int flags);
1925 #define LPRFL_EARLY_REPLY 1
1926 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
1927 char **bufs, int flags);
1928 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
1929 unsigned int newlen, int move_data);
1930 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
1931 int __lustre_unpack_msg(struct lustre_msg *m, int len);
1932 int lustre_msg_hdr_size(__u32 magic, int count);
1933 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
1934 int lustre_msg_size_v2(int count, __u32 *lengths);
1935 int lustre_packed_msg_size(struct lustre_msg *msg);
1936 int lustre_msg_early_size(void);
1937 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
1938 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
1939 int lustre_msg_buflen(struct lustre_msg *m, int n);
1940 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
1941 int lustre_msg_bufcount(struct lustre_msg *m);
1942 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
1943 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
1944 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
1945 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
1946 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
1947 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
1948 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
1949 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
1950 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
1951 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
1952 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
1953 __u32 lustre_msg_get_type(struct lustre_msg *msg);
1954 __u32 lustre_msg_get_version(struct lustre_msg *msg);
1955 void lustre_msg_add_version(struct lustre_msg *msg, int version);
1956 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
1957 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
1958 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
1959 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
1960 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
1961 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
1962 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
1963 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
1964 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
1965 int lustre_msg_get_status(struct lustre_msg *msg);
1966 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
1967 int lustre_msg_is_v1(struct lustre_msg *msg);
1968 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
1969 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
1970 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
1971 char *lustre_msg_get_jobid(struct lustre_msg *msg);
1972 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
1973 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
1974 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
1976 # warning "remove checksum compatibility support for b1_8"
1977 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
1979 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
1980 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
1981 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
1982 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
1983 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
1984 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
1985 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
1986 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
1987 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
1988 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
1989 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
1990 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
1991 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
1992 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
1993 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
1996 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
1997 unsigned int newlen, int move_data)
1999 LASSERT(req->rq_reply_state);
2000 LASSERT(req->rq_repmsg);
2001 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2006 /** Change request phase of \a req to \a new_phase */
2008 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2010 if (req->rq_phase == new_phase)
2013 if (new_phase == RQ_PHASE_UNREGISTERING) {
2014 req->rq_next_phase = req->rq_phase;
2016 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2019 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2021 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2024 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2025 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2027 req->rq_phase = new_phase;
2031 * Returns true if request \a req got early reply and hard deadline is not met
2034 ptlrpc_client_early(struct ptlrpc_request *req)
2036 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2037 req->rq_reply_deadline > cfs_time_current_sec())
2039 return req->rq_early;
2043 * Returns true if we got real reply from server for this request
2046 ptlrpc_client_replied(struct ptlrpc_request *req)
2048 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2049 req->rq_reply_deadline > cfs_time_current_sec())
2051 return req->rq_replied;
2054 /** Returns true if request \a req is in process of receiving server reply */
2056 ptlrpc_client_recv(struct ptlrpc_request *req)
2058 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2059 req->rq_reply_deadline > cfs_time_current_sec())
2061 return req->rq_receiving_reply;
2065 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2069 cfs_spin_lock(&req->rq_lock);
2070 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2071 req->rq_reply_deadline > cfs_time_current_sec()) {
2072 cfs_spin_unlock(&req->rq_lock);
2075 rc = req->rq_receiving_reply || req->rq_must_unlink;
2076 cfs_spin_unlock(&req->rq_lock);
2081 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2083 if (req->rq_set == NULL)
2084 cfs_waitq_signal(&req->rq_reply_waitq);
2086 cfs_waitq_signal(&req->rq_set->set_waitq);
2090 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2092 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2093 cfs_atomic_inc(&rs->rs_refcount);
2097 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2099 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2100 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
2101 lustre_free_reply_state(rs);
2104 /* Should only be called once per req */
2105 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2107 if (req->rq_reply_state == NULL)
2108 return; /* shouldn't occur */
2109 ptlrpc_rs_decref(req->rq_reply_state);
2110 req->rq_reply_state = NULL;
2111 req->rq_repmsg = NULL;
2114 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2116 return lustre_msg_get_magic(req->rq_reqmsg);
2119 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2121 switch (req->rq_reqmsg->lm_magic) {
2122 case LUSTRE_MSG_MAGIC_V2:
2123 return req->rq_reqmsg->lm_repsize;
2125 LASSERTF(0, "incorrect message magic: %08x\n",
2126 req->rq_reqmsg->lm_magic);
2131 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2133 if (req->rq_delay_limit != 0 &&
2134 cfs_time_before(cfs_time_add(req->rq_queued_time,
2135 cfs_time_seconds(req->rq_delay_limit)),
2136 cfs_time_current())) {
2142 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2144 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2145 cfs_spin_lock(&req->rq_lock);
2146 req->rq_no_resend = 1;
2147 cfs_spin_unlock(&req->rq_lock);
2149 return req->rq_no_resend;
2153 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2155 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2157 return svcpt->scp_service->srv_watchdog_factor *
2158 max_t(int, at, obd_timeout);
2161 static inline struct ptlrpc_service *
2162 ptlrpc_req2svc(struct ptlrpc_request *req)
2164 LASSERT(req->rq_rqbd != NULL);
2165 return req->rq_rqbd->rqbd_svcpt->scp_service;
2168 /* ldlm/ldlm_lib.c */
2170 * Target client logic
2173 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2174 int client_obd_cleanup(struct obd_device *obddev);
2175 int client_connect_import(const struct lu_env *env,
2176 struct obd_export **exp, struct obd_device *obd,
2177 struct obd_uuid *cluuid, struct obd_connect_data *,
2179 int client_disconnect_export(struct obd_export *exp);
2180 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2182 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2183 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2184 struct obd_uuid *uuid);
2185 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2186 void client_destroy_import(struct obd_import *imp);
2189 #ifdef HAVE_SERVER_SUPPORT
2190 int server_disconnect_export(struct obd_export *exp);
2193 /* ptlrpc/pinger.c */
2195 * Pinger API (client side only)
2198 enum timeout_event {
2201 struct timeout_item;
2202 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2203 int ptlrpc_pinger_add_import(struct obd_import *imp);
2204 int ptlrpc_pinger_del_import(struct obd_import *imp);
2205 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
2206 timeout_cb_t cb, void *data,
2207 cfs_list_t *obd_list);
2208 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
2209 enum timeout_event event);
2210 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2211 int ptlrpc_obd_ping(struct obd_device *obd);
2212 cfs_time_t ptlrpc_suspend_wakeup_time(void);
2214 void ping_evictor_start(void);
2215 void ping_evictor_stop(void);
2217 #define ping_evictor_start() do {} while (0)
2218 #define ping_evictor_stop() do {} while (0)
2220 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
2223 /* ptlrpc daemon bind policy */
2225 /* all ptlrpcd threads are free mode */
2226 PDB_POLICY_NONE = 1,
2227 /* all ptlrpcd threads are bound mode */
2228 PDB_POLICY_FULL = 2,
2229 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
2230 PDB_POLICY_PAIR = 3,
2231 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
2232 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
2233 * If kernel supports NUMA, pthrpcd threads are binded and
2234 * grouped by NUMA node */
2235 PDB_POLICY_NEIGHBOR = 4,
2238 /* ptlrpc daemon load policy
2239 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
2240 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
2241 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
2242 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
2243 * depends on which is scheduled firstly, to accelerate the RPC processing. */
2245 /* on the same CPU core as the caller */
2246 PDL_POLICY_SAME = 1,
2247 /* within the same CPU partition, but not the same core as the caller */
2248 PDL_POLICY_LOCAL = 2,
2249 /* round-robin on all CPU cores, but not the same core as the caller */
2250 PDL_POLICY_ROUND = 3,
2251 /* the specified CPU core is preferred, but not enforced */
2252 PDL_POLICY_PREFERRED = 4,
2255 /* ptlrpc/ptlrpcd.c */
2256 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2257 void ptlrpcd_wake(struct ptlrpc_request *req);
2258 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
2259 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2260 int ptlrpcd_addref(void);
2261 void ptlrpcd_decref(void);
2263 /* ptlrpc/lproc_ptlrpc.c */
2265 * procfs output related functions
2268 const char* ll_opcode2str(__u32 opcode);
2270 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2271 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2272 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2274 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2275 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2276 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2280 /* ptlrpc/llog_server.c */
2281 int llog_origin_handle_open(struct ptlrpc_request *req);
2282 int llog_origin_handle_destroy(struct ptlrpc_request *req);
2283 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2284 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2285 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2286 int llog_origin_handle_close(struct ptlrpc_request *req);
2287 int llog_origin_handle_cancel(struct ptlrpc_request *req);
2289 /* ptlrpc/llog_client.c */
2290 extern struct llog_operations llog_client_ops;