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 will 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_CLIENT_NBUFS 1
272 #define LDLM_SERVER_NBUFS 64
273 #define LDLM_BUFSIZE (8 * 1024)
274 #define LDLM_MAXREQSIZE (5 * 1024)
275 #define LDLM_MAXREPSIZE (1024)
278 * MDS threads constants:
280 * Please see examples in "Thread Constants", MDS threads number will be at
281 * the comparable level of old versions, unless the server has many cores.
283 #ifndef MDS_MAX_THREADS
284 #define MDS_MAX_THREADS 1024
285 #define MDS_MAX_OTHR_THREADS 256
287 #else /* MDS_MAX_THREADS */
288 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
289 #undef MDS_MAX_THREADS
290 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
292 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
295 /* default service */
296 #define MDS_THR_FACTOR 8
297 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
298 #define MDS_NTHRS_MAX MDS_MAX_THREADS
299 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
301 /* read-page service */
302 #define MDS_RDPG_THR_FACTOR 4
303 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
304 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
305 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
307 /* these should be removed when we remove setattr service in the future */
308 #define MDS_SETA_THR_FACTOR 4
309 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
310 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
311 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
313 /* non-affinity threads */
314 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
315 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
319 * Assume file name length = FNAME_MAX = 256 (true for ext3).
320 * path name length = PATH_MAX = 4096
321 * LOV MD size max = EA_MAX = 24 * 2000
322 * (NB: 24 is size of lov_ost_data)
323 * LOV LOGCOOKIE size max = 32 * 2000
324 * (NB: 32 is size of llog_cookie)
325 * symlink: FNAME_MAX + PATH_MAX <- largest
326 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
327 * rename: FNAME_MAX + FNAME_MAX
328 * open: FNAME_MAX + EA_MAX
330 * MDS_MAXREQSIZE ~= 4736 bytes =
331 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
332 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
334 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
335 * except in the open case where there are a large number of OSTs in a LOV.
337 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
338 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
341 * MDS incoming request with LOV EA
342 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
344 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
345 362 + LOV_MAX_STRIPE_COUNT * 24)
347 * MDS outgoing reply with LOV EA
349 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
350 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
352 * but 2.4 or later MDS will never send reply with llog_cookie to any
353 * version client. This macro is defined for server side reply buffer size.
355 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
358 * The update request includes all of updates from the create, which might
359 * include linkea (4K maxim), together with other updates, we set it to 9K:
360 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
362 #define MDS_OUT_MAXREQSIZE (9 * 1024)
363 #define MDS_OUT_MAXREPSIZE MDS_MAXREPSIZE
365 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
366 #define MDS_BUFSIZE max_t(int, MDS_MAXREQSIZE + 1024, 8 * 1024)
369 * MDS_LOV_BUFSIZE should be at least max_reqsize (with LOV EA) +
370 * max sptlrpc payload size, however, we need to allocate a much larger buffer
371 * for it because LNet requires each MD(rqbd) has at least MDS_LOVE_MAXREQSIZE
372 * bytes left to avoid dropping of maximum-sized incoming request.
373 * So if MDS_LOV_BUFSIZE is only a little larger than MDS_LOV_MAXREQSIZE,
374 * then it can only fit in one request even there are 48K bytes left in
375 * a rqbd, and memory utilization is very low.
377 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
378 * reused until all requests fit in it have been processed and released,
379 * which means one long blocked request can prevent the rqbd be reused.
380 * Now we set request buffer size to 128K, so even each rqbd is unlinked
381 * from LNet with unused 48K, buffer utilization will be about 62%.
382 * Please check LU-2432 for details.
384 /** MDS_LOV_BUFSIZE = max_reqsize (w/ LOV EA) + max sptlrpc payload size */
385 #define MDS_LOV_BUFSIZE max_t(int, MDS_LOV_MAXREQSIZE + 1024, \
389 * MDS_OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
390 * about 10K, for the same reason as MDS_LOV_BUFSIZE, we also give some
391 * extra bytes to each request buffer to improve buffer utilization rate.
393 #define MDS_OUT_BUFSIZE max_t(int, MDS_OUT_MAXREQSIZE + 1024, \
396 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
397 #define FLD_MAXREQSIZE (160)
399 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
400 #define FLD_MAXREPSIZE (152)
401 #define FLD_BUFSIZE (1 << 12)
404 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
406 #define SEQ_MAXREQSIZE (160)
408 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
409 #define SEQ_MAXREPSIZE (152)
410 #define SEQ_BUFSIZE (1 << 12)
412 /** MGS threads must be >= 3, see bug 22458 comment #28 */
413 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
414 #define MGS_NTHRS_MAX 32
417 #define MGS_BUFSIZE (8 * 1024)
418 #define MGS_MAXREQSIZE (7 * 1024)
419 #define MGS_MAXREPSIZE (9 * 1024)
422 * OSS threads constants:
424 * Given 8 as factor and 64 as base threads number
427 * On 8-core server configured to 2 partitions, we will have
428 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
431 * On 32-core machine configured to 4 partitions, we will have
432 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
433 * will be 112 * 4 = 448.
436 * On 64-core machine configured to 4 partitions, we will have
437 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
438 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
439 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
440 * for each partition.
442 * So we can see that with these constants, threads number wil be at the
443 * similar level of old versions, unless the server has many cores.
445 /* depress threads factor for VM with small memory size */
446 #define OSS_THR_FACTOR min_t(int, 8, \
447 CFS_NUM_CACHEPAGES >> (28 - CFS_PAGE_SHIFT))
448 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
449 #define OSS_NTHRS_BASE 64
450 #define OSS_NTHRS_MAX 512
452 /* threads for handling "create" request */
453 #define OSS_CR_THR_FACTOR 1
454 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
455 #define OSS_CR_NTHRS_BASE 8
456 #define OSS_CR_NTHRS_MAX 64
460 * lustre_msg + obdo + obd_ioobj + DT_MAX_BRW_PAGES * niobuf_remote
462 * - single object with 16 pages is 512 bytes
463 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
464 * - Must be a multiple of 1024
466 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + sizeof(struct obdo) + \
467 sizeof(struct obd_ioobj) + DT_MAX_BRW_PAGES * \
468 sizeof(struct niobuf_remote))
469 #define OST_MAXREQSIZE (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1)
471 #define OST_MAXREPSIZE (9 * 1024)
474 #define OST_BUFSIZE (OST_MAXREQSIZE + 1024)
476 /* Macro to hide a typecast. */
477 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
480 * Structure to single define portal connection.
482 struct ptlrpc_connection {
483 /** linkage for connections hash table */
484 cfs_hlist_node_t c_hash;
485 /** Our own lnet nid for this connection */
487 /** Remote side nid for this connection */
488 lnet_process_id_t c_peer;
489 /** UUID of the other side */
490 struct obd_uuid c_remote_uuid;
491 /** reference counter for this connection */
492 cfs_atomic_t c_refcount;
495 /** Client definition for PortalRPC */
496 struct ptlrpc_client {
497 /** What lnet portal does this client send messages to by default */
498 __u32 cli_request_portal;
499 /** What portal do we expect replies on */
500 __u32 cli_reply_portal;
501 /** Name of the client */
505 /** state flags of requests */
506 /* XXX only ones left are those used by the bulk descs as well! */
507 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
508 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
510 #define REQ_MAX_ACK_LOCKS 8
512 union ptlrpc_async_args {
514 * Scratchpad for passing args to completion interpreter. Users
515 * cast to the struct of their choosing, and CLASSERT that this is
516 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
517 * a pointer to it here. The pointer_arg ensures this struct is at
518 * least big enough for that.
520 void *pointer_arg[11];
524 struct ptlrpc_request_set;
525 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
526 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
529 * Definition of request set structure.
530 * Request set is a list of requests (not necessary to the same target) that
531 * once populated with RPCs could be sent in parallel.
532 * There are two kinds of request sets. General purpose and with dedicated
533 * serving thread. Example of the latter is ptlrpcd set.
534 * For general purpose sets once request set started sending it is impossible
535 * to add new requests to such set.
536 * Provides a way to call "completion callbacks" when all requests in the set
539 struct ptlrpc_request_set {
540 cfs_atomic_t set_refcount;
541 /** number of in queue requests */
542 cfs_atomic_t set_new_count;
543 /** number of uncompleted requests */
544 cfs_atomic_t set_remaining;
545 /** wait queue to wait on for request events */
546 cfs_waitq_t set_waitq;
547 cfs_waitq_t *set_wakeup_ptr;
548 /** List of requests in the set */
549 cfs_list_t set_requests;
551 * List of completion callbacks to be called when the set is completed
552 * This is only used if \a set_interpret is NULL.
553 * Links struct ptlrpc_set_cbdata.
555 cfs_list_t set_cblist;
556 /** Completion callback, if only one. */
557 set_interpreter_func set_interpret;
558 /** opaq argument passed to completion \a set_interpret callback. */
561 * Lock for \a set_new_requests manipulations
562 * locked so that any old caller can communicate requests to
563 * the set holder who can then fold them into the lock-free set
565 spinlock_t set_new_req_lock;
566 /** List of new yet unsent requests. Only used with ptlrpcd now. */
567 cfs_list_t set_new_requests;
569 /** rq_status of requests that have been freed already */
571 /** Additional fields used by the flow control extension */
572 /** Maximum number of RPCs in flight */
573 int set_max_inflight;
574 /** Callback function used to generate RPCs */
575 set_producer_func set_producer;
576 /** opaq argument passed to the producer callback */
577 void *set_producer_arg;
581 * Description of a single ptrlrpc_set callback
583 struct ptlrpc_set_cbdata {
584 /** List linkage item */
586 /** Pointer to interpreting function */
587 set_interpreter_func psc_interpret;
588 /** Opaq argument to pass to the callback */
592 struct ptlrpc_bulk_desc;
593 struct ptlrpc_service_part;
594 struct ptlrpc_service;
597 * ptlrpc callback & work item stuff
599 struct ptlrpc_cb_id {
600 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
601 void *cbid_arg; /* additional arg */
604 /** Maximum number of locks to fit into reply state */
605 #define RS_MAX_LOCKS 8
609 * Structure to define reply state on the server
610 * Reply state holds various reply message information. Also for "difficult"
611 * replies (rep-ack case) we store the state after sending reply and wait
612 * for the client to acknowledge the reception. In these cases locks could be
613 * added to the state for replay/failover consistency guarantees.
615 struct ptlrpc_reply_state {
616 /** Callback description */
617 struct ptlrpc_cb_id rs_cb_id;
618 /** Linkage for list of all reply states in a system */
620 /** Linkage for list of all reply states on same export */
621 cfs_list_t rs_exp_list;
622 /** Linkage for list of all reply states for same obd */
623 cfs_list_t rs_obd_list;
625 cfs_list_t rs_debug_list;
627 /** A spinlock to protect the reply state flags */
629 /** Reply state flags */
630 unsigned long rs_difficult:1; /* ACK/commit stuff */
631 unsigned long rs_no_ack:1; /* no ACK, even for
632 difficult requests */
633 unsigned long rs_scheduled:1; /* being handled? */
634 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
635 unsigned long rs_handled:1; /* been handled yet? */
636 unsigned long rs_on_net:1; /* reply_out_callback pending? */
637 unsigned long rs_prealloc:1; /* rs from prealloc list */
638 unsigned long rs_committed:1;/* the transaction was committed
639 and the rs was dispatched
640 by ptlrpc_commit_replies */
641 /** Size of the state */
645 /** Transaction number */
649 struct obd_export *rs_export;
650 struct ptlrpc_service_part *rs_svcpt;
651 /** Lnet metadata handle for the reply */
652 lnet_handle_md_t rs_md_h;
653 cfs_atomic_t rs_refcount;
655 /** Context for the sevice thread */
656 struct ptlrpc_svc_ctx *rs_svc_ctx;
657 /** Reply buffer (actually sent to the client), encoded if needed */
658 struct lustre_msg *rs_repbuf; /* wrapper */
659 /** Size of the reply buffer */
660 int rs_repbuf_len; /* wrapper buf length */
661 /** Size of the reply message */
662 int rs_repdata_len; /* wrapper msg length */
664 * Actual reply message. Its content is encrupted (if needed) to
665 * produce reply buffer for actual sending. In simple case
666 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
668 struct lustre_msg *rs_msg; /* reply message */
670 /** Number of locks awaiting client ACK */
672 /** Handles of locks awaiting client reply ACK */
673 struct lustre_handle rs_locks[RS_MAX_LOCKS];
674 /** Lock modes of locks in \a rs_locks */
675 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
678 struct ptlrpc_thread;
682 RQ_PHASE_NEW = 0xebc0de00,
683 RQ_PHASE_RPC = 0xebc0de01,
684 RQ_PHASE_BULK = 0xebc0de02,
685 RQ_PHASE_INTERPRET = 0xebc0de03,
686 RQ_PHASE_COMPLETE = 0xebc0de04,
687 RQ_PHASE_UNREGISTERING = 0xebc0de05,
688 RQ_PHASE_UNDEFINED = 0xebc0de06
691 /** Type of request interpreter call-back */
692 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
693 struct ptlrpc_request *req,
697 * Definition of request pool structure.
698 * The pool is used to store empty preallocated requests for the case
699 * when we would actually need to send something without performing
700 * any allocations (to avoid e.g. OOM).
702 struct ptlrpc_request_pool {
703 /** Locks the list */
705 /** list of ptlrpc_request structs */
706 cfs_list_t prp_req_list;
707 /** Maximum message size that would fit into a rquest from this pool */
709 /** Function to allocate more requests for this pool */
710 void (*prp_populate)(struct ptlrpc_request_pool *, int);
719 * \defgroup nrs Network Request Scheduler
722 struct ptlrpc_nrs_policy;
723 struct ptlrpc_nrs_resource;
724 struct ptlrpc_nrs_request;
727 * NRS control operations.
729 * These are common for all policies.
731 enum ptlrpc_nrs_ctl {
733 * Activate the policy.
735 PTLRPC_NRS_CTL_START,
737 * Reserved for multiple primary policies, which may be a possibility
742 * Recycle resources for inactive policies.
744 PTLRPC_NRS_CTL_SHRINK,
746 * Not a valid opcode.
748 PTLRPC_NRS_CTL_INVALID,
750 * Policies can start using opcodes from this value and onwards for
751 * their own purposes; the assigned value itself is arbitrary.
753 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
757 * NRS policy operations.
759 * These determine the behaviour of a policy, and are called in response to
762 struct ptlrpc_nrs_pol_ops {
764 * Called during policy registration; this operation is optional.
766 * \param[in] policy The policy being initialized
768 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
770 * Called during policy unregistration; this operation is optional.
772 * \param[in] policy The policy being unregistered/finalized
774 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
776 * Called when activating a policy via lprocfs; policies allocate and
777 * initialize their resources here; this operation is optional.
779 * \param[in] policy The policy being started
781 * \see nrs_policy_start_locked()
783 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
785 * Called when deactivating a policy via lprocfs; policies deallocate
786 * their resources here; this operation is optional
788 * \param[in] policy The policy being stopped
790 * \see nrs_policy_stop_final()
792 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
794 * Used for policy-specific operations; i.e. not generic ones like
795 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
796 * to an ioctl; this operation is optional.
798 * \param[in] policy The policy carrying out operation \a opc
799 * \param[in] opc The command operation being carried out
800 * \param[in,out] arg An generic buffer for communication between the
801 * user and the control operation
806 * \see ptlrpc_nrs_policy_control()
808 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
809 enum ptlrpc_nrs_ctl opc, void *arg);
812 * Called when obtaining references to the resources of the resource
813 * hierarchy for a request that has arrived for handling at the PTLRPC
814 * service. Policies should return -ve for requests they do not wish
815 * to handle. This operation is mandatory.
817 * \param[in] policy The policy we're getting resources for.
818 * \param[in] nrq The request we are getting resources for.
819 * \param[in] parent The parent resource of the resource being
820 * requested; set to NULL if none.
821 * \param[out] resp The resource is to be returned here; the
822 * fallback policy in an NRS head should
823 * \e always return a non-NULL pointer value.
824 * \param[in] moving_req When set, signifies that this is an attempt
825 * to obtain resources for a request being moved
826 * to the high-priority NRS head by
827 * ldlm_lock_reorder_req().
828 * This implies two things:
829 * 1. We are under obd_export::exp_rpc_lock and
830 * so should not sleep.
831 * 2. We should not perform non-idempotent or can
832 * skip performing idempotent operations that
833 * were carried out when resources were first
834 * taken for the request when it was initialized
835 * in ptlrpc_nrs_req_initialize().
837 * \retval 0, +ve The level of the returned resource in the resource
838 * hierarchy; currently only 0 (for a non-leaf resource)
839 * and 1 (for a leaf resource) are supported by the
843 * \see ptlrpc_nrs_req_initialize()
844 * \see ptlrpc_nrs_hpreq_add_nolock()
845 * \see ptlrpc_nrs_req_hp_move()
847 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
848 struct ptlrpc_nrs_request *nrq,
849 struct ptlrpc_nrs_resource *parent,
850 struct ptlrpc_nrs_resource **resp,
853 * Called when releasing references taken for resources in the resource
854 * hierarchy for the request; this operation is optional.
856 * \param[in] policy The policy the resource belongs to
857 * \param[in] res The resource to be freed
859 * \see ptlrpc_nrs_req_finalize()
860 * \see ptlrpc_nrs_hpreq_add_nolock()
861 * \see ptlrpc_nrs_req_hp_move()
863 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
864 struct ptlrpc_nrs_resource *res);
867 * Obtain a request for handling from the policy via polling; this
868 * operation is mandatory.
870 * \param[in] policy The policy to poll
872 * \retval NULL No erquest available for handling
873 * \retval valid-pointer The request polled for handling
875 * \see ptlrpc_nrs_req_poll_nolock()
877 struct ptlrpc_nrs_request *
878 (*op_req_poll) (struct ptlrpc_nrs_policy *policy);
880 * Called when attempting to add a request to a policy for later
881 * handling; this operation is mandatory.
883 * \param[in] policy The policy on which to enqueue \a nrq
884 * \param[in] nrq The request to enqueue
889 * \see ptlrpc_nrs_req_add_nolock()
891 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
892 struct ptlrpc_nrs_request *nrq);
894 * Removes a request from the policy's set of pending requests. Normally
895 * called after a request has been polled successfully from the policy
896 * for handling; this operation is mandatory.
898 * \param[in] policy The policy the request \a nrq belongs to
899 * \param[in] nrq The request to dequeue
901 * \see ptlrpc_nrs_req_del_nolock()
903 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
904 struct ptlrpc_nrs_request *nrq);
906 * Called before carrying out the request; should not block. Could be
907 * used for job/resource control; this operation is optional.
909 * \param[in] policy The policy which is starting to handle request
911 * \param[in] nrq The request
913 * \pre spin_is_locked(&svcpt->scp_req_lock)
915 * \see ptlrpc_nrs_req_start_nolock()
917 void (*op_req_start) (struct ptlrpc_nrs_policy *policy,
918 struct ptlrpc_nrs_request *nrq);
920 * Called after the request being carried out. Could be used for
921 * job/resource control; this operation is optional.
923 * \param[in] policy The policy which is stopping to handle request
925 * \param[in] nrq The request
927 * \pre spin_is_locked(&svcpt->scp_req_lock)
929 * \see ptlrpc_nrs_req_stop_nolock()
931 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
932 struct ptlrpc_nrs_request *nrq);
934 * Registers the policy's lprocfs interface with a PTLRPC service.
936 * \param[in] svc The service
941 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
943 * Unegisters the policy's lprocfs interface with a PTLRPC service.
945 * \param[in] svc The service
947 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
953 enum nrs_policy_flags {
955 * Fallback policy, use this flag only on a single supported policy per
956 * service. Do not use this flag for policies registering using
957 * ptlrpc_nrs_policy_register() (i.e. ones that are not in
958 * \e nrs_pols_builtin).
960 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
962 * Start policy immediately after registering.
964 PTLRPC_NRS_FL_REG_START = (1 << 1),
966 * This is a polciy registering externally with NRS core, via
967 * ptlrpc_nrs_policy_register(), (i.e. one that is not in
968 * \e nrs_pols_builtin. Used to avoid ptlrpc_nrs_policy_register()
969 * racing with a policy start operation issued by the user via lprocfs.
971 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
977 * Denotes whether an NRS instance is for handling normal or high-priority
978 * RPCs, or whether an operation pertains to one or both of the NRS instances
981 enum ptlrpc_nrs_queue_type {
982 PTLRPC_NRS_QUEUE_REG,
984 PTLRPC_NRS_QUEUE_BOTH,
990 * A PTLRPC service has at least one NRS head instance for handling normal
991 * priority RPCs, and may optionally have a second NRS head instance for
992 * handling high-priority RPCs. Each NRS head maintains a list of available
993 * policies, of which one and only one policy is acting as the fallback policy,
994 * and optionally a different policy may be acting as the primary policy. For
995 * all RPCs handled by this NRS head instance, NRS core will first attempt to
996 * enqueue the RPC using the primary policy (if any). The fallback policy is
997 * used in the following cases:
998 * - when there was no primary policy in the
999 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1001 * - when the primary policy that was at the
1002 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1003 * RPC was initialized, denoted it did not wish, or for some other reason was
1004 * not able to handle the request, by returning a non-valid NRS resource
1006 * - when the primary policy that was at the
1007 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1008 * RPC was initialized, fails later during the request enqueueing stage.
1010 * \see nrs_resource_get_safe()
1011 * \see nrs_request_enqueue()
1014 spinlock_t nrs_lock;
1015 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1017 * Linkage into nrs_core_heads_list
1019 cfs_list_t nrs_heads;
1021 * List of registered policies
1023 cfs_list_t nrs_policy_list;
1025 * List of policies with queued requests. Policies that have any
1026 * outstanding requests are queued here, and this list is queried
1027 * in a round-robin manner from NRS core when obtaining a request
1028 * for handling. This ensures that requests from policies that at some
1029 * point transition away from the
1030 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1032 cfs_list_t nrs_policy_queued;
1034 * Service partition for this NRS head
1036 struct ptlrpc_service_part *nrs_svcpt;
1038 * Primary policy, which is the preferred policy for handling RPCs
1040 struct ptlrpc_nrs_policy *nrs_policy_primary;
1042 * Fallback policy, which is the backup policy for handling RPCs
1044 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1046 * This NRS head handles either HP or regular requests
1048 enum ptlrpc_nrs_queue_type nrs_queue_type;
1050 * # queued requests from all policies in this NRS head
1052 unsigned long nrs_req_queued;
1054 * # scheduled requests from all policies in this NRS head
1056 unsigned long nrs_req_started;
1058 * # policies on this NRS
1059 * TODO: Can we avoid having this?
1061 unsigned nrs_num_pols;
1063 * This NRS head is in progress of starting a policy
1065 unsigned nrs_policy_starting:1;
1067 * In progress of shutting down the whole NRS head; used during
1070 unsigned nrs_stopping:1;
1073 #define NRS_POL_NAME_MAX 16
1076 * NRS policy registering descriptor
1078 * Is used to hold a description of a policy that can be passed to NRS core in
1079 * order to register the policy with NRS heads in different PTLRPC services.
1081 struct ptlrpc_nrs_pol_desc {
1083 * Human-readable policy name
1085 char pd_name[NRS_POL_NAME_MAX];
1087 * NRS operations for this policy
1089 struct ptlrpc_nrs_pol_ops *pd_ops;
1091 * Service Compatibility function; this determines whether a policy is
1092 * adequate for handling RPCs of a particular PTLRPC service.
1094 * XXX:This should give the same result during policy
1095 * registration and unregistration, and for all partitions of a
1096 * service; so the result should not depend on temporal service
1097 * or other properties, that may influence the result.
1099 bool (*pd_compat) (struct ptlrpc_service *svc,
1100 const struct ptlrpc_nrs_pol_desc *desc);
1102 * Optionally set for policies that support a single ptlrpc service,
1103 * i.e. ones that have \a pd_compat set to nrs_policy_compat_one()
1105 char *pd_compat_svc_name;
1107 * Bitmask of nrs_policy_flags
1111 * Link into nrs_core::nrs_policies
1119 * Policies transition from one state to the other during their lifetime
1121 enum ptlrpc_nrs_pol_state {
1123 * Not a valid policy state.
1125 NRS_POL_STATE_INVALID,
1127 * For now, this state is used exclusively for policies that register
1128 * externally to NRS core, i.e. ones that do so via
1129 * ptlrpc_nrs_policy_register() and are not part of nrs_pols_builtin;
1130 * it is used to prevent a race condition between the policy registering
1131 * with more than one service partition while service is operational,
1132 * and the user starting the policy via lprocfs.
1134 * \see nrs_pol_make_avail()
1136 NRS_POL_STATE_UNAVAIL,
1138 * Policies are at this state either at the start of their life, or
1139 * transition here when the user selects a different policy to act
1140 * as the primary one.
1142 NRS_POL_STATE_STOPPED,
1144 * Policy is progress of stopping
1146 NRS_POL_STATE_STOPPING,
1148 * Policy is in progress of starting
1150 NRS_POL_STATE_STARTING,
1152 * A policy is in this state in two cases:
1153 * - it is the fallback policy, which is always in this state.
1154 * - it has been activated by the user; i.e. it is the primary policy,
1156 NRS_POL_STATE_STARTED,
1160 * NRS policy information
1162 * Used for obtaining information for the status of a policy via lprocfs
1164 struct ptlrpc_nrs_pol_info {
1168 char pi_name[NRS_POL_NAME_MAX];
1170 * Current policy state
1172 enum ptlrpc_nrs_pol_state pi_state;
1174 * # RPCs enqueued for later dispatching by the policy
1178 * # RPCs started for dispatch by the policy
1180 long pi_req_started;
1182 * Is this a fallback policy?
1184 unsigned pi_fallback:1;
1190 * There is one instance of this for each policy in each NRS head of each
1191 * PTLRPC service partition.
1193 struct ptlrpc_nrs_policy {
1195 * Linkage into the NRS head's list of policies,
1196 * ptlrpc_nrs:nrs_policy_list
1198 cfs_list_t pol_list;
1200 * Linkage into the NRS head's list of policies with enqueued
1201 * requests ptlrpc_nrs:nrs_policy_queued
1203 cfs_list_t pol_list_queued;
1205 * Current state of this policy
1207 enum ptlrpc_nrs_pol_state pol_state;
1209 * Bitmask of nrs_policy_flags
1213 * # RPCs enqueued for later dispatching by the policy
1215 long pol_req_queued;
1217 * # RPCs started for dispatch by the policy
1219 long pol_req_started;
1221 * Usage Reference count taken on the policy instance
1225 * The NRS head this policy has been created at
1227 struct ptlrpc_nrs *pol_nrs;
1229 * NRS operations for this policy; points to ptlrpc_nrs_pol_desc::pd_ops
1231 struct ptlrpc_nrs_pol_ops *pol_ops;
1233 * Private policy data; varies by policy type
1237 * Human-readable policy name; point to ptlrpc_nrs_pol_desc::pd_name
1245 * Resources are embedded into two types of NRS entities:
1246 * - Inside NRS policies, in the policy's private data in
1247 * ptlrpc_nrs_policy::pol_private
1248 * - In objects that act as prime-level scheduling entities in different NRS
1249 * policies; e.g. on a policy that performs round robin or similar order
1250 * scheduling across client NIDs, there would be one NRS resource per unique
1251 * client NID. On a policy which performs round robin scheduling across
1252 * backend filesystem objects, there would be one resource associated with
1253 * each of the backend filesystem objects partaking in the scheduling
1254 * performed by the policy.
1256 * NRS resources share a parent-child relationship, in which resources embedded
1257 * in policy instances are the parent entities, with all scheduling entities
1258 * a policy schedules across being the children, thus forming a simple resource
1259 * hierarchy. This hierarchy may be extended with one or more levels in the
1260 * future if the ability to have more than one primary policy is added.
1262 * Upon request initialization, references to the then active NRS policies are
1263 * taken and used to later handle the dispatching of the request with one of
1266 * \see nrs_resource_get_safe()
1267 * \see ptlrpc_nrs_req_add()
1269 struct ptlrpc_nrs_resource {
1271 * This NRS resource's parent; is NULL for resources embedded in NRS
1272 * policy instances; i.e. those are top-level ones.
1274 struct ptlrpc_nrs_resource *res_parent;
1276 * The policy associated with this resource.
1278 struct ptlrpc_nrs_policy *res_policy;
1291 * This policy is a logical wrapper around previous, non-NRS functionality.
1292 * It dispatches RPCs in the same order as they arrive from the network. This
1293 * policy is currently used as the fallback policy, and the only enabled policy
1294 * on all NRS heads of all PTLRPC service partitions.
1299 * Private data structure for the FIFO policy
1301 struct nrs_fifo_head {
1303 * Resource object for policy instance.
1305 struct ptlrpc_nrs_resource fh_res;
1307 * List of queued requests.
1311 * For debugging purposes.
1316 struct nrs_fifo_req {
1317 /** request header, must be the first member of structure */
1327 * Instances of this object exist embedded within ptlrpc_request; the main
1328 * purpose of this object is to hold references to the request's resources
1329 * for the lifetime of the request, and to hold properties that policies use
1330 * use for determining the request's scheduling priority.
1332 struct ptlrpc_nrs_request {
1334 * The request's resource hierarchy.
1336 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1338 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1339 * policy that was used to enqueue the request.
1341 * \see nrs_request_enqueue()
1343 unsigned nr_res_idx;
1344 unsigned nr_initialized:1;
1345 unsigned nr_enqueued:1;
1346 unsigned nr_dequeued:1;
1347 unsigned nr_started:1;
1348 unsigned nr_finalized:1;
1349 cfs_binheap_node_t nr_node;
1352 * Policy-specific fields, used for determining a request's scheduling
1353 * priority, and other supporting functionality.
1357 * Fields for the FIFO policy
1359 struct nrs_fifo_req fifo;
1362 * Externally-registering policies may want to use this to allocate
1363 * their own request properties.
1371 * Basic request prioritization operations structure.
1372 * The whole idea is centered around locks and RPCs that might affect locks.
1373 * When a lock is contended we try to give priority to RPCs that might lead
1374 * to fastest release of that lock.
1375 * Currently only implemented for OSTs only in a way that makes all
1376 * IO and truncate RPCs that are coming from a locked region where a lock is
1377 * contended a priority over other requests.
1379 struct ptlrpc_hpreq_ops {
1381 * Check if the lock handle of the given lock is the same as
1382 * taken from the request.
1384 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1386 * Check if the request is a high priority one.
1388 int (*hpreq_check)(struct ptlrpc_request *);
1390 * Called after the request has been handled.
1392 void (*hpreq_fini)(struct ptlrpc_request *);
1396 * Represents remote procedure call.
1398 * This is a staple structure used by everybody wanting to send a request
1401 struct ptlrpc_request {
1402 /* Request type: one of PTL_RPC_MSG_* */
1404 /** Result of request processing */
1407 * Linkage item through which this request is included into
1408 * sending/delayed lists on client and into rqbd list on server
1412 * Server side list of incoming unserved requests sorted by arrival
1413 * time. Traversed from time to time to notice about to expire
1414 * requests and sent back "early replies" to clients to let them
1415 * know server is alive and well, just very busy to service their
1418 cfs_list_t rq_timed_list;
1419 /** server-side history, used for debuging purposes. */
1420 cfs_list_t rq_history_list;
1421 /** server-side per-export list */
1422 cfs_list_t rq_exp_list;
1423 /** server-side hp handlers */
1424 struct ptlrpc_hpreq_ops *rq_ops;
1426 /** initial thread servicing this request */
1427 struct ptlrpc_thread *rq_svc_thread;
1429 /** history sequence # */
1430 __u64 rq_history_seq;
1434 /** stub for NRS request */
1435 struct ptlrpc_nrs_request rq_nrq;
1437 /** the index of service's srv_at_array into which request is linked */
1439 /** Lock to protect request flags and some other important bits, like
1443 /** client-side flags are serialized by rq_lock */
1444 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1445 rq_timedout:1, rq_resend:1, rq_restart:1,
1447 * when ->rq_replay is set, request is kept by the client even
1448 * after server commits corresponding transaction. This is
1449 * used for operations that require sequence of multiple
1450 * requests to be replayed. The only example currently is file
1451 * open/close. When last request in such a sequence is
1452 * committed, ->rq_replay is cleared on all requests in the
1456 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1457 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1458 rq_early:1, rq_must_unlink:1,
1459 rq_memalloc:1, /* req originated from "kswapd" */
1460 /* server-side flags */
1461 rq_packed_final:1, /* packed final reply */
1462 rq_hp:1, /* high priority RPC */
1463 rq_at_linked:1, /* link into service's srv_at_array */
1464 rq_reply_truncate:1,
1466 /* whether the "rq_set" is a valid one */
1468 rq_generation_set:1,
1469 /* do not resend request on -EINPROGRESS */
1470 rq_no_retry_einprogress:1,
1471 /* allow the req to be sent if the import is in recovery
1475 unsigned int rq_nr_resend;
1477 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1478 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1479 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
1480 server-side refcounf for multiple replies */
1482 /** Portal to which this request would be sent */
1483 short rq_request_portal; /* XXX FIXME bug 249 */
1484 /** Portal where to wait for reply and where reply would be sent */
1485 short rq_reply_portal; /* XXX FIXME bug 249 */
1489 * !rq_truncate : # reply bytes actually received,
1490 * rq_truncate : required repbuf_len for resend
1492 int rq_nob_received;
1493 /** Request length */
1497 /** Request message - what client sent */
1498 struct lustre_msg *rq_reqmsg;
1499 /** Reply message - server response */
1500 struct lustre_msg *rq_repmsg;
1501 /** Transaction number */
1506 * List item to for replay list. Not yet commited requests get linked
1508 * Also see \a rq_replay comment above.
1510 cfs_list_t rq_replay_list;
1513 * security and encryption data
1515 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1516 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1517 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
1519 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1520 enum lustre_sec_part rq_sp_from;
1522 /* client/server security flags */
1524 rq_ctx_init:1, /* context initiation */
1525 rq_ctx_fini:1, /* context destroy */
1526 rq_bulk_read:1, /* request bulk read */
1527 rq_bulk_write:1, /* request bulk write */
1528 /* server authentication flags */
1529 rq_auth_gss:1, /* authenticated by gss */
1530 rq_auth_remote:1, /* authed as remote user */
1531 rq_auth_usr_root:1, /* authed as root */
1532 rq_auth_usr_mdt:1, /* authed as mdt */
1533 rq_auth_usr_ost:1, /* authed as ost */
1534 /* security tfm flags */
1537 /* doesn't expect reply FIXME */
1539 rq_pill_init:1; /* pill initialized */
1541 uid_t rq_auth_uid; /* authed uid */
1542 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1544 /* (server side), pointed directly into req buffer */
1545 struct ptlrpc_user_desc *rq_user_desc;
1547 /* various buffer pointers */
1548 struct lustre_msg *rq_reqbuf; /* req wrapper */
1549 char *rq_repbuf; /* rep buffer */
1550 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1551 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1552 int rq_reqbuf_len; /* req wrapper buf len */
1553 int rq_reqdata_len; /* req wrapper msg len */
1554 int rq_repbuf_len; /* rep buffer len */
1555 int rq_repdata_len; /* rep wrapper msg len */
1556 int rq_clrbuf_len; /* only in priv mode */
1557 int rq_clrdata_len; /* only in priv mode */
1559 /** early replies go to offset 0, regular replies go after that */
1560 unsigned int rq_reply_off;
1564 /** Fields that help to see if request and reply were swabbed or not */
1565 __u32 rq_req_swab_mask;
1566 __u32 rq_rep_swab_mask;
1568 /** What was import generation when this request was sent */
1569 int rq_import_generation;
1570 enum lustre_imp_state rq_send_state;
1572 /** how many early replies (for stats) */
1575 /** client+server request */
1576 lnet_handle_md_t rq_req_md_h;
1577 struct ptlrpc_cb_id rq_req_cbid;
1578 /** optional time limit for send attempts */
1579 cfs_duration_t rq_delay_limit;
1580 /** time request was first queued */
1581 cfs_time_t rq_queued_time;
1583 /* server-side... */
1584 /** request arrival time */
1585 struct timeval rq_arrival_time;
1586 /** separated reply state */
1587 struct ptlrpc_reply_state *rq_reply_state;
1588 /** incoming request buffer */
1589 struct ptlrpc_request_buffer_desc *rq_rqbd;
1591 /** client-only incoming reply */
1592 lnet_handle_md_t rq_reply_md_h;
1593 cfs_waitq_t rq_reply_waitq;
1594 struct ptlrpc_cb_id rq_reply_cbid;
1598 /** Peer description (the other side) */
1599 lnet_process_id_t rq_peer;
1600 /** Server-side, export on which request was received */
1601 struct obd_export *rq_export;
1602 /** Client side, import where request is being sent */
1603 struct obd_import *rq_import;
1605 /** Replay callback, called after request is replayed at recovery */
1606 void (*rq_replay_cb)(struct ptlrpc_request *);
1608 * Commit callback, called when request is committed and about to be
1611 void (*rq_commit_cb)(struct ptlrpc_request *);
1612 /** Opaq data for replay and commit callbacks. */
1615 /** For bulk requests on client only: bulk descriptor */
1616 struct ptlrpc_bulk_desc *rq_bulk;
1618 /** client outgoing req */
1620 * when request/reply sent (secs), or time when request should be sent
1623 /** time for request really sent out */
1624 time_t rq_real_sent;
1626 /** when request must finish. volatile
1627 * so that servers' early reply updates to the deadline aren't
1628 * kept in per-cpu cache */
1629 volatile time_t rq_deadline;
1630 /** when req reply unlink must finish. */
1631 time_t rq_reply_deadline;
1632 /** when req bulk unlink must finish. */
1633 time_t rq_bulk_deadline;
1635 * service time estimate (secs)
1636 * If the requestsis not served by this time, it is marked as timed out.
1640 /** Multi-rpc bits */
1641 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1642 cfs_waitq_t rq_set_waitq;
1643 /** Link item for request set lists */
1644 cfs_list_t rq_set_chain;
1645 /** Link back to the request set */
1646 struct ptlrpc_request_set *rq_set;
1647 /** Async completion handler, called when reply is received */
1648 ptlrpc_interpterer_t rq_interpret_reply;
1649 /** Async completion context */
1650 union ptlrpc_async_args rq_async_args;
1652 /** Pool if request is from preallocated list */
1653 struct ptlrpc_request_pool *rq_pool;
1655 struct lu_context rq_session;
1656 struct lu_context rq_recov_session;
1658 /** request format description */
1659 struct req_capsule rq_pill;
1663 * Call completion handler for rpc if any, return it's status or original
1664 * rc if there was no handler defined for this request.
1666 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1667 struct ptlrpc_request *req, int rc)
1669 if (req->rq_interpret_reply != NULL) {
1670 req->rq_status = req->rq_interpret_reply(env, req,
1671 &req->rq_async_args,
1673 return req->rq_status;
1681 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_desc *desc);
1682 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_desc *desc);
1683 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1684 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1685 struct ptlrpc_nrs_pol_info *info);
1688 * Can the request be moved from the regular NRS head to the high-priority NRS
1689 * head (of the same PTLRPC service partition), if any?
1691 * For a reliable result, this should be checked under svcpt->scp_req lock.
1694 ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1696 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1699 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1700 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1701 * to make sure it has not been scheduled yet (analogous to previous
1702 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1704 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1709 * Returns 1 if request buffer at offset \a index was already swabbed
1711 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1713 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1714 return req->rq_req_swab_mask & (1 << index);
1718 * Returns 1 if request reply buffer at offset \a index was already swabbed
1720 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1722 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1723 return req->rq_rep_swab_mask & (1 << index);
1727 * Returns 1 if request needs to be swabbed into local cpu byteorder
1729 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1731 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1735 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1737 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1739 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1743 * Mark request buffer at offset \a index that it was already swabbed
1745 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1747 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1748 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1749 req->rq_req_swab_mask |= 1 << index;
1753 * Mark request reply buffer at offset \a index that it was already swabbed
1755 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1757 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1758 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1759 req->rq_rep_swab_mask |= 1 << index;
1763 * Convert numerical request phase value \a phase into text string description
1765 static inline const char *
1766 ptlrpc_phase2str(enum rq_phase phase)
1775 case RQ_PHASE_INTERPRET:
1777 case RQ_PHASE_COMPLETE:
1779 case RQ_PHASE_UNREGISTERING:
1780 return "Unregistering";
1787 * Convert numerical request phase of the request \a req into text stringi
1790 static inline const char *
1791 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1793 return ptlrpc_phase2str(req->rq_phase);
1797 * Debugging functions and helpers to print request structure into debug log
1800 /* Spare the preprocessor, spoil the bugs. */
1801 #define FLAG(field, str) (field ? str : "")
1803 /** Convert bit flags into a string */
1804 #define DEBUG_REQ_FLAGS(req) \
1805 ptlrpc_rqphase2str(req), \
1806 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1807 FLAG(req->rq_err, "E"), \
1808 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1809 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1810 FLAG(req->rq_no_resend, "N"), \
1811 FLAG(req->rq_waiting, "W"), \
1812 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1813 FLAG(req->rq_committed, "M")
1815 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1817 void _debug_req(struct ptlrpc_request *req,
1818 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1819 __attribute__ ((format (printf, 3, 4)));
1822 * Helper that decides if we need to print request accordig to current debug
1825 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1827 CFS_CHECK_STACK(msgdata, mask, cdls); \
1829 if (((mask) & D_CANTMASK) != 0 || \
1830 ((libcfs_debug & (mask)) != 0 && \
1831 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1832 _debug_req((req), msgdata, fmt, ##a); \
1836 * This is the debug print function you need to use to print request sturucture
1837 * content into lustre debug log.
1838 * for most callers (level is a constant) this is resolved at compile time */
1839 #define DEBUG_REQ(level, req, fmt, args...) \
1841 if ((level) & (D_ERROR | D_WARNING)) { \
1842 static cfs_debug_limit_state_t cdls; \
1843 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1844 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1846 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1847 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1853 * Structure that defines a single page of a bulk transfer
1855 struct ptlrpc_bulk_page {
1856 /** Linkage to list of pages in a bulk */
1859 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1862 /** offset within a page */
1864 /** The page itself */
1865 struct page *bp_page;
1868 #define BULK_GET_SOURCE 0
1869 #define BULK_PUT_SINK 1
1870 #define BULK_GET_SINK 2
1871 #define BULK_PUT_SOURCE 3
1874 * Definition of bulk descriptor.
1875 * Bulks are special "Two phase" RPCs where initial request message
1876 * is sent first and it is followed bt a transfer (o receiving) of a large
1877 * amount of data to be settled into pages referenced from the bulk descriptors.
1878 * Bulks transfers (the actual data following the small requests) are done
1879 * on separate LNet portals.
1880 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1881 * Another user is readpage for MDT.
1883 struct ptlrpc_bulk_desc {
1884 /** completed with failure */
1885 unsigned long bd_failure:1;
1886 /** {put,get}{source,sink} */
1887 unsigned long bd_type:2;
1889 unsigned long bd_registered:1;
1890 /** For serialization with callback */
1892 /** Import generation when request for this bulk was sent */
1893 int bd_import_generation;
1894 /** LNet portal for this bulk */
1896 /** Server side - export this bulk created for */
1897 struct obd_export *bd_export;
1898 /** Client side - import this bulk was sent on */
1899 struct obd_import *bd_import;
1900 /** Back pointer to the request */
1901 struct ptlrpc_request *bd_req;
1902 cfs_waitq_t bd_waitq; /* server side only WQ */
1903 int bd_iov_count; /* # entries in bd_iov */
1904 int bd_max_iov; /* allocated size of bd_iov */
1905 int bd_nob; /* # bytes covered */
1906 int bd_nob_transferred; /* # bytes GOT/PUT */
1910 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1911 lnet_nid_t bd_sender; /* stash event::sender */
1912 int bd_md_count; /* # valid entries in bd_mds */
1913 int bd_md_max_brw; /* max entries in bd_mds */
1914 /** array of associated MDs */
1915 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
1917 #if defined(__KERNEL__)
1919 * encrypt iov, size is either 0 or bd_iov_count.
1921 lnet_kiov_t *bd_enc_iov;
1923 lnet_kiov_t bd_iov[0];
1925 lnet_md_iovec_t bd_iov[0];
1930 SVC_STOPPED = 1 << 0,
1931 SVC_STOPPING = 1 << 1,
1932 SVC_STARTING = 1 << 2,
1933 SVC_RUNNING = 1 << 3,
1935 SVC_SIGNAL = 1 << 5,
1938 #define PTLRPC_THR_NAME_LEN 32
1940 * Definition of server service thread structure
1942 struct ptlrpc_thread {
1944 * List of active threads in svc->srv_threads
1948 * thread-private data (preallocated memory)
1953 * service thread index, from ptlrpc_start_threads
1957 * service thread pid
1961 * put watchdog in the structure per thread b=14840
1963 struct lc_watchdog *t_watchdog;
1965 * the svc this thread belonged to b=18582
1967 struct ptlrpc_service_part *t_svcpt;
1968 cfs_waitq_t t_ctl_waitq;
1969 struct lu_env *t_env;
1970 char t_name[PTLRPC_THR_NAME_LEN];
1973 static inline int thread_is_init(struct ptlrpc_thread *thread)
1975 return thread->t_flags == 0;
1978 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1980 return !!(thread->t_flags & SVC_STOPPED);
1983 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1985 return !!(thread->t_flags & SVC_STOPPING);
1988 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1990 return !!(thread->t_flags & SVC_STARTING);
1993 static inline int thread_is_running(struct ptlrpc_thread *thread)
1995 return !!(thread->t_flags & SVC_RUNNING);
1998 static inline int thread_is_event(struct ptlrpc_thread *thread)
2000 return !!(thread->t_flags & SVC_EVENT);
2003 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2005 return !!(thread->t_flags & SVC_SIGNAL);
2008 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2010 thread->t_flags &= ~flags;
2013 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2015 thread->t_flags = flags;
2018 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2020 thread->t_flags |= flags;
2023 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2026 if (thread->t_flags & flags) {
2027 thread->t_flags &= ~flags;
2034 * Request buffer descriptor structure.
2035 * This is a structure that contains one posted request buffer for service.
2036 * Once data land into a buffer, event callback creates actual request and
2037 * notifies wakes one of the service threads to process new incoming request.
2038 * More than one request can fit into the buffer.
2040 struct ptlrpc_request_buffer_desc {
2041 /** Link item for rqbds on a service */
2042 cfs_list_t rqbd_list;
2043 /** History of requests for this buffer */
2044 cfs_list_t rqbd_reqs;
2045 /** Back pointer to service for which this buffer is registered */
2046 struct ptlrpc_service_part *rqbd_svcpt;
2047 /** LNet descriptor */
2048 lnet_handle_md_t rqbd_md_h;
2050 /** The buffer itself */
2052 struct ptlrpc_cb_id rqbd_cbid;
2054 * This "embedded" request structure is only used for the
2055 * last request to fit into the buffer
2057 struct ptlrpc_request rqbd_req;
2060 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2062 struct ptlrpc_service_ops {
2064 * if non-NULL called during thread creation (ptlrpc_start_thread())
2065 * to initialize service specific per-thread state.
2067 int (*so_thr_init)(struct ptlrpc_thread *thr);
2069 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2070 * destruct state created by ->srv_init().
2072 void (*so_thr_done)(struct ptlrpc_thread *thr);
2074 * Handler function for incoming requests for this service
2076 int (*so_req_handler)(struct ptlrpc_request *req);
2078 * function to determine priority of the request, it's called
2079 * on every new request
2081 int (*so_hpreq_handler)(struct ptlrpc_request *);
2083 * service-specific print fn
2085 void (*so_req_printer)(void *, struct ptlrpc_request *);
2088 #ifndef __cfs_cacheline_aligned
2089 /* NB: put it here for reducing patche dependence */
2090 # define __cfs_cacheline_aligned
2094 * How many high priority requests to serve before serving one normal
2097 #define PTLRPC_SVC_HP_RATIO 10
2100 * Definition of PortalRPC service.
2101 * The service is listening on a particular portal (like tcp port)
2102 * and perform actions for a specific server like IO service for OST
2103 * or general metadata service for MDS.
2105 struct ptlrpc_service {
2106 /** serialize /proc operations */
2107 spinlock_t srv_lock;
2108 /** most often accessed fields */
2109 /** chain thru all services */
2110 cfs_list_t srv_list;
2111 /** service operations table */
2112 struct ptlrpc_service_ops srv_ops;
2113 /** only statically allocated strings here; we don't clean them */
2115 /** only statically allocated strings here; we don't clean them */
2116 char *srv_thread_name;
2117 /** service thread list */
2118 cfs_list_t srv_threads;
2119 /** threads # should be created for each partition on initializing */
2120 int srv_nthrs_cpt_init;
2121 /** limit of threads number for each partition */
2122 int srv_nthrs_cpt_limit;
2123 /** Root of /proc dir tree for this service */
2124 cfs_proc_dir_entry_t *srv_procroot;
2125 /** Pointer to statistic data for this service */
2126 struct lprocfs_stats *srv_stats;
2127 /** # hp per lp reqs to handle */
2128 int srv_hpreq_ratio;
2129 /** biggest request to receive */
2130 int srv_max_req_size;
2131 /** biggest reply to send */
2132 int srv_max_reply_size;
2133 /** size of individual buffers */
2135 /** # buffers to allocate in 1 group */
2136 int srv_nbuf_per_group;
2137 /** Local portal on which to receive requests */
2138 __u32 srv_req_portal;
2139 /** Portal on the client to send replies to */
2140 __u32 srv_rep_portal;
2142 * Tags for lu_context associated with this thread, see struct
2146 /** soft watchdog timeout multiplier */
2147 int srv_watchdog_factor;
2148 /** under unregister_service */
2149 unsigned srv_is_stopping:1;
2151 /** max # request buffers in history per partition */
2152 int srv_hist_nrqbds_cpt_max;
2153 /** number of CPTs this service bound on */
2155 /** CPTs array this service bound on */
2157 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2159 /** CPT table this service is running over */
2160 struct cfs_cpt_table *srv_cptable;
2162 * partition data for ptlrpc service
2164 struct ptlrpc_service_part *srv_parts[0];
2168 * Definition of PortalRPC service partition data.
2169 * Although a service only has one instance of it right now, but we
2170 * will have multiple instances very soon (instance per CPT).
2172 * it has four locks:
2174 * serialize operations on rqbd and requests waiting for preprocess
2176 * serialize operations active requests sent to this portal
2178 * serialize adaptive timeout stuff
2180 * serialize operations on RS list (reply states)
2182 * We don't have any use-case to take two or more locks at the same time
2183 * for now, so there is no lock order issue.
2185 struct ptlrpc_service_part {
2186 /** back reference to owner */
2187 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2188 /* CPT id, reserved */
2190 /** always increasing number */
2192 /** # of starting threads */
2193 int scp_nthrs_starting;
2194 /** # of stopping threads, reserved for shrinking threads */
2195 int scp_nthrs_stopping;
2196 /** # running threads */
2197 int scp_nthrs_running;
2198 /** service threads list */
2199 cfs_list_t scp_threads;
2202 * serialize the following fields, used for protecting
2203 * rqbd list and incoming requests waiting for preprocess,
2204 * threads starting & stopping are also protected by this lock.
2206 spinlock_t scp_lock __cfs_cacheline_aligned;
2207 /** total # req buffer descs allocated */
2208 int scp_nrqbds_total;
2209 /** # posted request buffers for receiving */
2210 int scp_nrqbds_posted;
2211 /** in progress of allocating rqbd */
2212 int scp_rqbd_allocating;
2213 /** # incoming reqs */
2214 int scp_nreqs_incoming;
2215 /** request buffers to be reposted */
2216 cfs_list_t scp_rqbd_idle;
2217 /** req buffers receiving */
2218 cfs_list_t scp_rqbd_posted;
2219 /** incoming reqs */
2220 cfs_list_t scp_req_incoming;
2221 /** timeout before re-posting reqs, in tick */
2222 cfs_duration_t scp_rqbd_timeout;
2224 * all threads sleep on this. This wait-queue is signalled when new
2225 * incoming request arrives and when difficult reply has to be handled.
2227 cfs_waitq_t scp_waitq;
2229 /** request history */
2230 cfs_list_t scp_hist_reqs;
2231 /** request buffer history */
2232 cfs_list_t scp_hist_rqbds;
2233 /** # request buffers in history */
2234 int scp_hist_nrqbds;
2235 /** sequence number for request */
2237 /** highest seq culled from history */
2238 __u64 scp_hist_seq_culled;
2241 * serialize the following fields, used for processing requests
2242 * sent to this portal
2244 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2245 /** # reqs in either of the NRS heads below */
2246 /** # reqs being served */
2247 int scp_nreqs_active;
2248 /** # HPreqs being served */
2249 int scp_nhreqs_active;
2250 /** # hp requests handled */
2253 /** NRS head for regular requests */
2254 struct ptlrpc_nrs scp_nrs_reg;
2255 /** NRS head for HP requests; this is only valid for services that can
2256 * handle HP requests */
2257 struct ptlrpc_nrs *scp_nrs_hp;
2262 * serialize the following fields, used for changes on
2265 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2266 /** estimated rpc service time */
2267 struct adaptive_timeout scp_at_estimate;
2268 /** reqs waiting for replies */
2269 struct ptlrpc_at_array scp_at_array;
2270 /** early reply timer */
2271 cfs_timer_t scp_at_timer;
2273 cfs_time_t scp_at_checktime;
2274 /** check early replies */
2275 unsigned scp_at_check;
2279 * serialize the following fields, used for processing
2280 * replies for this portal
2282 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2283 /** all the active replies */
2284 cfs_list_t scp_rep_active;
2286 /** replies waiting for service */
2287 cfs_list_t scp_rep_queue;
2289 /** List of free reply_states */
2290 cfs_list_t scp_rep_idle;
2291 /** waitq to run, when adding stuff to srv_free_rs_list */
2292 cfs_waitq_t scp_rep_waitq;
2293 /** # 'difficult' replies */
2294 cfs_atomic_t scp_nreps_difficult;
2297 #define ptlrpc_service_for_each_part(part, i, svc) \
2299 i < (svc)->srv_ncpts && \
2300 (svc)->srv_parts != NULL && \
2301 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2304 * Declaration of ptlrpcd control structure
2306 struct ptlrpcd_ctl {
2308 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2310 unsigned long pc_flags;
2312 * Thread lock protecting structure fields.
2318 struct completion pc_starting;
2322 struct completion pc_finishing;
2324 * Thread requests set.
2326 struct ptlrpc_request_set *pc_set;
2328 * Thread name used in cfs_daemonize()
2332 * Environment for request interpreters to run in.
2334 struct lu_env pc_env;
2336 * Index of ptlrpcd thread in the array.
2340 * Number of the ptlrpcd's partners.
2344 * Pointer to the array of partners' ptlrpcd_ctl structure.
2346 struct ptlrpcd_ctl **pc_partners;
2348 * Record the partner index to be processed next.
2353 * Async rpcs flag to make sure that ptlrpcd_check() is called only
2358 * Currently not used.
2362 * User-space async rpcs callback.
2364 void *pc_wait_callback;
2366 * User-space check idle rpcs callback.
2368 void *pc_idle_callback;
2372 /* Bits for pc_flags */
2373 enum ptlrpcd_ctl_flags {
2375 * Ptlrpc thread start flag.
2377 LIOD_START = 1 << 0,
2379 * Ptlrpc thread stop flag.
2383 * Ptlrpc thread force flag (only stop force so far).
2384 * This will cause aborting any inflight rpcs handled
2385 * by thread if LIOD_STOP is specified.
2387 LIOD_FORCE = 1 << 2,
2389 * This is a recovery ptlrpc thread.
2391 LIOD_RECOVERY = 1 << 3,
2393 * The ptlrpcd is bound to some CPU core.
2402 * Service compatibility function; policy is compatible with all services.
2404 * \param[in] svc The service the policy is attempting to register with.
2405 * \param[in] desc The policy descriptor
2407 * \retval true The policy is compatible with the NRS head
2409 * \see ptlrpc_nrs_pol_desc::pd_compat()
2412 nrs_policy_compat_all(struct ptlrpc_service *svc,
2413 const struct ptlrpc_nrs_pol_desc *desc)
2419 * Service compatibility function; policy is compatible with only a specific
2420 * service which is identified by its human-readable name at
2421 * ptlrpc_service::srv_name.
2423 * \param[in] svc The service the policy is attempting to register with.
2424 * \param[in] desc The policy descriptor
2426 * \retval false The policy is not compatible with the NRS head
2427 * \retval true The policy is compatible with the NRS head
2429 * \see ptlrpc_nrs_pol_desc::pd_compat()
2432 nrs_policy_compat_one(struct ptlrpc_service *svc,
2433 const struct ptlrpc_nrs_pol_desc *desc)
2435 LASSERT(desc->pd_compat_svc_name != NULL);
2436 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2441 /* ptlrpc/events.c */
2442 extern lnet_handle_eq_t ptlrpc_eq_h;
2443 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2444 lnet_process_id_t *peer, lnet_nid_t *self);
2446 * These callbacks are invoked by LNet when something happened to
2450 extern void request_out_callback(lnet_event_t *ev);
2451 extern void reply_in_callback(lnet_event_t *ev);
2452 extern void client_bulk_callback(lnet_event_t *ev);
2453 extern void request_in_callback(lnet_event_t *ev);
2454 extern void reply_out_callback(lnet_event_t *ev);
2455 #ifdef HAVE_SERVER_SUPPORT
2456 extern void server_bulk_callback(lnet_event_t *ev);
2460 /* ptlrpc/connection.c */
2461 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2463 struct obd_uuid *uuid);
2464 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2465 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2466 int ptlrpc_connection_init(void);
2467 void ptlrpc_connection_fini(void);
2468 extern lnet_pid_t ptl_get_pid(void);
2470 /* ptlrpc/niobuf.c */
2472 * Actual interfacing with LNet to put/get/register/unregister stuff
2475 #ifdef HAVE_SERVER_SUPPORT
2476 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2477 unsigned npages, unsigned max_brw,
2478 unsigned type, unsigned portal);
2479 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2480 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2482 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2486 LASSERT(desc != NULL);
2488 spin_lock(&desc->bd_lock);
2489 rc = desc->bd_md_count;
2490 spin_unlock(&desc->bd_lock);
2495 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2496 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2498 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2500 struct ptlrpc_bulk_desc *desc;
2503 LASSERT(req != NULL);
2504 desc = req->rq_bulk;
2506 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2507 req->rq_bulk_deadline > cfs_time_current_sec())
2513 spin_lock(&desc->bd_lock);
2514 rc = desc->bd_md_count;
2515 spin_unlock(&desc->bd_lock);
2519 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2520 #define PTLRPC_REPLY_EARLY 0x02
2521 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2522 int ptlrpc_reply(struct ptlrpc_request *req);
2523 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2524 int ptlrpc_error(struct ptlrpc_request *req);
2525 void ptlrpc_resend_req(struct ptlrpc_request *request);
2526 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2527 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2528 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2531 /* ptlrpc/client.c */
2533 * Client-side portals API. Everything to send requests, receive replies,
2534 * request queues, request management, etc.
2537 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2538 struct ptlrpc_client *);
2539 void ptlrpc_cleanup_client(struct obd_import *imp);
2540 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2542 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2543 int ptlrpc_replay_req(struct ptlrpc_request *req);
2544 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2545 void ptlrpc_restart_req(struct ptlrpc_request *req);
2546 void ptlrpc_abort_inflight(struct obd_import *imp);
2547 void ptlrpc_cleanup_imp(struct obd_import *imp);
2548 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2550 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2551 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2553 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2554 set_interpreter_func fn, void *data);
2555 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2556 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2557 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2558 int ptlrpc_expired_set(void *data);
2559 void ptlrpc_interrupted_set(void *data);
2560 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2561 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2562 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2563 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2564 struct ptlrpc_request *req);
2566 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2567 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2569 struct ptlrpc_request_pool *
2570 ptlrpc_init_rq_pool(int, int,
2571 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2573 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2574 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2575 const struct req_format *format);
2576 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2577 struct ptlrpc_request_pool *,
2578 const struct req_format *format);
2579 void ptlrpc_request_free(struct ptlrpc_request *request);
2580 int ptlrpc_request_pack(struct ptlrpc_request *request,
2581 __u32 version, int opcode);
2582 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2583 const struct req_format *format,
2584 __u32 version, int opcode);
2585 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2586 __u32 version, int opcode, char **bufs,
2587 struct ptlrpc_cli_ctx *ctx);
2588 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2589 int opcode, int count, __u32 *lengths,
2591 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2592 __u32 version, int opcode,
2593 int count, __u32 *lengths, char **bufs,
2594 struct ptlrpc_request_pool *pool);
2595 void ptlrpc_req_finished(struct ptlrpc_request *request);
2596 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2597 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2598 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2599 unsigned npages, unsigned max_brw,
2600 unsigned type, unsigned portal);
2601 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2602 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2604 __ptlrpc_free_bulk(bulk, 1);
2606 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2608 __ptlrpc_free_bulk(bulk, 0);
2610 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2611 cfs_page_t *page, int pageoffset, int len, int);
2612 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2613 cfs_page_t *page, int pageoffset,
2616 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2619 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2620 cfs_page_t *page, int pageoffset,
2623 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2626 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2627 struct obd_import *imp);
2628 __u64 ptlrpc_next_xid(void);
2629 __u64 ptlrpc_sample_next_xid(void);
2630 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2632 /* Set of routines to run a function in ptlrpcd context */
2633 void *ptlrpcd_alloc_work(struct obd_import *imp,
2634 int (*cb)(const struct lu_env *, void *), void *data);
2635 void ptlrpcd_destroy_work(void *handler);
2636 int ptlrpcd_queue_work(void *handler);
2639 struct ptlrpc_service_buf_conf {
2640 /* nbufs is buffers # to allocate when growing the pool */
2641 unsigned int bc_nbufs;
2642 /* buffer size to post */
2643 unsigned int bc_buf_size;
2644 /* portal to listed for requests on */
2645 unsigned int bc_req_portal;
2646 /* portal of where to send replies to */
2647 unsigned int bc_rep_portal;
2648 /* maximum request size to be accepted for this service */
2649 unsigned int bc_req_max_size;
2650 /* maximum reply size this service can ever send */
2651 unsigned int bc_rep_max_size;
2654 struct ptlrpc_service_thr_conf {
2655 /* threadname should be 8 characters or less - 6 will be added on */
2657 /* threads increasing factor for each CPU */
2658 unsigned int tc_thr_factor;
2659 /* service threads # to start on each partition while initializing */
2660 unsigned int tc_nthrs_init;
2662 * low water of threads # upper-limit on each partition while running,
2663 * service availability may be impacted if threads number is lower
2664 * than this value. It can be ZERO if the service doesn't require
2665 * CPU affinity or there is only one partition.
2667 unsigned int tc_nthrs_base;
2668 /* "soft" limit for total threads number */
2669 unsigned int tc_nthrs_max;
2670 /* user specified threads number, it will be validated due to
2671 * other members of this structure. */
2672 unsigned int tc_nthrs_user;
2673 /* set NUMA node affinity for service threads */
2674 unsigned int tc_cpu_affinity;
2675 /* Tags for lu_context associated with service thread */
2679 struct ptlrpc_service_cpt_conf {
2680 struct cfs_cpt_table *cc_cptable;
2681 /* string pattern to describe CPTs for a service */
2685 struct ptlrpc_service_conf {
2688 /* soft watchdog timeout multiplifier to print stuck service traces */
2689 unsigned int psc_watchdog_factor;
2690 /* buffer information */
2691 struct ptlrpc_service_buf_conf psc_buf;
2692 /* thread information */
2693 struct ptlrpc_service_thr_conf psc_thr;
2694 /* CPU partition information */
2695 struct ptlrpc_service_cpt_conf psc_cpt;
2696 /* function table */
2697 struct ptlrpc_service_ops psc_ops;
2700 /* ptlrpc/service.c */
2702 * Server-side services API. Register/unregister service, request state
2703 * management, service thread management
2707 void ptlrpc_save_lock(struct ptlrpc_request *req,
2708 struct lustre_handle *lock, int mode, int no_ack);
2709 void ptlrpc_commit_replies(struct obd_export *exp);
2710 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2711 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2712 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2713 struct ptlrpc_service *ptlrpc_register_service(
2714 struct ptlrpc_service_conf *conf,
2715 struct proc_dir_entry *proc_entry);
2716 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2718 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2719 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2720 int liblustre_check_services(void *arg);
2721 void ptlrpc_daemonize(char *name);
2722 int ptlrpc_service_health_check(struct ptlrpc_service *);
2723 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2726 int ptlrpc_hr_init(void);
2727 void ptlrpc_hr_fini(void);
2729 # define ptlrpc_hr_init() (0)
2730 # define ptlrpc_hr_fini() do {} while(0)
2735 /* ptlrpc/import.c */
2740 int ptlrpc_connect_import(struct obd_import *imp);
2741 int ptlrpc_init_import(struct obd_import *imp);
2742 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2743 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2744 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2747 /* ptlrpc/pack_generic.c */
2748 int ptlrpc_reconnect_import(struct obd_import *imp);
2752 * ptlrpc msg buffer and swab interface
2756 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2758 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2760 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2761 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2763 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2764 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2766 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2767 __u32 *lens, char **bufs);
2768 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2770 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2771 __u32 *lens, char **bufs, int flags);
2772 #define LPRFL_EARLY_REPLY 1
2773 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2774 char **bufs, int flags);
2775 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2776 unsigned int newlen, int move_data);
2777 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2778 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2779 int lustre_msg_hdr_size(__u32 magic, int count);
2780 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2781 int lustre_msg_size_v2(int count, __u32 *lengths);
2782 int lustre_packed_msg_size(struct lustre_msg *msg);
2783 int lustre_msg_early_size(void);
2784 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2785 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2786 int lustre_msg_buflen(struct lustre_msg *m, int n);
2787 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2788 int lustre_msg_bufcount(struct lustre_msg *m);
2789 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2790 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2791 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2792 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2793 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2794 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2795 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2796 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2797 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2798 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2799 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2800 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2801 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2802 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2803 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2804 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2805 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2806 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2807 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2808 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2809 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2810 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2811 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2812 int lustre_msg_get_status(struct lustre_msg *msg);
2813 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2814 int lustre_msg_is_v1(struct lustre_msg *msg);
2815 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2816 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2817 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2818 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2819 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2820 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2821 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2823 # warning "remove checksum compatibility support for b1_8"
2824 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2826 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2827 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2828 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2829 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2830 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2831 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2832 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2833 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2834 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2835 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2836 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2837 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2838 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2839 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2840 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2843 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2844 unsigned int newlen, int move_data)
2846 LASSERT(req->rq_reply_state);
2847 LASSERT(req->rq_repmsg);
2848 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2853 /** Change request phase of \a req to \a new_phase */
2855 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2857 if (req->rq_phase == new_phase)
2860 if (new_phase == RQ_PHASE_UNREGISTERING) {
2861 req->rq_next_phase = req->rq_phase;
2863 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2866 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2868 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2871 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2872 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2874 req->rq_phase = new_phase;
2878 * Returns true if request \a req got early reply and hard deadline is not met
2881 ptlrpc_client_early(struct ptlrpc_request *req)
2883 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2884 req->rq_reply_deadline > cfs_time_current_sec())
2886 return req->rq_early;
2890 * Returns true if we got real reply from server for this request
2893 ptlrpc_client_replied(struct ptlrpc_request *req)
2895 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2896 req->rq_reply_deadline > cfs_time_current_sec())
2898 return req->rq_replied;
2901 /** Returns true if request \a req is in process of receiving server reply */
2903 ptlrpc_client_recv(struct ptlrpc_request *req)
2905 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2906 req->rq_reply_deadline > cfs_time_current_sec())
2908 return req->rq_receiving_reply;
2912 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2916 spin_lock(&req->rq_lock);
2917 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2918 req->rq_reply_deadline > cfs_time_current_sec()) {
2919 spin_unlock(&req->rq_lock);
2922 rc = req->rq_receiving_reply || req->rq_must_unlink;
2923 spin_unlock(&req->rq_lock);
2928 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2930 if (req->rq_set == NULL)
2931 cfs_waitq_signal(&req->rq_reply_waitq);
2933 cfs_waitq_signal(&req->rq_set->set_waitq);
2937 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2939 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2940 cfs_atomic_inc(&rs->rs_refcount);
2944 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2946 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2947 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
2948 lustre_free_reply_state(rs);
2951 /* Should only be called once per req */
2952 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2954 if (req->rq_reply_state == NULL)
2955 return; /* shouldn't occur */
2956 ptlrpc_rs_decref(req->rq_reply_state);
2957 req->rq_reply_state = NULL;
2958 req->rq_repmsg = NULL;
2961 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2963 return lustre_msg_get_magic(req->rq_reqmsg);
2966 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2968 switch (req->rq_reqmsg->lm_magic) {
2969 case LUSTRE_MSG_MAGIC_V2:
2970 return req->rq_reqmsg->lm_repsize;
2972 LASSERTF(0, "incorrect message magic: %08x\n",
2973 req->rq_reqmsg->lm_magic);
2978 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2980 if (req->rq_delay_limit != 0 &&
2981 cfs_time_before(cfs_time_add(req->rq_queued_time,
2982 cfs_time_seconds(req->rq_delay_limit)),
2983 cfs_time_current())) {
2989 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2991 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2992 spin_lock(&req->rq_lock);
2993 req->rq_no_resend = 1;
2994 spin_unlock(&req->rq_lock);
2996 return req->rq_no_resend;
3000 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3002 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3004 return svcpt->scp_service->srv_watchdog_factor *
3005 max_t(int, at, obd_timeout);
3008 static inline struct ptlrpc_service *
3009 ptlrpc_req2svc(struct ptlrpc_request *req)
3011 LASSERT(req->rq_rqbd != NULL);
3012 return req->rq_rqbd->rqbd_svcpt->scp_service;
3015 /* ldlm/ldlm_lib.c */
3017 * Target client logic
3020 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3021 int client_obd_cleanup(struct obd_device *obddev);
3022 int client_connect_import(const struct lu_env *env,
3023 struct obd_export **exp, struct obd_device *obd,
3024 struct obd_uuid *cluuid, struct obd_connect_data *,
3026 int client_disconnect_export(struct obd_export *exp);
3027 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3029 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3030 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3031 struct obd_uuid *uuid);
3032 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3033 void client_destroy_import(struct obd_import *imp);
3036 #ifdef HAVE_SERVER_SUPPORT
3037 int server_disconnect_export(struct obd_export *exp);
3040 /* ptlrpc/pinger.c */
3042 * Pinger API (client side only)
3045 extern int suppress_pings;
3046 enum timeout_event {
3049 struct timeout_item;
3050 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3051 int ptlrpc_pinger_add_import(struct obd_import *imp);
3052 int ptlrpc_pinger_del_import(struct obd_import *imp);
3053 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3054 timeout_cb_t cb, void *data,
3055 cfs_list_t *obd_list);
3056 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
3057 enum timeout_event event);
3058 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3059 int ptlrpc_obd_ping(struct obd_device *obd);
3060 cfs_time_t ptlrpc_suspend_wakeup_time(void);
3062 void ping_evictor_start(void);
3063 void ping_evictor_stop(void);
3065 #define ping_evictor_start() do {} while (0)
3066 #define ping_evictor_stop() do {} while (0)
3068 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
3071 /* ptlrpc daemon bind policy */
3073 /* all ptlrpcd threads are free mode */
3074 PDB_POLICY_NONE = 1,
3075 /* all ptlrpcd threads are bound mode */
3076 PDB_POLICY_FULL = 2,
3077 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3078 PDB_POLICY_PAIR = 3,
3079 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3080 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3081 * If kernel supports NUMA, pthrpcd threads are binded and
3082 * grouped by NUMA node */
3083 PDB_POLICY_NEIGHBOR = 4,
3086 /* ptlrpc daemon load policy
3087 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3088 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3089 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3090 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3091 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3093 /* on the same CPU core as the caller */
3094 PDL_POLICY_SAME = 1,
3095 /* within the same CPU partition, but not the same core as the caller */
3096 PDL_POLICY_LOCAL = 2,
3097 /* round-robin on all CPU cores, but not the same core as the caller */
3098 PDL_POLICY_ROUND = 3,
3099 /* the specified CPU core is preferred, but not enforced */
3100 PDL_POLICY_PREFERRED = 4,
3103 /* ptlrpc/ptlrpcd.c */
3104 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3105 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3106 void ptlrpcd_wake(struct ptlrpc_request *req);
3107 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3108 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3109 int ptlrpcd_addref(void);
3110 void ptlrpcd_decref(void);
3112 /* ptlrpc/lproc_ptlrpc.c */
3114 * procfs output related functions
3117 const char* ll_opcode2str(__u32 opcode);
3119 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3120 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3121 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3123 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3124 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3125 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3129 /* ptlrpc/llog_server.c */
3130 int llog_origin_handle_open(struct ptlrpc_request *req);
3131 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3132 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3133 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3134 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3135 int llog_origin_handle_close(struct ptlrpc_request *req);
3136 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3138 /* ptlrpc/llog_client.c */
3139 extern struct llog_operations llog_client_ops;