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) 2010, 2012, Intel Corporation.
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 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
94 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
95 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
96 #define DT_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
97 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
99 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
101 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
102 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
104 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE))
105 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE"
107 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU)
108 # error "PTLRPC_MAX_BRW_SIZE too big"
110 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV)
111 # error "PTLRPC_MAX_BRW_PAGES too big"
113 #endif /* __KERNEL__ */
115 #define PTLRPC_NTHRS_INIT 2
120 * Constants determine how memory is used to buffer incoming service requests.
122 * ?_NBUFS # buffers to allocate when growing the pool
123 * ?_BUFSIZE # bytes in a single request buffer
124 * ?_MAXREQSIZE # maximum request service will receive
126 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
127 * of ?_NBUFS is added to the pool.
129 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
130 * considered full when less than ?_MAXREQSIZE is left in them.
135 * Constants determine how threads are created for ptlrpc service.
137 * ?_NTHRS_INIT # threads to create for each service partition on
138 * initializing. If it's non-affinity service and
139 * there is only one partition, it's the overall #
140 * threads for the service while initializing.
141 * ?_NTHRS_BASE # threads should be created at least for each
142 * ptlrpc partition to keep the service healthy.
143 * It's the low-water mark of threads upper-limit
144 * for each partition.
145 * ?_THR_FACTOR # threads can be added on threads upper-limit for
146 * each CPU core. This factor is only for reference,
147 * we might decrease value of factor if number of cores
148 * per CPT is above a limit.
149 * ?_NTHRS_MAX # overall threads can be created for a service,
150 * it's a soft limit because if service is running
151 * on machine with hundreds of cores and tens of
152 * CPU partitions, we need to guarantee each partition
153 * has ?_NTHRS_BASE threads, which means total threads
154 * will be ?_NTHRS_BASE * number_of_cpts which can
155 * exceed ?_NTHRS_MAX.
159 * #define MDS_NTHRS_INIT 2
160 * #define MDS_NTHRS_BASE 64
161 * #define MDS_NTHRS_FACTOR 8
162 * #define MDS_NTHRS_MAX 1024
165 * ---------------------------------------------------------------------
166 * Server(A) has 16 cores, user configured it to 4 partitions so each
167 * partition has 4 cores, then actual number of service threads on each
169 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
171 * Total number of threads for the service is:
172 * 96 * partitions(4) = 384
175 * ---------------------------------------------------------------------
176 * Server(B) has 32 cores, user configured it to 4 partitions so each
177 * partition has 8 cores, then actual number of service threads on each
179 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
181 * Total number of threads for the service is:
182 * 128 * partitions(4) = 512
185 * ---------------------------------------------------------------------
186 * Server(B) has 96 cores, user configured it to 8 partitions so each
187 * partition has 12 cores, then actual number of service threads on each
189 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
191 * Total number of threads for the service is:
192 * 160 * partitions(8) = 1280
194 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
195 * as upper limit of threads number for each partition:
196 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
199 * ---------------------------------------------------------------------
200 * Server(C) have a thousand of cores and user configured it to 32 partitions
201 * MDS_NTHRS_BASE(64) * 32 = 2048
203 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
204 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
205 * to keep service healthy, so total number of threads will just be 2048.
207 * NB: we don't suggest to choose server with that many cores because backend
208 * filesystem itself, buffer cache, or underlying network stack might
209 * have some SMP scalability issues at that large scale.
211 * If user already has a fat machine with hundreds or thousands of cores,
212 * there are two choices for configuration:
213 * a) create CPU table from subset of all CPUs and run Lustre on
215 * b) bind service threads on a few partitions, see modparameters of
216 * MDS and OSS for details
218 * NB: these calculations (and examples below) are simplified to help
219 * understanding, the real implementation is a little more complex,
220 * please see ptlrpc_server_nthreads_check() for details.
225 * LDLM threads constants:
227 * Given 8 as factor and 24 as base threads number
230 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
233 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
234 * threads for each partition and total threads number will be 112.
237 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
238 * threads for each partition to keep service healthy, so total threads
239 * number should be 24 * 8 = 192.
241 * So with these constants, threads number wil be at the similar level
242 * of old versions, unless target machine has over a hundred cores
244 #define LDLM_THR_FACTOR 8
245 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
246 #define LDLM_NTHRS_BASE 24
247 #define LDLM_NTHRS_MAX (cfs_num_online_cpus() == 1 ? 64 : 128)
249 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
250 #define LDLM_NBUFS (64 * cfs_num_online_cpus())
251 #define LDLM_BUFSIZE (8 * 1024)
252 #define LDLM_MAXREQSIZE (5 * 1024)
253 #define LDLM_MAXREPSIZE (1024)
256 * MDS threads constants:
258 * Please see examples in "Thread Constants", MDS threads number will be at
259 * the comparable level of old versions, unless the server has many cores.
261 #ifndef MDS_MAX_THREADS
262 #define MDS_MAX_THREADS 1024
263 #define MDS_MAX_OTHR_THREADS 256
265 #else /* MDS_MAX_THREADS */
266 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
267 #undef MDS_MAX_THREADS
268 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
270 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
273 /* default service */
274 #define MDS_THR_FACTOR 8
275 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
276 #define MDS_NTHRS_MAX MDS_MAX_THREADS
277 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
279 /* read-page service */
280 #define MDS_RDPG_THR_FACTOR 4
281 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
282 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
283 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
285 /* these should be removed when we remove setattr service in the future */
286 #define MDS_SETA_THR_FACTOR 4
287 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
288 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
289 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
291 /* non-affinity threads */
292 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
293 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
295 #define MDS_NBUFS (64 * cfs_num_online_cpus())
297 * Assume file name length = FNAME_MAX = 256 (true for ext3).
298 * path name length = PATH_MAX = 4096
299 * LOV MD size max = EA_MAX = 48000 (2000 stripes)
300 * symlink: FNAME_MAX + PATH_MAX <- largest
301 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
302 * rename: FNAME_MAX + FNAME_MAX
303 * open: FNAME_MAX + EA_MAX
305 * MDS_MAXREQSIZE ~= 4736 bytes =
306 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
307 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
308 * or, for mds_close() and mds_reint_unlink() on a many-OST filesystem:
309 * = 9210 bytes = lustre_msg + mdt_body + 160 * (easize + cookiesize)
311 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
312 * except in the open case where there are a large number of OSTs in a LOV.
314 #define MDS_MAXREPSIZE max(10 * 1024, 362 + LOV_MAX_STRIPE_COUNT * 56)
315 #define MDS_MAXREQSIZE MDS_MAXREPSIZE
317 /** MDS_BUFSIZE = max_reqsize + max sptlrpc payload size */
318 #define MDS_BUFSIZE (MDS_MAXREQSIZE + 1024)
320 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
321 #define FLD_MAXREQSIZE (160)
323 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
324 #define FLD_MAXREPSIZE (152)
327 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
329 #define SEQ_MAXREQSIZE (160)
331 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
332 #define SEQ_MAXREPSIZE (152)
334 /** MGS threads must be >= 3, see bug 22458 comment #28 */
335 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
336 #define MGS_NTHRS_MAX 32
338 #define MGS_NBUFS (64 * cfs_num_online_cpus())
339 #define MGS_BUFSIZE (8 * 1024)
340 #define MGS_MAXREQSIZE (7 * 1024)
341 #define MGS_MAXREPSIZE (9 * 1024)
344 * OSS threads constants:
346 * Given 8 as factor and 64 as base threads number
349 * On 8-core server configured to 2 partitions, we will have
350 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
353 * On 32-core machine configured to 4 partitions, we will have
354 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
355 * will be 112 * 4 = 448.
358 * On 64-core machine configured to 4 partitions, we will have
359 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
360 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
361 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
362 * for each partition.
364 * So we can see that with these constants, threads number wil be at the
365 * similar level of old versions, unless the server has many cores.
367 /* depress threads factor for VM with small memory size */
368 #define OSS_THR_FACTOR min_t(int, 8, \
369 CFS_NUM_CACHEPAGES >> (28 - CFS_PAGE_SHIFT))
370 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
371 #define OSS_NTHRS_BASE 64
372 #define OSS_NTHRS_MAX 512
374 /* threads for handling "create" request */
375 #define OSS_CR_THR_FACTOR 1
376 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
377 #define OSS_CR_NTHRS_BASE 8
378 #define OSS_CR_NTHRS_MAX 64
380 #define OST_NBUFS (64 * cfs_num_online_cpus())
381 #define OST_BUFSIZE (8 * 1024)
384 * OST_MAXREQSIZE ~= 4768 bytes =
385 * lustre_msg + obdo + 16 * obd_ioobj + 256 * niobuf_remote
387 * - single object with 16 pages is 512 bytes
388 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
390 #define OST_MAXREQSIZE (5 * 1024)
391 #define OST_MAXREPSIZE (9 * 1024)
393 /* Macro to hide a typecast. */
394 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
397 * Structure to single define portal connection.
399 struct ptlrpc_connection {
400 /** linkage for connections hash table */
401 cfs_hlist_node_t c_hash;
402 /** Our own lnet nid for this connection */
404 /** Remote side nid for this connection */
405 lnet_process_id_t c_peer;
406 /** UUID of the other side */
407 struct obd_uuid c_remote_uuid;
408 /** reference counter for this connection */
409 cfs_atomic_t c_refcount;
412 /** Client definition for PortalRPC */
413 struct ptlrpc_client {
414 /** What lnet portal does this client send messages to by default */
415 __u32 cli_request_portal;
416 /** What portal do we expect replies on */
417 __u32 cli_reply_portal;
418 /** Name of the client */
422 /** state flags of requests */
423 /* XXX only ones left are those used by the bulk descs as well! */
424 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
425 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
427 #define REQ_MAX_ACK_LOCKS 8
429 union ptlrpc_async_args {
431 * Scratchpad for passing args to completion interpreter. Users
432 * cast to the struct of their choosing, and CLASSERT that this is
433 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
434 * a pointer to it here. The pointer_arg ensures this struct is at
435 * least big enough for that.
437 void *pointer_arg[11];
441 struct ptlrpc_request_set;
442 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
443 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
446 * Definition of request set structure.
447 * Request set is a list of requests (not necessary to the same target) that
448 * once populated with RPCs could be sent in parallel.
449 * There are two kinds of request sets. General purpose and with dedicated
450 * serving thread. Example of the latter is ptlrpcd set.
451 * For general purpose sets once request set started sending it is impossible
452 * to add new requests to such set.
453 * Provides a way to call "completion callbacks" when all requests in the set
456 struct ptlrpc_request_set {
457 cfs_atomic_t set_refcount;
458 /** number of in queue requests */
459 cfs_atomic_t set_new_count;
460 /** number of uncompleted requests */
461 cfs_atomic_t set_remaining;
462 /** wait queue to wait on for request events */
463 cfs_waitq_t set_waitq;
464 cfs_waitq_t *set_wakeup_ptr;
465 /** List of requests in the set */
466 cfs_list_t set_requests;
468 * List of completion callbacks to be called when the set is completed
469 * This is only used if \a set_interpret is NULL.
470 * Links struct ptlrpc_set_cbdata.
472 cfs_list_t set_cblist;
473 /** Completion callback, if only one. */
474 set_interpreter_func set_interpret;
475 /** opaq argument passed to completion \a set_interpret callback. */
478 * Lock for \a set_new_requests manipulations
479 * locked so that any old caller can communicate requests to
480 * the set holder who can then fold them into the lock-free set
482 spinlock_t set_new_req_lock;
483 /** List of new yet unsent requests. Only used with ptlrpcd now. */
484 cfs_list_t set_new_requests;
486 /** rq_status of requests that have been freed already */
488 /** Additional fields used by the flow control extension */
489 /** Maximum number of RPCs in flight */
490 int set_max_inflight;
491 /** Callback function used to generate RPCs */
492 set_producer_func set_producer;
493 /** opaq argument passed to the producer callback */
494 void *set_producer_arg;
498 * Description of a single ptrlrpc_set callback
500 struct ptlrpc_set_cbdata {
501 /** List linkage item */
503 /** Pointer to interpreting function */
504 set_interpreter_func psc_interpret;
505 /** Opaq argument to pass to the callback */
509 struct ptlrpc_bulk_desc;
510 struct ptlrpc_service_part;
511 struct ptlrpc_service;
514 * ptlrpc callback & work item stuff
516 struct ptlrpc_cb_id {
517 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
518 void *cbid_arg; /* additional arg */
521 /** Maximum number of locks to fit into reply state */
522 #define RS_MAX_LOCKS 8
526 * Structure to define reply state on the server
527 * Reply state holds various reply message information. Also for "difficult"
528 * replies (rep-ack case) we store the state after sending reply and wait
529 * for the client to acknowledge the reception. In these cases locks could be
530 * added to the state for replay/failover consistency guarantees.
532 struct ptlrpc_reply_state {
533 /** Callback description */
534 struct ptlrpc_cb_id rs_cb_id;
535 /** Linkage for list of all reply states in a system */
537 /** Linkage for list of all reply states on same export */
538 cfs_list_t rs_exp_list;
539 /** Linkage for list of all reply states for same obd */
540 cfs_list_t rs_obd_list;
542 cfs_list_t rs_debug_list;
544 /** A spinlock to protect the reply state flags */
546 /** Reply state flags */
547 unsigned long rs_difficult:1; /* ACK/commit stuff */
548 unsigned long rs_no_ack:1; /* no ACK, even for
549 difficult requests */
550 unsigned long rs_scheduled:1; /* being handled? */
551 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
552 unsigned long rs_handled:1; /* been handled yet? */
553 unsigned long rs_on_net:1; /* reply_out_callback pending? */
554 unsigned long rs_prealloc:1; /* rs from prealloc list */
555 unsigned long rs_committed:1;/* the transaction was committed
556 and the rs was dispatched
557 by ptlrpc_commit_replies */
558 /** Size of the state */
562 /** Transaction number */
566 struct obd_export *rs_export;
567 struct ptlrpc_service_part *rs_svcpt;
568 /** Lnet metadata handle for the reply */
569 lnet_handle_md_t rs_md_h;
570 cfs_atomic_t rs_refcount;
572 /** Context for the sevice thread */
573 struct ptlrpc_svc_ctx *rs_svc_ctx;
574 /** Reply buffer (actually sent to the client), encoded if needed */
575 struct lustre_msg *rs_repbuf; /* wrapper */
576 /** Size of the reply buffer */
577 int rs_repbuf_len; /* wrapper buf length */
578 /** Size of the reply message */
579 int rs_repdata_len; /* wrapper msg length */
581 * Actual reply message. Its content is encrupted (if needed) to
582 * produce reply buffer for actual sending. In simple case
583 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
585 struct lustre_msg *rs_msg; /* reply message */
587 /** Number of locks awaiting client ACK */
589 /** Handles of locks awaiting client reply ACK */
590 struct lustre_handle rs_locks[RS_MAX_LOCKS];
591 /** Lock modes of locks in \a rs_locks */
592 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
595 struct ptlrpc_thread;
599 RQ_PHASE_NEW = 0xebc0de00,
600 RQ_PHASE_RPC = 0xebc0de01,
601 RQ_PHASE_BULK = 0xebc0de02,
602 RQ_PHASE_INTERPRET = 0xebc0de03,
603 RQ_PHASE_COMPLETE = 0xebc0de04,
604 RQ_PHASE_UNREGISTERING = 0xebc0de05,
605 RQ_PHASE_UNDEFINED = 0xebc0de06
608 /** Type of request interpreter call-back */
609 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
610 struct ptlrpc_request *req,
614 * Definition of request pool structure.
615 * The pool is used to store empty preallocated requests for the case
616 * when we would actually need to send something without performing
617 * any allocations (to avoid e.g. OOM).
619 struct ptlrpc_request_pool {
620 /** Locks the list */
622 /** list of ptlrpc_request structs */
623 cfs_list_t prp_req_list;
624 /** Maximum message size that would fit into a rquest from this pool */
626 /** Function to allocate more requests for this pool */
627 void (*prp_populate)(struct ptlrpc_request_pool *, int);
636 * \defgroup nrs Network Request Scheduler
639 struct ptlrpc_nrs_policy;
640 struct ptlrpc_nrs_resource;
641 struct ptlrpc_nrs_request;
644 * NRS control operations.
646 * These are common for all policies.
648 enum ptlrpc_nrs_ctl {
650 * Activate the policy.
652 PTLRPC_NRS_CTL_START,
654 * Reserved for multiple primary policies, which may be a possibility
659 * Recycle resources for inactive policies.
661 PTLRPC_NRS_CTL_SHRINK,
663 * Not a valid opcode.
665 PTLRPC_NRS_CTL_INVALID,
667 * Policies can start using opcodes from this value and onwards for
668 * their own purposes; the assigned value itself is arbitrary.
670 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
674 * NRS policy operations.
676 * These determine the behaviour of a policy, and are called in response to
679 struct ptlrpc_nrs_pol_ops {
681 * Called during policy registration; this operation is optional.
683 * \param[in] policy The policy being initialized
685 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
687 * Called during policy unregistration; this operation is optional.
689 * \param[in] policy The policy being unregistered/finalized
691 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
693 * Called when activating a policy via lprocfs; policies allocate and
694 * initialize their resources here; this operation is optional.
696 * \param[in] policy The policy being started
698 * \see nrs_policy_start_locked()
700 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
702 * Called when deactivating a policy via lprocfs; policies deallocate
703 * their resources here; this operation is optional
705 * \param[in] policy The policy being stopped
707 * \see nrs_policy_stop_final()
709 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
711 * Used for policy-specific operations; i.e. not generic ones like
712 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
713 * to an ioctl; this operation is optional.
715 * \param[in] policy The policy carrying out operation \a opc
716 * \param[in] opc The command operation being carried out
717 * \param[in,out] arg An generic buffer for communication between the
718 * user and the control operation
723 * \see ptlrpc_nrs_policy_control()
725 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
726 enum ptlrpc_nrs_ctl opc, void *arg);
729 * Called when obtaining references to the resources of the resource
730 * hierarchy for a request that has arrived for handling at the PTLRPC
731 * service. Policies should return -ve for requests they do not wish
732 * to handle. This operation is mandatory.
734 * \param[in] policy The policy we're getting resources for.
735 * \param[in] nrq The request we are getting resources for.
736 * \param[in] parent The parent resource of the resource being
737 * requested; set to NULL if none.
738 * \param[out] resp The resource is to be returned here; the
739 * fallback policy in an NRS head should
740 * \e always return a non-NULL pointer value.
741 * \param[in] moving_req When set, signifies that this is an attempt
742 * to obtain resources for a request being moved
743 * to the high-priority NRS head by
744 * ldlm_lock_reorder_req().
745 * This implies two things:
746 * 1. We are under obd_export::exp_rpc_lock and
747 * so should not sleep.
748 * 2. We should not perform non-idempotent or can
749 * skip performing idempotent operations that
750 * were carried out when resources were first
751 * taken for the request when it was initialized
752 * in ptlrpc_nrs_req_initialize().
754 * \retval 0, +ve The level of the returned resource in the resource
755 * hierarchy; currently only 0 (for a non-leaf resource)
756 * and 1 (for a leaf resource) are supported by the
760 * \see ptlrpc_nrs_req_initialize()
761 * \see ptlrpc_nrs_hpreq_add_nolock()
762 * \see ptlrpc_nrs_req_hp_move()
764 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
765 struct ptlrpc_nrs_request *nrq,
766 struct ptlrpc_nrs_resource *parent,
767 struct ptlrpc_nrs_resource **resp,
770 * Called when releasing references taken for resources in the resource
771 * hierarchy for the request; this operation is optional.
773 * \param[in] policy The policy the resource belongs to
774 * \param[in] res The resource to be freed
776 * \see ptlrpc_nrs_req_finalize()
777 * \see ptlrpc_nrs_hpreq_add_nolock()
778 * \see ptlrpc_nrs_req_hp_move()
780 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
781 struct ptlrpc_nrs_resource *res);
784 * Obtain a request for handling from the policy via polling; this
785 * operation is mandatory.
787 * \param[in] policy The policy to poll
789 * \retval NULL No erquest available for handling
790 * \retval valid-pointer The request polled for handling
792 * \see ptlrpc_nrs_req_poll_nolock()
794 struct ptlrpc_nrs_request *
795 (*op_req_poll) (struct ptlrpc_nrs_policy *policy);
797 * Called when attempting to add a request to a policy for later
798 * handling; this operation is mandatory.
800 * \param[in] policy The policy on which to enqueue \a nrq
801 * \param[in] nrq The request to enqueue
806 * \see ptlrpc_nrs_req_add_nolock()
808 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
809 struct ptlrpc_nrs_request *nrq);
811 * Removes a request from the policy's set of pending requests. Normally
812 * called after a request has been polled successfully from the policy
813 * for handling; this operation is mandatory.
815 * \param[in] policy The policy the request \a nrq belongs to
816 * \param[in] nrq The request to dequeue
818 * \see ptlrpc_nrs_req_del_nolock()
820 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
821 struct ptlrpc_nrs_request *nrq);
823 * Called before carrying out the request; should not block. Could be
824 * used for job/resource control; this operation is optional.
826 * \param[in] policy The policy which is starting to handle request
828 * \param[in] nrq The request
830 * \pre spin_is_locked(&svcpt->scp_req_lock)
832 * \see ptlrpc_nrs_req_start_nolock()
834 void (*op_req_start) (struct ptlrpc_nrs_policy *policy,
835 struct ptlrpc_nrs_request *nrq);
837 * Called after the request being carried out. Could be used for
838 * job/resource control; this operation is optional.
840 * \param[in] policy The policy which is stopping to handle request
842 * \param[in] nrq The request
844 * \pre spin_is_locked(&svcpt->scp_req_lock)
846 * \see ptlrpc_nrs_req_stop_nolock()
848 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
849 struct ptlrpc_nrs_request *nrq);
851 * Registers the policy's lprocfs interface with a PTLRPC service.
853 * \param[in] svc The service
858 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
860 * Unegisters the policy's lprocfs interface with a PTLRPC service.
862 * \param[in] svc The service
864 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
870 enum nrs_policy_flags {
872 * Fallback policy, use this flag only on a single supported policy per
873 * service. Do not use this flag for policies registering using
874 * ptlrpc_nrs_policy_register() (i.e. ones that are not in
875 * \e nrs_pols_builtin).
877 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
879 * Start policy immediately after registering.
881 PTLRPC_NRS_FL_REG_START = (1 << 1),
883 * This is a polciy registering externally with NRS core, via
884 * ptlrpc_nrs_policy_register(), (i.e. one that is not in
885 * \e nrs_pols_builtin. Used to avoid ptlrpc_nrs_policy_register()
886 * racing with a policy start operation issued by the user via lprocfs.
888 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
894 * Denotes whether an NRS instance is for handling normal or high-priority
895 * RPCs, or whether an operation pertains to one or both of the NRS instances
898 enum ptlrpc_nrs_queue_type {
899 PTLRPC_NRS_QUEUE_REG,
901 PTLRPC_NRS_QUEUE_BOTH,
907 * A PTLRPC service has at least one NRS head instance for handling normal
908 * priority RPCs, and may optionally have a second NRS head instance for
909 * handling high-priority RPCs. Each NRS head maintains a list of available
910 * policies, of which one and only one policy is acting as the fallback policy,
911 * and optionally a different policy may be acting as the primary policy. For
912 * all RPCs handled by this NRS head instance, NRS core will first attempt to
913 * enqueue the RPC using the primary policy (if any). The fallback policy is
914 * used in the following cases:
915 * - when there was no primary policy in the
916 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
918 * - when the primary policy that was at the
919 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
920 * RPC was initialized, denoted it did not wish, or for some other reason was
921 * not able to handle the request, by returning a non-valid NRS resource
923 * - when the primary policy that was at the
924 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
925 * RPC was initialized, fails later during the request enqueueing stage.
927 * \see nrs_resource_get_safe()
928 * \see nrs_request_enqueue()
932 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
934 * Linkage into nrs_core_heads_list
936 cfs_list_t nrs_heads;
938 * List of registered policies
940 cfs_list_t nrs_policy_list;
942 * List of policies with queued requests. Policies that have any
943 * outstanding requests are queued here, and this list is queried
944 * in a round-robin manner from NRS core when obtaining a request
945 * for handling. This ensures that requests from policies that at some
946 * point transition away from the
947 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
949 cfs_list_t nrs_policy_queued;
951 * Service partition for this NRS head
953 struct ptlrpc_service_part *nrs_svcpt;
955 * Primary policy, which is the preferred policy for handling RPCs
957 struct ptlrpc_nrs_policy *nrs_policy_primary;
959 * Fallback policy, which is the backup policy for handling RPCs
961 struct ptlrpc_nrs_policy *nrs_policy_fallback;
963 * This NRS head handles either HP or regular requests
965 enum ptlrpc_nrs_queue_type nrs_queue_type;
967 * # queued requests from all policies in this NRS head
969 unsigned long nrs_req_queued;
971 * # scheduled requests from all policies in this NRS head
973 unsigned long nrs_req_started;
975 * # policies on this NRS
976 * TODO: Can we avoid having this?
978 unsigned nrs_num_pols;
980 * This NRS head is in progress of starting a policy
982 unsigned nrs_policy_starting:1;
984 * In progress of shutting down the whole NRS head; used during
987 unsigned nrs_stopping:1;
990 #define NRS_POL_NAME_MAX 16
993 * NRS policy registering descriptor
995 * Is used to hold a description of a policy that can be passed to NRS core in
996 * order to register the policy with NRS heads in different PTLRPC services.
998 struct ptlrpc_nrs_pol_desc {
1000 * Human-readable policy name
1002 char pd_name[NRS_POL_NAME_MAX];
1004 * NRS operations for this policy
1006 struct ptlrpc_nrs_pol_ops *pd_ops;
1008 * Service Compatibility function; this determines whether a policy is
1009 * adequate for handling RPCs of a particular PTLRPC service.
1011 * XXX:This should give the same result during policy
1012 * registration and unregistration, and for all partitions of a
1013 * service; so the result should not depend on temporal service
1014 * or other properties, that may influence the result.
1016 bool (*pd_compat) (struct ptlrpc_service *svc,
1017 const struct ptlrpc_nrs_pol_desc *desc);
1019 * Optionally set for policies that support a single ptlrpc service,
1020 * i.e. ones that have \a pd_compat set to nrs_policy_compat_one()
1022 char *pd_compat_svc_name;
1024 * Bitmask of nrs_policy_flags
1028 * Link into nrs_core::nrs_policies
1036 * Policies transition from one state to the other during their lifetime
1038 enum ptlrpc_nrs_pol_state {
1040 * Not a valid policy state.
1042 NRS_POL_STATE_INVALID,
1044 * For now, this state is used exclusively for policies that register
1045 * externally to NRS core, i.e. ones that do so via
1046 * ptlrpc_nrs_policy_register() and are not part of nrs_pols_builtin;
1047 * it is used to prevent a race condition between the policy registering
1048 * with more than one service partition while service is operational,
1049 * and the user starting the policy via lprocfs.
1051 * \see nrs_pol_make_avail()
1053 NRS_POL_STATE_UNAVAIL,
1055 * Policies are at this state either at the start of their life, or
1056 * transition here when the user selects a different policy to act
1057 * as the primary one.
1059 NRS_POL_STATE_STOPPED,
1061 * Policy is progress of stopping
1063 NRS_POL_STATE_STOPPING,
1065 * Policy is in progress of starting
1067 NRS_POL_STATE_STARTING,
1069 * A policy is in this state in two cases:
1070 * - it is the fallback policy, which is always in this state.
1071 * - it has been activated by the user; i.e. it is the primary policy,
1073 NRS_POL_STATE_STARTED,
1077 * NRS policy information
1079 * Used for obtaining information for the status of a policy via lprocfs
1081 struct ptlrpc_nrs_pol_info {
1085 char pi_name[NRS_POL_NAME_MAX];
1087 * Current policy state
1089 enum ptlrpc_nrs_pol_state pi_state;
1091 * # RPCs enqueued for later dispatching by the policy
1095 * # RPCs started for dispatch by the policy
1097 long pi_req_started;
1099 * Is this a fallback policy?
1101 unsigned pi_fallback:1;
1107 * There is one instance of this for each policy in each NRS head of each
1108 * PTLRPC service partition.
1110 struct ptlrpc_nrs_policy {
1112 * Linkage into the NRS head's list of policies,
1113 * ptlrpc_nrs:nrs_policy_list
1115 cfs_list_t pol_list;
1117 * Linkage into the NRS head's list of policies with enqueued
1118 * requests ptlrpc_nrs:nrs_policy_queued
1120 cfs_list_t pol_list_queued;
1122 * Current state of this policy
1124 enum ptlrpc_nrs_pol_state pol_state;
1126 * Bitmask of nrs_policy_flags
1130 * # RPCs enqueued for later dispatching by the policy
1132 long pol_req_queued;
1134 * # RPCs started for dispatch by the policy
1136 long pol_req_started;
1138 * Usage Reference count taken on the policy instance
1142 * The NRS head this policy has been created at
1144 struct ptlrpc_nrs *pol_nrs;
1146 * NRS operations for this policy; points to ptlrpc_nrs_pol_desc::pd_ops
1148 struct ptlrpc_nrs_pol_ops *pol_ops;
1150 * Private policy data; varies by policy type
1154 * Human-readable policy name; point to ptlrpc_nrs_pol_desc::pd_name
1162 * Resources are embedded into two types of NRS entities:
1163 * - Inside NRS policies, in the policy's private data in
1164 * ptlrpc_nrs_policy::pol_private
1165 * - In objects that act as prime-level scheduling entities in different NRS
1166 * policies; e.g. on a policy that performs round robin or similar order
1167 * scheduling across client NIDs, there would be one NRS resource per unique
1168 * client NID. On a policy which performs round robin scheduling across
1169 * backend filesystem objects, there would be one resource associated with
1170 * each of the backend filesystem objects partaking in the scheduling
1171 * performed by the policy.
1173 * NRS resources share a parent-child relationship, in which resources embedded
1174 * in policy instances are the parent entities, with all scheduling entities
1175 * a policy schedules across being the children, thus forming a simple resource
1176 * hierarchy. This hierarchy may be extended with one or more levels in the
1177 * future if the ability to have more than one primary policy is added.
1179 * Upon request initialization, references to the then active NRS policies are
1180 * taken and used to later handle the dispatching of the request with one of
1183 * \see nrs_resource_get_safe()
1184 * \see ptlrpc_nrs_req_add()
1186 struct ptlrpc_nrs_resource {
1188 * This NRS resource's parent; is NULL for resources embedded in NRS
1189 * policy instances; i.e. those are top-level ones.
1191 struct ptlrpc_nrs_resource *res_parent;
1193 * The policy associated with this resource.
1195 struct ptlrpc_nrs_policy *res_policy;
1208 * This policy is a logical wrapper around previous, non-NRS functionality.
1209 * It dispatches RPCs in the same order as they arrive from the network. This
1210 * policy is currently used as the fallback policy, and the only enabled policy
1211 * on all NRS heads of all PTLRPC service partitions.
1216 * Private data structure for the FIFO policy
1218 struct nrs_fifo_head {
1220 * Resource object for policy instance.
1222 struct ptlrpc_nrs_resource fh_res;
1224 * List of queued requests.
1228 * For debugging purposes.
1233 struct nrs_fifo_req {
1234 /** request header, must be the first member of structure */
1244 * Instances of this object exist embedded within ptlrpc_request; the main
1245 * purpose of this object is to hold references to the request's resources
1246 * for the lifetime of the request, and to hold properties that policies use
1247 * use for determining the request's scheduling priority.
1249 struct ptlrpc_nrs_request {
1251 * The request's resource hierarchy.
1253 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1255 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1256 * policy that was used to enqueue the request.
1258 * \see nrs_request_enqueue()
1260 unsigned nr_res_idx;
1261 unsigned nr_initialized:1;
1262 unsigned nr_enqueued:1;
1263 unsigned nr_dequeued:1;
1264 unsigned nr_started:1;
1265 unsigned nr_finalized:1;
1266 cfs_binheap_node_t nr_node;
1269 * Policy-specific fields, used for determining a request's scheduling
1270 * priority, and other supporting functionality.
1274 * Fields for the FIFO policy
1276 struct nrs_fifo_req fifo;
1279 * Externally-registering policies may want to use this to allocate
1280 * their own request properties.
1288 * Basic request prioritization operations structure.
1289 * The whole idea is centered around locks and RPCs that might affect locks.
1290 * When a lock is contended we try to give priority to RPCs that might lead
1291 * to fastest release of that lock.
1292 * Currently only implemented for OSTs only in a way that makes all
1293 * IO and truncate RPCs that are coming from a locked region where a lock is
1294 * contended a priority over other requests.
1296 struct ptlrpc_hpreq_ops {
1298 * Check if the lock handle of the given lock is the same as
1299 * taken from the request.
1301 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1303 * Check if the request is a high priority one.
1305 int (*hpreq_check)(struct ptlrpc_request *);
1307 * Called after the request has been handled.
1309 void (*hpreq_fini)(struct ptlrpc_request *);
1313 * Represents remote procedure call.
1315 * This is a staple structure used by everybody wanting to send a request
1318 struct ptlrpc_request {
1319 /* Request type: one of PTL_RPC_MSG_* */
1321 /** Result of request processing */
1324 * Linkage item through which this request is included into
1325 * sending/delayed lists on client and into rqbd list on server
1329 * Server side list of incoming unserved requests sorted by arrival
1330 * time. Traversed from time to time to notice about to expire
1331 * requests and sent back "early replies" to clients to let them
1332 * know server is alive and well, just very busy to service their
1335 cfs_list_t rq_timed_list;
1336 /** server-side history, used for debuging purposes. */
1337 cfs_list_t rq_history_list;
1338 /** server-side per-export list */
1339 cfs_list_t rq_exp_list;
1340 /** server-side hp handlers */
1341 struct ptlrpc_hpreq_ops *rq_ops;
1343 /** initial thread servicing this request */
1344 struct ptlrpc_thread *rq_svc_thread;
1346 /** history sequence # */
1347 __u64 rq_history_seq;
1351 /** stub for NRS request */
1352 struct ptlrpc_nrs_request rq_nrq;
1354 /** the index of service's srv_at_array into which request is linked */
1356 /** Lock to protect request flags and some other important bits, like
1360 /** client-side flags are serialized by rq_lock */
1361 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1362 rq_timedout:1, rq_resend:1, rq_restart:1,
1364 * when ->rq_replay is set, request is kept by the client even
1365 * after server commits corresponding transaction. This is
1366 * used for operations that require sequence of multiple
1367 * requests to be replayed. The only example currently is file
1368 * open/close. When last request in such a sequence is
1369 * committed, ->rq_replay is cleared on all requests in the
1373 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1374 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1375 rq_early:1, rq_must_unlink:1,
1376 rq_memalloc:1, /* req originated from "kswapd" */
1377 /* server-side flags */
1378 rq_packed_final:1, /* packed final reply */
1379 rq_hp:1, /* high priority RPC */
1380 rq_at_linked:1, /* link into service's srv_at_array */
1381 rq_reply_truncate:1,
1383 /* whether the "rq_set" is a valid one */
1385 rq_generation_set:1,
1386 /* do not resend request on -EINPROGRESS */
1387 rq_no_retry_einprogress:1,
1388 /* allow the req to be sent if the import is in recovery
1392 unsigned int rq_nr_resend;
1394 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1395 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1396 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
1397 server-side refcounf for multiple replies */
1399 /** Portal to which this request would be sent */
1400 short rq_request_portal; /* XXX FIXME bug 249 */
1401 /** Portal where to wait for reply and where reply would be sent */
1402 short rq_reply_portal; /* XXX FIXME bug 249 */
1406 * !rq_truncate : # reply bytes actually received,
1407 * rq_truncate : required repbuf_len for resend
1409 int rq_nob_received;
1410 /** Request length */
1414 /** Request message - what client sent */
1415 struct lustre_msg *rq_reqmsg;
1416 /** Reply message - server response */
1417 struct lustre_msg *rq_repmsg;
1418 /** Transaction number */
1423 * List item to for replay list. Not yet commited requests get linked
1425 * Also see \a rq_replay comment above.
1427 cfs_list_t rq_replay_list;
1430 * security and encryption data
1432 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1433 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1434 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
1436 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1437 enum lustre_sec_part rq_sp_from;
1439 /* client/server security flags */
1441 rq_ctx_init:1, /* context initiation */
1442 rq_ctx_fini:1, /* context destroy */
1443 rq_bulk_read:1, /* request bulk read */
1444 rq_bulk_write:1, /* request bulk write */
1445 /* server authentication flags */
1446 rq_auth_gss:1, /* authenticated by gss */
1447 rq_auth_remote:1, /* authed as remote user */
1448 rq_auth_usr_root:1, /* authed as root */
1449 rq_auth_usr_mdt:1, /* authed as mdt */
1450 rq_auth_usr_ost:1, /* authed as ost */
1451 /* security tfm flags */
1454 /* doesn't expect reply FIXME */
1456 rq_pill_init:1; /* pill initialized */
1458 uid_t rq_auth_uid; /* authed uid */
1459 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1461 /* (server side), pointed directly into req buffer */
1462 struct ptlrpc_user_desc *rq_user_desc;
1464 /* various buffer pointers */
1465 struct lustre_msg *rq_reqbuf; /* req wrapper */
1466 char *rq_repbuf; /* rep buffer */
1467 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1468 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1469 int rq_reqbuf_len; /* req wrapper buf len */
1470 int rq_reqdata_len; /* req wrapper msg len */
1471 int rq_repbuf_len; /* rep buffer len */
1472 int rq_repdata_len; /* rep wrapper msg len */
1473 int rq_clrbuf_len; /* only in priv mode */
1474 int rq_clrdata_len; /* only in priv mode */
1476 /** early replies go to offset 0, regular replies go after that */
1477 unsigned int rq_reply_off;
1481 /** Fields that help to see if request and reply were swabbed or not */
1482 __u32 rq_req_swab_mask;
1483 __u32 rq_rep_swab_mask;
1485 /** What was import generation when this request was sent */
1486 int rq_import_generation;
1487 enum lustre_imp_state rq_send_state;
1489 /** how many early replies (for stats) */
1492 /** client+server request */
1493 lnet_handle_md_t rq_req_md_h;
1494 struct ptlrpc_cb_id rq_req_cbid;
1495 /** optional time limit for send attempts */
1496 cfs_duration_t rq_delay_limit;
1497 /** time request was first queued */
1498 cfs_time_t rq_queued_time;
1500 /* server-side... */
1501 /** request arrival time */
1502 struct timeval rq_arrival_time;
1503 /** separated reply state */
1504 struct ptlrpc_reply_state *rq_reply_state;
1505 /** incoming request buffer */
1506 struct ptlrpc_request_buffer_desc *rq_rqbd;
1508 /** client-only incoming reply */
1509 lnet_handle_md_t rq_reply_md_h;
1510 cfs_waitq_t rq_reply_waitq;
1511 struct ptlrpc_cb_id rq_reply_cbid;
1515 /** Peer description (the other side) */
1516 lnet_process_id_t rq_peer;
1517 /** Server-side, export on which request was received */
1518 struct obd_export *rq_export;
1519 /** Client side, import where request is being sent */
1520 struct obd_import *rq_import;
1522 /** Replay callback, called after request is replayed at recovery */
1523 void (*rq_replay_cb)(struct ptlrpc_request *);
1525 * Commit callback, called when request is committed and about to be
1528 void (*rq_commit_cb)(struct ptlrpc_request *);
1529 /** Opaq data for replay and commit callbacks. */
1532 /** For bulk requests on client only: bulk descriptor */
1533 struct ptlrpc_bulk_desc *rq_bulk;
1535 /** client outgoing req */
1537 * when request/reply sent (secs), or time when request should be sent
1540 /** time for request really sent out */
1541 time_t rq_real_sent;
1543 /** when request must finish. volatile
1544 * so that servers' early reply updates to the deadline aren't
1545 * kept in per-cpu cache */
1546 volatile time_t rq_deadline;
1547 /** when req reply unlink must finish. */
1548 time_t rq_reply_deadline;
1549 /** when req bulk unlink must finish. */
1550 time_t rq_bulk_deadline;
1552 * service time estimate (secs)
1553 * If the requestsis not served by this time, it is marked as timed out.
1557 /** Multi-rpc bits */
1558 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1559 cfs_waitq_t rq_set_waitq;
1560 /** Link item for request set lists */
1561 cfs_list_t rq_set_chain;
1562 /** Link back to the request set */
1563 struct ptlrpc_request_set *rq_set;
1564 /** Async completion handler, called when reply is received */
1565 ptlrpc_interpterer_t rq_interpret_reply;
1566 /** Async completion context */
1567 union ptlrpc_async_args rq_async_args;
1569 /** Pool if request is from preallocated list */
1570 struct ptlrpc_request_pool *rq_pool;
1572 struct lu_context rq_session;
1573 struct lu_context rq_recov_session;
1575 /** request format description */
1576 struct req_capsule rq_pill;
1580 * Call completion handler for rpc if any, return it's status or original
1581 * rc if there was no handler defined for this request.
1583 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1584 struct ptlrpc_request *req, int rc)
1586 if (req->rq_interpret_reply != NULL) {
1587 req->rq_status = req->rq_interpret_reply(env, req,
1588 &req->rq_async_args,
1590 return req->rq_status;
1598 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_desc *desc);
1599 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_desc *desc);
1600 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1601 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1602 struct ptlrpc_nrs_pol_info *info);
1605 * Can the request be moved from the regular NRS head to the high-priority NRS
1606 * head (of the same PTLRPC service partition), if any?
1608 * For a reliable result, this should be checked under svcpt->scp_req lock.
1611 ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1613 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1616 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1617 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1618 * to make sure it has not been scheduled yet (analogous to previous
1619 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1621 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1626 * Returns 1 if request buffer at offset \a index was already swabbed
1628 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1630 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1631 return req->rq_req_swab_mask & (1 << index);
1635 * Returns 1 if request reply buffer at offset \a index was already swabbed
1637 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1639 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1640 return req->rq_rep_swab_mask & (1 << index);
1644 * Returns 1 if request needs to be swabbed into local cpu byteorder
1646 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1648 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1652 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1654 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1656 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1660 * Mark request buffer at offset \a index that it was already swabbed
1662 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1664 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1665 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1666 req->rq_req_swab_mask |= 1 << index;
1670 * Mark request reply buffer at offset \a index that it was already swabbed
1672 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1674 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1675 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1676 req->rq_rep_swab_mask |= 1 << index;
1680 * Convert numerical request phase value \a phase into text string description
1682 static inline const char *
1683 ptlrpc_phase2str(enum rq_phase phase)
1692 case RQ_PHASE_INTERPRET:
1694 case RQ_PHASE_COMPLETE:
1696 case RQ_PHASE_UNREGISTERING:
1697 return "Unregistering";
1704 * Convert numerical request phase of the request \a req into text stringi
1707 static inline const char *
1708 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1710 return ptlrpc_phase2str(req->rq_phase);
1714 * Debugging functions and helpers to print request structure into debug log
1717 /* Spare the preprocessor, spoil the bugs. */
1718 #define FLAG(field, str) (field ? str : "")
1720 /** Convert bit flags into a string */
1721 #define DEBUG_REQ_FLAGS(req) \
1722 ptlrpc_rqphase2str(req), \
1723 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1724 FLAG(req->rq_err, "E"), \
1725 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1726 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1727 FLAG(req->rq_no_resend, "N"), \
1728 FLAG(req->rq_waiting, "W"), \
1729 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1730 FLAG(req->rq_committed, "M")
1732 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1734 void _debug_req(struct ptlrpc_request *req,
1735 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1736 __attribute__ ((format (printf, 3, 4)));
1739 * Helper that decides if we need to print request accordig to current debug
1742 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1744 CFS_CHECK_STACK(msgdata, mask, cdls); \
1746 if (((mask) & D_CANTMASK) != 0 || \
1747 ((libcfs_debug & (mask)) != 0 && \
1748 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1749 _debug_req((req), msgdata, fmt, ##a); \
1753 * This is the debug print function you need to use to print request sturucture
1754 * content into lustre debug log.
1755 * for most callers (level is a constant) this is resolved at compile time */
1756 #define DEBUG_REQ(level, req, fmt, args...) \
1758 if ((level) & (D_ERROR | D_WARNING)) { \
1759 static cfs_debug_limit_state_t cdls; \
1760 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1761 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1763 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1764 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1770 * Structure that defines a single page of a bulk transfer
1772 struct ptlrpc_bulk_page {
1773 /** Linkage to list of pages in a bulk */
1776 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1779 /** offset within a page */
1781 /** The page itself */
1782 struct page *bp_page;
1785 #define BULK_GET_SOURCE 0
1786 #define BULK_PUT_SINK 1
1787 #define BULK_GET_SINK 2
1788 #define BULK_PUT_SOURCE 3
1791 * Definition of buk descriptor.
1792 * Bulks are special "Two phase" RPCs where initial request message
1793 * is sent first and it is followed bt a transfer (o receiving) of a large
1794 * amount of data to be settled into pages referenced from the bulk descriptors.
1795 * Bulks transfers (the actual data following the small requests) are done
1796 * on separate LNet portals.
1797 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1798 * Another user is readpage for MDT.
1800 struct ptlrpc_bulk_desc {
1801 /** completed successfully */
1802 unsigned long bd_success:1;
1803 /** accessible to the network (network io potentially in progress) */
1804 unsigned long bd_network_rw:1;
1805 /** {put,get}{source,sink} */
1806 unsigned long bd_type:2;
1808 unsigned long bd_registered:1;
1809 /** For serialization with callback */
1811 /** Import generation when request for this bulk was sent */
1812 int bd_import_generation;
1813 /** Server side - export this bulk created for */
1814 struct obd_export *bd_export;
1815 /** Client side - import this bulk was sent on */
1816 struct obd_import *bd_import;
1817 /** LNet portal for this bulk */
1819 /** Back pointer to the request */
1820 struct ptlrpc_request *bd_req;
1821 cfs_waitq_t bd_waitq; /* server side only WQ */
1822 int bd_iov_count; /* # entries in bd_iov */
1823 int bd_max_iov; /* allocated size of bd_iov */
1824 int bd_nob; /* # bytes covered */
1825 int bd_nob_transferred; /* # bytes GOT/PUT */
1829 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1830 lnet_handle_md_t bd_md_h; /* associated MD */
1831 lnet_nid_t bd_sender; /* stash event::sender */
1833 #if defined(__KERNEL__)
1835 * encrypt iov, size is either 0 or bd_iov_count.
1837 lnet_kiov_t *bd_enc_iov;
1839 lnet_kiov_t bd_iov[0];
1841 lnet_md_iovec_t bd_iov[0];
1846 SVC_STOPPED = 1 << 0,
1847 SVC_STOPPING = 1 << 1,
1848 SVC_STARTING = 1 << 2,
1849 SVC_RUNNING = 1 << 3,
1851 SVC_SIGNAL = 1 << 5,
1854 #define PTLRPC_THR_NAME_LEN 32
1856 * Definition of server service thread structure
1858 struct ptlrpc_thread {
1860 * List of active threads in svc->srv_threads
1864 * thread-private data (preallocated memory)
1869 * service thread index, from ptlrpc_start_threads
1873 * service thread pid
1877 * put watchdog in the structure per thread b=14840
1879 struct lc_watchdog *t_watchdog;
1881 * the svc this thread belonged to b=18582
1883 struct ptlrpc_service_part *t_svcpt;
1884 cfs_waitq_t t_ctl_waitq;
1885 struct lu_env *t_env;
1886 char t_name[PTLRPC_THR_NAME_LEN];
1889 static inline int thread_is_init(struct ptlrpc_thread *thread)
1891 return thread->t_flags == 0;
1894 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1896 return !!(thread->t_flags & SVC_STOPPED);
1899 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1901 return !!(thread->t_flags & SVC_STOPPING);
1904 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1906 return !!(thread->t_flags & SVC_STARTING);
1909 static inline int thread_is_running(struct ptlrpc_thread *thread)
1911 return !!(thread->t_flags & SVC_RUNNING);
1914 static inline int thread_is_event(struct ptlrpc_thread *thread)
1916 return !!(thread->t_flags & SVC_EVENT);
1919 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1921 return !!(thread->t_flags & SVC_SIGNAL);
1924 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1926 thread->t_flags &= ~flags;
1929 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1931 thread->t_flags = flags;
1934 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1936 thread->t_flags |= flags;
1939 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1942 if (thread->t_flags & flags) {
1943 thread->t_flags &= ~flags;
1950 * Request buffer descriptor structure.
1951 * This is a structure that contains one posted request buffer for service.
1952 * Once data land into a buffer, event callback creates actual request and
1953 * notifies wakes one of the service threads to process new incoming request.
1954 * More than one request can fit into the buffer.
1956 struct ptlrpc_request_buffer_desc {
1957 /** Link item for rqbds on a service */
1958 cfs_list_t rqbd_list;
1959 /** History of requests for this buffer */
1960 cfs_list_t rqbd_reqs;
1961 /** Back pointer to service for which this buffer is registered */
1962 struct ptlrpc_service_part *rqbd_svcpt;
1963 /** LNet descriptor */
1964 lnet_handle_md_t rqbd_md_h;
1966 /** The buffer itself */
1968 struct ptlrpc_cb_id rqbd_cbid;
1970 * This "embedded" request structure is only used for the
1971 * last request to fit into the buffer
1973 struct ptlrpc_request rqbd_req;
1976 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1978 struct ptlrpc_service_ops {
1980 * if non-NULL called during thread creation (ptlrpc_start_thread())
1981 * to initialize service specific per-thread state.
1983 int (*so_thr_init)(struct ptlrpc_thread *thr);
1985 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1986 * destruct state created by ->srv_init().
1988 void (*so_thr_done)(struct ptlrpc_thread *thr);
1990 * Handler function for incoming requests for this service
1992 int (*so_req_handler)(struct ptlrpc_request *req);
1994 * function to determine priority of the request, it's called
1995 * on every new request
1997 int (*so_hpreq_handler)(struct ptlrpc_request *);
1999 * service-specific print fn
2001 void (*so_req_printer)(void *, struct ptlrpc_request *);
2004 #ifndef __cfs_cacheline_aligned
2005 /* NB: put it here for reducing patche dependence */
2006 # define __cfs_cacheline_aligned
2010 * How many high priority requests to serve before serving one normal
2013 #define PTLRPC_SVC_HP_RATIO 10
2016 * Definition of PortalRPC service.
2017 * The service is listening on a particular portal (like tcp port)
2018 * and perform actions for a specific server like IO service for OST
2019 * or general metadata service for MDS.
2021 struct ptlrpc_service {
2022 /** serialize /proc operations */
2023 spinlock_t srv_lock;
2024 /** most often accessed fields */
2025 /** chain thru all services */
2026 cfs_list_t srv_list;
2027 /** service operations table */
2028 struct ptlrpc_service_ops srv_ops;
2029 /** only statically allocated strings here; we don't clean them */
2031 /** only statically allocated strings here; we don't clean them */
2032 char *srv_thread_name;
2033 /** service thread list */
2034 cfs_list_t srv_threads;
2035 /** threads # should be created for each partition on initializing */
2036 int srv_nthrs_cpt_init;
2037 /** limit of threads number for each partition */
2038 int srv_nthrs_cpt_limit;
2039 /** Root of /proc dir tree for this service */
2040 cfs_proc_dir_entry_t *srv_procroot;
2041 /** Pointer to statistic data for this service */
2042 struct lprocfs_stats *srv_stats;
2043 /** # hp per lp reqs to handle */
2044 int srv_hpreq_ratio;
2045 /** biggest request to receive */
2046 int srv_max_req_size;
2047 /** biggest reply to send */
2048 int srv_max_reply_size;
2049 /** size of individual buffers */
2051 /** # buffers to allocate in 1 group */
2052 int srv_nbuf_per_group;
2053 /** Local portal on which to receive requests */
2054 __u32 srv_req_portal;
2055 /** Portal on the client to send replies to */
2056 __u32 srv_rep_portal;
2058 * Tags for lu_context associated with this thread, see struct
2062 /** soft watchdog timeout multiplier */
2063 int srv_watchdog_factor;
2064 /** under unregister_service */
2065 unsigned srv_is_stopping:1;
2067 /** max # request buffers in history per partition */
2068 int srv_hist_nrqbds_cpt_max;
2069 /** number of CPTs this service bound on */
2071 /** CPTs array this service bound on */
2073 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2075 /** CPT table this service is running over */
2076 struct cfs_cpt_table *srv_cptable;
2078 * partition data for ptlrpc service
2080 struct ptlrpc_service_part *srv_parts[0];
2084 * Definition of PortalRPC service partition data.
2085 * Although a service only has one instance of it right now, but we
2086 * will have multiple instances very soon (instance per CPT).
2088 * it has four locks:
2090 * serialize operations on rqbd and requests waiting for preprocess
2092 * serialize operations active requests sent to this portal
2094 * serialize adaptive timeout stuff
2096 * serialize operations on RS list (reply states)
2098 * We don't have any use-case to take two or more locks at the same time
2099 * for now, so there is no lock order issue.
2101 struct ptlrpc_service_part {
2102 /** back reference to owner */
2103 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2104 /* CPT id, reserved */
2106 /** always increasing number */
2108 /** # of starting threads */
2109 int scp_nthrs_starting;
2110 /** # of stopping threads, reserved for shrinking threads */
2111 int scp_nthrs_stopping;
2112 /** # running threads */
2113 int scp_nthrs_running;
2114 /** service threads list */
2115 cfs_list_t scp_threads;
2118 * serialize the following fields, used for protecting
2119 * rqbd list and incoming requests waiting for preprocess,
2120 * threads starting & stopping are also protected by this lock.
2122 spinlock_t scp_lock __cfs_cacheline_aligned;
2123 /** total # req buffer descs allocated */
2124 int scp_nrqbds_total;
2125 /** # posted request buffers for receiving */
2126 int scp_nrqbds_posted;
2127 /** in progress of allocating rqbd */
2128 int scp_rqbd_allocating;
2129 /** # incoming reqs */
2130 int scp_nreqs_incoming;
2131 /** request buffers to be reposted */
2132 cfs_list_t scp_rqbd_idle;
2133 /** req buffers receiving */
2134 cfs_list_t scp_rqbd_posted;
2135 /** incoming reqs */
2136 cfs_list_t scp_req_incoming;
2137 /** timeout before re-posting reqs, in tick */
2138 cfs_duration_t scp_rqbd_timeout;
2140 * all threads sleep on this. This wait-queue is signalled when new
2141 * incoming request arrives and when difficult reply has to be handled.
2143 cfs_waitq_t scp_waitq;
2145 /** request history */
2146 cfs_list_t scp_hist_reqs;
2147 /** request buffer history */
2148 cfs_list_t scp_hist_rqbds;
2149 /** # request buffers in history */
2150 int scp_hist_nrqbds;
2151 /** sequence number for request */
2153 /** highest seq culled from history */
2154 __u64 scp_hist_seq_culled;
2157 * serialize the following fields, used for processing requests
2158 * sent to this portal
2160 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2161 /** # reqs in either of the NRS heads below */
2162 /** # reqs being served */
2163 int scp_nreqs_active;
2164 /** # HPreqs being served */
2165 int scp_nhreqs_active;
2166 /** # hp requests handled */
2169 /** NRS head for regular requests */
2170 struct ptlrpc_nrs scp_nrs_reg;
2171 /** NRS head for HP requests; this is only valid for services that can
2172 * handle HP requests */
2173 struct ptlrpc_nrs *scp_nrs_hp;
2178 * serialize the following fields, used for changes on
2181 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2182 /** estimated rpc service time */
2183 struct adaptive_timeout scp_at_estimate;
2184 /** reqs waiting for replies */
2185 struct ptlrpc_at_array scp_at_array;
2186 /** early reply timer */
2187 cfs_timer_t scp_at_timer;
2189 cfs_time_t scp_at_checktime;
2190 /** check early replies */
2191 unsigned scp_at_check;
2195 * serialize the following fields, used for processing
2196 * replies for this portal
2198 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2199 /** all the active replies */
2200 cfs_list_t scp_rep_active;
2202 /** replies waiting for service */
2203 cfs_list_t scp_rep_queue;
2205 /** List of free reply_states */
2206 cfs_list_t scp_rep_idle;
2207 /** waitq to run, when adding stuff to srv_free_rs_list */
2208 cfs_waitq_t scp_rep_waitq;
2209 /** # 'difficult' replies */
2210 cfs_atomic_t scp_nreps_difficult;
2213 #define ptlrpc_service_for_each_part(part, i, svc) \
2215 i < (svc)->srv_ncpts && \
2216 (svc)->srv_parts != NULL && \
2217 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2220 * Declaration of ptlrpcd control structure
2222 struct ptlrpcd_ctl {
2224 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2226 unsigned long pc_flags;
2228 * Thread lock protecting structure fields.
2234 struct completion pc_starting;
2238 struct completion pc_finishing;
2240 * Thread requests set.
2242 struct ptlrpc_request_set *pc_set;
2244 * Thread name used in cfs_daemonize()
2248 * Environment for request interpreters to run in.
2250 struct lu_env pc_env;
2252 * Index of ptlrpcd thread in the array.
2256 * Number of the ptlrpcd's partners.
2260 * Pointer to the array of partners' ptlrpcd_ctl structure.
2262 struct ptlrpcd_ctl **pc_partners;
2264 * Record the partner index to be processed next.
2269 * Async rpcs flag to make sure that ptlrpcd_check() is called only
2274 * Currently not used.
2278 * User-space async rpcs callback.
2280 void *pc_wait_callback;
2282 * User-space check idle rpcs callback.
2284 void *pc_idle_callback;
2288 /* Bits for pc_flags */
2289 enum ptlrpcd_ctl_flags {
2291 * Ptlrpc thread start flag.
2293 LIOD_START = 1 << 0,
2295 * Ptlrpc thread stop flag.
2299 * Ptlrpc thread force flag (only stop force so far).
2300 * This will cause aborting any inflight rpcs handled
2301 * by thread if LIOD_STOP is specified.
2303 LIOD_FORCE = 1 << 2,
2305 * This is a recovery ptlrpc thread.
2307 LIOD_RECOVERY = 1 << 3,
2309 * The ptlrpcd is bound to some CPU core.
2318 * Service compatibility function; policy is compatible with all services.
2320 * \param[in] svc The service the policy is attempting to register with.
2321 * \param[in] desc The policy descriptor
2323 * \retval true The policy is compatible with the NRS head
2325 * \see ptlrpc_nrs_pol_desc::pd_compat()
2328 nrs_policy_compat_all(struct ptlrpc_service *svc,
2329 const struct ptlrpc_nrs_pol_desc *desc)
2335 * Service compatibility function; policy is compatible with only a specific
2336 * service which is identified by its human-readable name at
2337 * ptlrpc_service::srv_name.
2339 * \param[in] svc The service the policy is attempting to register with.
2340 * \param[in] desc The policy descriptor
2342 * \retval false The policy is not compatible with the NRS head
2343 * \retval true The policy is compatible with the NRS head
2345 * \see ptlrpc_nrs_pol_desc::pd_compat()
2348 nrs_policy_compat_one(struct ptlrpc_service *svc,
2349 const struct ptlrpc_nrs_pol_desc *desc)
2351 LASSERT(desc->pd_compat_svc_name != NULL);
2352 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2357 /* ptlrpc/events.c */
2358 extern lnet_handle_eq_t ptlrpc_eq_h;
2359 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2360 lnet_process_id_t *peer, lnet_nid_t *self);
2362 * These callbacks are invoked by LNet when something happened to
2366 extern void request_out_callback(lnet_event_t *ev);
2367 extern void reply_in_callback(lnet_event_t *ev);
2368 extern void client_bulk_callback(lnet_event_t *ev);
2369 extern void request_in_callback(lnet_event_t *ev);
2370 extern void reply_out_callback(lnet_event_t *ev);
2371 #ifdef HAVE_SERVER_SUPPORT
2372 extern void server_bulk_callback(lnet_event_t *ev);
2376 /* ptlrpc/connection.c */
2377 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2379 struct obd_uuid *uuid);
2380 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2381 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2382 int ptlrpc_connection_init(void);
2383 void ptlrpc_connection_fini(void);
2384 extern lnet_pid_t ptl_get_pid(void);
2386 /* ptlrpc/niobuf.c */
2388 * Actual interfacing with LNet to put/get/register/unregister stuff
2391 #ifdef HAVE_SERVER_SUPPORT
2392 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2393 int npages, int type, int portal);
2394 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2395 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2397 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2401 LASSERT(desc != NULL);
2403 spin_lock(&desc->bd_lock);
2404 rc = desc->bd_network_rw;
2405 spin_unlock(&desc->bd_lock);
2410 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2411 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2413 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2415 struct ptlrpc_bulk_desc *desc;
2418 LASSERT(req != NULL);
2419 desc = req->rq_bulk;
2421 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2422 req->rq_bulk_deadline > cfs_time_current_sec())
2428 spin_lock(&desc->bd_lock);
2429 rc = desc->bd_network_rw;
2430 spin_unlock(&desc->bd_lock);
2434 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2435 #define PTLRPC_REPLY_EARLY 0x02
2436 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2437 int ptlrpc_reply(struct ptlrpc_request *req);
2438 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2439 int ptlrpc_error(struct ptlrpc_request *req);
2440 void ptlrpc_resend_req(struct ptlrpc_request *request);
2441 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2442 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2443 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2446 /* ptlrpc/client.c */
2448 * Client-side portals API. Everything to send requests, receive replies,
2449 * request queues, request management, etc.
2452 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2453 struct ptlrpc_client *);
2454 void ptlrpc_cleanup_client(struct obd_import *imp);
2455 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2457 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2458 int ptlrpc_replay_req(struct ptlrpc_request *req);
2459 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2460 void ptlrpc_restart_req(struct ptlrpc_request *req);
2461 void ptlrpc_abort_inflight(struct obd_import *imp);
2462 void ptlrpc_cleanup_imp(struct obd_import *imp);
2463 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2465 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2466 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2468 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2469 set_interpreter_func fn, void *data);
2470 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2471 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2472 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2473 int ptlrpc_expired_set(void *data);
2474 void ptlrpc_interrupted_set(void *data);
2475 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2476 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2477 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2478 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2479 struct ptlrpc_request *req);
2481 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2482 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2484 struct ptlrpc_request_pool *
2485 ptlrpc_init_rq_pool(int, int,
2486 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2488 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2489 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2490 const struct req_format *format);
2491 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2492 struct ptlrpc_request_pool *,
2493 const struct req_format *format);
2494 void ptlrpc_request_free(struct ptlrpc_request *request);
2495 int ptlrpc_request_pack(struct ptlrpc_request *request,
2496 __u32 version, int opcode);
2497 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2498 const struct req_format *format,
2499 __u32 version, int opcode);
2500 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2501 __u32 version, int opcode, char **bufs,
2502 struct ptlrpc_cli_ctx *ctx);
2503 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2504 int opcode, int count, __u32 *lengths,
2506 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2507 __u32 version, int opcode,
2508 int count, __u32 *lengths, char **bufs,
2509 struct ptlrpc_request_pool *pool);
2510 void ptlrpc_req_finished(struct ptlrpc_request *request);
2511 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2512 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2513 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2514 int npages, int type, int portal);
2515 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2516 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2518 __ptlrpc_free_bulk(bulk, 1);
2520 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2522 __ptlrpc_free_bulk(bulk, 0);
2524 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2525 cfs_page_t *page, int pageoffset, int len, int);
2526 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2527 cfs_page_t *page, int pageoffset,
2530 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2533 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2534 cfs_page_t *page, int pageoffset,
2537 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2540 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2541 struct obd_import *imp);
2542 __u64 ptlrpc_next_xid(void);
2543 __u64 ptlrpc_sample_next_xid(void);
2544 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2546 /* Set of routines to run a function in ptlrpcd context */
2547 void *ptlrpcd_alloc_work(struct obd_import *imp,
2548 int (*cb)(const struct lu_env *, void *), void *data);
2549 void ptlrpcd_destroy_work(void *handler);
2550 int ptlrpcd_queue_work(void *handler);
2553 struct ptlrpc_service_buf_conf {
2554 /* nbufs is how many buffers to post */
2555 unsigned int bc_nbufs;
2556 /* buffer size to post */
2557 unsigned int bc_buf_size;
2558 /* portal to listed for requests on */
2559 unsigned int bc_req_portal;
2560 /* portal of where to send replies to */
2561 unsigned int bc_rep_portal;
2562 /* maximum request size to be accepted for this service */
2563 unsigned int bc_req_max_size;
2564 /* maximum reply size this service can ever send */
2565 unsigned int bc_rep_max_size;
2568 struct ptlrpc_service_thr_conf {
2569 /* threadname should be 8 characters or less - 6 will be added on */
2571 /* threads increasing factor for each CPU */
2572 unsigned int tc_thr_factor;
2573 /* service threads # to start on each partition while initializing */
2574 unsigned int tc_nthrs_init;
2576 * low water of threads # upper-limit on each partition while running,
2577 * service availability may be impacted if threads number is lower
2578 * than this value. It can be ZERO if the service doesn't require
2579 * CPU affinity or there is only one partition.
2581 unsigned int tc_nthrs_base;
2582 /* "soft" limit for total threads number */
2583 unsigned int tc_nthrs_max;
2584 /* user specified threads number, it will be validated due to
2585 * other members of this structure. */
2586 unsigned int tc_nthrs_user;
2587 /* set NUMA node affinity for service threads */
2588 unsigned int tc_cpu_affinity;
2589 /* Tags for lu_context associated with service thread */
2593 struct ptlrpc_service_cpt_conf {
2594 struct cfs_cpt_table *cc_cptable;
2595 /* string pattern to describe CPTs for a service */
2599 struct ptlrpc_service_conf {
2602 /* soft watchdog timeout multiplifier to print stuck service traces */
2603 unsigned int psc_watchdog_factor;
2604 /* buffer information */
2605 struct ptlrpc_service_buf_conf psc_buf;
2606 /* thread information */
2607 struct ptlrpc_service_thr_conf psc_thr;
2608 /* CPU partition information */
2609 struct ptlrpc_service_cpt_conf psc_cpt;
2610 /* function table */
2611 struct ptlrpc_service_ops psc_ops;
2614 /* ptlrpc/service.c */
2616 * Server-side services API. Register/unregister service, request state
2617 * management, service thread management
2621 void ptlrpc_save_lock(struct ptlrpc_request *req,
2622 struct lustre_handle *lock, int mode, int no_ack);
2623 void ptlrpc_commit_replies(struct obd_export *exp);
2624 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2625 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2626 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2627 struct ptlrpc_service *ptlrpc_register_service(
2628 struct ptlrpc_service_conf *conf,
2629 struct proc_dir_entry *proc_entry);
2630 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2632 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2633 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2634 int liblustre_check_services(void *arg);
2635 void ptlrpc_daemonize(char *name);
2636 int ptlrpc_service_health_check(struct ptlrpc_service *);
2637 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2640 int ptlrpc_hr_init(void);
2641 void ptlrpc_hr_fini(void);
2643 # define ptlrpc_hr_init() (0)
2644 # define ptlrpc_hr_fini() do {} while(0)
2649 /* ptlrpc/import.c */
2654 int ptlrpc_connect_import(struct obd_import *imp);
2655 int ptlrpc_init_import(struct obd_import *imp);
2656 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2657 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2658 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2661 /* ptlrpc/pack_generic.c */
2662 int ptlrpc_reconnect_import(struct obd_import *imp);
2666 * ptlrpc msg buffer and swab interface
2670 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2672 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2674 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2675 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2677 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2678 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2680 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2681 __u32 *lens, char **bufs);
2682 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2684 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2685 __u32 *lens, char **bufs, int flags);
2686 #define LPRFL_EARLY_REPLY 1
2687 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2688 char **bufs, int flags);
2689 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2690 unsigned int newlen, int move_data);
2691 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2692 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2693 int lustre_msg_hdr_size(__u32 magic, int count);
2694 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2695 int lustre_msg_size_v2(int count, __u32 *lengths);
2696 int lustre_packed_msg_size(struct lustre_msg *msg);
2697 int lustre_msg_early_size(void);
2698 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2699 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2700 int lustre_msg_buflen(struct lustre_msg *m, int n);
2701 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2702 int lustre_msg_bufcount(struct lustre_msg *m);
2703 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2704 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2705 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2706 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2707 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2708 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2709 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2710 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2711 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2712 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2713 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2714 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2715 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2716 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2717 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2718 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2719 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2720 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2721 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2722 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2723 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2724 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2725 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2726 int lustre_msg_get_status(struct lustre_msg *msg);
2727 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2728 int lustre_msg_is_v1(struct lustre_msg *msg);
2729 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2730 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2731 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2732 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2733 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2734 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2735 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2737 # warning "remove checksum compatibility support for b1_8"
2738 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2740 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2741 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2742 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2743 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2744 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2745 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2746 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2747 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2748 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2749 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2750 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2751 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2752 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2753 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2754 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2757 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2758 unsigned int newlen, int move_data)
2760 LASSERT(req->rq_reply_state);
2761 LASSERT(req->rq_repmsg);
2762 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2767 /** Change request phase of \a req to \a new_phase */
2769 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2771 if (req->rq_phase == new_phase)
2774 if (new_phase == RQ_PHASE_UNREGISTERING) {
2775 req->rq_next_phase = req->rq_phase;
2777 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2780 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2782 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2785 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2786 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2788 req->rq_phase = new_phase;
2792 * Returns true if request \a req got early reply and hard deadline is not met
2795 ptlrpc_client_early(struct ptlrpc_request *req)
2797 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2798 req->rq_reply_deadline > cfs_time_current_sec())
2800 return req->rq_early;
2804 * Returns true if we got real reply from server for this request
2807 ptlrpc_client_replied(struct ptlrpc_request *req)
2809 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2810 req->rq_reply_deadline > cfs_time_current_sec())
2812 return req->rq_replied;
2815 /** Returns true if request \a req is in process of receiving server reply */
2817 ptlrpc_client_recv(struct ptlrpc_request *req)
2819 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2820 req->rq_reply_deadline > cfs_time_current_sec())
2822 return req->rq_receiving_reply;
2826 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2830 spin_lock(&req->rq_lock);
2831 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2832 req->rq_reply_deadline > cfs_time_current_sec()) {
2833 spin_unlock(&req->rq_lock);
2836 rc = req->rq_receiving_reply || req->rq_must_unlink;
2837 spin_unlock(&req->rq_lock);
2842 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2844 if (req->rq_set == NULL)
2845 cfs_waitq_signal(&req->rq_reply_waitq);
2847 cfs_waitq_signal(&req->rq_set->set_waitq);
2851 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2853 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2854 cfs_atomic_inc(&rs->rs_refcount);
2858 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2860 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2861 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
2862 lustre_free_reply_state(rs);
2865 /* Should only be called once per req */
2866 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2868 if (req->rq_reply_state == NULL)
2869 return; /* shouldn't occur */
2870 ptlrpc_rs_decref(req->rq_reply_state);
2871 req->rq_reply_state = NULL;
2872 req->rq_repmsg = NULL;
2875 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2877 return lustre_msg_get_magic(req->rq_reqmsg);
2880 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2882 switch (req->rq_reqmsg->lm_magic) {
2883 case LUSTRE_MSG_MAGIC_V2:
2884 return req->rq_reqmsg->lm_repsize;
2886 LASSERTF(0, "incorrect message magic: %08x\n",
2887 req->rq_reqmsg->lm_magic);
2892 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2894 if (req->rq_delay_limit != 0 &&
2895 cfs_time_before(cfs_time_add(req->rq_queued_time,
2896 cfs_time_seconds(req->rq_delay_limit)),
2897 cfs_time_current())) {
2903 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2905 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2906 spin_lock(&req->rq_lock);
2907 req->rq_no_resend = 1;
2908 spin_unlock(&req->rq_lock);
2910 return req->rq_no_resend;
2914 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2916 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2918 return svcpt->scp_service->srv_watchdog_factor *
2919 max_t(int, at, obd_timeout);
2922 static inline struct ptlrpc_service *
2923 ptlrpc_req2svc(struct ptlrpc_request *req)
2925 LASSERT(req->rq_rqbd != NULL);
2926 return req->rq_rqbd->rqbd_svcpt->scp_service;
2929 /* ldlm/ldlm_lib.c */
2931 * Target client logic
2934 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2935 int client_obd_cleanup(struct obd_device *obddev);
2936 int client_connect_import(const struct lu_env *env,
2937 struct obd_export **exp, struct obd_device *obd,
2938 struct obd_uuid *cluuid, struct obd_connect_data *,
2940 int client_disconnect_export(struct obd_export *exp);
2941 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2943 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2944 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2945 struct obd_uuid *uuid);
2946 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2947 void client_destroy_import(struct obd_import *imp);
2950 #ifdef HAVE_SERVER_SUPPORT
2951 int server_disconnect_export(struct obd_export *exp);
2954 /* ptlrpc/pinger.c */
2956 * Pinger API (client side only)
2959 extern int suppress_pings;
2960 enum timeout_event {
2963 struct timeout_item;
2964 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2965 int ptlrpc_pinger_add_import(struct obd_import *imp);
2966 int ptlrpc_pinger_del_import(struct obd_import *imp);
2967 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
2968 timeout_cb_t cb, void *data,
2969 cfs_list_t *obd_list);
2970 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
2971 enum timeout_event event);
2972 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2973 int ptlrpc_obd_ping(struct obd_device *obd);
2974 cfs_time_t ptlrpc_suspend_wakeup_time(void);
2976 void ping_evictor_start(void);
2977 void ping_evictor_stop(void);
2979 #define ping_evictor_start() do {} while (0)
2980 #define ping_evictor_stop() do {} while (0)
2982 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
2985 /* ptlrpc daemon bind policy */
2987 /* all ptlrpcd threads are free mode */
2988 PDB_POLICY_NONE = 1,
2989 /* all ptlrpcd threads are bound mode */
2990 PDB_POLICY_FULL = 2,
2991 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
2992 PDB_POLICY_PAIR = 3,
2993 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
2994 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
2995 * If kernel supports NUMA, pthrpcd threads are binded and
2996 * grouped by NUMA node */
2997 PDB_POLICY_NEIGHBOR = 4,
3000 /* ptlrpc daemon load policy
3001 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3002 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3003 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3004 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3005 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3007 /* on the same CPU core as the caller */
3008 PDL_POLICY_SAME = 1,
3009 /* within the same CPU partition, but not the same core as the caller */
3010 PDL_POLICY_LOCAL = 2,
3011 /* round-robin on all CPU cores, but not the same core as the caller */
3012 PDL_POLICY_ROUND = 3,
3013 /* the specified CPU core is preferred, but not enforced */
3014 PDL_POLICY_PREFERRED = 4,
3017 /* ptlrpc/ptlrpcd.c */
3018 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3019 void ptlrpcd_wake(struct ptlrpc_request *req);
3020 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3021 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3022 int ptlrpcd_addref(void);
3023 void ptlrpcd_decref(void);
3025 /* ptlrpc/lproc_ptlrpc.c */
3027 * procfs output related functions
3030 const char* ll_opcode2str(__u32 opcode);
3032 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3033 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3034 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3036 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3037 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3038 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3042 /* ptlrpc/llog_server.c */
3043 int llog_origin_handle_open(struct ptlrpc_request *req);
3044 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3045 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3046 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3047 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3048 int llog_origin_handle_close(struct ptlrpc_request *req);
3049 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3051 /* ptlrpc/llog_client.c */
3052 extern struct llog_operations llog_client_ops;