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 * Not a valid opcode.
772 PTLRPC_NRS_CTL_INVALID,
774 * Activate the policy.
776 PTLRPC_NRS_CTL_START,
778 * Reserved for multiple primary policies, which may be a possibility
783 * Policies can start using opcodes from this value and onwards for
784 * their own purposes; the assigned value itself is arbitrary.
786 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
790 * NRS policy operations.
792 * These determine the behaviour of a policy, and are called in response to
795 struct ptlrpc_nrs_pol_ops {
797 * Called during policy registration; this operation is optional.
799 * \param[in,out] policy The policy being initialized
801 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
803 * Called during policy unregistration; this operation is optional.
805 * \param[in,out] policy The policy being unregistered/finalized
807 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
809 * Called when activating a policy via lprocfs; policies allocate and
810 * initialize their resources here; this operation is optional.
812 * \param[in,out] policy The policy being started
814 * \see nrs_policy_start_locked()
816 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
818 * Called when deactivating a policy via lprocfs; policies deallocate
819 * their resources here; this operation is optional
821 * \param[in,out] policy The policy being stopped
823 * \see nrs_policy_stop0()
825 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
827 * Used for policy-specific operations; i.e. not generic ones like
828 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
829 * to an ioctl; this operation is optional.
831 * \param[in,out] policy The policy carrying out operation \a opc
832 * \param[in] opc The command operation being carried out
833 * \param[in,out] arg An generic buffer for communication between the
834 * user and the control operation
839 * \see ptlrpc_nrs_policy_control()
841 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
842 enum ptlrpc_nrs_ctl opc, void *arg);
845 * Called when obtaining references to the resources of the resource
846 * hierarchy for a request that has arrived for handling at the PTLRPC
847 * service. Policies should return -ve for requests they do not wish
848 * to handle. This operation is mandatory.
850 * \param[in,out] policy The policy we're getting resources for.
851 * \param[in,out] nrq The request we are getting resources for.
852 * \param[in] parent The parent resource of the resource being
853 * requested; set to NULL if none.
854 * \param[out] resp The resource is to be returned here; the
855 * fallback policy in an NRS head should
856 * \e always return a non-NULL pointer value.
857 * \param[in] moving_req When set, signifies that this is an attempt
858 * to obtain resources for a request being moved
859 * to the high-priority NRS head by
860 * ldlm_lock_reorder_req().
861 * This implies two things:
862 * 1. We are under obd_export::exp_rpc_lock and
863 * so should not sleep.
864 * 2. We should not perform non-idempotent or can
865 * skip performing idempotent operations that
866 * were carried out when resources were first
867 * taken for the request when it was initialized
868 * in ptlrpc_nrs_req_initialize().
870 * \retval 0, +ve The level of the returned resource in the resource
871 * hierarchy; currently only 0 (for a non-leaf resource)
872 * and 1 (for a leaf resource) are supported by the
876 * \see ptlrpc_nrs_req_initialize()
877 * \see ptlrpc_nrs_hpreq_add_nolock()
878 * \see ptlrpc_nrs_req_hp_move()
880 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
881 struct ptlrpc_nrs_request *nrq,
882 const struct ptlrpc_nrs_resource *parent,
883 struct ptlrpc_nrs_resource **resp,
886 * Called when releasing references taken for resources in the resource
887 * hierarchy for the request; this operation is optional.
889 * \param[in,out] policy The policy the resource belongs to
890 * \param[in] res The resource to be freed
892 * \see ptlrpc_nrs_req_finalize()
893 * \see ptlrpc_nrs_hpreq_add_nolock()
894 * \see ptlrpc_nrs_req_hp_move()
896 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
897 const struct ptlrpc_nrs_resource *res);
900 * Obtains a request for handling from the policy, and optionally
901 * removes the request from the policy; this operation is mandatory.
903 * \param[in,out] policy The policy to poll
904 * \param[in] peek When set, signifies that we just want to
905 * examine the request, and not handle it, so the
906 * request is not removed from the policy.
907 * \param[in] force When set, it will force a policy to return a
908 * request if it has one queued.
910 * \retval NULL No request available for handling
911 * \retval valid-pointer The request polled for handling
913 * \see ptlrpc_nrs_req_get_nolock()
915 struct ptlrpc_nrs_request *
916 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
919 * Called when attempting to add a request to a policy for later
920 * handling; this operation is mandatory.
922 * \param[in,out] policy The policy on which to enqueue \a nrq
923 * \param[in,out] nrq The request to enqueue
928 * \see ptlrpc_nrs_req_add_nolock()
930 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
931 struct ptlrpc_nrs_request *nrq);
933 * Removes a request from the policy's set of pending requests. Normally
934 * called after a request has been polled successfully from the policy
935 * for handling; this operation is mandatory.
937 * \param[in,out] policy The policy the request \a nrq belongs to
938 * \param[in,out] nrq The request to dequeue
940 * \see ptlrpc_nrs_req_del_nolock()
942 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
943 struct ptlrpc_nrs_request *nrq);
945 * Called after the request being carried out. Could be used for
946 * job/resource control; this operation is optional.
948 * \param[in,out] policy The policy which is stopping to handle request
950 * \param[in,out] nrq The request
952 * \pre spin_is_locked(&svcpt->scp_req_lock)
954 * \see ptlrpc_nrs_req_stop_nolock()
956 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
957 struct ptlrpc_nrs_request *nrq);
959 * Registers the policy's lprocfs interface with a PTLRPC service.
961 * \param[in] svc The service
966 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
968 * Unegisters the policy's lprocfs interface with a PTLRPC service.
970 * In cases of failed policy registration in
971 * \e ptlrpc_nrs_policy_register(), this function may be called for a
972 * service which has not registered the policy successfully, so
973 * implementations of this method should make sure their operations are
974 * safe in such cases.
976 * \param[in] svc The service
978 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
984 enum nrs_policy_flags {
986 * Fallback policy, use this flag only on a single supported policy per
987 * service. The flag cannot be used on policies that use
988 * \e PTLRPC_NRS_FL_REG_EXTERN
990 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
992 * Start policy immediately after registering.
994 PTLRPC_NRS_FL_REG_START = (1 << 1),
996 * This is a policy registering from a module different to the one NRS
997 * core ships in (currently ptlrpc).
999 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1005 * Denotes whether an NRS instance is for handling normal or high-priority
1006 * RPCs, or whether an operation pertains to one or both of the NRS instances
1009 enum ptlrpc_nrs_queue_type {
1010 PTLRPC_NRS_QUEUE_REG = (1 << 0),
1011 PTLRPC_NRS_QUEUE_HP = (1 << 1),
1012 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
1018 * A PTLRPC service has at least one NRS head instance for handling normal
1019 * priority RPCs, and may optionally have a second NRS head instance for
1020 * handling high-priority RPCs. Each NRS head maintains a list of available
1021 * policies, of which one and only one policy is acting as the fallback policy,
1022 * and optionally a different policy may be acting as the primary policy. For
1023 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1024 * enqueue the RPC using the primary policy (if any). The fallback policy is
1025 * used in the following cases:
1026 * - when there was no primary policy in the
1027 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1029 * - when the primary policy that was at the
1030 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1031 * RPC was initialized, denoted it did not wish, or for some other reason was
1032 * not able to handle the request, by returning a non-valid NRS resource
1034 * - when the primary policy that was at the
1035 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1036 * RPC was initialized, fails later during the request enqueueing stage.
1038 * \see nrs_resource_get_safe()
1039 * \see nrs_request_enqueue()
1042 spinlock_t nrs_lock;
1043 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1045 * List of registered policies
1047 cfs_list_t nrs_policy_list;
1049 * List of policies with queued requests. Policies that have any
1050 * outstanding requests are queued here, and this list is queried
1051 * in a round-robin manner from NRS core when obtaining a request
1052 * for handling. This ensures that requests from policies that at some
1053 * point transition away from the
1054 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1056 cfs_list_t nrs_policy_queued;
1058 * Service partition for this NRS head
1060 struct ptlrpc_service_part *nrs_svcpt;
1062 * Primary policy, which is the preferred policy for handling RPCs
1064 struct ptlrpc_nrs_policy *nrs_policy_primary;
1066 * Fallback policy, which is the backup policy for handling RPCs
1068 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1070 * This NRS head handles either HP or regular requests
1072 enum ptlrpc_nrs_queue_type nrs_queue_type;
1074 * # queued requests from all policies in this NRS head
1076 unsigned long nrs_req_queued;
1078 * # scheduled requests from all policies in this NRS head
1080 unsigned long nrs_req_started;
1082 * # policies on this NRS
1084 unsigned nrs_num_pols;
1086 * This NRS head is in progress of starting a policy
1088 unsigned nrs_policy_starting:1;
1090 * In progress of shutting down the whole NRS head; used during
1093 unsigned nrs_stopping:1;
1096 #define NRS_POL_NAME_MAX 16
1098 struct ptlrpc_nrs_pol_desc;
1101 * Service compatibility predicate; this determines whether a policy is adequate
1102 * for handling RPCs of a particular PTLRPC service.
1104 * XXX:This should give the same result during policy registration and
1105 * unregistration, and for all partitions of a service; so the result should not
1106 * depend on temporal service or other properties, that may influence the
1109 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
1110 const struct ptlrpc_nrs_pol_desc *desc);
1112 struct ptlrpc_nrs_pol_conf {
1114 * Human-readable policy name
1116 char nc_name[NRS_POL_NAME_MAX];
1118 * NRS operations for this policy
1120 const struct ptlrpc_nrs_pol_ops *nc_ops;
1122 * Service compatibility predicate
1124 nrs_pol_desc_compat_t nc_compat;
1126 * Set for policies that support a single ptlrpc service, i.e. ones that
1127 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
1128 * depicts the name of the single service that such policies are
1131 const char *nc_compat_svc_name;
1133 * Owner module for this policy descriptor; policies registering from a
1134 * different module to the one the NRS framework is held within
1135 * (currently ptlrpc), should set this field to THIS_MODULE.
1137 cfs_module_t *nc_owner;
1139 * Policy registration flags; a bitmast of \e nrs_policy_flags
1145 * NRS policy registering descriptor
1147 * Is used to hold a description of a policy that can be passed to NRS core in
1148 * order to register the policy with NRS heads in different PTLRPC services.
1150 struct ptlrpc_nrs_pol_desc {
1152 * Human-readable policy name
1154 char pd_name[NRS_POL_NAME_MAX];
1156 * Link into nrs_core::nrs_policies
1160 * NRS operations for this policy
1162 const struct ptlrpc_nrs_pol_ops *pd_ops;
1164 * Service compatibility predicate
1166 nrs_pol_desc_compat_t pd_compat;
1168 * Set for policies that are compatible with only one PTLRPC service.
1170 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
1172 const char *pd_compat_svc_name;
1174 * Owner module for this policy descriptor.
1176 * We need to hold a reference to the module whenever we might make use
1177 * of any of the module's contents, i.e.
1178 * - If one or more instances of the policy are at a state where they
1179 * might be handling a request, i.e.
1180 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1181 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
1182 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
1183 * is taken on the module when
1184 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
1185 * becomes 0, so that we hold only one reference to the module maximum
1188 * We do not need to hold a reference to the module, even though we
1189 * might use code and data from the module, in the following cases:
1190 * - During external policy registration, because this should happen in
1191 * the module's init() function, in which case the module is safe from
1192 * removal because a reference is being held on the module by the
1193 * kernel, and iirc kmod (and I guess module-init-tools also) will
1194 * serialize any racing processes properly anyway.
1195 * - During external policy unregistration, because this should happen
1196 * in a module's exit() function, and any attempts to start a policy
1197 * instance would need to take a reference on the module, and this is
1198 * not possible once we have reached the point where the exit()
1199 * handler is called.
1200 * - During service registration and unregistration, as service setup
1201 * and cleanup, and policy registration, unregistration and policy
1202 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
1203 * as long as users adhere to the convention of registering policies
1204 * in init() and unregistering them in module exit() functions, there
1205 * should not be a race between these operations.
1206 * - During any policy-specific lprocfs operations, because a reference
1207 * is held by the kernel on a proc entry that has been entered by a
1208 * syscall, so as long as proc entries are removed during unregistration time,
1209 * then unregistration and lprocfs operations will be properly
1212 cfs_module_t *pd_owner;
1214 * Bitmask of \e nrs_policy_flags
1218 * # of references on this descriptor
1220 cfs_atomic_t pd_refs;
1226 * Policies transition from one state to the other during their lifetime
1228 enum ptlrpc_nrs_pol_state {
1230 * Not a valid policy state.
1232 NRS_POL_STATE_INVALID,
1234 * Policies are at this state either at the start of their life, or
1235 * transition here when the user selects a different policy to act
1236 * as the primary one.
1238 NRS_POL_STATE_STOPPED,
1240 * Policy is progress of stopping
1242 NRS_POL_STATE_STOPPING,
1244 * Policy is in progress of starting
1246 NRS_POL_STATE_STARTING,
1248 * A policy is in this state in two cases:
1249 * - it is the fallback policy, which is always in this state.
1250 * - it has been activated by the user; i.e. it is the primary policy,
1252 NRS_POL_STATE_STARTED,
1256 * NRS policy information
1258 * Used for obtaining information for the status of a policy via lprocfs
1260 struct ptlrpc_nrs_pol_info {
1264 char pi_name[NRS_POL_NAME_MAX];
1266 * Current policy state
1268 enum ptlrpc_nrs_pol_state pi_state;
1270 * # RPCs enqueued for later dispatching by the policy
1274 * # RPCs started for dispatch by the policy
1276 long pi_req_started;
1278 * Is this a fallback policy?
1280 unsigned pi_fallback:1;
1286 * There is one instance of this for each policy in each NRS head of each
1287 * PTLRPC service partition.
1289 struct ptlrpc_nrs_policy {
1291 * Linkage into the NRS head's list of policies,
1292 * ptlrpc_nrs:nrs_policy_list
1294 cfs_list_t pol_list;
1296 * Linkage into the NRS head's list of policies with enqueued
1297 * requests ptlrpc_nrs:nrs_policy_queued
1299 cfs_list_t pol_list_queued;
1301 * Current state of this policy
1303 enum ptlrpc_nrs_pol_state pol_state;
1305 * Bitmask of nrs_policy_flags
1309 * # RPCs enqueued for later dispatching by the policy
1311 long pol_req_queued;
1313 * # RPCs started for dispatch by the policy
1315 long pol_req_started;
1317 * Usage Reference count taken on the policy instance
1321 * The NRS head this policy has been created at
1323 struct ptlrpc_nrs *pol_nrs;
1325 * Private policy data; varies by policy type
1329 * Policy descriptor for this policy instance.
1331 struct ptlrpc_nrs_pol_desc *pol_desc;
1337 * Resources are embedded into two types of NRS entities:
1338 * - Inside NRS policies, in the policy's private data in
1339 * ptlrpc_nrs_policy::pol_private
1340 * - In objects that act as prime-level scheduling entities in different NRS
1341 * policies; e.g. on a policy that performs round robin or similar order
1342 * scheduling across client NIDs, there would be one NRS resource per unique
1343 * client NID. On a policy which performs round robin scheduling across
1344 * backend filesystem objects, there would be one resource associated with
1345 * each of the backend filesystem objects partaking in the scheduling
1346 * performed by the policy.
1348 * NRS resources share a parent-child relationship, in which resources embedded
1349 * in policy instances are the parent entities, with all scheduling entities
1350 * a policy schedules across being the children, thus forming a simple resource
1351 * hierarchy. This hierarchy may be extended with one or more levels in the
1352 * future if the ability to have more than one primary policy is added.
1354 * Upon request initialization, references to the then active NRS policies are
1355 * taken and used to later handle the dispatching of the request with one of
1358 * \see nrs_resource_get_safe()
1359 * \see ptlrpc_nrs_req_add()
1361 struct ptlrpc_nrs_resource {
1363 * This NRS resource's parent; is NULL for resources embedded in NRS
1364 * policy instances; i.e. those are top-level ones.
1366 struct ptlrpc_nrs_resource *res_parent;
1368 * The policy associated with this resource.
1370 struct ptlrpc_nrs_policy *res_policy;
1383 * This policy is a logical wrapper around previous, non-NRS functionality.
1384 * It dispatches RPCs in the same order as they arrive from the network. This
1385 * policy is currently used as the fallback policy, and the only enabled policy
1386 * on all NRS heads of all PTLRPC service partitions.
1391 * Private data structure for the FIFO policy
1393 struct nrs_fifo_head {
1395 * Resource object for policy instance.
1397 struct ptlrpc_nrs_resource fh_res;
1399 * List of queued requests.
1403 * For debugging purposes.
1408 struct nrs_fifo_req {
1418 * Instances of this object exist embedded within ptlrpc_request; the main
1419 * purpose of this object is to hold references to the request's resources
1420 * for the lifetime of the request, and to hold properties that policies use
1421 * use for determining the request's scheduling priority.
1423 struct ptlrpc_nrs_request {
1425 * The request's resource hierarchy.
1427 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1429 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1430 * policy that was used to enqueue the request.
1432 * \see nrs_request_enqueue()
1434 unsigned nr_res_idx;
1435 unsigned nr_initialized:1;
1436 unsigned nr_enqueued:1;
1437 unsigned nr_started:1;
1438 unsigned nr_finalized:1;
1439 cfs_binheap_node_t nr_node;
1442 * Policy-specific fields, used for determining a request's scheduling
1443 * priority, and other supporting functionality.
1447 * Fields for the FIFO policy
1449 struct nrs_fifo_req fifo;
1452 * Externally-registering policies may want to use this to allocate
1453 * their own request properties.
1461 * Basic request prioritization operations structure.
1462 * The whole idea is centered around locks and RPCs that might affect locks.
1463 * When a lock is contended we try to give priority to RPCs that might lead
1464 * to fastest release of that lock.
1465 * Currently only implemented for OSTs only in a way that makes all
1466 * IO and truncate RPCs that are coming from a locked region where a lock is
1467 * contended a priority over other requests.
1469 struct ptlrpc_hpreq_ops {
1471 * Check if the lock handle of the given lock is the same as
1472 * taken from the request.
1474 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1476 * Check if the request is a high priority one.
1478 int (*hpreq_check)(struct ptlrpc_request *);
1480 * Called after the request has been handled.
1482 void (*hpreq_fini)(struct ptlrpc_request *);
1486 * Represents remote procedure call.
1488 * This is a staple structure used by everybody wanting to send a request
1491 struct ptlrpc_request {
1492 /* Request type: one of PTL_RPC_MSG_* */
1494 /** Result of request processing */
1497 * Linkage item through which this request is included into
1498 * sending/delayed lists on client and into rqbd list on server
1502 * Server side list of incoming unserved requests sorted by arrival
1503 * time. Traversed from time to time to notice about to expire
1504 * requests and sent back "early replies" to clients to let them
1505 * know server is alive and well, just very busy to service their
1508 cfs_list_t rq_timed_list;
1509 /** server-side history, used for debuging purposes. */
1510 cfs_list_t rq_history_list;
1511 /** server-side per-export list */
1512 cfs_list_t rq_exp_list;
1513 /** server-side hp handlers */
1514 struct ptlrpc_hpreq_ops *rq_ops;
1516 /** initial thread servicing this request */
1517 struct ptlrpc_thread *rq_svc_thread;
1519 /** history sequence # */
1520 __u64 rq_history_seq;
1524 /** stub for NRS request */
1525 struct ptlrpc_nrs_request rq_nrq;
1527 /** the index of service's srv_at_array into which request is linked */
1529 /** Lock to protect request flags and some other important bits, like
1533 /** client-side flags are serialized by rq_lock */
1534 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1535 rq_timedout:1, rq_resend:1, rq_restart:1,
1537 * when ->rq_replay is set, request is kept by the client even
1538 * after server commits corresponding transaction. This is
1539 * used for operations that require sequence of multiple
1540 * requests to be replayed. The only example currently is file
1541 * open/close. When last request in such a sequence is
1542 * committed, ->rq_replay is cleared on all requests in the
1546 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1547 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1548 rq_early:1, rq_must_unlink:1,
1549 rq_memalloc:1, /* req originated from "kswapd" */
1550 /* server-side flags */
1551 rq_packed_final:1, /* packed final reply */
1552 rq_hp:1, /* high priority RPC */
1553 rq_at_linked:1, /* link into service's srv_at_array */
1554 rq_reply_truncate:1,
1556 /* whether the "rq_set" is a valid one */
1558 rq_generation_set:1,
1559 /* do not resend request on -EINPROGRESS */
1560 rq_no_retry_einprogress:1,
1561 /* allow the req to be sent if the import is in recovery
1565 unsigned int rq_nr_resend;
1567 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1568 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1569 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
1570 server-side refcounf for multiple replies */
1572 /** Portal to which this request would be sent */
1573 short rq_request_portal; /* XXX FIXME bug 249 */
1574 /** Portal where to wait for reply and where reply would be sent */
1575 short rq_reply_portal; /* XXX FIXME bug 249 */
1579 * !rq_truncate : # reply bytes actually received,
1580 * rq_truncate : required repbuf_len for resend
1582 int rq_nob_received;
1583 /** Request length */
1587 /** Request message - what client sent */
1588 struct lustre_msg *rq_reqmsg;
1589 /** Reply message - server response */
1590 struct lustre_msg *rq_repmsg;
1591 /** Transaction number */
1596 * List item to for replay list. Not yet commited requests get linked
1598 * Also see \a rq_replay comment above.
1600 cfs_list_t rq_replay_list;
1603 * security and encryption data
1605 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1606 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1607 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
1609 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1610 enum lustre_sec_part rq_sp_from;
1612 /* client/server security flags */
1614 rq_ctx_init:1, /* context initiation */
1615 rq_ctx_fini:1, /* context destroy */
1616 rq_bulk_read:1, /* request bulk read */
1617 rq_bulk_write:1, /* request bulk write */
1618 /* server authentication flags */
1619 rq_auth_gss:1, /* authenticated by gss */
1620 rq_auth_remote:1, /* authed as remote user */
1621 rq_auth_usr_root:1, /* authed as root */
1622 rq_auth_usr_mdt:1, /* authed as mdt */
1623 rq_auth_usr_ost:1, /* authed as ost */
1624 /* security tfm flags */
1627 /* doesn't expect reply FIXME */
1629 rq_pill_init:1; /* pill initialized */
1631 uid_t rq_auth_uid; /* authed uid */
1632 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1634 /* (server side), pointed directly into req buffer */
1635 struct ptlrpc_user_desc *rq_user_desc;
1637 /* various buffer pointers */
1638 struct lustre_msg *rq_reqbuf; /* req wrapper */
1639 char *rq_repbuf; /* rep buffer */
1640 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1641 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1642 int rq_reqbuf_len; /* req wrapper buf len */
1643 int rq_reqdata_len; /* req wrapper msg len */
1644 int rq_repbuf_len; /* rep buffer len */
1645 int rq_repdata_len; /* rep wrapper msg len */
1646 int rq_clrbuf_len; /* only in priv mode */
1647 int rq_clrdata_len; /* only in priv mode */
1649 /** early replies go to offset 0, regular replies go after that */
1650 unsigned int rq_reply_off;
1654 /** Fields that help to see if request and reply were swabbed or not */
1655 __u32 rq_req_swab_mask;
1656 __u32 rq_rep_swab_mask;
1658 /** What was import generation when this request was sent */
1659 int rq_import_generation;
1660 enum lustre_imp_state rq_send_state;
1662 /** how many early replies (for stats) */
1665 /** client+server request */
1666 lnet_handle_md_t rq_req_md_h;
1667 struct ptlrpc_cb_id rq_req_cbid;
1668 /** optional time limit for send attempts */
1669 cfs_duration_t rq_delay_limit;
1670 /** time request was first queued */
1671 cfs_time_t rq_queued_time;
1673 /* server-side... */
1674 /** request arrival time */
1675 struct timeval rq_arrival_time;
1676 /** separated reply state */
1677 struct ptlrpc_reply_state *rq_reply_state;
1678 /** incoming request buffer */
1679 struct ptlrpc_request_buffer_desc *rq_rqbd;
1681 /** client-only incoming reply */
1682 lnet_handle_md_t rq_reply_md_h;
1683 cfs_waitq_t rq_reply_waitq;
1684 struct ptlrpc_cb_id rq_reply_cbid;
1688 /** Peer description (the other side) */
1689 lnet_process_id_t rq_peer;
1690 /** Server-side, export on which request was received */
1691 struct obd_export *rq_export;
1692 /** Client side, import where request is being sent */
1693 struct obd_import *rq_import;
1695 /** Replay callback, called after request is replayed at recovery */
1696 void (*rq_replay_cb)(struct ptlrpc_request *);
1698 * Commit callback, called when request is committed and about to be
1701 void (*rq_commit_cb)(struct ptlrpc_request *);
1702 /** Opaq data for replay and commit callbacks. */
1705 /** For bulk requests on client only: bulk descriptor */
1706 struct ptlrpc_bulk_desc *rq_bulk;
1708 /** client outgoing req */
1710 * when request/reply sent (secs), or time when request should be sent
1713 /** time for request really sent out */
1714 time_t rq_real_sent;
1716 /** when request must finish. volatile
1717 * so that servers' early reply updates to the deadline aren't
1718 * kept in per-cpu cache */
1719 volatile time_t rq_deadline;
1720 /** when req reply unlink must finish. */
1721 time_t rq_reply_deadline;
1722 /** when req bulk unlink must finish. */
1723 time_t rq_bulk_deadline;
1725 * service time estimate (secs)
1726 * If the requestsis not served by this time, it is marked as timed out.
1730 /** Multi-rpc bits */
1731 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1732 cfs_waitq_t rq_set_waitq;
1733 /** Link item for request set lists */
1734 cfs_list_t rq_set_chain;
1735 /** Link back to the request set */
1736 struct ptlrpc_request_set *rq_set;
1737 /** Async completion handler, called when reply is received */
1738 ptlrpc_interpterer_t rq_interpret_reply;
1739 /** Async completion context */
1740 union ptlrpc_async_args rq_async_args;
1742 /** Pool if request is from preallocated list */
1743 struct ptlrpc_request_pool *rq_pool;
1745 struct lu_context rq_session;
1746 struct lu_context rq_recov_session;
1748 /** request format description */
1749 struct req_capsule rq_pill;
1753 * Call completion handler for rpc if any, return it's status or original
1754 * rc if there was no handler defined for this request.
1756 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1757 struct ptlrpc_request *req, int rc)
1759 if (req->rq_interpret_reply != NULL) {
1760 req->rq_status = req->rq_interpret_reply(env, req,
1761 &req->rq_async_args,
1763 return req->rq_status;
1771 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
1772 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
1773 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1774 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1775 struct ptlrpc_nrs_pol_info *info);
1778 * Can the request be moved from the regular NRS head to the high-priority NRS
1779 * head (of the same PTLRPC service partition), if any?
1781 * For a reliable result, this should be checked under svcpt->scp_req lock.
1783 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1785 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1788 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1789 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1790 * to make sure it has not been scheduled yet (analogous to previous
1791 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1793 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1798 * Returns 1 if request buffer at offset \a index was already swabbed
1800 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1802 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1803 return req->rq_req_swab_mask & (1 << index);
1807 * Returns 1 if request reply buffer at offset \a index was already swabbed
1809 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1811 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1812 return req->rq_rep_swab_mask & (1 << index);
1816 * Returns 1 if request needs to be swabbed into local cpu byteorder
1818 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1820 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1824 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1826 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1828 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1832 * Mark request buffer at offset \a index that it was already swabbed
1834 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1836 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1837 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1838 req->rq_req_swab_mask |= 1 << index;
1842 * Mark request reply buffer at offset \a index that it was already swabbed
1844 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1846 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1847 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1848 req->rq_rep_swab_mask |= 1 << index;
1852 * Convert numerical request phase value \a phase into text string description
1854 static inline const char *
1855 ptlrpc_phase2str(enum rq_phase phase)
1864 case RQ_PHASE_INTERPRET:
1866 case RQ_PHASE_COMPLETE:
1868 case RQ_PHASE_UNREGISTERING:
1869 return "Unregistering";
1876 * Convert numerical request phase of the request \a req into text stringi
1879 static inline const char *
1880 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1882 return ptlrpc_phase2str(req->rq_phase);
1886 * Debugging functions and helpers to print request structure into debug log
1889 /* Spare the preprocessor, spoil the bugs. */
1890 #define FLAG(field, str) (field ? str : "")
1892 /** Convert bit flags into a string */
1893 #define DEBUG_REQ_FLAGS(req) \
1894 ptlrpc_rqphase2str(req), \
1895 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1896 FLAG(req->rq_err, "E"), \
1897 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1898 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1899 FLAG(req->rq_no_resend, "N"), \
1900 FLAG(req->rq_waiting, "W"), \
1901 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1902 FLAG(req->rq_committed, "M")
1904 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1906 void _debug_req(struct ptlrpc_request *req,
1907 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1908 __attribute__ ((format (printf, 3, 4)));
1911 * Helper that decides if we need to print request accordig to current debug
1914 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1916 CFS_CHECK_STACK(msgdata, mask, cdls); \
1918 if (((mask) & D_CANTMASK) != 0 || \
1919 ((libcfs_debug & (mask)) != 0 && \
1920 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1921 _debug_req((req), msgdata, fmt, ##a); \
1925 * This is the debug print function you need to use to print request sturucture
1926 * content into lustre debug log.
1927 * for most callers (level is a constant) this is resolved at compile time */
1928 #define DEBUG_REQ(level, req, fmt, args...) \
1930 if ((level) & (D_ERROR | D_WARNING)) { \
1931 static cfs_debug_limit_state_t cdls; \
1932 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1933 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1935 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1936 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1942 * Structure that defines a single page of a bulk transfer
1944 struct ptlrpc_bulk_page {
1945 /** Linkage to list of pages in a bulk */
1948 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1951 /** offset within a page */
1953 /** The page itself */
1954 struct page *bp_page;
1957 #define BULK_GET_SOURCE 0
1958 #define BULK_PUT_SINK 1
1959 #define BULK_GET_SINK 2
1960 #define BULK_PUT_SOURCE 3
1963 * Definition of bulk descriptor.
1964 * Bulks are special "Two phase" RPCs where initial request message
1965 * is sent first and it is followed bt a transfer (o receiving) of a large
1966 * amount of data to be settled into pages referenced from the bulk descriptors.
1967 * Bulks transfers (the actual data following the small requests) are done
1968 * on separate LNet portals.
1969 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1970 * Another user is readpage for MDT.
1972 struct ptlrpc_bulk_desc {
1973 /** completed with failure */
1974 unsigned long bd_failure:1;
1975 /** {put,get}{source,sink} */
1976 unsigned long bd_type:2;
1978 unsigned long bd_registered:1;
1979 /** For serialization with callback */
1981 /** Import generation when request for this bulk was sent */
1982 int bd_import_generation;
1983 /** LNet portal for this bulk */
1985 /** Server side - export this bulk created for */
1986 struct obd_export *bd_export;
1987 /** Client side - import this bulk was sent on */
1988 struct obd_import *bd_import;
1989 /** Back pointer to the request */
1990 struct ptlrpc_request *bd_req;
1991 cfs_waitq_t bd_waitq; /* server side only WQ */
1992 int bd_iov_count; /* # entries in bd_iov */
1993 int bd_max_iov; /* allocated size of bd_iov */
1994 int bd_nob; /* # bytes covered */
1995 int bd_nob_transferred; /* # bytes GOT/PUT */
1999 struct ptlrpc_cb_id bd_cbid; /* network callback info */
2000 lnet_nid_t bd_sender; /* stash event::sender */
2001 int bd_md_count; /* # valid entries in bd_mds */
2002 int bd_md_max_brw; /* max entries in bd_mds */
2003 /** array of associated MDs */
2004 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
2006 #if defined(__KERNEL__)
2008 * encrypt iov, size is either 0 or bd_iov_count.
2010 lnet_kiov_t *bd_enc_iov;
2012 lnet_kiov_t bd_iov[0];
2014 lnet_md_iovec_t bd_iov[0];
2019 SVC_STOPPED = 1 << 0,
2020 SVC_STOPPING = 1 << 1,
2021 SVC_STARTING = 1 << 2,
2022 SVC_RUNNING = 1 << 3,
2024 SVC_SIGNAL = 1 << 5,
2027 #define PTLRPC_THR_NAME_LEN 32
2029 * Definition of server service thread structure
2031 struct ptlrpc_thread {
2033 * List of active threads in svc->srv_threads
2037 * thread-private data (preallocated memory)
2042 * service thread index, from ptlrpc_start_threads
2046 * service thread pid
2050 * put watchdog in the structure per thread b=14840
2052 struct lc_watchdog *t_watchdog;
2054 * the svc this thread belonged to b=18582
2056 struct ptlrpc_service_part *t_svcpt;
2057 cfs_waitq_t t_ctl_waitq;
2058 struct lu_env *t_env;
2059 char t_name[PTLRPC_THR_NAME_LEN];
2062 static inline int thread_is_init(struct ptlrpc_thread *thread)
2064 return thread->t_flags == 0;
2067 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2069 return !!(thread->t_flags & SVC_STOPPED);
2072 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2074 return !!(thread->t_flags & SVC_STOPPING);
2077 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2079 return !!(thread->t_flags & SVC_STARTING);
2082 static inline int thread_is_running(struct ptlrpc_thread *thread)
2084 return !!(thread->t_flags & SVC_RUNNING);
2087 static inline int thread_is_event(struct ptlrpc_thread *thread)
2089 return !!(thread->t_flags & SVC_EVENT);
2092 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2094 return !!(thread->t_flags & SVC_SIGNAL);
2097 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2099 thread->t_flags &= ~flags;
2102 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2104 thread->t_flags = flags;
2107 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2109 thread->t_flags |= flags;
2112 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2115 if (thread->t_flags & flags) {
2116 thread->t_flags &= ~flags;
2123 * Request buffer descriptor structure.
2124 * This is a structure that contains one posted request buffer for service.
2125 * Once data land into a buffer, event callback creates actual request and
2126 * notifies wakes one of the service threads to process new incoming request.
2127 * More than one request can fit into the buffer.
2129 struct ptlrpc_request_buffer_desc {
2130 /** Link item for rqbds on a service */
2131 cfs_list_t rqbd_list;
2132 /** History of requests for this buffer */
2133 cfs_list_t rqbd_reqs;
2134 /** Back pointer to service for which this buffer is registered */
2135 struct ptlrpc_service_part *rqbd_svcpt;
2136 /** LNet descriptor */
2137 lnet_handle_md_t rqbd_md_h;
2139 /** The buffer itself */
2141 struct ptlrpc_cb_id rqbd_cbid;
2143 * This "embedded" request structure is only used for the
2144 * last request to fit into the buffer
2146 struct ptlrpc_request rqbd_req;
2149 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2151 struct ptlrpc_service_ops {
2153 * if non-NULL called during thread creation (ptlrpc_start_thread())
2154 * to initialize service specific per-thread state.
2156 int (*so_thr_init)(struct ptlrpc_thread *thr);
2158 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2159 * destruct state created by ->srv_init().
2161 void (*so_thr_done)(struct ptlrpc_thread *thr);
2163 * Handler function for incoming requests for this service
2165 int (*so_req_handler)(struct ptlrpc_request *req);
2167 * function to determine priority of the request, it's called
2168 * on every new request
2170 int (*so_hpreq_handler)(struct ptlrpc_request *);
2172 * service-specific print fn
2174 void (*so_req_printer)(void *, struct ptlrpc_request *);
2177 #ifndef __cfs_cacheline_aligned
2178 /* NB: put it here for reducing patche dependence */
2179 # define __cfs_cacheline_aligned
2183 * How many high priority requests to serve before serving one normal
2186 #define PTLRPC_SVC_HP_RATIO 10
2189 * Definition of PortalRPC service.
2190 * The service is listening on a particular portal (like tcp port)
2191 * and perform actions for a specific server like IO service for OST
2192 * or general metadata service for MDS.
2194 struct ptlrpc_service {
2195 /** serialize /proc operations */
2196 spinlock_t srv_lock;
2197 /** most often accessed fields */
2198 /** chain thru all services */
2199 cfs_list_t srv_list;
2200 /** service operations table */
2201 struct ptlrpc_service_ops srv_ops;
2202 /** only statically allocated strings here; we don't clean them */
2204 /** only statically allocated strings here; we don't clean them */
2205 char *srv_thread_name;
2206 /** service thread list */
2207 cfs_list_t srv_threads;
2208 /** threads # should be created for each partition on initializing */
2209 int srv_nthrs_cpt_init;
2210 /** limit of threads number for each partition */
2211 int srv_nthrs_cpt_limit;
2212 /** Root of /proc dir tree for this service */
2213 cfs_proc_dir_entry_t *srv_procroot;
2214 /** Pointer to statistic data for this service */
2215 struct lprocfs_stats *srv_stats;
2216 /** # hp per lp reqs to handle */
2217 int srv_hpreq_ratio;
2218 /** biggest request to receive */
2219 int srv_max_req_size;
2220 /** biggest reply to send */
2221 int srv_max_reply_size;
2222 /** size of individual buffers */
2224 /** # buffers to allocate in 1 group */
2225 int srv_nbuf_per_group;
2226 /** Local portal on which to receive requests */
2227 __u32 srv_req_portal;
2228 /** Portal on the client to send replies to */
2229 __u32 srv_rep_portal;
2231 * Tags for lu_context associated with this thread, see struct
2235 /** soft watchdog timeout multiplier */
2236 int srv_watchdog_factor;
2237 /** under unregister_service */
2238 unsigned srv_is_stopping:1;
2240 /** max # request buffers in history per partition */
2241 int srv_hist_nrqbds_cpt_max;
2242 /** number of CPTs this service bound on */
2244 /** CPTs array this service bound on */
2246 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2248 /** CPT table this service is running over */
2249 struct cfs_cpt_table *srv_cptable;
2251 * partition data for ptlrpc service
2253 struct ptlrpc_service_part *srv_parts[0];
2257 * Definition of PortalRPC service partition data.
2258 * Although a service only has one instance of it right now, but we
2259 * will have multiple instances very soon (instance per CPT).
2261 * it has four locks:
2263 * serialize operations on rqbd and requests waiting for preprocess
2265 * serialize operations active requests sent to this portal
2267 * serialize adaptive timeout stuff
2269 * serialize operations on RS list (reply states)
2271 * We don't have any use-case to take two or more locks at the same time
2272 * for now, so there is no lock order issue.
2274 struct ptlrpc_service_part {
2275 /** back reference to owner */
2276 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2277 /* CPT id, reserved */
2279 /** always increasing number */
2281 /** # of starting threads */
2282 int scp_nthrs_starting;
2283 /** # of stopping threads, reserved for shrinking threads */
2284 int scp_nthrs_stopping;
2285 /** # running threads */
2286 int scp_nthrs_running;
2287 /** service threads list */
2288 cfs_list_t scp_threads;
2291 * serialize the following fields, used for protecting
2292 * rqbd list and incoming requests waiting for preprocess,
2293 * threads starting & stopping are also protected by this lock.
2295 spinlock_t scp_lock __cfs_cacheline_aligned;
2296 /** total # req buffer descs allocated */
2297 int scp_nrqbds_total;
2298 /** # posted request buffers for receiving */
2299 int scp_nrqbds_posted;
2300 /** in progress of allocating rqbd */
2301 int scp_rqbd_allocating;
2302 /** # incoming reqs */
2303 int scp_nreqs_incoming;
2304 /** request buffers to be reposted */
2305 cfs_list_t scp_rqbd_idle;
2306 /** req buffers receiving */
2307 cfs_list_t scp_rqbd_posted;
2308 /** incoming reqs */
2309 cfs_list_t scp_req_incoming;
2310 /** timeout before re-posting reqs, in tick */
2311 cfs_duration_t scp_rqbd_timeout;
2313 * all threads sleep on this. This wait-queue is signalled when new
2314 * incoming request arrives and when difficult reply has to be handled.
2316 cfs_waitq_t scp_waitq;
2318 /** request history */
2319 cfs_list_t scp_hist_reqs;
2320 /** request buffer history */
2321 cfs_list_t scp_hist_rqbds;
2322 /** # request buffers in history */
2323 int scp_hist_nrqbds;
2324 /** sequence number for request */
2326 /** highest seq culled from history */
2327 __u64 scp_hist_seq_culled;
2330 * serialize the following fields, used for processing requests
2331 * sent to this portal
2333 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2334 /** # reqs in either of the NRS heads below */
2335 /** # reqs being served */
2336 int scp_nreqs_active;
2337 /** # HPreqs being served */
2338 int scp_nhreqs_active;
2339 /** # hp requests handled */
2342 /** NRS head for regular requests */
2343 struct ptlrpc_nrs scp_nrs_reg;
2344 /** NRS head for HP requests; this is only valid for services that can
2345 * handle HP requests */
2346 struct ptlrpc_nrs *scp_nrs_hp;
2351 * serialize the following fields, used for changes on
2354 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2355 /** estimated rpc service time */
2356 struct adaptive_timeout scp_at_estimate;
2357 /** reqs waiting for replies */
2358 struct ptlrpc_at_array scp_at_array;
2359 /** early reply timer */
2360 cfs_timer_t scp_at_timer;
2362 cfs_time_t scp_at_checktime;
2363 /** check early replies */
2364 unsigned scp_at_check;
2368 * serialize the following fields, used for processing
2369 * replies for this portal
2371 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2372 /** all the active replies */
2373 cfs_list_t scp_rep_active;
2375 /** replies waiting for service */
2376 cfs_list_t scp_rep_queue;
2378 /** List of free reply_states */
2379 cfs_list_t scp_rep_idle;
2380 /** waitq to run, when adding stuff to srv_free_rs_list */
2381 cfs_waitq_t scp_rep_waitq;
2382 /** # 'difficult' replies */
2383 cfs_atomic_t scp_nreps_difficult;
2386 #define ptlrpc_service_for_each_part(part, i, svc) \
2388 i < (svc)->srv_ncpts && \
2389 (svc)->srv_parts != NULL && \
2390 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2393 * Declaration of ptlrpcd control structure
2395 struct ptlrpcd_ctl {
2397 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2399 unsigned long pc_flags;
2401 * Thread lock protecting structure fields.
2407 struct completion pc_starting;
2411 struct completion pc_finishing;
2413 * Thread requests set.
2415 struct ptlrpc_request_set *pc_set;
2417 * Thread name used in cfs_daemonize()
2421 * Environment for request interpreters to run in.
2423 struct lu_env pc_env;
2425 * Index of ptlrpcd thread in the array.
2429 * Number of the ptlrpcd's partners.
2433 * Pointer to the array of partners' ptlrpcd_ctl structure.
2435 struct ptlrpcd_ctl **pc_partners;
2437 * Record the partner index to be processed next.
2442 * Async rpcs flag to make sure that ptlrpcd_check() is called only
2447 * Currently not used.
2451 * User-space async rpcs callback.
2453 void *pc_wait_callback;
2455 * User-space check idle rpcs callback.
2457 void *pc_idle_callback;
2461 /* Bits for pc_flags */
2462 enum ptlrpcd_ctl_flags {
2464 * Ptlrpc thread start flag.
2466 LIOD_START = 1 << 0,
2468 * Ptlrpc thread stop flag.
2472 * Ptlrpc thread force flag (only stop force so far).
2473 * This will cause aborting any inflight rpcs handled
2474 * by thread if LIOD_STOP is specified.
2476 LIOD_FORCE = 1 << 2,
2478 * This is a recovery ptlrpc thread.
2480 LIOD_RECOVERY = 1 << 3,
2482 * The ptlrpcd is bound to some CPU core.
2491 * Service compatibility function; the policy is compatible with all services.
2493 * \param[in] svc The service the policy is attempting to register with.
2494 * \param[in] desc The policy descriptor
2496 * \retval true The policy is compatible with the service
2498 * \see ptlrpc_nrs_pol_desc::pd_compat()
2500 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2501 const struct ptlrpc_nrs_pol_desc *desc)
2507 * Service compatibility function; the policy is compatible with only a specific
2508 * service which is identified by its human-readable name at
2509 * ptlrpc_service::srv_name.
2511 * \param[in] svc The service the policy is attempting to register with.
2512 * \param[in] desc The policy descriptor
2514 * \retval false The policy is not compatible with the service
2515 * \retval true The policy is compatible with the service
2517 * \see ptlrpc_nrs_pol_desc::pd_compat()
2519 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2520 const struct ptlrpc_nrs_pol_desc *desc)
2522 LASSERT(desc->pd_compat_svc_name != NULL);
2523 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2528 /* ptlrpc/events.c */
2529 extern lnet_handle_eq_t ptlrpc_eq_h;
2530 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2531 lnet_process_id_t *peer, lnet_nid_t *self);
2533 * These callbacks are invoked by LNet when something happened to
2537 extern void request_out_callback(lnet_event_t *ev);
2538 extern void reply_in_callback(lnet_event_t *ev);
2539 extern void client_bulk_callback(lnet_event_t *ev);
2540 extern void request_in_callback(lnet_event_t *ev);
2541 extern void reply_out_callback(lnet_event_t *ev);
2542 #ifdef HAVE_SERVER_SUPPORT
2543 extern void server_bulk_callback(lnet_event_t *ev);
2547 /* ptlrpc/connection.c */
2548 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2550 struct obd_uuid *uuid);
2551 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2552 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2553 int ptlrpc_connection_init(void);
2554 void ptlrpc_connection_fini(void);
2555 extern lnet_pid_t ptl_get_pid(void);
2557 /* ptlrpc/niobuf.c */
2559 * Actual interfacing with LNet to put/get/register/unregister stuff
2562 #ifdef HAVE_SERVER_SUPPORT
2563 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2564 unsigned npages, unsigned max_brw,
2565 unsigned type, unsigned portal);
2566 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2567 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2569 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2573 LASSERT(desc != NULL);
2575 spin_lock(&desc->bd_lock);
2576 rc = desc->bd_md_count;
2577 spin_unlock(&desc->bd_lock);
2582 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2583 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2585 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2587 struct ptlrpc_bulk_desc *desc;
2590 LASSERT(req != NULL);
2591 desc = req->rq_bulk;
2593 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2594 req->rq_bulk_deadline > cfs_time_current_sec())
2600 spin_lock(&desc->bd_lock);
2601 rc = desc->bd_md_count;
2602 spin_unlock(&desc->bd_lock);
2606 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2607 #define PTLRPC_REPLY_EARLY 0x02
2608 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2609 int ptlrpc_reply(struct ptlrpc_request *req);
2610 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2611 int ptlrpc_error(struct ptlrpc_request *req);
2612 void ptlrpc_resend_req(struct ptlrpc_request *request);
2613 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2614 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2615 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2618 /* ptlrpc/client.c */
2620 * Client-side portals API. Everything to send requests, receive replies,
2621 * request queues, request management, etc.
2624 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2625 struct ptlrpc_client *);
2626 void ptlrpc_cleanup_client(struct obd_import *imp);
2627 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2629 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2630 int ptlrpc_replay_req(struct ptlrpc_request *req);
2631 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2632 void ptlrpc_restart_req(struct ptlrpc_request *req);
2633 void ptlrpc_abort_inflight(struct obd_import *imp);
2634 void ptlrpc_cleanup_imp(struct obd_import *imp);
2635 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2637 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2638 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2640 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2641 set_interpreter_func fn, void *data);
2642 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2643 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2644 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2645 int ptlrpc_expired_set(void *data);
2646 void ptlrpc_interrupted_set(void *data);
2647 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2648 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2649 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2650 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2651 struct ptlrpc_request *req);
2653 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2654 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2656 struct ptlrpc_request_pool *
2657 ptlrpc_init_rq_pool(int, int,
2658 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2660 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2661 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2662 const struct req_format *format);
2663 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2664 struct ptlrpc_request_pool *,
2665 const struct req_format *format);
2666 void ptlrpc_request_free(struct ptlrpc_request *request);
2667 int ptlrpc_request_pack(struct ptlrpc_request *request,
2668 __u32 version, int opcode);
2669 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2670 const struct req_format *format,
2671 __u32 version, int opcode);
2672 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2673 __u32 version, int opcode, char **bufs,
2674 struct ptlrpc_cli_ctx *ctx);
2675 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2676 int opcode, int count, __u32 *lengths,
2678 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2679 __u32 version, int opcode,
2680 int count, __u32 *lengths, char **bufs,
2681 struct ptlrpc_request_pool *pool);
2682 void ptlrpc_req_finished(struct ptlrpc_request *request);
2683 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2684 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2685 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2686 unsigned npages, unsigned max_brw,
2687 unsigned type, unsigned portal);
2688 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2689 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2691 __ptlrpc_free_bulk(bulk, 1);
2693 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2695 __ptlrpc_free_bulk(bulk, 0);
2697 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2698 cfs_page_t *page, int pageoffset, int len, int);
2699 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2700 cfs_page_t *page, int pageoffset,
2703 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2706 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2707 cfs_page_t *page, int pageoffset,
2710 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2713 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2714 struct obd_import *imp);
2715 __u64 ptlrpc_next_xid(void);
2716 __u64 ptlrpc_sample_next_xid(void);
2717 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2719 /* Set of routines to run a function in ptlrpcd context */
2720 void *ptlrpcd_alloc_work(struct obd_import *imp,
2721 int (*cb)(const struct lu_env *, void *), void *data);
2722 void ptlrpcd_destroy_work(void *handler);
2723 int ptlrpcd_queue_work(void *handler);
2726 struct ptlrpc_service_buf_conf {
2727 /* nbufs is buffers # to allocate when growing the pool */
2728 unsigned int bc_nbufs;
2729 /* buffer size to post */
2730 unsigned int bc_buf_size;
2731 /* portal to listed for requests on */
2732 unsigned int bc_req_portal;
2733 /* portal of where to send replies to */
2734 unsigned int bc_rep_portal;
2735 /* maximum request size to be accepted for this service */
2736 unsigned int bc_req_max_size;
2737 /* maximum reply size this service can ever send */
2738 unsigned int bc_rep_max_size;
2741 struct ptlrpc_service_thr_conf {
2742 /* threadname should be 8 characters or less - 6 will be added on */
2744 /* threads increasing factor for each CPU */
2745 unsigned int tc_thr_factor;
2746 /* service threads # to start on each partition while initializing */
2747 unsigned int tc_nthrs_init;
2749 * low water of threads # upper-limit on each partition while running,
2750 * service availability may be impacted if threads number is lower
2751 * than this value. It can be ZERO if the service doesn't require
2752 * CPU affinity or there is only one partition.
2754 unsigned int tc_nthrs_base;
2755 /* "soft" limit for total threads number */
2756 unsigned int tc_nthrs_max;
2757 /* user specified threads number, it will be validated due to
2758 * other members of this structure. */
2759 unsigned int tc_nthrs_user;
2760 /* set NUMA node affinity for service threads */
2761 unsigned int tc_cpu_affinity;
2762 /* Tags for lu_context associated with service thread */
2766 struct ptlrpc_service_cpt_conf {
2767 struct cfs_cpt_table *cc_cptable;
2768 /* string pattern to describe CPTs for a service */
2772 struct ptlrpc_service_conf {
2775 /* soft watchdog timeout multiplifier to print stuck service traces */
2776 unsigned int psc_watchdog_factor;
2777 /* buffer information */
2778 struct ptlrpc_service_buf_conf psc_buf;
2779 /* thread information */
2780 struct ptlrpc_service_thr_conf psc_thr;
2781 /* CPU partition information */
2782 struct ptlrpc_service_cpt_conf psc_cpt;
2783 /* function table */
2784 struct ptlrpc_service_ops psc_ops;
2787 /* ptlrpc/service.c */
2789 * Server-side services API. Register/unregister service, request state
2790 * management, service thread management
2794 void ptlrpc_save_lock(struct ptlrpc_request *req,
2795 struct lustre_handle *lock, int mode, int no_ack);
2796 void ptlrpc_commit_replies(struct obd_export *exp);
2797 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2798 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2799 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2800 struct ptlrpc_service *ptlrpc_register_service(
2801 struct ptlrpc_service_conf *conf,
2802 struct proc_dir_entry *proc_entry);
2803 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2805 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2806 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2807 int liblustre_check_services(void *arg);
2808 void ptlrpc_daemonize(char *name);
2809 int ptlrpc_service_health_check(struct ptlrpc_service *);
2810 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2813 int ptlrpc_hr_init(void);
2814 void ptlrpc_hr_fini(void);
2816 # define ptlrpc_hr_init() (0)
2817 # define ptlrpc_hr_fini() do {} while(0)
2822 /* ptlrpc/import.c */
2827 int ptlrpc_connect_import(struct obd_import *imp);
2828 int ptlrpc_init_import(struct obd_import *imp);
2829 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2830 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2831 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2834 /* ptlrpc/pack_generic.c */
2835 int ptlrpc_reconnect_import(struct obd_import *imp);
2839 * ptlrpc msg buffer and swab interface
2843 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2845 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2847 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2848 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2850 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2851 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2853 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2854 __u32 *lens, char **bufs);
2855 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2857 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2858 __u32 *lens, char **bufs, int flags);
2859 #define LPRFL_EARLY_REPLY 1
2860 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2861 char **bufs, int flags);
2862 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2863 unsigned int newlen, int move_data);
2864 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2865 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2866 int lustre_msg_hdr_size(__u32 magic, int count);
2867 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2868 int lustre_msg_size_v2(int count, __u32 *lengths);
2869 int lustre_packed_msg_size(struct lustre_msg *msg);
2870 int lustre_msg_early_size(void);
2871 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2872 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2873 int lustre_msg_buflen(struct lustre_msg *m, int n);
2874 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2875 int lustre_msg_bufcount(struct lustre_msg *m);
2876 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2877 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2878 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2879 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2880 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2881 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2882 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2883 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2884 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2885 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2886 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2887 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2888 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2889 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2890 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2891 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2892 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2893 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2894 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2895 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2896 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2897 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2898 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2899 int lustre_msg_get_status(struct lustre_msg *msg);
2900 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2901 int lustre_msg_is_v1(struct lustre_msg *msg);
2902 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2903 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2904 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2905 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2906 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2907 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2908 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2910 # warning "remove checksum compatibility support for b1_8"
2911 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2913 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2914 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2915 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2916 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2917 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2918 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2919 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2920 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2921 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2922 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2923 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2924 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2925 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2926 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2927 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2930 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2931 unsigned int newlen, int move_data)
2933 LASSERT(req->rq_reply_state);
2934 LASSERT(req->rq_repmsg);
2935 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2940 /** Change request phase of \a req to \a new_phase */
2942 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2944 if (req->rq_phase == new_phase)
2947 if (new_phase == RQ_PHASE_UNREGISTERING) {
2948 req->rq_next_phase = req->rq_phase;
2950 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2953 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2955 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2958 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2959 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2961 req->rq_phase = new_phase;
2965 * Returns true if request \a req got early reply and hard deadline is not met
2968 ptlrpc_client_early(struct ptlrpc_request *req)
2970 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2971 req->rq_reply_deadline > cfs_time_current_sec())
2973 return req->rq_early;
2977 * Returns true if we got real reply from server for this request
2980 ptlrpc_client_replied(struct ptlrpc_request *req)
2982 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2983 req->rq_reply_deadline > cfs_time_current_sec())
2985 return req->rq_replied;
2988 /** Returns true if request \a req is in process of receiving server reply */
2990 ptlrpc_client_recv(struct ptlrpc_request *req)
2992 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2993 req->rq_reply_deadline > cfs_time_current_sec())
2995 return req->rq_receiving_reply;
2999 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
3003 spin_lock(&req->rq_lock);
3004 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3005 req->rq_reply_deadline > cfs_time_current_sec()) {
3006 spin_unlock(&req->rq_lock);
3009 rc = req->rq_receiving_reply || req->rq_must_unlink;
3010 spin_unlock(&req->rq_lock);
3015 ptlrpc_client_wake_req(struct ptlrpc_request *req)
3017 if (req->rq_set == NULL)
3018 cfs_waitq_signal(&req->rq_reply_waitq);
3020 cfs_waitq_signal(&req->rq_set->set_waitq);
3024 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
3026 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
3027 cfs_atomic_inc(&rs->rs_refcount);
3031 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
3033 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
3034 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
3035 lustre_free_reply_state(rs);
3038 /* Should only be called once per req */
3039 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
3041 if (req->rq_reply_state == NULL)
3042 return; /* shouldn't occur */
3043 ptlrpc_rs_decref(req->rq_reply_state);
3044 req->rq_reply_state = NULL;
3045 req->rq_repmsg = NULL;
3048 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3050 return lustre_msg_get_magic(req->rq_reqmsg);
3053 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3055 switch (req->rq_reqmsg->lm_magic) {
3056 case LUSTRE_MSG_MAGIC_V2:
3057 return req->rq_reqmsg->lm_repsize;
3059 LASSERTF(0, "incorrect message magic: %08x\n",
3060 req->rq_reqmsg->lm_magic);
3065 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3067 if (req->rq_delay_limit != 0 &&
3068 cfs_time_before(cfs_time_add(req->rq_queued_time,
3069 cfs_time_seconds(req->rq_delay_limit)),
3070 cfs_time_current())) {
3076 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3078 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3079 spin_lock(&req->rq_lock);
3080 req->rq_no_resend = 1;
3081 spin_unlock(&req->rq_lock);
3083 return req->rq_no_resend;
3087 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3089 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3091 return svcpt->scp_service->srv_watchdog_factor *
3092 max_t(int, at, obd_timeout);
3095 static inline struct ptlrpc_service *
3096 ptlrpc_req2svc(struct ptlrpc_request *req)
3098 LASSERT(req->rq_rqbd != NULL);
3099 return req->rq_rqbd->rqbd_svcpt->scp_service;
3102 /* ldlm/ldlm_lib.c */
3104 * Target client logic
3107 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3108 int client_obd_cleanup(struct obd_device *obddev);
3109 int client_connect_import(const struct lu_env *env,
3110 struct obd_export **exp, struct obd_device *obd,
3111 struct obd_uuid *cluuid, struct obd_connect_data *,
3113 int client_disconnect_export(struct obd_export *exp);
3114 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3116 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3117 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3118 struct obd_uuid *uuid);
3119 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3120 void client_destroy_import(struct obd_import *imp);
3123 #ifdef HAVE_SERVER_SUPPORT
3124 int server_disconnect_export(struct obd_export *exp);
3127 /* ptlrpc/pinger.c */
3129 * Pinger API (client side only)
3132 extern int suppress_pings;
3133 enum timeout_event {
3136 struct timeout_item;
3137 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3138 int ptlrpc_pinger_add_import(struct obd_import *imp);
3139 int ptlrpc_pinger_del_import(struct obd_import *imp);
3140 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3141 timeout_cb_t cb, void *data,
3142 cfs_list_t *obd_list);
3143 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
3144 enum timeout_event event);
3145 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3146 int ptlrpc_obd_ping(struct obd_device *obd);
3147 cfs_time_t ptlrpc_suspend_wakeup_time(void);
3149 void ping_evictor_start(void);
3150 void ping_evictor_stop(void);
3152 #define ping_evictor_start() do {} while (0)
3153 #define ping_evictor_stop() do {} while (0)
3155 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
3158 /* ptlrpc daemon bind policy */
3160 /* all ptlrpcd threads are free mode */
3161 PDB_POLICY_NONE = 1,
3162 /* all ptlrpcd threads are bound mode */
3163 PDB_POLICY_FULL = 2,
3164 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3165 PDB_POLICY_PAIR = 3,
3166 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3167 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3168 * If kernel supports NUMA, pthrpcd threads are binded and
3169 * grouped by NUMA node */
3170 PDB_POLICY_NEIGHBOR = 4,
3173 /* ptlrpc daemon load policy
3174 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3175 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3176 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3177 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3178 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3180 /* on the same CPU core as the caller */
3181 PDL_POLICY_SAME = 1,
3182 /* within the same CPU partition, but not the same core as the caller */
3183 PDL_POLICY_LOCAL = 2,
3184 /* round-robin on all CPU cores, but not the same core as the caller */
3185 PDL_POLICY_ROUND = 3,
3186 /* the specified CPU core is preferred, but not enforced */
3187 PDL_POLICY_PREFERRED = 4,
3190 /* ptlrpc/ptlrpcd.c */
3191 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3192 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3193 void ptlrpcd_wake(struct ptlrpc_request *req);
3194 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3195 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3196 int ptlrpcd_addref(void);
3197 void ptlrpcd_decref(void);
3199 /* ptlrpc/lproc_ptlrpc.c */
3201 * procfs output related functions
3204 const char* ll_opcode2str(__u32 opcode);
3206 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3207 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3208 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3210 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3211 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3212 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3216 /* ptlrpc/llog_server.c */
3217 int llog_origin_handle_open(struct ptlrpc_request *req);
3218 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3219 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3220 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3221 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3222 int llog_origin_handle_close(struct ptlrpc_request *req);
3223 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3225 /* ptlrpc/llog_client.c */
3226 extern struct llog_operations llog_client_ops;