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
320 * Assume file name length = FNAME_MAX = 256 (true for ext3).
321 * path name length = PATH_MAX = 4096
322 * LOV MD size max = EA_MAX = 24 * 2000
323 * (NB: 24 is size of lov_ost_data)
324 * LOV LOGCOOKIE size max = 32 * 2000
325 * (NB: 32 is size of llog_cookie)
326 * symlink: FNAME_MAX + PATH_MAX <- largest
327 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
328 * rename: FNAME_MAX + FNAME_MAX
329 * open: FNAME_MAX + EA_MAX
331 * MDS_MAXREQSIZE ~= 4736 bytes =
332 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
333 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
335 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
336 * except in the open case where there are a large number of OSTs in a LOV.
338 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
339 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
342 * MDS incoming request with LOV EA
343 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
345 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
346 362 + LOV_MAX_STRIPE_COUNT * 24)
348 * MDS outgoing reply with LOV EA
350 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
351 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
353 * but 2.4 or later MDS will never send reply with llog_cookie to any
354 * version client. This macro is defined for server side reply buffer size.
356 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
359 * This is the size of a maximum REINT_SETXATTR request:
361 * lustre_msg 56 (32 + 4 x 5 + 4)
363 * mdt_rec_setxattr 136
365 * name 256 (XATTR_NAME_MAX)
366 * value 65536 (XATTR_SIZE_MAX)
368 #define MDS_EA_MAXREQSIZE 66288
371 * These are the maximum request and reply sizes (rounded up to 1 KB
372 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
374 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
375 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
376 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
379 * The update request includes all of updates from the create, which might
380 * include linkea (4K maxim), together with other updates, we set it to 9K:
381 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
383 #define MDS_OUT_MAXREQSIZE (9 * 1024)
384 #define MDS_OUT_MAXREPSIZE MDS_MAXREPSIZE
386 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
387 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
391 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
392 * However, we need to allocate a much larger buffer for it because LNet
393 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
394 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
395 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
396 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
397 * utilization is very low.
399 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
400 * reused until all requests fit in it have been processed and released,
401 * which means one long blocked request can prevent the rqbd be reused.
402 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
403 * from LNet with unused 65 KB, buffer utilization will be about 59%.
404 * Please check LU-2432 for details.
406 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
410 * MDS_OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
411 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
412 * extra bytes to each request buffer to improve buffer utilization rate.
414 #define MDS_OUT_BUFSIZE max(MDS_OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
417 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
418 #define FLD_MAXREQSIZE (160)
420 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
421 #define FLD_MAXREPSIZE (152)
422 #define FLD_BUFSIZE (1 << 12)
425 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
427 #define SEQ_MAXREQSIZE (160)
429 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
430 #define SEQ_MAXREPSIZE (152)
431 #define SEQ_BUFSIZE (1 << 12)
433 /** MGS threads must be >= 3, see bug 22458 comment #28 */
434 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
435 #define MGS_NTHRS_MAX 32
438 #define MGS_BUFSIZE (8 * 1024)
439 #define MGS_MAXREQSIZE (7 * 1024)
440 #define MGS_MAXREPSIZE (9 * 1024)
443 * OSS threads constants:
445 * Given 8 as factor and 64 as base threads number
448 * On 8-core server configured to 2 partitions, we will have
449 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
452 * On 32-core machine configured to 4 partitions, we will have
453 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
454 * will be 112 * 4 = 448.
457 * On 64-core machine configured to 4 partitions, we will have
458 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
459 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
460 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
461 * for each partition.
463 * So we can see that with these constants, threads number wil be at the
464 * similar level of old versions, unless the server has many cores.
466 /* depress threads factor for VM with small memory size */
467 #define OSS_THR_FACTOR min_t(int, 8, \
468 CFS_NUM_CACHEPAGES >> (28 - CFS_PAGE_SHIFT))
469 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
470 #define OSS_NTHRS_BASE 64
471 #define OSS_NTHRS_MAX 512
473 /* threads for handling "create" request */
474 #define OSS_CR_THR_FACTOR 1
475 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
476 #define OSS_CR_NTHRS_BASE 8
477 #define OSS_CR_NTHRS_MAX 64
480 * OST_IO_MAXREQSIZE ~=
481 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
482 * DT_MAX_BRW_PAGES * niobuf_remote
484 * - single object with 16 pages is 512 bytes
485 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
486 * - Must be a multiple of 1024
487 * - actual size is about 18K
489 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
490 sizeof(struct ptlrpc_body) + \
491 sizeof(struct obdo) + \
492 sizeof(struct obd_ioobj) + \
493 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
495 * FIEMAP request can be 4K+ for now
497 #define OST_MAXREQSIZE (5 * 1024)
498 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
499 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
501 #define OST_MAXREPSIZE (9 * 1024)
502 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
505 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
506 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
508 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
509 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
511 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
513 /* Macro to hide a typecast. */
514 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
517 * Structure to single define portal connection.
519 struct ptlrpc_connection {
520 /** linkage for connections hash table */
521 cfs_hlist_node_t c_hash;
522 /** Our own lnet nid for this connection */
524 /** Remote side nid for this connection */
525 lnet_process_id_t c_peer;
526 /** UUID of the other side */
527 struct obd_uuid c_remote_uuid;
528 /** reference counter for this connection */
529 cfs_atomic_t c_refcount;
532 /** Client definition for PortalRPC */
533 struct ptlrpc_client {
534 /** What lnet portal does this client send messages to by default */
535 __u32 cli_request_portal;
536 /** What portal do we expect replies on */
537 __u32 cli_reply_portal;
538 /** Name of the client */
542 /** state flags of requests */
543 /* XXX only ones left are those used by the bulk descs as well! */
544 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
545 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
547 #define REQ_MAX_ACK_LOCKS 8
549 union ptlrpc_async_args {
551 * Scratchpad for passing args to completion interpreter. Users
552 * cast to the struct of their choosing, and CLASSERT that this is
553 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
554 * a pointer to it here. The pointer_arg ensures this struct is at
555 * least big enough for that.
557 void *pointer_arg[11];
561 struct ptlrpc_request_set;
562 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
563 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
566 * Definition of request set structure.
567 * Request set is a list of requests (not necessary to the same target) that
568 * once populated with RPCs could be sent in parallel.
569 * There are two kinds of request sets. General purpose and with dedicated
570 * serving thread. Example of the latter is ptlrpcd set.
571 * For general purpose sets once request set started sending it is impossible
572 * to add new requests to such set.
573 * Provides a way to call "completion callbacks" when all requests in the set
576 struct ptlrpc_request_set {
577 cfs_atomic_t set_refcount;
578 /** number of in queue requests */
579 cfs_atomic_t set_new_count;
580 /** number of uncompleted requests */
581 cfs_atomic_t set_remaining;
582 /** wait queue to wait on for request events */
583 cfs_waitq_t set_waitq;
584 cfs_waitq_t *set_wakeup_ptr;
585 /** List of requests in the set */
586 cfs_list_t set_requests;
588 * List of completion callbacks to be called when the set is completed
589 * This is only used if \a set_interpret is NULL.
590 * Links struct ptlrpc_set_cbdata.
592 cfs_list_t set_cblist;
593 /** Completion callback, if only one. */
594 set_interpreter_func set_interpret;
595 /** opaq argument passed to completion \a set_interpret callback. */
598 * Lock for \a set_new_requests manipulations
599 * locked so that any old caller can communicate requests to
600 * the set holder who can then fold them into the lock-free set
602 spinlock_t set_new_req_lock;
603 /** List of new yet unsent requests. Only used with ptlrpcd now. */
604 cfs_list_t set_new_requests;
606 /** rq_status of requests that have been freed already */
608 /** Additional fields used by the flow control extension */
609 /** Maximum number of RPCs in flight */
610 int set_max_inflight;
611 /** Callback function used to generate RPCs */
612 set_producer_func set_producer;
613 /** opaq argument passed to the producer callback */
614 void *set_producer_arg;
618 * Description of a single ptrlrpc_set callback
620 struct ptlrpc_set_cbdata {
621 /** List linkage item */
623 /** Pointer to interpreting function */
624 set_interpreter_func psc_interpret;
625 /** Opaq argument to pass to the callback */
629 struct ptlrpc_bulk_desc;
630 struct ptlrpc_service_part;
631 struct ptlrpc_service;
634 * ptlrpc callback & work item stuff
636 struct ptlrpc_cb_id {
637 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
638 void *cbid_arg; /* additional arg */
641 /** Maximum number of locks to fit into reply state */
642 #define RS_MAX_LOCKS 8
646 * Structure to define reply state on the server
647 * Reply state holds various reply message information. Also for "difficult"
648 * replies (rep-ack case) we store the state after sending reply and wait
649 * for the client to acknowledge the reception. In these cases locks could be
650 * added to the state for replay/failover consistency guarantees.
652 struct ptlrpc_reply_state {
653 /** Callback description */
654 struct ptlrpc_cb_id rs_cb_id;
655 /** Linkage for list of all reply states in a system */
657 /** Linkage for list of all reply states on same export */
658 cfs_list_t rs_exp_list;
659 /** Linkage for list of all reply states for same obd */
660 cfs_list_t rs_obd_list;
662 cfs_list_t rs_debug_list;
664 /** A spinlock to protect the reply state flags */
666 /** Reply state flags */
667 unsigned long rs_difficult:1; /* ACK/commit stuff */
668 unsigned long rs_no_ack:1; /* no ACK, even for
669 difficult requests */
670 unsigned long rs_scheduled:1; /* being handled? */
671 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
672 unsigned long rs_handled:1; /* been handled yet? */
673 unsigned long rs_on_net:1; /* reply_out_callback pending? */
674 unsigned long rs_prealloc:1; /* rs from prealloc list */
675 unsigned long rs_committed:1;/* the transaction was committed
676 and the rs was dispatched
677 by ptlrpc_commit_replies */
678 /** Size of the state */
682 /** Transaction number */
686 struct obd_export *rs_export;
687 struct ptlrpc_service_part *rs_svcpt;
688 /** Lnet metadata handle for the reply */
689 lnet_handle_md_t rs_md_h;
690 cfs_atomic_t rs_refcount;
692 /** Context for the sevice thread */
693 struct ptlrpc_svc_ctx *rs_svc_ctx;
694 /** Reply buffer (actually sent to the client), encoded if needed */
695 struct lustre_msg *rs_repbuf; /* wrapper */
696 /** Size of the reply buffer */
697 int rs_repbuf_len; /* wrapper buf length */
698 /** Size of the reply message */
699 int rs_repdata_len; /* wrapper msg length */
701 * Actual reply message. Its content is encrupted (if needed) to
702 * produce reply buffer for actual sending. In simple case
703 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
705 struct lustre_msg *rs_msg; /* reply message */
707 /** Number of locks awaiting client ACK */
709 /** Handles of locks awaiting client reply ACK */
710 struct lustre_handle rs_locks[RS_MAX_LOCKS];
711 /** Lock modes of locks in \a rs_locks */
712 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
715 struct ptlrpc_thread;
719 RQ_PHASE_NEW = 0xebc0de00,
720 RQ_PHASE_RPC = 0xebc0de01,
721 RQ_PHASE_BULK = 0xebc0de02,
722 RQ_PHASE_INTERPRET = 0xebc0de03,
723 RQ_PHASE_COMPLETE = 0xebc0de04,
724 RQ_PHASE_UNREGISTERING = 0xebc0de05,
725 RQ_PHASE_UNDEFINED = 0xebc0de06
728 /** Type of request interpreter call-back */
729 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
730 struct ptlrpc_request *req,
734 * Definition of request pool structure.
735 * The pool is used to store empty preallocated requests for the case
736 * when we would actually need to send something without performing
737 * any allocations (to avoid e.g. OOM).
739 struct ptlrpc_request_pool {
740 /** Locks the list */
742 /** list of ptlrpc_request structs */
743 cfs_list_t prp_req_list;
744 /** Maximum message size that would fit into a rquest from this pool */
746 /** Function to allocate more requests for this pool */
747 void (*prp_populate)(struct ptlrpc_request_pool *, int);
756 * \defgroup nrs Network Request Scheduler
759 struct ptlrpc_nrs_policy;
760 struct ptlrpc_nrs_resource;
761 struct ptlrpc_nrs_request;
764 * NRS control operations.
766 * These are common for all policies.
768 enum ptlrpc_nrs_ctl {
770 * Activate the policy.
772 PTLRPC_NRS_CTL_START,
774 * Reserved for multiple primary policies, which may be a possibility
779 * Recycle resources for inactive policies.
781 PTLRPC_NRS_CTL_SHRINK,
783 * Not a valid opcode.
785 PTLRPC_NRS_CTL_INVALID,
787 * Policies can start using opcodes from this value and onwards for
788 * their own purposes; the assigned value itself is arbitrary.
790 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
794 * NRS policy operations.
796 * These determine the behaviour of a policy, and are called in response to
799 struct ptlrpc_nrs_pol_ops {
801 * Called during policy registration; this operation is optional.
803 * \param[in] policy The policy being initialized
805 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
807 * Called during policy unregistration; this operation is optional.
809 * \param[in] policy The policy being unregistered/finalized
811 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
813 * Called when activating a policy via lprocfs; policies allocate and
814 * initialize their resources here; this operation is optional.
816 * \param[in] policy The policy being started
818 * \see nrs_policy_start_locked()
820 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
822 * Called when deactivating a policy via lprocfs; policies deallocate
823 * their resources here; this operation is optional
825 * \param[in] policy The policy being stopped
827 * \see nrs_policy_stop_final()
829 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
831 * Used for policy-specific operations; i.e. not generic ones like
832 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
833 * to an ioctl; this operation is optional.
835 * \param[in] policy The policy carrying out operation \a opc
836 * \param[in] opc The command operation being carried out
837 * \param[in,out] arg An generic buffer for communication between the
838 * user and the control operation
843 * \see ptlrpc_nrs_policy_control()
845 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
846 enum ptlrpc_nrs_ctl opc, void *arg);
849 * Called when obtaining references to the resources of the resource
850 * hierarchy for a request that has arrived for handling at the PTLRPC
851 * service. Policies should return -ve for requests they do not wish
852 * to handle. This operation is mandatory.
854 * \param[in] policy The policy we're getting resources for.
855 * \param[in] nrq The request we are getting resources for.
856 * \param[in] parent The parent resource of the resource being
857 * requested; set to NULL if none.
858 * \param[out] resp The resource is to be returned here; the
859 * fallback policy in an NRS head should
860 * \e always return a non-NULL pointer value.
861 * \param[in] moving_req When set, signifies that this is an attempt
862 * to obtain resources for a request being moved
863 * to the high-priority NRS head by
864 * ldlm_lock_reorder_req().
865 * This implies two things:
866 * 1. We are under obd_export::exp_rpc_lock and
867 * so should not sleep.
868 * 2. We should not perform non-idempotent or can
869 * skip performing idempotent operations that
870 * were carried out when resources were first
871 * taken for the request when it was initialized
872 * in ptlrpc_nrs_req_initialize().
874 * \retval 0, +ve The level of the returned resource in the resource
875 * hierarchy; currently only 0 (for a non-leaf resource)
876 * and 1 (for a leaf resource) are supported by the
880 * \see ptlrpc_nrs_req_initialize()
881 * \see ptlrpc_nrs_hpreq_add_nolock()
882 * \see ptlrpc_nrs_req_hp_move()
884 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
885 struct ptlrpc_nrs_request *nrq,
886 struct ptlrpc_nrs_resource *parent,
887 struct ptlrpc_nrs_resource **resp,
890 * Called when releasing references taken for resources in the resource
891 * hierarchy for the request; this operation is optional.
893 * \param[in] policy The policy the resource belongs to
894 * \param[in] res The resource to be freed
896 * \see ptlrpc_nrs_req_finalize()
897 * \see ptlrpc_nrs_hpreq_add_nolock()
898 * \see ptlrpc_nrs_req_hp_move()
900 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
901 struct ptlrpc_nrs_resource *res);
904 * Obtain a request for handling from the policy via polling; this
905 * operation is mandatory.
907 * \param[in] policy The policy to poll
909 * \retval NULL No erquest available for handling
910 * \retval valid-pointer The request polled for handling
912 * \see ptlrpc_nrs_req_poll_nolock()
914 struct ptlrpc_nrs_request *
915 (*op_req_poll) (struct ptlrpc_nrs_policy *policy);
917 * Called when attempting to add a request to a policy for later
918 * handling; this operation is mandatory.
920 * \param[in] policy The policy on which to enqueue \a nrq
921 * \param[in] nrq The request to enqueue
926 * \see ptlrpc_nrs_req_add_nolock()
928 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
929 struct ptlrpc_nrs_request *nrq);
931 * Removes a request from the policy's set of pending requests. Normally
932 * called after a request has been polled successfully from the policy
933 * for handling; this operation is mandatory.
935 * \param[in] policy The policy the request \a nrq belongs to
936 * \param[in] nrq The request to dequeue
938 * \see ptlrpc_nrs_req_del_nolock()
940 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
941 struct ptlrpc_nrs_request *nrq);
943 * Called before carrying out the request; should not block. Could be
944 * used for job/resource control; this operation is optional.
946 * \param[in] policy The policy which is starting to handle request
948 * \param[in] nrq The request
950 * \pre spin_is_locked(&svcpt->scp_req_lock)
952 * \see ptlrpc_nrs_req_start_nolock()
954 void (*op_req_start) (struct ptlrpc_nrs_policy *policy,
955 struct ptlrpc_nrs_request *nrq);
957 * Called after the request being carried out. Could be used for
958 * job/resource control; this operation is optional.
960 * \param[in] policy The policy which is stopping to handle request
962 * \param[in] nrq The request
964 * \pre spin_is_locked(&svcpt->scp_req_lock)
966 * \see ptlrpc_nrs_req_stop_nolock()
968 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
969 struct ptlrpc_nrs_request *nrq);
971 * Registers the policy's lprocfs interface with a PTLRPC service.
973 * \param[in] svc The service
978 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
980 * Unegisters the policy's lprocfs interface with a PTLRPC service.
982 * \param[in] svc The service
984 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
990 enum nrs_policy_flags {
992 * Fallback policy, use this flag only on a single supported policy per
993 * service. Do not use this flag for policies registering using
994 * ptlrpc_nrs_policy_register() (i.e. ones that are not in
995 * \e nrs_pols_builtin).
997 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
999 * Start policy immediately after registering.
1001 PTLRPC_NRS_FL_REG_START = (1 << 1),
1003 * This is a polciy registering externally with NRS core, via
1004 * ptlrpc_nrs_policy_register(), (i.e. one that is not in
1005 * \e nrs_pols_builtin. Used to avoid ptlrpc_nrs_policy_register()
1006 * racing with a policy start operation issued by the user via lprocfs.
1008 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1014 * Denotes whether an NRS instance is for handling normal or high-priority
1015 * RPCs, or whether an operation pertains to one or both of the NRS instances
1018 enum ptlrpc_nrs_queue_type {
1019 PTLRPC_NRS_QUEUE_REG,
1020 PTLRPC_NRS_QUEUE_HP,
1021 PTLRPC_NRS_QUEUE_BOTH,
1027 * A PTLRPC service has at least one NRS head instance for handling normal
1028 * priority RPCs, and may optionally have a second NRS head instance for
1029 * handling high-priority RPCs. Each NRS head maintains a list of available
1030 * policies, of which one and only one policy is acting as the fallback policy,
1031 * and optionally a different policy may be acting as the primary policy. For
1032 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1033 * enqueue the RPC using the primary policy (if any). The fallback policy is
1034 * used in the following cases:
1035 * - when there was no primary policy in the
1036 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1038 * - when the primary policy that was at the
1039 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1040 * RPC was initialized, denoted it did not wish, or for some other reason was
1041 * not able to handle the request, by returning a non-valid NRS resource
1043 * - when the primary policy that was at the
1044 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1045 * RPC was initialized, fails later during the request enqueueing stage.
1047 * \see nrs_resource_get_safe()
1048 * \see nrs_request_enqueue()
1051 spinlock_t nrs_lock;
1052 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1054 * Linkage into nrs_core_heads_list
1056 cfs_list_t nrs_heads;
1058 * List of registered policies
1060 cfs_list_t nrs_policy_list;
1062 * List of policies with queued requests. Policies that have any
1063 * outstanding requests are queued here, and this list is queried
1064 * in a round-robin manner from NRS core when obtaining a request
1065 * for handling. This ensures that requests from policies that at some
1066 * point transition away from the
1067 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1069 cfs_list_t nrs_policy_queued;
1071 * Service partition for this NRS head
1073 struct ptlrpc_service_part *nrs_svcpt;
1075 * Primary policy, which is the preferred policy for handling RPCs
1077 struct ptlrpc_nrs_policy *nrs_policy_primary;
1079 * Fallback policy, which is the backup policy for handling RPCs
1081 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1083 * This NRS head handles either HP or regular requests
1085 enum ptlrpc_nrs_queue_type nrs_queue_type;
1087 * # queued requests from all policies in this NRS head
1089 unsigned long nrs_req_queued;
1091 * # scheduled requests from all policies in this NRS head
1093 unsigned long nrs_req_started;
1095 * # policies on this NRS
1096 * TODO: Can we avoid having this?
1098 unsigned nrs_num_pols;
1100 * This NRS head is in progress of starting a policy
1102 unsigned nrs_policy_starting:1;
1104 * In progress of shutting down the whole NRS head; used during
1107 unsigned nrs_stopping:1;
1110 #define NRS_POL_NAME_MAX 16
1113 * NRS policy registering descriptor
1115 * Is used to hold a description of a policy that can be passed to NRS core in
1116 * order to register the policy with NRS heads in different PTLRPC services.
1118 struct ptlrpc_nrs_pol_desc {
1120 * Human-readable policy name
1122 char pd_name[NRS_POL_NAME_MAX];
1124 * NRS operations for this policy
1126 struct ptlrpc_nrs_pol_ops *pd_ops;
1128 * Service Compatibility function; this determines whether a policy is
1129 * adequate for handling RPCs of a particular PTLRPC service.
1131 * XXX:This should give the same result during policy
1132 * registration and unregistration, and for all partitions of a
1133 * service; so the result should not depend on temporal service
1134 * or other properties, that may influence the result.
1136 bool (*pd_compat) (struct ptlrpc_service *svc,
1137 const struct ptlrpc_nrs_pol_desc *desc);
1139 * Optionally set for policies that support a single ptlrpc service,
1140 * i.e. ones that have \a pd_compat set to nrs_policy_compat_one()
1142 char *pd_compat_svc_name;
1144 * Bitmask of nrs_policy_flags
1148 * Link into nrs_core::nrs_policies
1156 * Policies transition from one state to the other during their lifetime
1158 enum ptlrpc_nrs_pol_state {
1160 * Not a valid policy state.
1162 NRS_POL_STATE_INVALID,
1164 * For now, this state is used exclusively for policies that register
1165 * externally to NRS core, i.e. ones that do so via
1166 * ptlrpc_nrs_policy_register() and are not part of nrs_pols_builtin;
1167 * it is used to prevent a race condition between the policy registering
1168 * with more than one service partition while service is operational,
1169 * and the user starting the policy via lprocfs.
1171 * \see nrs_pol_make_avail()
1173 NRS_POL_STATE_UNAVAIL,
1175 * Policies are at this state either at the start of their life, or
1176 * transition here when the user selects a different policy to act
1177 * as the primary one.
1179 NRS_POL_STATE_STOPPED,
1181 * Policy is progress of stopping
1183 NRS_POL_STATE_STOPPING,
1185 * Policy is in progress of starting
1187 NRS_POL_STATE_STARTING,
1189 * A policy is in this state in two cases:
1190 * - it is the fallback policy, which is always in this state.
1191 * - it has been activated by the user; i.e. it is the primary policy,
1193 NRS_POL_STATE_STARTED,
1197 * NRS policy information
1199 * Used for obtaining information for the status of a policy via lprocfs
1201 struct ptlrpc_nrs_pol_info {
1205 char pi_name[NRS_POL_NAME_MAX];
1207 * Current policy state
1209 enum ptlrpc_nrs_pol_state pi_state;
1211 * # RPCs enqueued for later dispatching by the policy
1215 * # RPCs started for dispatch by the policy
1217 long pi_req_started;
1219 * Is this a fallback policy?
1221 unsigned pi_fallback:1;
1227 * There is one instance of this for each policy in each NRS head of each
1228 * PTLRPC service partition.
1230 struct ptlrpc_nrs_policy {
1232 * Linkage into the NRS head's list of policies,
1233 * ptlrpc_nrs:nrs_policy_list
1235 cfs_list_t pol_list;
1237 * Linkage into the NRS head's list of policies with enqueued
1238 * requests ptlrpc_nrs:nrs_policy_queued
1240 cfs_list_t pol_list_queued;
1242 * Current state of this policy
1244 enum ptlrpc_nrs_pol_state pol_state;
1246 * Bitmask of nrs_policy_flags
1250 * # RPCs enqueued for later dispatching by the policy
1252 long pol_req_queued;
1254 * # RPCs started for dispatch by the policy
1256 long pol_req_started;
1258 * Usage Reference count taken on the policy instance
1262 * The NRS head this policy has been created at
1264 struct ptlrpc_nrs *pol_nrs;
1266 * NRS operations for this policy; points to ptlrpc_nrs_pol_desc::pd_ops
1268 struct ptlrpc_nrs_pol_ops *pol_ops;
1270 * Private policy data; varies by policy type
1274 * Human-readable policy name; point to ptlrpc_nrs_pol_desc::pd_name
1282 * Resources are embedded into two types of NRS entities:
1283 * - Inside NRS policies, in the policy's private data in
1284 * ptlrpc_nrs_policy::pol_private
1285 * - In objects that act as prime-level scheduling entities in different NRS
1286 * policies; e.g. on a policy that performs round robin or similar order
1287 * scheduling across client NIDs, there would be one NRS resource per unique
1288 * client NID. On a policy which performs round robin scheduling across
1289 * backend filesystem objects, there would be one resource associated with
1290 * each of the backend filesystem objects partaking in the scheduling
1291 * performed by the policy.
1293 * NRS resources share a parent-child relationship, in which resources embedded
1294 * in policy instances are the parent entities, with all scheduling entities
1295 * a policy schedules across being the children, thus forming a simple resource
1296 * hierarchy. This hierarchy may be extended with one or more levels in the
1297 * future if the ability to have more than one primary policy is added.
1299 * Upon request initialization, references to the then active NRS policies are
1300 * taken and used to later handle the dispatching of the request with one of
1303 * \see nrs_resource_get_safe()
1304 * \see ptlrpc_nrs_req_add()
1306 struct ptlrpc_nrs_resource {
1308 * This NRS resource's parent; is NULL for resources embedded in NRS
1309 * policy instances; i.e. those are top-level ones.
1311 struct ptlrpc_nrs_resource *res_parent;
1313 * The policy associated with this resource.
1315 struct ptlrpc_nrs_policy *res_policy;
1328 * This policy is a logical wrapper around previous, non-NRS functionality.
1329 * It dispatches RPCs in the same order as they arrive from the network. This
1330 * policy is currently used as the fallback policy, and the only enabled policy
1331 * on all NRS heads of all PTLRPC service partitions.
1336 * Private data structure for the FIFO policy
1338 struct nrs_fifo_head {
1340 * Resource object for policy instance.
1342 struct ptlrpc_nrs_resource fh_res;
1344 * List of queued requests.
1348 * For debugging purposes.
1353 struct nrs_fifo_req {
1354 /** request header, must be the first member of structure */
1364 * Instances of this object exist embedded within ptlrpc_request; the main
1365 * purpose of this object is to hold references to the request's resources
1366 * for the lifetime of the request, and to hold properties that policies use
1367 * use for determining the request's scheduling priority.
1369 struct ptlrpc_nrs_request {
1371 * The request's resource hierarchy.
1373 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1375 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1376 * policy that was used to enqueue the request.
1378 * \see nrs_request_enqueue()
1380 unsigned nr_res_idx;
1381 unsigned nr_initialized:1;
1382 unsigned nr_enqueued:1;
1383 unsigned nr_dequeued:1;
1384 unsigned nr_started:1;
1385 unsigned nr_finalized:1;
1386 cfs_binheap_node_t nr_node;
1389 * Policy-specific fields, used for determining a request's scheduling
1390 * priority, and other supporting functionality.
1394 * Fields for the FIFO policy
1396 struct nrs_fifo_req fifo;
1399 * Externally-registering policies may want to use this to allocate
1400 * their own request properties.
1408 * Basic request prioritization operations structure.
1409 * The whole idea is centered around locks and RPCs that might affect locks.
1410 * When a lock is contended we try to give priority to RPCs that might lead
1411 * to fastest release of that lock.
1412 * Currently only implemented for OSTs only in a way that makes all
1413 * IO and truncate RPCs that are coming from a locked region where a lock is
1414 * contended a priority over other requests.
1416 struct ptlrpc_hpreq_ops {
1418 * Check if the lock handle of the given lock is the same as
1419 * taken from the request.
1421 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1423 * Check if the request is a high priority one.
1425 int (*hpreq_check)(struct ptlrpc_request *);
1427 * Called after the request has been handled.
1429 void (*hpreq_fini)(struct ptlrpc_request *);
1433 * Represents remote procedure call.
1435 * This is a staple structure used by everybody wanting to send a request
1438 struct ptlrpc_request {
1439 /* Request type: one of PTL_RPC_MSG_* */
1441 /** Result of request processing */
1444 * Linkage item through which this request is included into
1445 * sending/delayed lists on client and into rqbd list on server
1449 * Server side list of incoming unserved requests sorted by arrival
1450 * time. Traversed from time to time to notice about to expire
1451 * requests and sent back "early replies" to clients to let them
1452 * know server is alive and well, just very busy to service their
1455 cfs_list_t rq_timed_list;
1456 /** server-side history, used for debuging purposes. */
1457 cfs_list_t rq_history_list;
1458 /** server-side per-export list */
1459 cfs_list_t rq_exp_list;
1460 /** server-side hp handlers */
1461 struct ptlrpc_hpreq_ops *rq_ops;
1463 /** initial thread servicing this request */
1464 struct ptlrpc_thread *rq_svc_thread;
1466 /** history sequence # */
1467 __u64 rq_history_seq;
1471 /** stub for NRS request */
1472 struct ptlrpc_nrs_request rq_nrq;
1474 /** the index of service's srv_at_array into which request is linked */
1476 /** Lock to protect request flags and some other important bits, like
1480 /** client-side flags are serialized by rq_lock */
1481 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1482 rq_timedout:1, rq_resend:1, rq_restart:1,
1484 * when ->rq_replay is set, request is kept by the client even
1485 * after server commits corresponding transaction. This is
1486 * used for operations that require sequence of multiple
1487 * requests to be replayed. The only example currently is file
1488 * open/close. When last request in such a sequence is
1489 * committed, ->rq_replay is cleared on all requests in the
1493 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1494 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1495 rq_early:1, rq_must_unlink:1,
1496 rq_memalloc:1, /* req originated from "kswapd" */
1497 /* server-side flags */
1498 rq_packed_final:1, /* packed final reply */
1499 rq_hp:1, /* high priority RPC */
1500 rq_at_linked:1, /* link into service's srv_at_array */
1501 rq_reply_truncate:1,
1503 /* whether the "rq_set" is a valid one */
1505 rq_generation_set:1,
1506 /* do not resend request on -EINPROGRESS */
1507 rq_no_retry_einprogress:1,
1508 /* allow the req to be sent if the import is in recovery
1512 unsigned int rq_nr_resend;
1514 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1515 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1516 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
1517 server-side refcounf for multiple replies */
1519 /** Portal to which this request would be sent */
1520 short rq_request_portal; /* XXX FIXME bug 249 */
1521 /** Portal where to wait for reply and where reply would be sent */
1522 short rq_reply_portal; /* XXX FIXME bug 249 */
1526 * !rq_truncate : # reply bytes actually received,
1527 * rq_truncate : required repbuf_len for resend
1529 int rq_nob_received;
1530 /** Request length */
1534 /** Request message - what client sent */
1535 struct lustre_msg *rq_reqmsg;
1536 /** Reply message - server response */
1537 struct lustre_msg *rq_repmsg;
1538 /** Transaction number */
1543 * List item to for replay list. Not yet commited requests get linked
1545 * Also see \a rq_replay comment above.
1547 cfs_list_t rq_replay_list;
1550 * security and encryption data
1552 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1553 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1554 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
1556 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1557 enum lustre_sec_part rq_sp_from;
1559 /* client/server security flags */
1561 rq_ctx_init:1, /* context initiation */
1562 rq_ctx_fini:1, /* context destroy */
1563 rq_bulk_read:1, /* request bulk read */
1564 rq_bulk_write:1, /* request bulk write */
1565 /* server authentication flags */
1566 rq_auth_gss:1, /* authenticated by gss */
1567 rq_auth_remote:1, /* authed as remote user */
1568 rq_auth_usr_root:1, /* authed as root */
1569 rq_auth_usr_mdt:1, /* authed as mdt */
1570 rq_auth_usr_ost:1, /* authed as ost */
1571 /* security tfm flags */
1574 /* doesn't expect reply FIXME */
1576 rq_pill_init:1; /* pill initialized */
1578 uid_t rq_auth_uid; /* authed uid */
1579 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1581 /* (server side), pointed directly into req buffer */
1582 struct ptlrpc_user_desc *rq_user_desc;
1584 /* various buffer pointers */
1585 struct lustre_msg *rq_reqbuf; /* req wrapper */
1586 char *rq_repbuf; /* rep buffer */
1587 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1588 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1589 int rq_reqbuf_len; /* req wrapper buf len */
1590 int rq_reqdata_len; /* req wrapper msg len */
1591 int rq_repbuf_len; /* rep buffer len */
1592 int rq_repdata_len; /* rep wrapper msg len */
1593 int rq_clrbuf_len; /* only in priv mode */
1594 int rq_clrdata_len; /* only in priv mode */
1596 /** early replies go to offset 0, regular replies go after that */
1597 unsigned int rq_reply_off;
1601 /** Fields that help to see if request and reply were swabbed or not */
1602 __u32 rq_req_swab_mask;
1603 __u32 rq_rep_swab_mask;
1605 /** What was import generation when this request was sent */
1606 int rq_import_generation;
1607 enum lustre_imp_state rq_send_state;
1609 /** how many early replies (for stats) */
1612 /** client+server request */
1613 lnet_handle_md_t rq_req_md_h;
1614 struct ptlrpc_cb_id rq_req_cbid;
1615 /** optional time limit for send attempts */
1616 cfs_duration_t rq_delay_limit;
1617 /** time request was first queued */
1618 cfs_time_t rq_queued_time;
1620 /* server-side... */
1621 /** request arrival time */
1622 struct timeval rq_arrival_time;
1623 /** separated reply state */
1624 struct ptlrpc_reply_state *rq_reply_state;
1625 /** incoming request buffer */
1626 struct ptlrpc_request_buffer_desc *rq_rqbd;
1628 /** client-only incoming reply */
1629 lnet_handle_md_t rq_reply_md_h;
1630 cfs_waitq_t rq_reply_waitq;
1631 struct ptlrpc_cb_id rq_reply_cbid;
1635 /** Peer description (the other side) */
1636 lnet_process_id_t rq_peer;
1637 /** Server-side, export on which request was received */
1638 struct obd_export *rq_export;
1639 /** Client side, import where request is being sent */
1640 struct obd_import *rq_import;
1642 /** Replay callback, called after request is replayed at recovery */
1643 void (*rq_replay_cb)(struct ptlrpc_request *);
1645 * Commit callback, called when request is committed and about to be
1648 void (*rq_commit_cb)(struct ptlrpc_request *);
1649 /** Opaq data for replay and commit callbacks. */
1652 /** For bulk requests on client only: bulk descriptor */
1653 struct ptlrpc_bulk_desc *rq_bulk;
1655 /** client outgoing req */
1657 * when request/reply sent (secs), or time when request should be sent
1660 /** time for request really sent out */
1661 time_t rq_real_sent;
1663 /** when request must finish. volatile
1664 * so that servers' early reply updates to the deadline aren't
1665 * kept in per-cpu cache */
1666 volatile time_t rq_deadline;
1667 /** when req reply unlink must finish. */
1668 time_t rq_reply_deadline;
1669 /** when req bulk unlink must finish. */
1670 time_t rq_bulk_deadline;
1672 * service time estimate (secs)
1673 * If the requestsis not served by this time, it is marked as timed out.
1677 /** Multi-rpc bits */
1678 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1679 cfs_waitq_t rq_set_waitq;
1680 /** Link item for request set lists */
1681 cfs_list_t rq_set_chain;
1682 /** Link back to the request set */
1683 struct ptlrpc_request_set *rq_set;
1684 /** Async completion handler, called when reply is received */
1685 ptlrpc_interpterer_t rq_interpret_reply;
1686 /** Async completion context */
1687 union ptlrpc_async_args rq_async_args;
1689 /** Pool if request is from preallocated list */
1690 struct ptlrpc_request_pool *rq_pool;
1692 struct lu_context rq_session;
1693 struct lu_context rq_recov_session;
1695 /** request format description */
1696 struct req_capsule rq_pill;
1700 * Call completion handler for rpc if any, return it's status or original
1701 * rc if there was no handler defined for this request.
1703 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1704 struct ptlrpc_request *req, int rc)
1706 if (req->rq_interpret_reply != NULL) {
1707 req->rq_status = req->rq_interpret_reply(env, req,
1708 &req->rq_async_args,
1710 return req->rq_status;
1718 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_desc *desc);
1719 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_desc *desc);
1720 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1721 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1722 struct ptlrpc_nrs_pol_info *info);
1725 * Can the request be moved from the regular NRS head to the high-priority NRS
1726 * head (of the same PTLRPC service partition), if any?
1728 * For a reliable result, this should be checked under svcpt->scp_req lock.
1731 ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1733 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1736 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1737 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1738 * to make sure it has not been scheduled yet (analogous to previous
1739 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1741 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1746 * Returns 1 if request buffer at offset \a index was already swabbed
1748 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1750 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1751 return req->rq_req_swab_mask & (1 << index);
1755 * Returns 1 if request reply buffer at offset \a index was already swabbed
1757 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1759 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1760 return req->rq_rep_swab_mask & (1 << index);
1764 * Returns 1 if request needs to be swabbed into local cpu byteorder
1766 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1768 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1772 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1774 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1776 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1780 * Mark request buffer at offset \a index that it was already swabbed
1782 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1784 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1785 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1786 req->rq_req_swab_mask |= 1 << index;
1790 * Mark request reply buffer at offset \a index that it was already swabbed
1792 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1794 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1795 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1796 req->rq_rep_swab_mask |= 1 << index;
1800 * Convert numerical request phase value \a phase into text string description
1802 static inline const char *
1803 ptlrpc_phase2str(enum rq_phase phase)
1812 case RQ_PHASE_INTERPRET:
1814 case RQ_PHASE_COMPLETE:
1816 case RQ_PHASE_UNREGISTERING:
1817 return "Unregistering";
1824 * Convert numerical request phase of the request \a req into text stringi
1827 static inline const char *
1828 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1830 return ptlrpc_phase2str(req->rq_phase);
1834 * Debugging functions and helpers to print request structure into debug log
1837 /* Spare the preprocessor, spoil the bugs. */
1838 #define FLAG(field, str) (field ? str : "")
1840 /** Convert bit flags into a string */
1841 #define DEBUG_REQ_FLAGS(req) \
1842 ptlrpc_rqphase2str(req), \
1843 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1844 FLAG(req->rq_err, "E"), \
1845 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1846 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1847 FLAG(req->rq_no_resend, "N"), \
1848 FLAG(req->rq_waiting, "W"), \
1849 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1850 FLAG(req->rq_committed, "M")
1852 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1854 void _debug_req(struct ptlrpc_request *req,
1855 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1856 __attribute__ ((format (printf, 3, 4)));
1859 * Helper that decides if we need to print request accordig to current debug
1862 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1864 CFS_CHECK_STACK(msgdata, mask, cdls); \
1866 if (((mask) & D_CANTMASK) != 0 || \
1867 ((libcfs_debug & (mask)) != 0 && \
1868 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1869 _debug_req((req), msgdata, fmt, ##a); \
1873 * This is the debug print function you need to use to print request sturucture
1874 * content into lustre debug log.
1875 * for most callers (level is a constant) this is resolved at compile time */
1876 #define DEBUG_REQ(level, req, fmt, args...) \
1878 if ((level) & (D_ERROR | D_WARNING)) { \
1879 static cfs_debug_limit_state_t cdls; \
1880 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1881 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1883 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1884 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1890 * Structure that defines a single page of a bulk transfer
1892 struct ptlrpc_bulk_page {
1893 /** Linkage to list of pages in a bulk */
1896 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1899 /** offset within a page */
1901 /** The page itself */
1902 struct page *bp_page;
1905 #define BULK_GET_SOURCE 0
1906 #define BULK_PUT_SINK 1
1907 #define BULK_GET_SINK 2
1908 #define BULK_PUT_SOURCE 3
1911 * Definition of bulk descriptor.
1912 * Bulks are special "Two phase" RPCs where initial request message
1913 * is sent first and it is followed bt a transfer (o receiving) of a large
1914 * amount of data to be settled into pages referenced from the bulk descriptors.
1915 * Bulks transfers (the actual data following the small requests) are done
1916 * on separate LNet portals.
1917 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1918 * Another user is readpage for MDT.
1920 struct ptlrpc_bulk_desc {
1921 /** completed with failure */
1922 unsigned long bd_failure:1;
1923 /** {put,get}{source,sink} */
1924 unsigned long bd_type:2;
1926 unsigned long bd_registered:1;
1927 /** For serialization with callback */
1929 /** Import generation when request for this bulk was sent */
1930 int bd_import_generation;
1931 /** LNet portal for this bulk */
1933 /** Server side - export this bulk created for */
1934 struct obd_export *bd_export;
1935 /** Client side - import this bulk was sent on */
1936 struct obd_import *bd_import;
1937 /** Back pointer to the request */
1938 struct ptlrpc_request *bd_req;
1939 cfs_waitq_t bd_waitq; /* server side only WQ */
1940 int bd_iov_count; /* # entries in bd_iov */
1941 int bd_max_iov; /* allocated size of bd_iov */
1942 int bd_nob; /* # bytes covered */
1943 int bd_nob_transferred; /* # bytes GOT/PUT */
1947 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1948 lnet_nid_t bd_sender; /* stash event::sender */
1949 int bd_md_count; /* # valid entries in bd_mds */
1950 int bd_md_max_brw; /* max entries in bd_mds */
1951 /** array of associated MDs */
1952 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
1954 #if defined(__KERNEL__)
1956 * encrypt iov, size is either 0 or bd_iov_count.
1958 lnet_kiov_t *bd_enc_iov;
1960 lnet_kiov_t bd_iov[0];
1962 lnet_md_iovec_t bd_iov[0];
1967 SVC_STOPPED = 1 << 0,
1968 SVC_STOPPING = 1 << 1,
1969 SVC_STARTING = 1 << 2,
1970 SVC_RUNNING = 1 << 3,
1972 SVC_SIGNAL = 1 << 5,
1975 #define PTLRPC_THR_NAME_LEN 32
1977 * Definition of server service thread structure
1979 struct ptlrpc_thread {
1981 * List of active threads in svc->srv_threads
1985 * thread-private data (preallocated memory)
1990 * service thread index, from ptlrpc_start_threads
1994 * service thread pid
1998 * put watchdog in the structure per thread b=14840
2000 struct lc_watchdog *t_watchdog;
2002 * the svc this thread belonged to b=18582
2004 struct ptlrpc_service_part *t_svcpt;
2005 cfs_waitq_t t_ctl_waitq;
2006 struct lu_env *t_env;
2007 char t_name[PTLRPC_THR_NAME_LEN];
2010 static inline int thread_is_init(struct ptlrpc_thread *thread)
2012 return thread->t_flags == 0;
2015 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2017 return !!(thread->t_flags & SVC_STOPPED);
2020 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2022 return !!(thread->t_flags & SVC_STOPPING);
2025 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2027 return !!(thread->t_flags & SVC_STARTING);
2030 static inline int thread_is_running(struct ptlrpc_thread *thread)
2032 return !!(thread->t_flags & SVC_RUNNING);
2035 static inline int thread_is_event(struct ptlrpc_thread *thread)
2037 return !!(thread->t_flags & SVC_EVENT);
2040 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2042 return !!(thread->t_flags & SVC_SIGNAL);
2045 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2047 thread->t_flags &= ~flags;
2050 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2052 thread->t_flags = flags;
2055 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2057 thread->t_flags |= flags;
2060 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2063 if (thread->t_flags & flags) {
2064 thread->t_flags &= ~flags;
2071 * Request buffer descriptor structure.
2072 * This is a structure that contains one posted request buffer for service.
2073 * Once data land into a buffer, event callback creates actual request and
2074 * notifies wakes one of the service threads to process new incoming request.
2075 * More than one request can fit into the buffer.
2077 struct ptlrpc_request_buffer_desc {
2078 /** Link item for rqbds on a service */
2079 cfs_list_t rqbd_list;
2080 /** History of requests for this buffer */
2081 cfs_list_t rqbd_reqs;
2082 /** Back pointer to service for which this buffer is registered */
2083 struct ptlrpc_service_part *rqbd_svcpt;
2084 /** LNet descriptor */
2085 lnet_handle_md_t rqbd_md_h;
2087 /** The buffer itself */
2089 struct ptlrpc_cb_id rqbd_cbid;
2091 * This "embedded" request structure is only used for the
2092 * last request to fit into the buffer
2094 struct ptlrpc_request rqbd_req;
2097 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2099 struct ptlrpc_service_ops {
2101 * if non-NULL called during thread creation (ptlrpc_start_thread())
2102 * to initialize service specific per-thread state.
2104 int (*so_thr_init)(struct ptlrpc_thread *thr);
2106 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2107 * destruct state created by ->srv_init().
2109 void (*so_thr_done)(struct ptlrpc_thread *thr);
2111 * Handler function for incoming requests for this service
2113 int (*so_req_handler)(struct ptlrpc_request *req);
2115 * function to determine priority of the request, it's called
2116 * on every new request
2118 int (*so_hpreq_handler)(struct ptlrpc_request *);
2120 * service-specific print fn
2122 void (*so_req_printer)(void *, struct ptlrpc_request *);
2125 #ifndef __cfs_cacheline_aligned
2126 /* NB: put it here for reducing patche dependence */
2127 # define __cfs_cacheline_aligned
2131 * How many high priority requests to serve before serving one normal
2134 #define PTLRPC_SVC_HP_RATIO 10
2137 * Definition of PortalRPC service.
2138 * The service is listening on a particular portal (like tcp port)
2139 * and perform actions for a specific server like IO service for OST
2140 * or general metadata service for MDS.
2142 struct ptlrpc_service {
2143 /** serialize /proc operations */
2144 spinlock_t srv_lock;
2145 /** most often accessed fields */
2146 /** chain thru all services */
2147 cfs_list_t srv_list;
2148 /** service operations table */
2149 struct ptlrpc_service_ops srv_ops;
2150 /** only statically allocated strings here; we don't clean them */
2152 /** only statically allocated strings here; we don't clean them */
2153 char *srv_thread_name;
2154 /** service thread list */
2155 cfs_list_t srv_threads;
2156 /** threads # should be created for each partition on initializing */
2157 int srv_nthrs_cpt_init;
2158 /** limit of threads number for each partition */
2159 int srv_nthrs_cpt_limit;
2160 /** Root of /proc dir tree for this service */
2161 cfs_proc_dir_entry_t *srv_procroot;
2162 /** Pointer to statistic data for this service */
2163 struct lprocfs_stats *srv_stats;
2164 /** # hp per lp reqs to handle */
2165 int srv_hpreq_ratio;
2166 /** biggest request to receive */
2167 int srv_max_req_size;
2168 /** biggest reply to send */
2169 int srv_max_reply_size;
2170 /** size of individual buffers */
2172 /** # buffers to allocate in 1 group */
2173 int srv_nbuf_per_group;
2174 /** Local portal on which to receive requests */
2175 __u32 srv_req_portal;
2176 /** Portal on the client to send replies to */
2177 __u32 srv_rep_portal;
2179 * Tags for lu_context associated with this thread, see struct
2183 /** soft watchdog timeout multiplier */
2184 int srv_watchdog_factor;
2185 /** under unregister_service */
2186 unsigned srv_is_stopping:1;
2188 /** max # request buffers in history per partition */
2189 int srv_hist_nrqbds_cpt_max;
2190 /** number of CPTs this service bound on */
2192 /** CPTs array this service bound on */
2194 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2196 /** CPT table this service is running over */
2197 struct cfs_cpt_table *srv_cptable;
2199 * partition data for ptlrpc service
2201 struct ptlrpc_service_part *srv_parts[0];
2205 * Definition of PortalRPC service partition data.
2206 * Although a service only has one instance of it right now, but we
2207 * will have multiple instances very soon (instance per CPT).
2209 * it has four locks:
2211 * serialize operations on rqbd and requests waiting for preprocess
2213 * serialize operations active requests sent to this portal
2215 * serialize adaptive timeout stuff
2217 * serialize operations on RS list (reply states)
2219 * We don't have any use-case to take two or more locks at the same time
2220 * for now, so there is no lock order issue.
2222 struct ptlrpc_service_part {
2223 /** back reference to owner */
2224 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2225 /* CPT id, reserved */
2227 /** always increasing number */
2229 /** # of starting threads */
2230 int scp_nthrs_starting;
2231 /** # of stopping threads, reserved for shrinking threads */
2232 int scp_nthrs_stopping;
2233 /** # running threads */
2234 int scp_nthrs_running;
2235 /** service threads list */
2236 cfs_list_t scp_threads;
2239 * serialize the following fields, used for protecting
2240 * rqbd list and incoming requests waiting for preprocess,
2241 * threads starting & stopping are also protected by this lock.
2243 spinlock_t scp_lock __cfs_cacheline_aligned;
2244 /** total # req buffer descs allocated */
2245 int scp_nrqbds_total;
2246 /** # posted request buffers for receiving */
2247 int scp_nrqbds_posted;
2248 /** in progress of allocating rqbd */
2249 int scp_rqbd_allocating;
2250 /** # incoming reqs */
2251 int scp_nreqs_incoming;
2252 /** request buffers to be reposted */
2253 cfs_list_t scp_rqbd_idle;
2254 /** req buffers receiving */
2255 cfs_list_t scp_rqbd_posted;
2256 /** incoming reqs */
2257 cfs_list_t scp_req_incoming;
2258 /** timeout before re-posting reqs, in tick */
2259 cfs_duration_t scp_rqbd_timeout;
2261 * all threads sleep on this. This wait-queue is signalled when new
2262 * incoming request arrives and when difficult reply has to be handled.
2264 cfs_waitq_t scp_waitq;
2266 /** request history */
2267 cfs_list_t scp_hist_reqs;
2268 /** request buffer history */
2269 cfs_list_t scp_hist_rqbds;
2270 /** # request buffers in history */
2271 int scp_hist_nrqbds;
2272 /** sequence number for request */
2274 /** highest seq culled from history */
2275 __u64 scp_hist_seq_culled;
2278 * serialize the following fields, used for processing requests
2279 * sent to this portal
2281 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2282 /** # reqs in either of the NRS heads below */
2283 /** # reqs being served */
2284 int scp_nreqs_active;
2285 /** # HPreqs being served */
2286 int scp_nhreqs_active;
2287 /** # hp requests handled */
2290 /** NRS head for regular requests */
2291 struct ptlrpc_nrs scp_nrs_reg;
2292 /** NRS head for HP requests; this is only valid for services that can
2293 * handle HP requests */
2294 struct ptlrpc_nrs *scp_nrs_hp;
2299 * serialize the following fields, used for changes on
2302 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2303 /** estimated rpc service time */
2304 struct adaptive_timeout scp_at_estimate;
2305 /** reqs waiting for replies */
2306 struct ptlrpc_at_array scp_at_array;
2307 /** early reply timer */
2308 cfs_timer_t scp_at_timer;
2310 cfs_time_t scp_at_checktime;
2311 /** check early replies */
2312 unsigned scp_at_check;
2316 * serialize the following fields, used for processing
2317 * replies for this portal
2319 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2320 /** all the active replies */
2321 cfs_list_t scp_rep_active;
2323 /** replies waiting for service */
2324 cfs_list_t scp_rep_queue;
2326 /** List of free reply_states */
2327 cfs_list_t scp_rep_idle;
2328 /** waitq to run, when adding stuff to srv_free_rs_list */
2329 cfs_waitq_t scp_rep_waitq;
2330 /** # 'difficult' replies */
2331 cfs_atomic_t scp_nreps_difficult;
2334 #define ptlrpc_service_for_each_part(part, i, svc) \
2336 i < (svc)->srv_ncpts && \
2337 (svc)->srv_parts != NULL && \
2338 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2341 * Declaration of ptlrpcd control structure
2343 struct ptlrpcd_ctl {
2345 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2347 unsigned long pc_flags;
2349 * Thread lock protecting structure fields.
2355 struct completion pc_starting;
2359 struct completion pc_finishing;
2361 * Thread requests set.
2363 struct ptlrpc_request_set *pc_set;
2365 * Thread name used in cfs_daemonize()
2369 * Environment for request interpreters to run in.
2371 struct lu_env pc_env;
2373 * Index of ptlrpcd thread in the array.
2377 * Number of the ptlrpcd's partners.
2381 * Pointer to the array of partners' ptlrpcd_ctl structure.
2383 struct ptlrpcd_ctl **pc_partners;
2385 * Record the partner index to be processed next.
2390 * Async rpcs flag to make sure that ptlrpcd_check() is called only
2395 * Currently not used.
2399 * User-space async rpcs callback.
2401 void *pc_wait_callback;
2403 * User-space check idle rpcs callback.
2405 void *pc_idle_callback;
2409 /* Bits for pc_flags */
2410 enum ptlrpcd_ctl_flags {
2412 * Ptlrpc thread start flag.
2414 LIOD_START = 1 << 0,
2416 * Ptlrpc thread stop flag.
2420 * Ptlrpc thread force flag (only stop force so far).
2421 * This will cause aborting any inflight rpcs handled
2422 * by thread if LIOD_STOP is specified.
2424 LIOD_FORCE = 1 << 2,
2426 * This is a recovery ptlrpc thread.
2428 LIOD_RECOVERY = 1 << 3,
2430 * The ptlrpcd is bound to some CPU core.
2439 * Service compatibility function; policy is compatible with all services.
2441 * \param[in] svc The service the policy is attempting to register with.
2442 * \param[in] desc The policy descriptor
2444 * \retval true The policy is compatible with the NRS head
2446 * \see ptlrpc_nrs_pol_desc::pd_compat()
2449 nrs_policy_compat_all(struct ptlrpc_service *svc,
2450 const struct ptlrpc_nrs_pol_desc *desc)
2456 * Service compatibility function; policy is compatible with only a specific
2457 * service which is identified by its human-readable name at
2458 * ptlrpc_service::srv_name.
2460 * \param[in] svc The service the policy is attempting to register with.
2461 * \param[in] desc The policy descriptor
2463 * \retval false The policy is not compatible with the NRS head
2464 * \retval true The policy is compatible with the NRS head
2466 * \see ptlrpc_nrs_pol_desc::pd_compat()
2469 nrs_policy_compat_one(struct ptlrpc_service *svc,
2470 const struct ptlrpc_nrs_pol_desc *desc)
2472 LASSERT(desc->pd_compat_svc_name != NULL);
2473 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2478 /* ptlrpc/events.c */
2479 extern lnet_handle_eq_t ptlrpc_eq_h;
2480 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2481 lnet_process_id_t *peer, lnet_nid_t *self);
2483 * These callbacks are invoked by LNet when something happened to
2487 extern void request_out_callback(lnet_event_t *ev);
2488 extern void reply_in_callback(lnet_event_t *ev);
2489 extern void client_bulk_callback(lnet_event_t *ev);
2490 extern void request_in_callback(lnet_event_t *ev);
2491 extern void reply_out_callback(lnet_event_t *ev);
2492 #ifdef HAVE_SERVER_SUPPORT
2493 extern void server_bulk_callback(lnet_event_t *ev);
2497 /* ptlrpc/connection.c */
2498 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2500 struct obd_uuid *uuid);
2501 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2502 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2503 int ptlrpc_connection_init(void);
2504 void ptlrpc_connection_fini(void);
2505 extern lnet_pid_t ptl_get_pid(void);
2507 /* ptlrpc/niobuf.c */
2509 * Actual interfacing with LNet to put/get/register/unregister stuff
2512 #ifdef HAVE_SERVER_SUPPORT
2513 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2514 unsigned npages, unsigned max_brw,
2515 unsigned type, unsigned portal);
2516 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2517 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2519 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2523 LASSERT(desc != NULL);
2525 spin_lock(&desc->bd_lock);
2526 rc = desc->bd_md_count;
2527 spin_unlock(&desc->bd_lock);
2532 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2533 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2535 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2537 struct ptlrpc_bulk_desc *desc;
2540 LASSERT(req != NULL);
2541 desc = req->rq_bulk;
2543 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2544 req->rq_bulk_deadline > cfs_time_current_sec())
2550 spin_lock(&desc->bd_lock);
2551 rc = desc->bd_md_count;
2552 spin_unlock(&desc->bd_lock);
2556 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2557 #define PTLRPC_REPLY_EARLY 0x02
2558 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2559 int ptlrpc_reply(struct ptlrpc_request *req);
2560 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2561 int ptlrpc_error(struct ptlrpc_request *req);
2562 void ptlrpc_resend_req(struct ptlrpc_request *request);
2563 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2564 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2565 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2568 /* ptlrpc/client.c */
2570 * Client-side portals API. Everything to send requests, receive replies,
2571 * request queues, request management, etc.
2574 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2575 struct ptlrpc_client *);
2576 void ptlrpc_cleanup_client(struct obd_import *imp);
2577 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2579 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2580 int ptlrpc_replay_req(struct ptlrpc_request *req);
2581 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2582 void ptlrpc_restart_req(struct ptlrpc_request *req);
2583 void ptlrpc_abort_inflight(struct obd_import *imp);
2584 void ptlrpc_cleanup_imp(struct obd_import *imp);
2585 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2587 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2588 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2590 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2591 set_interpreter_func fn, void *data);
2592 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2593 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2594 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2595 int ptlrpc_expired_set(void *data);
2596 void ptlrpc_interrupted_set(void *data);
2597 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2598 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2599 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2600 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2601 struct ptlrpc_request *req);
2603 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2604 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2606 struct ptlrpc_request_pool *
2607 ptlrpc_init_rq_pool(int, int,
2608 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2610 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2611 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2612 const struct req_format *format);
2613 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2614 struct ptlrpc_request_pool *,
2615 const struct req_format *format);
2616 void ptlrpc_request_free(struct ptlrpc_request *request);
2617 int ptlrpc_request_pack(struct ptlrpc_request *request,
2618 __u32 version, int opcode);
2619 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2620 const struct req_format *format,
2621 __u32 version, int opcode);
2622 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2623 __u32 version, int opcode, char **bufs,
2624 struct ptlrpc_cli_ctx *ctx);
2625 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2626 int opcode, int count, __u32 *lengths,
2628 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2629 __u32 version, int opcode,
2630 int count, __u32 *lengths, char **bufs,
2631 struct ptlrpc_request_pool *pool);
2632 void ptlrpc_req_finished(struct ptlrpc_request *request);
2633 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2634 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2635 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2636 unsigned npages, unsigned max_brw,
2637 unsigned type, unsigned portal);
2638 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2639 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2641 __ptlrpc_free_bulk(bulk, 1);
2643 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2645 __ptlrpc_free_bulk(bulk, 0);
2647 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2648 cfs_page_t *page, int pageoffset, int len, int);
2649 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2650 cfs_page_t *page, int pageoffset,
2653 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2656 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2657 cfs_page_t *page, int pageoffset,
2660 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2663 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2664 struct obd_import *imp);
2665 __u64 ptlrpc_next_xid(void);
2666 __u64 ptlrpc_sample_next_xid(void);
2667 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2669 /* Set of routines to run a function in ptlrpcd context */
2670 void *ptlrpcd_alloc_work(struct obd_import *imp,
2671 int (*cb)(const struct lu_env *, void *), void *data);
2672 void ptlrpcd_destroy_work(void *handler);
2673 int ptlrpcd_queue_work(void *handler);
2676 struct ptlrpc_service_buf_conf {
2677 /* nbufs is buffers # to allocate when growing the pool */
2678 unsigned int bc_nbufs;
2679 /* buffer size to post */
2680 unsigned int bc_buf_size;
2681 /* portal to listed for requests on */
2682 unsigned int bc_req_portal;
2683 /* portal of where to send replies to */
2684 unsigned int bc_rep_portal;
2685 /* maximum request size to be accepted for this service */
2686 unsigned int bc_req_max_size;
2687 /* maximum reply size this service can ever send */
2688 unsigned int bc_rep_max_size;
2691 struct ptlrpc_service_thr_conf {
2692 /* threadname should be 8 characters or less - 6 will be added on */
2694 /* threads increasing factor for each CPU */
2695 unsigned int tc_thr_factor;
2696 /* service threads # to start on each partition while initializing */
2697 unsigned int tc_nthrs_init;
2699 * low water of threads # upper-limit on each partition while running,
2700 * service availability may be impacted if threads number is lower
2701 * than this value. It can be ZERO if the service doesn't require
2702 * CPU affinity or there is only one partition.
2704 unsigned int tc_nthrs_base;
2705 /* "soft" limit for total threads number */
2706 unsigned int tc_nthrs_max;
2707 /* user specified threads number, it will be validated due to
2708 * other members of this structure. */
2709 unsigned int tc_nthrs_user;
2710 /* set NUMA node affinity for service threads */
2711 unsigned int tc_cpu_affinity;
2712 /* Tags for lu_context associated with service thread */
2716 struct ptlrpc_service_cpt_conf {
2717 struct cfs_cpt_table *cc_cptable;
2718 /* string pattern to describe CPTs for a service */
2722 struct ptlrpc_service_conf {
2725 /* soft watchdog timeout multiplifier to print stuck service traces */
2726 unsigned int psc_watchdog_factor;
2727 /* buffer information */
2728 struct ptlrpc_service_buf_conf psc_buf;
2729 /* thread information */
2730 struct ptlrpc_service_thr_conf psc_thr;
2731 /* CPU partition information */
2732 struct ptlrpc_service_cpt_conf psc_cpt;
2733 /* function table */
2734 struct ptlrpc_service_ops psc_ops;
2737 /* ptlrpc/service.c */
2739 * Server-side services API. Register/unregister service, request state
2740 * management, service thread management
2744 void ptlrpc_save_lock(struct ptlrpc_request *req,
2745 struct lustre_handle *lock, int mode, int no_ack);
2746 void ptlrpc_commit_replies(struct obd_export *exp);
2747 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2748 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2749 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2750 struct ptlrpc_service *ptlrpc_register_service(
2751 struct ptlrpc_service_conf *conf,
2752 struct proc_dir_entry *proc_entry);
2753 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2755 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2756 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2757 int liblustre_check_services(void *arg);
2758 void ptlrpc_daemonize(char *name);
2759 int ptlrpc_service_health_check(struct ptlrpc_service *);
2760 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2763 int ptlrpc_hr_init(void);
2764 void ptlrpc_hr_fini(void);
2766 # define ptlrpc_hr_init() (0)
2767 # define ptlrpc_hr_fini() do {} while(0)
2772 /* ptlrpc/import.c */
2777 int ptlrpc_connect_import(struct obd_import *imp);
2778 int ptlrpc_init_import(struct obd_import *imp);
2779 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2780 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2781 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2784 /* ptlrpc/pack_generic.c */
2785 int ptlrpc_reconnect_import(struct obd_import *imp);
2789 * ptlrpc msg buffer and swab interface
2793 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2795 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2797 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2798 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2800 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2801 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2803 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2804 __u32 *lens, char **bufs);
2805 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2807 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2808 __u32 *lens, char **bufs, int flags);
2809 #define LPRFL_EARLY_REPLY 1
2810 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2811 char **bufs, int flags);
2812 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2813 unsigned int newlen, int move_data);
2814 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2815 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2816 int lustre_msg_hdr_size(__u32 magic, int count);
2817 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2818 int lustre_msg_size_v2(int count, __u32 *lengths);
2819 int lustre_packed_msg_size(struct lustre_msg *msg);
2820 int lustre_msg_early_size(void);
2821 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2822 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2823 int lustre_msg_buflen(struct lustre_msg *m, int n);
2824 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2825 int lustre_msg_bufcount(struct lustre_msg *m);
2826 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2827 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2828 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2829 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2830 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2831 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2832 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2833 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2834 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2835 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2836 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2837 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2838 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2839 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2840 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2841 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2842 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2843 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2844 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2845 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2846 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2847 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2848 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2849 int lustre_msg_get_status(struct lustre_msg *msg);
2850 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2851 int lustre_msg_is_v1(struct lustre_msg *msg);
2852 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2853 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2854 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2855 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2856 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2857 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2858 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2860 # warning "remove checksum compatibility support for b1_8"
2861 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2863 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2864 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2865 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2866 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2867 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2868 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2869 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2870 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2871 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2872 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2873 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2874 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2875 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2876 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2877 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2880 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2881 unsigned int newlen, int move_data)
2883 LASSERT(req->rq_reply_state);
2884 LASSERT(req->rq_repmsg);
2885 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2890 /** Change request phase of \a req to \a new_phase */
2892 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2894 if (req->rq_phase == new_phase)
2897 if (new_phase == RQ_PHASE_UNREGISTERING) {
2898 req->rq_next_phase = req->rq_phase;
2900 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2903 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2905 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2908 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2909 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2911 req->rq_phase = new_phase;
2915 * Returns true if request \a req got early reply and hard deadline is not met
2918 ptlrpc_client_early(struct ptlrpc_request *req)
2920 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2921 req->rq_reply_deadline > cfs_time_current_sec())
2923 return req->rq_early;
2927 * Returns true if we got real reply from server for this request
2930 ptlrpc_client_replied(struct ptlrpc_request *req)
2932 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2933 req->rq_reply_deadline > cfs_time_current_sec())
2935 return req->rq_replied;
2938 /** Returns true if request \a req is in process of receiving server reply */
2940 ptlrpc_client_recv(struct ptlrpc_request *req)
2942 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2943 req->rq_reply_deadline > cfs_time_current_sec())
2945 return req->rq_receiving_reply;
2949 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2953 spin_lock(&req->rq_lock);
2954 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2955 req->rq_reply_deadline > cfs_time_current_sec()) {
2956 spin_unlock(&req->rq_lock);
2959 rc = req->rq_receiving_reply || req->rq_must_unlink;
2960 spin_unlock(&req->rq_lock);
2965 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2967 if (req->rq_set == NULL)
2968 cfs_waitq_signal(&req->rq_reply_waitq);
2970 cfs_waitq_signal(&req->rq_set->set_waitq);
2974 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2976 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2977 cfs_atomic_inc(&rs->rs_refcount);
2981 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2983 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2984 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
2985 lustre_free_reply_state(rs);
2988 /* Should only be called once per req */
2989 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2991 if (req->rq_reply_state == NULL)
2992 return; /* shouldn't occur */
2993 ptlrpc_rs_decref(req->rq_reply_state);
2994 req->rq_reply_state = NULL;
2995 req->rq_repmsg = NULL;
2998 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3000 return lustre_msg_get_magic(req->rq_reqmsg);
3003 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3005 switch (req->rq_reqmsg->lm_magic) {
3006 case LUSTRE_MSG_MAGIC_V2:
3007 return req->rq_reqmsg->lm_repsize;
3009 LASSERTF(0, "incorrect message magic: %08x\n",
3010 req->rq_reqmsg->lm_magic);
3015 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3017 if (req->rq_delay_limit != 0 &&
3018 cfs_time_before(cfs_time_add(req->rq_queued_time,
3019 cfs_time_seconds(req->rq_delay_limit)),
3020 cfs_time_current())) {
3026 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3028 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3029 spin_lock(&req->rq_lock);
3030 req->rq_no_resend = 1;
3031 spin_unlock(&req->rq_lock);
3033 return req->rq_no_resend;
3037 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3039 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3041 return svcpt->scp_service->srv_watchdog_factor *
3042 max_t(int, at, obd_timeout);
3045 static inline struct ptlrpc_service *
3046 ptlrpc_req2svc(struct ptlrpc_request *req)
3048 LASSERT(req->rq_rqbd != NULL);
3049 return req->rq_rqbd->rqbd_svcpt->scp_service;
3052 /* ldlm/ldlm_lib.c */
3054 * Target client logic
3057 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3058 int client_obd_cleanup(struct obd_device *obddev);
3059 int client_connect_import(const struct lu_env *env,
3060 struct obd_export **exp, struct obd_device *obd,
3061 struct obd_uuid *cluuid, struct obd_connect_data *,
3063 int client_disconnect_export(struct obd_export *exp);
3064 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3066 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3067 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3068 struct obd_uuid *uuid);
3069 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3070 void client_destroy_import(struct obd_import *imp);
3073 #ifdef HAVE_SERVER_SUPPORT
3074 int server_disconnect_export(struct obd_export *exp);
3077 /* ptlrpc/pinger.c */
3079 * Pinger API (client side only)
3082 extern int suppress_pings;
3083 enum timeout_event {
3086 struct timeout_item;
3087 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3088 int ptlrpc_pinger_add_import(struct obd_import *imp);
3089 int ptlrpc_pinger_del_import(struct obd_import *imp);
3090 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3091 timeout_cb_t cb, void *data,
3092 cfs_list_t *obd_list);
3093 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
3094 enum timeout_event event);
3095 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3096 int ptlrpc_obd_ping(struct obd_device *obd);
3097 cfs_time_t ptlrpc_suspend_wakeup_time(void);
3099 void ping_evictor_start(void);
3100 void ping_evictor_stop(void);
3102 #define ping_evictor_start() do {} while (0)
3103 #define ping_evictor_stop() do {} while (0)
3105 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
3108 /* ptlrpc daemon bind policy */
3110 /* all ptlrpcd threads are free mode */
3111 PDB_POLICY_NONE = 1,
3112 /* all ptlrpcd threads are bound mode */
3113 PDB_POLICY_FULL = 2,
3114 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3115 PDB_POLICY_PAIR = 3,
3116 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3117 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3118 * If kernel supports NUMA, pthrpcd threads are binded and
3119 * grouped by NUMA node */
3120 PDB_POLICY_NEIGHBOR = 4,
3123 /* ptlrpc daemon load policy
3124 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3125 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3126 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3127 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3128 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3130 /* on the same CPU core as the caller */
3131 PDL_POLICY_SAME = 1,
3132 /* within the same CPU partition, but not the same core as the caller */
3133 PDL_POLICY_LOCAL = 2,
3134 /* round-robin on all CPU cores, but not the same core as the caller */
3135 PDL_POLICY_ROUND = 3,
3136 /* the specified CPU core is preferred, but not enforced */
3137 PDL_POLICY_PREFERRED = 4,
3140 /* ptlrpc/ptlrpcd.c */
3141 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3142 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3143 void ptlrpcd_wake(struct ptlrpc_request *req);
3144 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3145 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3146 int ptlrpcd_addref(void);
3147 void ptlrpcd_decref(void);
3149 /* ptlrpc/lproc_ptlrpc.c */
3151 * procfs output related functions
3154 const char* ll_opcode2str(__u32 opcode);
3156 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3157 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3158 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3160 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3161 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3162 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3166 /* ptlrpc/llog_server.c */
3167 int llog_origin_handle_open(struct ptlrpc_request *req);
3168 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3169 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3170 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3171 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3172 int llog_origin_handle_close(struct ptlrpc_request *req);
3173 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3175 /* ptlrpc/llog_client.c */
3176 extern struct llog_operations llog_client_ops;