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 * Max # of bulk operations in one request.
87 * In order for the client and server to properly negotiate the maximum
88 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
89 * value. The client is free to limit the actual RPC size for any bulk
90 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
91 #define PTLRPC_BULK_OPS_BITS 2
92 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
94 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
95 * should not be used on the server at all. Otherwise, it imposes a
96 * protocol limitation on the maximum RPC size that can be used by any
97 * RPC sent to that server in the future. Instead, the server should
98 * use the negotiated per-client ocd_brw_size to determine the bulk
100 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
103 * Define maxima for bulk I/O.
105 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
106 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
107 * currently supported maximum between peers at connect via ocd_brw_size.
109 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
110 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
111 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
113 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
114 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
115 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
116 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
117 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
118 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
120 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
122 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
123 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
125 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE))
126 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE"
128 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
129 # error "PTLRPC_MAX_BRW_SIZE too big"
131 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
132 # error "PTLRPC_MAX_BRW_PAGES too big"
134 #endif /* __KERNEL__ */
136 #define PTLRPC_NTHRS_INIT 2
141 * Constants determine how memory is used to buffer incoming service requests.
143 * ?_NBUFS # buffers to allocate when growing the pool
144 * ?_BUFSIZE # bytes in a single request buffer
145 * ?_MAXREQSIZE # maximum request service will receive
147 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
148 * of ?_NBUFS is added to the pool.
150 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
151 * considered full when less than ?_MAXREQSIZE is left in them.
156 * Constants determine how threads are created for ptlrpc service.
158 * ?_NTHRS_INIT # threads to create for each service partition on
159 * initializing. If it's non-affinity service and
160 * there is only one partition, it's the overall #
161 * threads for the service while initializing.
162 * ?_NTHRS_BASE # threads should be created at least for each
163 * ptlrpc partition to keep the service healthy.
164 * It's the low-water mark of threads upper-limit
165 * for each partition.
166 * ?_THR_FACTOR # threads can be added on threads upper-limit for
167 * each CPU core. This factor is only for reference,
168 * we might decrease value of factor if number of cores
169 * per CPT is above a limit.
170 * ?_NTHRS_MAX # overall threads can be created for a service,
171 * it's a soft limit because if service is running
172 * on machine with hundreds of cores and tens of
173 * CPU partitions, we need to guarantee each partition
174 * has ?_NTHRS_BASE threads, which means total threads
175 * will be ?_NTHRS_BASE * number_of_cpts which can
176 * exceed ?_NTHRS_MAX.
180 * #define MDS_NTHRS_INIT 2
181 * #define MDS_NTHRS_BASE 64
182 * #define MDS_NTHRS_FACTOR 8
183 * #define MDS_NTHRS_MAX 1024
186 * ---------------------------------------------------------------------
187 * Server(A) has 16 cores, user configured it to 4 partitions so each
188 * partition has 4 cores, then actual number of service threads on each
190 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
192 * Total number of threads for the service is:
193 * 96 * partitions(4) = 384
196 * ---------------------------------------------------------------------
197 * Server(B) has 32 cores, user configured it to 4 partitions so each
198 * partition has 8 cores, then actual number of service threads on each
200 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
202 * Total number of threads for the service is:
203 * 128 * partitions(4) = 512
206 * ---------------------------------------------------------------------
207 * Server(B) has 96 cores, user configured it to 8 partitions so each
208 * partition has 12 cores, then actual number of service threads on each
210 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
212 * Total number of threads for the service is:
213 * 160 * partitions(8) = 1280
215 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
216 * as upper limit of threads number for each partition:
217 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
220 * ---------------------------------------------------------------------
221 * Server(C) have a thousand of cores and user configured it to 32 partitions
222 * MDS_NTHRS_BASE(64) * 32 = 2048
224 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
225 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
226 * to keep service healthy, so total number of threads will just be 2048.
228 * NB: we don't suggest to choose server with that many cores because backend
229 * filesystem itself, buffer cache, or underlying network stack might
230 * have some SMP scalability issues at that large scale.
232 * If user already has a fat machine with hundreds or thousands of cores,
233 * there are two choices for configuration:
234 * a) create CPU table from subset of all CPUs and run Lustre on
236 * b) bind service threads on a few partitions, see modparameters of
237 * MDS and OSS for details
239 * NB: these calculations (and examples below) are simplified to help
240 * understanding, the real implementation is a little more complex,
241 * please see ptlrpc_server_nthreads_check() for details.
246 * LDLM threads constants:
248 * Given 8 as factor and 24 as base threads number
251 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
254 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
255 * threads for each partition and total threads number will be 112.
258 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
259 * threads for each partition to keep service healthy, so total threads
260 * number should be 24 * 8 = 192.
262 * So with these constants, threads number wil be at the similar level
263 * of old versions, unless target machine has over a hundred cores
265 #define LDLM_THR_FACTOR 8
266 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
267 #define LDLM_NTHRS_BASE 24
268 #define LDLM_NTHRS_MAX (cfs_num_online_cpus() == 1 ? 64 : 128)
270 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
271 #define LDLM_NBUFS 64
272 #define LDLM_BUFSIZE (8 * 1024)
273 #define LDLM_MAXREQSIZE (5 * 1024)
274 #define LDLM_MAXREPSIZE (1024)
277 * MDS threads constants:
279 * Please see examples in "Thread Constants", MDS threads number will be at
280 * the comparable level of old versions, unless the server has many cores.
282 #ifndef MDS_MAX_THREADS
283 #define MDS_MAX_THREADS 1024
284 #define MDS_MAX_OTHR_THREADS 256
286 #else /* MDS_MAX_THREADS */
287 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
288 #undef MDS_MAX_THREADS
289 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
291 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
294 /* default service */
295 #define MDS_THR_FACTOR 8
296 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
297 #define MDS_NTHRS_MAX MDS_MAX_THREADS
298 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
300 /* read-page service */
301 #define MDS_RDPG_THR_FACTOR 4
302 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
303 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
304 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
306 /* these should be removed when we remove setattr service in the future */
307 #define MDS_SETA_THR_FACTOR 4
308 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
309 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
310 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
312 /* non-affinity threads */
313 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
314 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
318 * Assume file name length = FNAME_MAX = 256 (true for ext3).
319 * path name length = PATH_MAX = 4096
320 * LOV MD size max = EA_MAX = 24 * 2000
321 * (NB: 24 is size of lov_ost_data)
322 * LOV LOGCOOKIE size max = 32 * 2000
323 * (NB: 32 is size of llog_cookie)
324 * symlink: FNAME_MAX + PATH_MAX <- largest
325 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
326 * rename: FNAME_MAX + FNAME_MAX
327 * open: FNAME_MAX + EA_MAX
329 * MDS_MAXREQSIZE ~= 4736 bytes =
330 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
331 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
333 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
334 * except in the open case where there are a large number of OSTs in a LOV.
336 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
337 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
340 * MDS incoming request with LOV EA
341 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
343 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
344 362 + LOV_MAX_STRIPE_COUNT * 24)
346 * MDS outgoing reply with LOV EA
348 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
349 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
351 * but 2.4 or later MDS will never send reply with llog_cookie to any
352 * version client. This macro is defined for server side reply buffer size.
354 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
356 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
357 #define MDS_BUFSIZE (MDS_MAXREQSIZE + 1024)
359 * MDS_LOV_BUFSIZE should be at least max_reqsize (with LOV EA) +
360 * max sptlrpc payload size, however, we need to allocate a much larger buffer
361 * for it because LNet requires each MD(rqbd) has at least MDS_LOVE_MAXREQSIZE
362 * bytes left to avoid dropping of maximum-sized incoming request.
363 * So if MDS_LOV_BUFSIZE is only a little larger than MDS_LOV_MAXREQSIZE,
364 * then it can only fit in one request even there are 48K bytes left in
365 * a rqbd, and memory utilization is very low.
367 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
368 * reused until all requests fit in it have been processed and released,
369 * which means one long blocked request can prevent the rqbd be reused.
370 * Now we give extra 128K to buffer size, so even each rqbd is unlinked
371 * from LNet with unused 48K, buffer utilization will be about 72%.
372 * Please check LU-2432 for details.
374 /** MDS_LOV_BUFSIZE = max_reqsize (w/ LOV EA) + max sptlrpc payload size */
375 #define MDS_LOV_BUFSIZE (MDS_LOV_MAXREQSIZE + (1 << 17))
377 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
378 #define FLD_MAXREQSIZE (160)
380 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
381 #define FLD_MAXREPSIZE (152)
382 #define FLD_BUFSIZE (1 << 12)
385 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
387 #define SEQ_MAXREQSIZE (160)
389 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
390 #define SEQ_MAXREPSIZE (152)
391 #define SEQ_BUFSIZE (1 << 12)
393 /** MGS threads must be >= 3, see bug 22458 comment #28 */
394 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
395 #define MGS_NTHRS_MAX 32
398 #define MGS_BUFSIZE (8 * 1024)
399 #define MGS_MAXREQSIZE (7 * 1024)
400 #define MGS_MAXREPSIZE (9 * 1024)
403 * OSS threads constants:
405 * Given 8 as factor and 64 as base threads number
408 * On 8-core server configured to 2 partitions, we will have
409 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
412 * On 32-core machine configured to 4 partitions, we will have
413 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
414 * will be 112 * 4 = 448.
417 * On 64-core machine configured to 4 partitions, we will have
418 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
419 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
420 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
421 * for each partition.
423 * So we can see that with these constants, threads number wil be at the
424 * similar level of old versions, unless the server has many cores.
426 /* depress threads factor for VM with small memory size */
427 #define OSS_THR_FACTOR min_t(int, 8, \
428 CFS_NUM_CACHEPAGES >> (28 - CFS_PAGE_SHIFT))
429 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
430 #define OSS_NTHRS_BASE 64
431 #define OSS_NTHRS_MAX 512
433 /* threads for handling "create" request */
434 #define OSS_CR_THR_FACTOR 1
435 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
436 #define OSS_CR_NTHRS_BASE 8
437 #define OSS_CR_NTHRS_MAX 64
441 * lustre_msg + obdo + obd_ioobj + DT_MAX_BRW_PAGES * niobuf_remote
443 * - single object with 16 pages is 512 bytes
444 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
445 * - Must be a multiple of 1024
447 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + sizeof(struct obdo) + \
448 sizeof(struct obd_ioobj) + DT_MAX_BRW_PAGES * \
449 sizeof(struct niobuf_remote))
450 #define OST_MAXREQSIZE (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1)
452 #define OST_MAXREPSIZE (9 * 1024)
455 #define OST_BUFSIZE (OST_MAXREQSIZE + 1024)
457 /* Macro to hide a typecast. */
458 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
461 * Structure to single define portal connection.
463 struct ptlrpc_connection {
464 /** linkage for connections hash table */
465 cfs_hlist_node_t c_hash;
466 /** Our own lnet nid for this connection */
468 /** Remote side nid for this connection */
469 lnet_process_id_t c_peer;
470 /** UUID of the other side */
471 struct obd_uuid c_remote_uuid;
472 /** reference counter for this connection */
473 cfs_atomic_t c_refcount;
476 /** Client definition for PortalRPC */
477 struct ptlrpc_client {
478 /** What lnet portal does this client send messages to by default */
479 __u32 cli_request_portal;
480 /** What portal do we expect replies on */
481 __u32 cli_reply_portal;
482 /** Name of the client */
486 /** state flags of requests */
487 /* XXX only ones left are those used by the bulk descs as well! */
488 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
489 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
491 #define REQ_MAX_ACK_LOCKS 8
493 union ptlrpc_async_args {
495 * Scratchpad for passing args to completion interpreter. Users
496 * cast to the struct of their choosing, and CLASSERT that this is
497 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
498 * a pointer to it here. The pointer_arg ensures this struct is at
499 * least big enough for that.
501 void *pointer_arg[11];
505 struct ptlrpc_request_set;
506 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
507 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
510 * Definition of request set structure.
511 * Request set is a list of requests (not necessary to the same target) that
512 * once populated with RPCs could be sent in parallel.
513 * There are two kinds of request sets. General purpose and with dedicated
514 * serving thread. Example of the latter is ptlrpcd set.
515 * For general purpose sets once request set started sending it is impossible
516 * to add new requests to such set.
517 * Provides a way to call "completion callbacks" when all requests in the set
520 struct ptlrpc_request_set {
521 cfs_atomic_t set_refcount;
522 /** number of in queue requests */
523 cfs_atomic_t set_new_count;
524 /** number of uncompleted requests */
525 cfs_atomic_t set_remaining;
526 /** wait queue to wait on for request events */
527 cfs_waitq_t set_waitq;
528 cfs_waitq_t *set_wakeup_ptr;
529 /** List of requests in the set */
530 cfs_list_t set_requests;
532 * List of completion callbacks to be called when the set is completed
533 * This is only used if \a set_interpret is NULL.
534 * Links struct ptlrpc_set_cbdata.
536 cfs_list_t set_cblist;
537 /** Completion callback, if only one. */
538 set_interpreter_func set_interpret;
539 /** opaq argument passed to completion \a set_interpret callback. */
542 * Lock for \a set_new_requests manipulations
543 * locked so that any old caller can communicate requests to
544 * the set holder who can then fold them into the lock-free set
546 spinlock_t set_new_req_lock;
547 /** List of new yet unsent requests. Only used with ptlrpcd now. */
548 cfs_list_t set_new_requests;
550 /** rq_status of requests that have been freed already */
552 /** Additional fields used by the flow control extension */
553 /** Maximum number of RPCs in flight */
554 int set_max_inflight;
555 /** Callback function used to generate RPCs */
556 set_producer_func set_producer;
557 /** opaq argument passed to the producer callback */
558 void *set_producer_arg;
562 * Description of a single ptrlrpc_set callback
564 struct ptlrpc_set_cbdata {
565 /** List linkage item */
567 /** Pointer to interpreting function */
568 set_interpreter_func psc_interpret;
569 /** Opaq argument to pass to the callback */
573 struct ptlrpc_bulk_desc;
574 struct ptlrpc_service_part;
575 struct ptlrpc_service;
578 * ptlrpc callback & work item stuff
580 struct ptlrpc_cb_id {
581 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
582 void *cbid_arg; /* additional arg */
585 /** Maximum number of locks to fit into reply state */
586 #define RS_MAX_LOCKS 8
590 * Structure to define reply state on the server
591 * Reply state holds various reply message information. Also for "difficult"
592 * replies (rep-ack case) we store the state after sending reply and wait
593 * for the client to acknowledge the reception. In these cases locks could be
594 * added to the state for replay/failover consistency guarantees.
596 struct ptlrpc_reply_state {
597 /** Callback description */
598 struct ptlrpc_cb_id rs_cb_id;
599 /** Linkage for list of all reply states in a system */
601 /** Linkage for list of all reply states on same export */
602 cfs_list_t rs_exp_list;
603 /** Linkage for list of all reply states for same obd */
604 cfs_list_t rs_obd_list;
606 cfs_list_t rs_debug_list;
608 /** A spinlock to protect the reply state flags */
610 /** Reply state flags */
611 unsigned long rs_difficult:1; /* ACK/commit stuff */
612 unsigned long rs_no_ack:1; /* no ACK, even for
613 difficult requests */
614 unsigned long rs_scheduled:1; /* being handled? */
615 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
616 unsigned long rs_handled:1; /* been handled yet? */
617 unsigned long rs_on_net:1; /* reply_out_callback pending? */
618 unsigned long rs_prealloc:1; /* rs from prealloc list */
619 unsigned long rs_committed:1;/* the transaction was committed
620 and the rs was dispatched
621 by ptlrpc_commit_replies */
622 /** Size of the state */
626 /** Transaction number */
630 struct obd_export *rs_export;
631 struct ptlrpc_service_part *rs_svcpt;
632 /** Lnet metadata handle for the reply */
633 lnet_handle_md_t rs_md_h;
634 cfs_atomic_t rs_refcount;
636 /** Context for the sevice thread */
637 struct ptlrpc_svc_ctx *rs_svc_ctx;
638 /** Reply buffer (actually sent to the client), encoded if needed */
639 struct lustre_msg *rs_repbuf; /* wrapper */
640 /** Size of the reply buffer */
641 int rs_repbuf_len; /* wrapper buf length */
642 /** Size of the reply message */
643 int rs_repdata_len; /* wrapper msg length */
645 * Actual reply message. Its content is encrupted (if needed) to
646 * produce reply buffer for actual sending. In simple case
647 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
649 struct lustre_msg *rs_msg; /* reply message */
651 /** Number of locks awaiting client ACK */
653 /** Handles of locks awaiting client reply ACK */
654 struct lustre_handle rs_locks[RS_MAX_LOCKS];
655 /** Lock modes of locks in \a rs_locks */
656 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
659 struct ptlrpc_thread;
663 RQ_PHASE_NEW = 0xebc0de00,
664 RQ_PHASE_RPC = 0xebc0de01,
665 RQ_PHASE_BULK = 0xebc0de02,
666 RQ_PHASE_INTERPRET = 0xebc0de03,
667 RQ_PHASE_COMPLETE = 0xebc0de04,
668 RQ_PHASE_UNREGISTERING = 0xebc0de05,
669 RQ_PHASE_UNDEFINED = 0xebc0de06
672 /** Type of request interpreter call-back */
673 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
674 struct ptlrpc_request *req,
678 * Definition of request pool structure.
679 * The pool is used to store empty preallocated requests for the case
680 * when we would actually need to send something without performing
681 * any allocations (to avoid e.g. OOM).
683 struct ptlrpc_request_pool {
684 /** Locks the list */
686 /** list of ptlrpc_request structs */
687 cfs_list_t prp_req_list;
688 /** Maximum message size that would fit into a rquest from this pool */
690 /** Function to allocate more requests for this pool */
691 void (*prp_populate)(struct ptlrpc_request_pool *, int);
700 * \defgroup nrs Network Request Scheduler
703 struct ptlrpc_nrs_policy;
704 struct ptlrpc_nrs_resource;
705 struct ptlrpc_nrs_request;
708 * NRS control operations.
710 * These are common for all policies.
712 enum ptlrpc_nrs_ctl {
714 * Activate the policy.
716 PTLRPC_NRS_CTL_START,
718 * Reserved for multiple primary policies, which may be a possibility
723 * Recycle resources for inactive policies.
725 PTLRPC_NRS_CTL_SHRINK,
727 * Not a valid opcode.
729 PTLRPC_NRS_CTL_INVALID,
731 * Policies can start using opcodes from this value and onwards for
732 * their own purposes; the assigned value itself is arbitrary.
734 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
738 * NRS policy operations.
740 * These determine the behaviour of a policy, and are called in response to
743 struct ptlrpc_nrs_pol_ops {
745 * Called during policy registration; this operation is optional.
747 * \param[in] policy The policy being initialized
749 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
751 * Called during policy unregistration; this operation is optional.
753 * \param[in] policy The policy being unregistered/finalized
755 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
757 * Called when activating a policy via lprocfs; policies allocate and
758 * initialize their resources here; this operation is optional.
760 * \param[in] policy The policy being started
762 * \see nrs_policy_start_locked()
764 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
766 * Called when deactivating a policy via lprocfs; policies deallocate
767 * their resources here; this operation is optional
769 * \param[in] policy The policy being stopped
771 * \see nrs_policy_stop_final()
773 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
775 * Used for policy-specific operations; i.e. not generic ones like
776 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
777 * to an ioctl; this operation is optional.
779 * \param[in] policy The policy carrying out operation \a opc
780 * \param[in] opc The command operation being carried out
781 * \param[in,out] arg An generic buffer for communication between the
782 * user and the control operation
787 * \see ptlrpc_nrs_policy_control()
789 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
790 enum ptlrpc_nrs_ctl opc, void *arg);
793 * Called when obtaining references to the resources of the resource
794 * hierarchy for a request that has arrived for handling at the PTLRPC
795 * service. Policies should return -ve for requests they do not wish
796 * to handle. This operation is mandatory.
798 * \param[in] policy The policy we're getting resources for.
799 * \param[in] nrq The request we are getting resources for.
800 * \param[in] parent The parent resource of the resource being
801 * requested; set to NULL if none.
802 * \param[out] resp The resource is to be returned here; the
803 * fallback policy in an NRS head should
804 * \e always return a non-NULL pointer value.
805 * \param[in] moving_req When set, signifies that this is an attempt
806 * to obtain resources for a request being moved
807 * to the high-priority NRS head by
808 * ldlm_lock_reorder_req().
809 * This implies two things:
810 * 1. We are under obd_export::exp_rpc_lock and
811 * so should not sleep.
812 * 2. We should not perform non-idempotent or can
813 * skip performing idempotent operations that
814 * were carried out when resources were first
815 * taken for the request when it was initialized
816 * in ptlrpc_nrs_req_initialize().
818 * \retval 0, +ve The level of the returned resource in the resource
819 * hierarchy; currently only 0 (for a non-leaf resource)
820 * and 1 (for a leaf resource) are supported by the
824 * \see ptlrpc_nrs_req_initialize()
825 * \see ptlrpc_nrs_hpreq_add_nolock()
826 * \see ptlrpc_nrs_req_hp_move()
828 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
829 struct ptlrpc_nrs_request *nrq,
830 struct ptlrpc_nrs_resource *parent,
831 struct ptlrpc_nrs_resource **resp,
834 * Called when releasing references taken for resources in the resource
835 * hierarchy for the request; this operation is optional.
837 * \param[in] policy The policy the resource belongs to
838 * \param[in] res The resource to be freed
840 * \see ptlrpc_nrs_req_finalize()
841 * \see ptlrpc_nrs_hpreq_add_nolock()
842 * \see ptlrpc_nrs_req_hp_move()
844 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
845 struct ptlrpc_nrs_resource *res);
848 * Obtain a request for handling from the policy via polling; this
849 * operation is mandatory.
851 * \param[in] policy The policy to poll
853 * \retval NULL No erquest available for handling
854 * \retval valid-pointer The request polled for handling
856 * \see ptlrpc_nrs_req_poll_nolock()
858 struct ptlrpc_nrs_request *
859 (*op_req_poll) (struct ptlrpc_nrs_policy *policy);
861 * Called when attempting to add a request to a policy for later
862 * handling; this operation is mandatory.
864 * \param[in] policy The policy on which to enqueue \a nrq
865 * \param[in] nrq The request to enqueue
870 * \see ptlrpc_nrs_req_add_nolock()
872 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
873 struct ptlrpc_nrs_request *nrq);
875 * Removes a request from the policy's set of pending requests. Normally
876 * called after a request has been polled successfully from the policy
877 * for handling; this operation is mandatory.
879 * \param[in] policy The policy the request \a nrq belongs to
880 * \param[in] nrq The request to dequeue
882 * \see ptlrpc_nrs_req_del_nolock()
884 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
885 struct ptlrpc_nrs_request *nrq);
887 * Called before carrying out the request; should not block. Could be
888 * used for job/resource control; this operation is optional.
890 * \param[in] policy The policy which is starting to handle request
892 * \param[in] nrq The request
894 * \pre spin_is_locked(&svcpt->scp_req_lock)
896 * \see ptlrpc_nrs_req_start_nolock()
898 void (*op_req_start) (struct ptlrpc_nrs_policy *policy,
899 struct ptlrpc_nrs_request *nrq);
901 * Called after the request being carried out. Could be used for
902 * job/resource control; this operation is optional.
904 * \param[in] policy The policy which is stopping to handle request
906 * \param[in] nrq The request
908 * \pre spin_is_locked(&svcpt->scp_req_lock)
910 * \see ptlrpc_nrs_req_stop_nolock()
912 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
913 struct ptlrpc_nrs_request *nrq);
915 * Registers the policy's lprocfs interface with a PTLRPC service.
917 * \param[in] svc The service
922 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
924 * Unegisters the policy's lprocfs interface with a PTLRPC service.
926 * \param[in] svc The service
928 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
934 enum nrs_policy_flags {
936 * Fallback policy, use this flag only on a single supported policy per
937 * service. Do not use this flag for policies registering using
938 * ptlrpc_nrs_policy_register() (i.e. ones that are not in
939 * \e nrs_pols_builtin).
941 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
943 * Start policy immediately after registering.
945 PTLRPC_NRS_FL_REG_START = (1 << 1),
947 * This is a polciy registering externally with NRS core, via
948 * ptlrpc_nrs_policy_register(), (i.e. one that is not in
949 * \e nrs_pols_builtin. Used to avoid ptlrpc_nrs_policy_register()
950 * racing with a policy start operation issued by the user via lprocfs.
952 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
958 * Denotes whether an NRS instance is for handling normal or high-priority
959 * RPCs, or whether an operation pertains to one or both of the NRS instances
962 enum ptlrpc_nrs_queue_type {
963 PTLRPC_NRS_QUEUE_REG,
965 PTLRPC_NRS_QUEUE_BOTH,
971 * A PTLRPC service has at least one NRS head instance for handling normal
972 * priority RPCs, and may optionally have a second NRS head instance for
973 * handling high-priority RPCs. Each NRS head maintains a list of available
974 * policies, of which one and only one policy is acting as the fallback policy,
975 * and optionally a different policy may be acting as the primary policy. For
976 * all RPCs handled by this NRS head instance, NRS core will first attempt to
977 * enqueue the RPC using the primary policy (if any). The fallback policy is
978 * used in the following cases:
979 * - when there was no primary policy in the
980 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
982 * - when the primary policy that was at the
983 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
984 * RPC was initialized, denoted it did not wish, or for some other reason was
985 * not able to handle the request, by returning a non-valid NRS resource
987 * - when the primary policy that was at the
988 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
989 * RPC was initialized, fails later during the request enqueueing stage.
991 * \see nrs_resource_get_safe()
992 * \see nrs_request_enqueue()
996 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
998 * Linkage into nrs_core_heads_list
1000 cfs_list_t nrs_heads;
1002 * List of registered policies
1004 cfs_list_t nrs_policy_list;
1006 * List of policies with queued requests. Policies that have any
1007 * outstanding requests are queued here, and this list is queried
1008 * in a round-robin manner from NRS core when obtaining a request
1009 * for handling. This ensures that requests from policies that at some
1010 * point transition away from the
1011 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1013 cfs_list_t nrs_policy_queued;
1015 * Service partition for this NRS head
1017 struct ptlrpc_service_part *nrs_svcpt;
1019 * Primary policy, which is the preferred policy for handling RPCs
1021 struct ptlrpc_nrs_policy *nrs_policy_primary;
1023 * Fallback policy, which is the backup policy for handling RPCs
1025 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1027 * This NRS head handles either HP or regular requests
1029 enum ptlrpc_nrs_queue_type nrs_queue_type;
1031 * # queued requests from all policies in this NRS head
1033 unsigned long nrs_req_queued;
1035 * # scheduled requests from all policies in this NRS head
1037 unsigned long nrs_req_started;
1039 * # policies on this NRS
1040 * TODO: Can we avoid having this?
1042 unsigned nrs_num_pols;
1044 * This NRS head is in progress of starting a policy
1046 unsigned nrs_policy_starting:1;
1048 * In progress of shutting down the whole NRS head; used during
1051 unsigned nrs_stopping:1;
1054 #define NRS_POL_NAME_MAX 16
1057 * NRS policy registering descriptor
1059 * Is used to hold a description of a policy that can be passed to NRS core in
1060 * order to register the policy with NRS heads in different PTLRPC services.
1062 struct ptlrpc_nrs_pol_desc {
1064 * Human-readable policy name
1066 char pd_name[NRS_POL_NAME_MAX];
1068 * NRS operations for this policy
1070 struct ptlrpc_nrs_pol_ops *pd_ops;
1072 * Service Compatibility function; this determines whether a policy is
1073 * adequate for handling RPCs of a particular PTLRPC service.
1075 * XXX:This should give the same result during policy
1076 * registration and unregistration, and for all partitions of a
1077 * service; so the result should not depend on temporal service
1078 * or other properties, that may influence the result.
1080 bool (*pd_compat) (struct ptlrpc_service *svc,
1081 const struct ptlrpc_nrs_pol_desc *desc);
1083 * Optionally set for policies that support a single ptlrpc service,
1084 * i.e. ones that have \a pd_compat set to nrs_policy_compat_one()
1086 char *pd_compat_svc_name;
1088 * Bitmask of nrs_policy_flags
1092 * Link into nrs_core::nrs_policies
1100 * Policies transition from one state to the other during their lifetime
1102 enum ptlrpc_nrs_pol_state {
1104 * Not a valid policy state.
1106 NRS_POL_STATE_INVALID,
1108 * For now, this state is used exclusively for policies that register
1109 * externally to NRS core, i.e. ones that do so via
1110 * ptlrpc_nrs_policy_register() and are not part of nrs_pols_builtin;
1111 * it is used to prevent a race condition between the policy registering
1112 * with more than one service partition while service is operational,
1113 * and the user starting the policy via lprocfs.
1115 * \see nrs_pol_make_avail()
1117 NRS_POL_STATE_UNAVAIL,
1119 * Policies are at this state either at the start of their life, or
1120 * transition here when the user selects a different policy to act
1121 * as the primary one.
1123 NRS_POL_STATE_STOPPED,
1125 * Policy is progress of stopping
1127 NRS_POL_STATE_STOPPING,
1129 * Policy is in progress of starting
1131 NRS_POL_STATE_STARTING,
1133 * A policy is in this state in two cases:
1134 * - it is the fallback policy, which is always in this state.
1135 * - it has been activated by the user; i.e. it is the primary policy,
1137 NRS_POL_STATE_STARTED,
1141 * NRS policy information
1143 * Used for obtaining information for the status of a policy via lprocfs
1145 struct ptlrpc_nrs_pol_info {
1149 char pi_name[NRS_POL_NAME_MAX];
1151 * Current policy state
1153 enum ptlrpc_nrs_pol_state pi_state;
1155 * # RPCs enqueued for later dispatching by the policy
1159 * # RPCs started for dispatch by the policy
1161 long pi_req_started;
1163 * Is this a fallback policy?
1165 unsigned pi_fallback:1;
1171 * There is one instance of this for each policy in each NRS head of each
1172 * PTLRPC service partition.
1174 struct ptlrpc_nrs_policy {
1176 * Linkage into the NRS head's list of policies,
1177 * ptlrpc_nrs:nrs_policy_list
1179 cfs_list_t pol_list;
1181 * Linkage into the NRS head's list of policies with enqueued
1182 * requests ptlrpc_nrs:nrs_policy_queued
1184 cfs_list_t pol_list_queued;
1186 * Current state of this policy
1188 enum ptlrpc_nrs_pol_state pol_state;
1190 * Bitmask of nrs_policy_flags
1194 * # RPCs enqueued for later dispatching by the policy
1196 long pol_req_queued;
1198 * # RPCs started for dispatch by the policy
1200 long pol_req_started;
1202 * Usage Reference count taken on the policy instance
1206 * The NRS head this policy has been created at
1208 struct ptlrpc_nrs *pol_nrs;
1210 * NRS operations for this policy; points to ptlrpc_nrs_pol_desc::pd_ops
1212 struct ptlrpc_nrs_pol_ops *pol_ops;
1214 * Private policy data; varies by policy type
1218 * Human-readable policy name; point to ptlrpc_nrs_pol_desc::pd_name
1226 * Resources are embedded into two types of NRS entities:
1227 * - Inside NRS policies, in the policy's private data in
1228 * ptlrpc_nrs_policy::pol_private
1229 * - In objects that act as prime-level scheduling entities in different NRS
1230 * policies; e.g. on a policy that performs round robin or similar order
1231 * scheduling across client NIDs, there would be one NRS resource per unique
1232 * client NID. On a policy which performs round robin scheduling across
1233 * backend filesystem objects, there would be one resource associated with
1234 * each of the backend filesystem objects partaking in the scheduling
1235 * performed by the policy.
1237 * NRS resources share a parent-child relationship, in which resources embedded
1238 * in policy instances are the parent entities, with all scheduling entities
1239 * a policy schedules across being the children, thus forming a simple resource
1240 * hierarchy. This hierarchy may be extended with one or more levels in the
1241 * future if the ability to have more than one primary policy is added.
1243 * Upon request initialization, references to the then active NRS policies are
1244 * taken and used to later handle the dispatching of the request with one of
1247 * \see nrs_resource_get_safe()
1248 * \see ptlrpc_nrs_req_add()
1250 struct ptlrpc_nrs_resource {
1252 * This NRS resource's parent; is NULL for resources embedded in NRS
1253 * policy instances; i.e. those are top-level ones.
1255 struct ptlrpc_nrs_resource *res_parent;
1257 * The policy associated with this resource.
1259 struct ptlrpc_nrs_policy *res_policy;
1272 * This policy is a logical wrapper around previous, non-NRS functionality.
1273 * It dispatches RPCs in the same order as they arrive from the network. This
1274 * policy is currently used as the fallback policy, and the only enabled policy
1275 * on all NRS heads of all PTLRPC service partitions.
1280 * Private data structure for the FIFO policy
1282 struct nrs_fifo_head {
1284 * Resource object for policy instance.
1286 struct ptlrpc_nrs_resource fh_res;
1288 * List of queued requests.
1292 * For debugging purposes.
1297 struct nrs_fifo_req {
1298 /** request header, must be the first member of structure */
1308 * Instances of this object exist embedded within ptlrpc_request; the main
1309 * purpose of this object is to hold references to the request's resources
1310 * for the lifetime of the request, and to hold properties that policies use
1311 * use for determining the request's scheduling priority.
1313 struct ptlrpc_nrs_request {
1315 * The request's resource hierarchy.
1317 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1319 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1320 * policy that was used to enqueue the request.
1322 * \see nrs_request_enqueue()
1324 unsigned nr_res_idx;
1325 unsigned nr_initialized:1;
1326 unsigned nr_enqueued:1;
1327 unsigned nr_dequeued:1;
1328 unsigned nr_started:1;
1329 unsigned nr_finalized:1;
1330 cfs_binheap_node_t nr_node;
1333 * Policy-specific fields, used for determining a request's scheduling
1334 * priority, and other supporting functionality.
1338 * Fields for the FIFO policy
1340 struct nrs_fifo_req fifo;
1343 * Externally-registering policies may want to use this to allocate
1344 * their own request properties.
1352 * Basic request prioritization operations structure.
1353 * The whole idea is centered around locks and RPCs that might affect locks.
1354 * When a lock is contended we try to give priority to RPCs that might lead
1355 * to fastest release of that lock.
1356 * Currently only implemented for OSTs only in a way that makes all
1357 * IO and truncate RPCs that are coming from a locked region where a lock is
1358 * contended a priority over other requests.
1360 struct ptlrpc_hpreq_ops {
1362 * Check if the lock handle of the given lock is the same as
1363 * taken from the request.
1365 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1367 * Check if the request is a high priority one.
1369 int (*hpreq_check)(struct ptlrpc_request *);
1371 * Called after the request has been handled.
1373 void (*hpreq_fini)(struct ptlrpc_request *);
1377 * Represents remote procedure call.
1379 * This is a staple structure used by everybody wanting to send a request
1382 struct ptlrpc_request {
1383 /* Request type: one of PTL_RPC_MSG_* */
1385 /** Result of request processing */
1388 * Linkage item through which this request is included into
1389 * sending/delayed lists on client and into rqbd list on server
1393 * Server side list of incoming unserved requests sorted by arrival
1394 * time. Traversed from time to time to notice about to expire
1395 * requests and sent back "early replies" to clients to let them
1396 * know server is alive and well, just very busy to service their
1399 cfs_list_t rq_timed_list;
1400 /** server-side history, used for debuging purposes. */
1401 cfs_list_t rq_history_list;
1402 /** server-side per-export list */
1403 cfs_list_t rq_exp_list;
1404 /** server-side hp handlers */
1405 struct ptlrpc_hpreq_ops *rq_ops;
1407 /** initial thread servicing this request */
1408 struct ptlrpc_thread *rq_svc_thread;
1410 /** history sequence # */
1411 __u64 rq_history_seq;
1415 /** stub for NRS request */
1416 struct ptlrpc_nrs_request rq_nrq;
1418 /** the index of service's srv_at_array into which request is linked */
1420 /** Lock to protect request flags and some other important bits, like
1424 /** client-side flags are serialized by rq_lock */
1425 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1426 rq_timedout:1, rq_resend:1, rq_restart:1,
1428 * when ->rq_replay is set, request is kept by the client even
1429 * after server commits corresponding transaction. This is
1430 * used for operations that require sequence of multiple
1431 * requests to be replayed. The only example currently is file
1432 * open/close. When last request in such a sequence is
1433 * committed, ->rq_replay is cleared on all requests in the
1437 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1438 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1439 rq_early:1, rq_must_unlink:1,
1440 rq_memalloc:1, /* req originated from "kswapd" */
1441 /* server-side flags */
1442 rq_packed_final:1, /* packed final reply */
1443 rq_hp:1, /* high priority RPC */
1444 rq_at_linked:1, /* link into service's srv_at_array */
1445 rq_reply_truncate:1,
1447 /* whether the "rq_set" is a valid one */
1449 rq_generation_set:1,
1450 /* do not resend request on -EINPROGRESS */
1451 rq_no_retry_einprogress:1,
1452 /* allow the req to be sent if the import is in recovery
1456 unsigned int rq_nr_resend;
1458 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1459 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1460 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
1461 server-side refcounf for multiple replies */
1463 /** Portal to which this request would be sent */
1464 short rq_request_portal; /* XXX FIXME bug 249 */
1465 /** Portal where to wait for reply and where reply would be sent */
1466 short rq_reply_portal; /* XXX FIXME bug 249 */
1470 * !rq_truncate : # reply bytes actually received,
1471 * rq_truncate : required repbuf_len for resend
1473 int rq_nob_received;
1474 /** Request length */
1478 /** Request message - what client sent */
1479 struct lustre_msg *rq_reqmsg;
1480 /** Reply message - server response */
1481 struct lustre_msg *rq_repmsg;
1482 /** Transaction number */
1487 * List item to for replay list. Not yet commited requests get linked
1489 * Also see \a rq_replay comment above.
1491 cfs_list_t rq_replay_list;
1494 * security and encryption data
1496 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1497 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1498 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
1500 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1501 enum lustre_sec_part rq_sp_from;
1503 /* client/server security flags */
1505 rq_ctx_init:1, /* context initiation */
1506 rq_ctx_fini:1, /* context destroy */
1507 rq_bulk_read:1, /* request bulk read */
1508 rq_bulk_write:1, /* request bulk write */
1509 /* server authentication flags */
1510 rq_auth_gss:1, /* authenticated by gss */
1511 rq_auth_remote:1, /* authed as remote user */
1512 rq_auth_usr_root:1, /* authed as root */
1513 rq_auth_usr_mdt:1, /* authed as mdt */
1514 rq_auth_usr_ost:1, /* authed as ost */
1515 /* security tfm flags */
1518 /* doesn't expect reply FIXME */
1520 rq_pill_init:1; /* pill initialized */
1522 uid_t rq_auth_uid; /* authed uid */
1523 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1525 /* (server side), pointed directly into req buffer */
1526 struct ptlrpc_user_desc *rq_user_desc;
1528 /* various buffer pointers */
1529 struct lustre_msg *rq_reqbuf; /* req wrapper */
1530 char *rq_repbuf; /* rep buffer */
1531 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1532 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1533 int rq_reqbuf_len; /* req wrapper buf len */
1534 int rq_reqdata_len; /* req wrapper msg len */
1535 int rq_repbuf_len; /* rep buffer len */
1536 int rq_repdata_len; /* rep wrapper msg len */
1537 int rq_clrbuf_len; /* only in priv mode */
1538 int rq_clrdata_len; /* only in priv mode */
1540 /** early replies go to offset 0, regular replies go after that */
1541 unsigned int rq_reply_off;
1545 /** Fields that help to see if request and reply were swabbed or not */
1546 __u32 rq_req_swab_mask;
1547 __u32 rq_rep_swab_mask;
1549 /** What was import generation when this request was sent */
1550 int rq_import_generation;
1551 enum lustre_imp_state rq_send_state;
1553 /** how many early replies (for stats) */
1556 /** client+server request */
1557 lnet_handle_md_t rq_req_md_h;
1558 struct ptlrpc_cb_id rq_req_cbid;
1559 /** optional time limit for send attempts */
1560 cfs_duration_t rq_delay_limit;
1561 /** time request was first queued */
1562 cfs_time_t rq_queued_time;
1564 /* server-side... */
1565 /** request arrival time */
1566 struct timeval rq_arrival_time;
1567 /** separated reply state */
1568 struct ptlrpc_reply_state *rq_reply_state;
1569 /** incoming request buffer */
1570 struct ptlrpc_request_buffer_desc *rq_rqbd;
1572 /** client-only incoming reply */
1573 lnet_handle_md_t rq_reply_md_h;
1574 cfs_waitq_t rq_reply_waitq;
1575 struct ptlrpc_cb_id rq_reply_cbid;
1579 /** Peer description (the other side) */
1580 lnet_process_id_t rq_peer;
1581 /** Server-side, export on which request was received */
1582 struct obd_export *rq_export;
1583 /** Client side, import where request is being sent */
1584 struct obd_import *rq_import;
1586 /** Replay callback, called after request is replayed at recovery */
1587 void (*rq_replay_cb)(struct ptlrpc_request *);
1589 * Commit callback, called when request is committed and about to be
1592 void (*rq_commit_cb)(struct ptlrpc_request *);
1593 /** Opaq data for replay and commit callbacks. */
1596 /** For bulk requests on client only: bulk descriptor */
1597 struct ptlrpc_bulk_desc *rq_bulk;
1599 /** client outgoing req */
1601 * when request/reply sent (secs), or time when request should be sent
1604 /** time for request really sent out */
1605 time_t rq_real_sent;
1607 /** when request must finish. volatile
1608 * so that servers' early reply updates to the deadline aren't
1609 * kept in per-cpu cache */
1610 volatile time_t rq_deadline;
1611 /** when req reply unlink must finish. */
1612 time_t rq_reply_deadline;
1613 /** when req bulk unlink must finish. */
1614 time_t rq_bulk_deadline;
1616 * service time estimate (secs)
1617 * If the requestsis not served by this time, it is marked as timed out.
1621 /** Multi-rpc bits */
1622 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1623 cfs_waitq_t rq_set_waitq;
1624 /** Link item for request set lists */
1625 cfs_list_t rq_set_chain;
1626 /** Link back to the request set */
1627 struct ptlrpc_request_set *rq_set;
1628 /** Async completion handler, called when reply is received */
1629 ptlrpc_interpterer_t rq_interpret_reply;
1630 /** Async completion context */
1631 union ptlrpc_async_args rq_async_args;
1633 /** Pool if request is from preallocated list */
1634 struct ptlrpc_request_pool *rq_pool;
1636 struct lu_context rq_session;
1637 struct lu_context rq_recov_session;
1639 /** request format description */
1640 struct req_capsule rq_pill;
1644 * Call completion handler for rpc if any, return it's status or original
1645 * rc if there was no handler defined for this request.
1647 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1648 struct ptlrpc_request *req, int rc)
1650 if (req->rq_interpret_reply != NULL) {
1651 req->rq_status = req->rq_interpret_reply(env, req,
1652 &req->rq_async_args,
1654 return req->rq_status;
1662 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_desc *desc);
1663 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_desc *desc);
1664 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1665 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1666 struct ptlrpc_nrs_pol_info *info);
1669 * Can the request be moved from the regular NRS head to the high-priority NRS
1670 * head (of the same PTLRPC service partition), if any?
1672 * For a reliable result, this should be checked under svcpt->scp_req lock.
1675 ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1677 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1680 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1681 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1682 * to make sure it has not been scheduled yet (analogous to previous
1683 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1685 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1690 * Returns 1 if request buffer at offset \a index was already swabbed
1692 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1694 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1695 return req->rq_req_swab_mask & (1 << index);
1699 * Returns 1 if request reply buffer at offset \a index was already swabbed
1701 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1703 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1704 return req->rq_rep_swab_mask & (1 << index);
1708 * Returns 1 if request needs to be swabbed into local cpu byteorder
1710 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1712 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1716 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1718 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1720 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1724 * Mark request buffer at offset \a index that it was already swabbed
1726 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1728 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1729 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1730 req->rq_req_swab_mask |= 1 << index;
1734 * Mark request reply buffer at offset \a index that it was already swabbed
1736 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1738 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1739 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1740 req->rq_rep_swab_mask |= 1 << index;
1744 * Convert numerical request phase value \a phase into text string description
1746 static inline const char *
1747 ptlrpc_phase2str(enum rq_phase phase)
1756 case RQ_PHASE_INTERPRET:
1758 case RQ_PHASE_COMPLETE:
1760 case RQ_PHASE_UNREGISTERING:
1761 return "Unregistering";
1768 * Convert numerical request phase of the request \a req into text stringi
1771 static inline const char *
1772 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1774 return ptlrpc_phase2str(req->rq_phase);
1778 * Debugging functions and helpers to print request structure into debug log
1781 /* Spare the preprocessor, spoil the bugs. */
1782 #define FLAG(field, str) (field ? str : "")
1784 /** Convert bit flags into a string */
1785 #define DEBUG_REQ_FLAGS(req) \
1786 ptlrpc_rqphase2str(req), \
1787 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1788 FLAG(req->rq_err, "E"), \
1789 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1790 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1791 FLAG(req->rq_no_resend, "N"), \
1792 FLAG(req->rq_waiting, "W"), \
1793 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1794 FLAG(req->rq_committed, "M")
1796 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1798 void _debug_req(struct ptlrpc_request *req,
1799 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1800 __attribute__ ((format (printf, 3, 4)));
1803 * Helper that decides if we need to print request accordig to current debug
1806 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1808 CFS_CHECK_STACK(msgdata, mask, cdls); \
1810 if (((mask) & D_CANTMASK) != 0 || \
1811 ((libcfs_debug & (mask)) != 0 && \
1812 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1813 _debug_req((req), msgdata, fmt, ##a); \
1817 * This is the debug print function you need to use to print request sturucture
1818 * content into lustre debug log.
1819 * for most callers (level is a constant) this is resolved at compile time */
1820 #define DEBUG_REQ(level, req, fmt, args...) \
1822 if ((level) & (D_ERROR | D_WARNING)) { \
1823 static cfs_debug_limit_state_t cdls; \
1824 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1825 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1827 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1828 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1834 * Structure that defines a single page of a bulk transfer
1836 struct ptlrpc_bulk_page {
1837 /** Linkage to list of pages in a bulk */
1840 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1843 /** offset within a page */
1845 /** The page itself */
1846 struct page *bp_page;
1849 #define BULK_GET_SOURCE 0
1850 #define BULK_PUT_SINK 1
1851 #define BULK_GET_SINK 2
1852 #define BULK_PUT_SOURCE 3
1855 * Definition of bulk descriptor.
1856 * Bulks are special "Two phase" RPCs where initial request message
1857 * is sent first and it is followed bt a transfer (o receiving) of a large
1858 * amount of data to be settled into pages referenced from the bulk descriptors.
1859 * Bulks transfers (the actual data following the small requests) are done
1860 * on separate LNet portals.
1861 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1862 * Another user is readpage for MDT.
1864 struct ptlrpc_bulk_desc {
1865 /** completed with failure */
1866 unsigned long bd_failure:1;
1867 /** {put,get}{source,sink} */
1868 unsigned long bd_type:2;
1870 unsigned long bd_registered:1;
1871 /** For serialization with callback */
1873 /** Import generation when request for this bulk was sent */
1874 int bd_import_generation;
1875 /** LNet portal for this bulk */
1877 /** Server side - export this bulk created for */
1878 struct obd_export *bd_export;
1879 /** Client side - import this bulk was sent on */
1880 struct obd_import *bd_import;
1881 /** Back pointer to the request */
1882 struct ptlrpc_request *bd_req;
1883 cfs_waitq_t bd_waitq; /* server side only WQ */
1884 int bd_iov_count; /* # entries in bd_iov */
1885 int bd_max_iov; /* allocated size of bd_iov */
1886 int bd_nob; /* # bytes covered */
1887 int bd_nob_transferred; /* # bytes GOT/PUT */
1891 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1892 lnet_nid_t bd_sender; /* stash event::sender */
1893 int bd_md_count; /* # valid entries in bd_mds */
1894 int bd_md_max_brw; /* max entries in bd_mds */
1895 /** array of associated MDs */
1896 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
1898 #if defined(__KERNEL__)
1900 * encrypt iov, size is either 0 or bd_iov_count.
1902 lnet_kiov_t *bd_enc_iov;
1904 lnet_kiov_t bd_iov[0];
1906 lnet_md_iovec_t bd_iov[0];
1911 SVC_STOPPED = 1 << 0,
1912 SVC_STOPPING = 1 << 1,
1913 SVC_STARTING = 1 << 2,
1914 SVC_RUNNING = 1 << 3,
1916 SVC_SIGNAL = 1 << 5,
1919 #define PTLRPC_THR_NAME_LEN 32
1921 * Definition of server service thread structure
1923 struct ptlrpc_thread {
1925 * List of active threads in svc->srv_threads
1929 * thread-private data (preallocated memory)
1934 * service thread index, from ptlrpc_start_threads
1938 * service thread pid
1942 * put watchdog in the structure per thread b=14840
1944 struct lc_watchdog *t_watchdog;
1946 * the svc this thread belonged to b=18582
1948 struct ptlrpc_service_part *t_svcpt;
1949 cfs_waitq_t t_ctl_waitq;
1950 struct lu_env *t_env;
1951 char t_name[PTLRPC_THR_NAME_LEN];
1954 static inline int thread_is_init(struct ptlrpc_thread *thread)
1956 return thread->t_flags == 0;
1959 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1961 return !!(thread->t_flags & SVC_STOPPED);
1964 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1966 return !!(thread->t_flags & SVC_STOPPING);
1969 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1971 return !!(thread->t_flags & SVC_STARTING);
1974 static inline int thread_is_running(struct ptlrpc_thread *thread)
1976 return !!(thread->t_flags & SVC_RUNNING);
1979 static inline int thread_is_event(struct ptlrpc_thread *thread)
1981 return !!(thread->t_flags & SVC_EVENT);
1984 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1986 return !!(thread->t_flags & SVC_SIGNAL);
1989 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1991 thread->t_flags &= ~flags;
1994 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1996 thread->t_flags = flags;
1999 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2001 thread->t_flags |= flags;
2004 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2007 if (thread->t_flags & flags) {
2008 thread->t_flags &= ~flags;
2015 * Request buffer descriptor structure.
2016 * This is a structure that contains one posted request buffer for service.
2017 * Once data land into a buffer, event callback creates actual request and
2018 * notifies wakes one of the service threads to process new incoming request.
2019 * More than one request can fit into the buffer.
2021 struct ptlrpc_request_buffer_desc {
2022 /** Link item for rqbds on a service */
2023 cfs_list_t rqbd_list;
2024 /** History of requests for this buffer */
2025 cfs_list_t rqbd_reqs;
2026 /** Back pointer to service for which this buffer is registered */
2027 struct ptlrpc_service_part *rqbd_svcpt;
2028 /** LNet descriptor */
2029 lnet_handle_md_t rqbd_md_h;
2031 /** The buffer itself */
2033 struct ptlrpc_cb_id rqbd_cbid;
2035 * This "embedded" request structure is only used for the
2036 * last request to fit into the buffer
2038 struct ptlrpc_request rqbd_req;
2041 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2043 struct ptlrpc_service_ops {
2045 * if non-NULL called during thread creation (ptlrpc_start_thread())
2046 * to initialize service specific per-thread state.
2048 int (*so_thr_init)(struct ptlrpc_thread *thr);
2050 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2051 * destruct state created by ->srv_init().
2053 void (*so_thr_done)(struct ptlrpc_thread *thr);
2055 * Handler function for incoming requests for this service
2057 int (*so_req_handler)(struct ptlrpc_request *req);
2059 * function to determine priority of the request, it's called
2060 * on every new request
2062 int (*so_hpreq_handler)(struct ptlrpc_request *);
2064 * service-specific print fn
2066 void (*so_req_printer)(void *, struct ptlrpc_request *);
2069 #ifndef __cfs_cacheline_aligned
2070 /* NB: put it here for reducing patche dependence */
2071 # define __cfs_cacheline_aligned
2075 * How many high priority requests to serve before serving one normal
2078 #define PTLRPC_SVC_HP_RATIO 10
2081 * Definition of PortalRPC service.
2082 * The service is listening on a particular portal (like tcp port)
2083 * and perform actions for a specific server like IO service for OST
2084 * or general metadata service for MDS.
2086 struct ptlrpc_service {
2087 /** serialize /proc operations */
2088 spinlock_t srv_lock;
2089 /** most often accessed fields */
2090 /** chain thru all services */
2091 cfs_list_t srv_list;
2092 /** service operations table */
2093 struct ptlrpc_service_ops srv_ops;
2094 /** only statically allocated strings here; we don't clean them */
2096 /** only statically allocated strings here; we don't clean them */
2097 char *srv_thread_name;
2098 /** service thread list */
2099 cfs_list_t srv_threads;
2100 /** threads # should be created for each partition on initializing */
2101 int srv_nthrs_cpt_init;
2102 /** limit of threads number for each partition */
2103 int srv_nthrs_cpt_limit;
2104 /** Root of /proc dir tree for this service */
2105 cfs_proc_dir_entry_t *srv_procroot;
2106 /** Pointer to statistic data for this service */
2107 struct lprocfs_stats *srv_stats;
2108 /** # hp per lp reqs to handle */
2109 int srv_hpreq_ratio;
2110 /** biggest request to receive */
2111 int srv_max_req_size;
2112 /** biggest reply to send */
2113 int srv_max_reply_size;
2114 /** size of individual buffers */
2116 /** # buffers to allocate in 1 group */
2117 int srv_nbuf_per_group;
2118 /** Local portal on which to receive requests */
2119 __u32 srv_req_portal;
2120 /** Portal on the client to send replies to */
2121 __u32 srv_rep_portal;
2123 * Tags for lu_context associated with this thread, see struct
2127 /** soft watchdog timeout multiplier */
2128 int srv_watchdog_factor;
2129 /** under unregister_service */
2130 unsigned srv_is_stopping:1;
2132 /** max # request buffers in history per partition */
2133 int srv_hist_nrqbds_cpt_max;
2134 /** number of CPTs this service bound on */
2136 /** CPTs array this service bound on */
2138 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2140 /** CPT table this service is running over */
2141 struct cfs_cpt_table *srv_cptable;
2143 * partition data for ptlrpc service
2145 struct ptlrpc_service_part *srv_parts[0];
2149 * Definition of PortalRPC service partition data.
2150 * Although a service only has one instance of it right now, but we
2151 * will have multiple instances very soon (instance per CPT).
2153 * it has four locks:
2155 * serialize operations on rqbd and requests waiting for preprocess
2157 * serialize operations active requests sent to this portal
2159 * serialize adaptive timeout stuff
2161 * serialize operations on RS list (reply states)
2163 * We don't have any use-case to take two or more locks at the same time
2164 * for now, so there is no lock order issue.
2166 struct ptlrpc_service_part {
2167 /** back reference to owner */
2168 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2169 /* CPT id, reserved */
2171 /** always increasing number */
2173 /** # of starting threads */
2174 int scp_nthrs_starting;
2175 /** # of stopping threads, reserved for shrinking threads */
2176 int scp_nthrs_stopping;
2177 /** # running threads */
2178 int scp_nthrs_running;
2179 /** service threads list */
2180 cfs_list_t scp_threads;
2183 * serialize the following fields, used for protecting
2184 * rqbd list and incoming requests waiting for preprocess,
2185 * threads starting & stopping are also protected by this lock.
2187 spinlock_t scp_lock __cfs_cacheline_aligned;
2188 /** total # req buffer descs allocated */
2189 int scp_nrqbds_total;
2190 /** # posted request buffers for receiving */
2191 int scp_nrqbds_posted;
2192 /** in progress of allocating rqbd */
2193 int scp_rqbd_allocating;
2194 /** # incoming reqs */
2195 int scp_nreqs_incoming;
2196 /** request buffers to be reposted */
2197 cfs_list_t scp_rqbd_idle;
2198 /** req buffers receiving */
2199 cfs_list_t scp_rqbd_posted;
2200 /** incoming reqs */
2201 cfs_list_t scp_req_incoming;
2202 /** timeout before re-posting reqs, in tick */
2203 cfs_duration_t scp_rqbd_timeout;
2205 * all threads sleep on this. This wait-queue is signalled when new
2206 * incoming request arrives and when difficult reply has to be handled.
2208 cfs_waitq_t scp_waitq;
2210 /** request history */
2211 cfs_list_t scp_hist_reqs;
2212 /** request buffer history */
2213 cfs_list_t scp_hist_rqbds;
2214 /** # request buffers in history */
2215 int scp_hist_nrqbds;
2216 /** sequence number for request */
2218 /** highest seq culled from history */
2219 __u64 scp_hist_seq_culled;
2222 * serialize the following fields, used for processing requests
2223 * sent to this portal
2225 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2226 /** # reqs in either of the NRS heads below */
2227 /** # reqs being served */
2228 int scp_nreqs_active;
2229 /** # HPreqs being served */
2230 int scp_nhreqs_active;
2231 /** # hp requests handled */
2234 /** NRS head for regular requests */
2235 struct ptlrpc_nrs scp_nrs_reg;
2236 /** NRS head for HP requests; this is only valid for services that can
2237 * handle HP requests */
2238 struct ptlrpc_nrs *scp_nrs_hp;
2243 * serialize the following fields, used for changes on
2246 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2247 /** estimated rpc service time */
2248 struct adaptive_timeout scp_at_estimate;
2249 /** reqs waiting for replies */
2250 struct ptlrpc_at_array scp_at_array;
2251 /** early reply timer */
2252 cfs_timer_t scp_at_timer;
2254 cfs_time_t scp_at_checktime;
2255 /** check early replies */
2256 unsigned scp_at_check;
2260 * serialize the following fields, used for processing
2261 * replies for this portal
2263 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2264 /** all the active replies */
2265 cfs_list_t scp_rep_active;
2267 /** replies waiting for service */
2268 cfs_list_t scp_rep_queue;
2270 /** List of free reply_states */
2271 cfs_list_t scp_rep_idle;
2272 /** waitq to run, when adding stuff to srv_free_rs_list */
2273 cfs_waitq_t scp_rep_waitq;
2274 /** # 'difficult' replies */
2275 cfs_atomic_t scp_nreps_difficult;
2278 #define ptlrpc_service_for_each_part(part, i, svc) \
2280 i < (svc)->srv_ncpts && \
2281 (svc)->srv_parts != NULL && \
2282 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2285 * Declaration of ptlrpcd control structure
2287 struct ptlrpcd_ctl {
2289 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2291 unsigned long pc_flags;
2293 * Thread lock protecting structure fields.
2299 struct completion pc_starting;
2303 struct completion pc_finishing;
2305 * Thread requests set.
2307 struct ptlrpc_request_set *pc_set;
2309 * Thread name used in cfs_daemonize()
2313 * Environment for request interpreters to run in.
2315 struct lu_env pc_env;
2317 * Index of ptlrpcd thread in the array.
2321 * Number of the ptlrpcd's partners.
2325 * Pointer to the array of partners' ptlrpcd_ctl structure.
2327 struct ptlrpcd_ctl **pc_partners;
2329 * Record the partner index to be processed next.
2334 * Async rpcs flag to make sure that ptlrpcd_check() is called only
2339 * Currently not used.
2343 * User-space async rpcs callback.
2345 void *pc_wait_callback;
2347 * User-space check idle rpcs callback.
2349 void *pc_idle_callback;
2353 /* Bits for pc_flags */
2354 enum ptlrpcd_ctl_flags {
2356 * Ptlrpc thread start flag.
2358 LIOD_START = 1 << 0,
2360 * Ptlrpc thread stop flag.
2364 * Ptlrpc thread force flag (only stop force so far).
2365 * This will cause aborting any inflight rpcs handled
2366 * by thread if LIOD_STOP is specified.
2368 LIOD_FORCE = 1 << 2,
2370 * This is a recovery ptlrpc thread.
2372 LIOD_RECOVERY = 1 << 3,
2374 * The ptlrpcd is bound to some CPU core.
2383 * Service compatibility function; policy is compatible with all services.
2385 * \param[in] svc The service the policy is attempting to register with.
2386 * \param[in] desc The policy descriptor
2388 * \retval true The policy is compatible with the NRS head
2390 * \see ptlrpc_nrs_pol_desc::pd_compat()
2393 nrs_policy_compat_all(struct ptlrpc_service *svc,
2394 const struct ptlrpc_nrs_pol_desc *desc)
2400 * Service compatibility function; policy is compatible with only a specific
2401 * service which is identified by its human-readable name at
2402 * ptlrpc_service::srv_name.
2404 * \param[in] svc The service the policy is attempting to register with.
2405 * \param[in] desc The policy descriptor
2407 * \retval false The policy is not compatible with the NRS head
2408 * \retval true The policy is compatible with the NRS head
2410 * \see ptlrpc_nrs_pol_desc::pd_compat()
2413 nrs_policy_compat_one(struct ptlrpc_service *svc,
2414 const struct ptlrpc_nrs_pol_desc *desc)
2416 LASSERT(desc->pd_compat_svc_name != NULL);
2417 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2422 /* ptlrpc/events.c */
2423 extern lnet_handle_eq_t ptlrpc_eq_h;
2424 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2425 lnet_process_id_t *peer, lnet_nid_t *self);
2427 * These callbacks are invoked by LNet when something happened to
2431 extern void request_out_callback(lnet_event_t *ev);
2432 extern void reply_in_callback(lnet_event_t *ev);
2433 extern void client_bulk_callback(lnet_event_t *ev);
2434 extern void request_in_callback(lnet_event_t *ev);
2435 extern void reply_out_callback(lnet_event_t *ev);
2436 #ifdef HAVE_SERVER_SUPPORT
2437 extern void server_bulk_callback(lnet_event_t *ev);
2441 /* ptlrpc/connection.c */
2442 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2444 struct obd_uuid *uuid);
2445 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2446 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2447 int ptlrpc_connection_init(void);
2448 void ptlrpc_connection_fini(void);
2449 extern lnet_pid_t ptl_get_pid(void);
2451 /* ptlrpc/niobuf.c */
2453 * Actual interfacing with LNet to put/get/register/unregister stuff
2456 #ifdef HAVE_SERVER_SUPPORT
2457 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2458 unsigned npages, unsigned max_brw,
2459 unsigned type, unsigned portal);
2460 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2461 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2463 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2467 LASSERT(desc != NULL);
2469 spin_lock(&desc->bd_lock);
2470 rc = desc->bd_md_count;
2471 spin_unlock(&desc->bd_lock);
2476 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2477 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2479 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2481 struct ptlrpc_bulk_desc *desc;
2484 LASSERT(req != NULL);
2485 desc = req->rq_bulk;
2487 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2488 req->rq_bulk_deadline > cfs_time_current_sec())
2494 spin_lock(&desc->bd_lock);
2495 rc = desc->bd_md_count;
2496 spin_unlock(&desc->bd_lock);
2500 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2501 #define PTLRPC_REPLY_EARLY 0x02
2502 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2503 int ptlrpc_reply(struct ptlrpc_request *req);
2504 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2505 int ptlrpc_error(struct ptlrpc_request *req);
2506 void ptlrpc_resend_req(struct ptlrpc_request *request);
2507 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2508 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2509 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2512 /* ptlrpc/client.c */
2514 * Client-side portals API. Everything to send requests, receive replies,
2515 * request queues, request management, etc.
2518 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2519 struct ptlrpc_client *);
2520 void ptlrpc_cleanup_client(struct obd_import *imp);
2521 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2523 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2524 int ptlrpc_replay_req(struct ptlrpc_request *req);
2525 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2526 void ptlrpc_restart_req(struct ptlrpc_request *req);
2527 void ptlrpc_abort_inflight(struct obd_import *imp);
2528 void ptlrpc_cleanup_imp(struct obd_import *imp);
2529 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2531 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2532 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2534 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2535 set_interpreter_func fn, void *data);
2536 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2537 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2538 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2539 int ptlrpc_expired_set(void *data);
2540 void ptlrpc_interrupted_set(void *data);
2541 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2542 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2543 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2544 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2545 struct ptlrpc_request *req);
2547 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2548 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2550 struct ptlrpc_request_pool *
2551 ptlrpc_init_rq_pool(int, int,
2552 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2554 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2555 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2556 const struct req_format *format);
2557 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2558 struct ptlrpc_request_pool *,
2559 const struct req_format *format);
2560 void ptlrpc_request_free(struct ptlrpc_request *request);
2561 int ptlrpc_request_pack(struct ptlrpc_request *request,
2562 __u32 version, int opcode);
2563 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2564 const struct req_format *format,
2565 __u32 version, int opcode);
2566 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2567 __u32 version, int opcode, char **bufs,
2568 struct ptlrpc_cli_ctx *ctx);
2569 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2570 int opcode, int count, __u32 *lengths,
2572 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2573 __u32 version, int opcode,
2574 int count, __u32 *lengths, char **bufs,
2575 struct ptlrpc_request_pool *pool);
2576 void ptlrpc_req_finished(struct ptlrpc_request *request);
2577 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2578 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2579 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2580 unsigned npages, unsigned max_brw,
2581 unsigned type, unsigned portal);
2582 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2583 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2585 __ptlrpc_free_bulk(bulk, 1);
2587 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2589 __ptlrpc_free_bulk(bulk, 0);
2591 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2592 cfs_page_t *page, int pageoffset, int len, int);
2593 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2594 cfs_page_t *page, int pageoffset,
2597 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2600 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2601 cfs_page_t *page, int pageoffset,
2604 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2607 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2608 struct obd_import *imp);
2609 __u64 ptlrpc_next_xid(void);
2610 __u64 ptlrpc_sample_next_xid(void);
2611 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2613 /* Set of routines to run a function in ptlrpcd context */
2614 void *ptlrpcd_alloc_work(struct obd_import *imp,
2615 int (*cb)(const struct lu_env *, void *), void *data);
2616 void ptlrpcd_destroy_work(void *handler);
2617 int ptlrpcd_queue_work(void *handler);
2620 struct ptlrpc_service_buf_conf {
2621 /* nbufs is buffers # to allocate when growing the pool */
2622 unsigned int bc_nbufs;
2623 /* buffer size to post */
2624 unsigned int bc_buf_size;
2625 /* portal to listed for requests on */
2626 unsigned int bc_req_portal;
2627 /* portal of where to send replies to */
2628 unsigned int bc_rep_portal;
2629 /* maximum request size to be accepted for this service */
2630 unsigned int bc_req_max_size;
2631 /* maximum reply size this service can ever send */
2632 unsigned int bc_rep_max_size;
2635 struct ptlrpc_service_thr_conf {
2636 /* threadname should be 8 characters or less - 6 will be added on */
2638 /* threads increasing factor for each CPU */
2639 unsigned int tc_thr_factor;
2640 /* service threads # to start on each partition while initializing */
2641 unsigned int tc_nthrs_init;
2643 * low water of threads # upper-limit on each partition while running,
2644 * service availability may be impacted if threads number is lower
2645 * than this value. It can be ZERO if the service doesn't require
2646 * CPU affinity or there is only one partition.
2648 unsigned int tc_nthrs_base;
2649 /* "soft" limit for total threads number */
2650 unsigned int tc_nthrs_max;
2651 /* user specified threads number, it will be validated due to
2652 * other members of this structure. */
2653 unsigned int tc_nthrs_user;
2654 /* set NUMA node affinity for service threads */
2655 unsigned int tc_cpu_affinity;
2656 /* Tags for lu_context associated with service thread */
2660 struct ptlrpc_service_cpt_conf {
2661 struct cfs_cpt_table *cc_cptable;
2662 /* string pattern to describe CPTs for a service */
2666 struct ptlrpc_service_conf {
2669 /* soft watchdog timeout multiplifier to print stuck service traces */
2670 unsigned int psc_watchdog_factor;
2671 /* buffer information */
2672 struct ptlrpc_service_buf_conf psc_buf;
2673 /* thread information */
2674 struct ptlrpc_service_thr_conf psc_thr;
2675 /* CPU partition information */
2676 struct ptlrpc_service_cpt_conf psc_cpt;
2677 /* function table */
2678 struct ptlrpc_service_ops psc_ops;
2681 /* ptlrpc/service.c */
2683 * Server-side services API. Register/unregister service, request state
2684 * management, service thread management
2688 void ptlrpc_save_lock(struct ptlrpc_request *req,
2689 struct lustre_handle *lock, int mode, int no_ack);
2690 void ptlrpc_commit_replies(struct obd_export *exp);
2691 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2692 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2693 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2694 struct ptlrpc_service *ptlrpc_register_service(
2695 struct ptlrpc_service_conf *conf,
2696 struct proc_dir_entry *proc_entry);
2697 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2699 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2700 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2701 int liblustre_check_services(void *arg);
2702 void ptlrpc_daemonize(char *name);
2703 int ptlrpc_service_health_check(struct ptlrpc_service *);
2704 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2707 int ptlrpc_hr_init(void);
2708 void ptlrpc_hr_fini(void);
2710 # define ptlrpc_hr_init() (0)
2711 # define ptlrpc_hr_fini() do {} while(0)
2716 /* ptlrpc/import.c */
2721 int ptlrpc_connect_import(struct obd_import *imp);
2722 int ptlrpc_init_import(struct obd_import *imp);
2723 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2724 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2725 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2728 /* ptlrpc/pack_generic.c */
2729 int ptlrpc_reconnect_import(struct obd_import *imp);
2733 * ptlrpc msg buffer and swab interface
2737 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2739 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2741 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2742 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2744 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2745 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2747 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2748 __u32 *lens, char **bufs);
2749 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2751 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2752 __u32 *lens, char **bufs, int flags);
2753 #define LPRFL_EARLY_REPLY 1
2754 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2755 char **bufs, int flags);
2756 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2757 unsigned int newlen, int move_data);
2758 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2759 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2760 int lustre_msg_hdr_size(__u32 magic, int count);
2761 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2762 int lustre_msg_size_v2(int count, __u32 *lengths);
2763 int lustre_packed_msg_size(struct lustre_msg *msg);
2764 int lustre_msg_early_size(void);
2765 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2766 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2767 int lustre_msg_buflen(struct lustre_msg *m, int n);
2768 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2769 int lustre_msg_bufcount(struct lustre_msg *m);
2770 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2771 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2772 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2773 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2774 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2775 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2776 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2777 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2778 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2779 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2780 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2781 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2782 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2783 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2784 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2785 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2786 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2787 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2788 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2789 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2790 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2791 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2792 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2793 int lustre_msg_get_status(struct lustre_msg *msg);
2794 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2795 int lustre_msg_is_v1(struct lustre_msg *msg);
2796 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2797 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2798 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2799 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2800 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2801 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2802 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2804 # warning "remove checksum compatibility support for b1_8"
2805 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2807 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2808 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2809 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2810 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2811 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2812 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2813 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2814 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2815 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2816 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2817 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2818 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2819 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2820 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2821 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2824 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2825 unsigned int newlen, int move_data)
2827 LASSERT(req->rq_reply_state);
2828 LASSERT(req->rq_repmsg);
2829 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2834 /** Change request phase of \a req to \a new_phase */
2836 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2838 if (req->rq_phase == new_phase)
2841 if (new_phase == RQ_PHASE_UNREGISTERING) {
2842 req->rq_next_phase = req->rq_phase;
2844 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2847 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2849 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2852 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2853 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2855 req->rq_phase = new_phase;
2859 * Returns true if request \a req got early reply and hard deadline is not met
2862 ptlrpc_client_early(struct ptlrpc_request *req)
2864 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2865 req->rq_reply_deadline > cfs_time_current_sec())
2867 return req->rq_early;
2871 * Returns true if we got real reply from server for this request
2874 ptlrpc_client_replied(struct ptlrpc_request *req)
2876 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2877 req->rq_reply_deadline > cfs_time_current_sec())
2879 return req->rq_replied;
2882 /** Returns true if request \a req is in process of receiving server reply */
2884 ptlrpc_client_recv(struct ptlrpc_request *req)
2886 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2887 req->rq_reply_deadline > cfs_time_current_sec())
2889 return req->rq_receiving_reply;
2893 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2897 spin_lock(&req->rq_lock);
2898 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2899 req->rq_reply_deadline > cfs_time_current_sec()) {
2900 spin_unlock(&req->rq_lock);
2903 rc = req->rq_receiving_reply || req->rq_must_unlink;
2904 spin_unlock(&req->rq_lock);
2909 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2911 if (req->rq_set == NULL)
2912 cfs_waitq_signal(&req->rq_reply_waitq);
2914 cfs_waitq_signal(&req->rq_set->set_waitq);
2918 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2920 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2921 cfs_atomic_inc(&rs->rs_refcount);
2925 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2927 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2928 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
2929 lustre_free_reply_state(rs);
2932 /* Should only be called once per req */
2933 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2935 if (req->rq_reply_state == NULL)
2936 return; /* shouldn't occur */
2937 ptlrpc_rs_decref(req->rq_reply_state);
2938 req->rq_reply_state = NULL;
2939 req->rq_repmsg = NULL;
2942 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2944 return lustre_msg_get_magic(req->rq_reqmsg);
2947 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2949 switch (req->rq_reqmsg->lm_magic) {
2950 case LUSTRE_MSG_MAGIC_V2:
2951 return req->rq_reqmsg->lm_repsize;
2953 LASSERTF(0, "incorrect message magic: %08x\n",
2954 req->rq_reqmsg->lm_magic);
2959 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2961 if (req->rq_delay_limit != 0 &&
2962 cfs_time_before(cfs_time_add(req->rq_queued_time,
2963 cfs_time_seconds(req->rq_delay_limit)),
2964 cfs_time_current())) {
2970 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2972 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2973 spin_lock(&req->rq_lock);
2974 req->rq_no_resend = 1;
2975 spin_unlock(&req->rq_lock);
2977 return req->rq_no_resend;
2981 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2983 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2985 return svcpt->scp_service->srv_watchdog_factor *
2986 max_t(int, at, obd_timeout);
2989 static inline struct ptlrpc_service *
2990 ptlrpc_req2svc(struct ptlrpc_request *req)
2992 LASSERT(req->rq_rqbd != NULL);
2993 return req->rq_rqbd->rqbd_svcpt->scp_service;
2996 /* ldlm/ldlm_lib.c */
2998 * Target client logic
3001 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3002 int client_obd_cleanup(struct obd_device *obddev);
3003 int client_connect_import(const struct lu_env *env,
3004 struct obd_export **exp, struct obd_device *obd,
3005 struct obd_uuid *cluuid, struct obd_connect_data *,
3007 int client_disconnect_export(struct obd_export *exp);
3008 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3010 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3011 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3012 struct obd_uuid *uuid);
3013 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3014 void client_destroy_import(struct obd_import *imp);
3017 #ifdef HAVE_SERVER_SUPPORT
3018 int server_disconnect_export(struct obd_export *exp);
3021 /* ptlrpc/pinger.c */
3023 * Pinger API (client side only)
3026 extern int suppress_pings;
3027 enum timeout_event {
3030 struct timeout_item;
3031 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3032 int ptlrpc_pinger_add_import(struct obd_import *imp);
3033 int ptlrpc_pinger_del_import(struct obd_import *imp);
3034 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3035 timeout_cb_t cb, void *data,
3036 cfs_list_t *obd_list);
3037 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
3038 enum timeout_event event);
3039 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3040 int ptlrpc_obd_ping(struct obd_device *obd);
3041 cfs_time_t ptlrpc_suspend_wakeup_time(void);
3043 void ping_evictor_start(void);
3044 void ping_evictor_stop(void);
3046 #define ping_evictor_start() do {} while (0)
3047 #define ping_evictor_stop() do {} while (0)
3049 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
3052 /* ptlrpc daemon bind policy */
3054 /* all ptlrpcd threads are free mode */
3055 PDB_POLICY_NONE = 1,
3056 /* all ptlrpcd threads are bound mode */
3057 PDB_POLICY_FULL = 2,
3058 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3059 PDB_POLICY_PAIR = 3,
3060 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3061 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3062 * If kernel supports NUMA, pthrpcd threads are binded and
3063 * grouped by NUMA node */
3064 PDB_POLICY_NEIGHBOR = 4,
3067 /* ptlrpc daemon load policy
3068 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3069 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3070 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3071 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3072 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3074 /* on the same CPU core as the caller */
3075 PDL_POLICY_SAME = 1,
3076 /* within the same CPU partition, but not the same core as the caller */
3077 PDL_POLICY_LOCAL = 2,
3078 /* round-robin on all CPU cores, but not the same core as the caller */
3079 PDL_POLICY_ROUND = 3,
3080 /* the specified CPU core is preferred, but not enforced */
3081 PDL_POLICY_PREFERRED = 4,
3084 /* ptlrpc/ptlrpcd.c */
3085 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3086 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3087 void ptlrpcd_wake(struct ptlrpc_request *req);
3088 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3089 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3090 int ptlrpcd_addref(void);
3091 void ptlrpcd_decref(void);
3093 /* ptlrpc/lproc_ptlrpc.c */
3095 * procfs output related functions
3098 const char* ll_opcode2str(__u32 opcode);
3100 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3101 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3102 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3104 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3105 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3106 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3110 /* ptlrpc/llog_server.c */
3111 int llog_origin_handle_open(struct ptlrpc_request *req);
3112 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3113 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3114 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3115 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3116 int llog_origin_handle_close(struct ptlrpc_request *req);
3117 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3119 /* ptlrpc/llog_client.c */
3120 extern struct llog_operations llog_client_ops;