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, 2014, 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 #include <libcfs/libcfs.h>
59 #include <lnet/nidstr.h>
61 #include <lustre/lustre_idl.h>
62 #include <lustre_ha.h>
63 #include <lustre_sec.h>
64 #include <lustre_import.h>
65 #include <lprocfs_status.h>
66 #include <lu_object.h>
67 #include <lustre_req_layout.h>
68 #include <obd_support.h>
69 #include <lustre_ver.h>
71 /* MD flags we _always_ use */
72 #define PTLRPC_MD_OPTIONS 0
75 * Max # of bulk operations in one request.
76 * In order for the client and server to properly negotiate the maximum
77 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
78 * value. The client is free to limit the actual RPC size for any bulk
79 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
80 #define PTLRPC_BULK_OPS_BITS 2
81 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
83 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
84 * should not be used on the server at all. Otherwise, it imposes a
85 * protocol limitation on the maximum RPC size that can be used by any
86 * RPC sent to that server in the future. Instead, the server should
87 * use the negotiated per-client ocd_brw_size to determine the bulk
89 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
92 * Define maxima for bulk I/O.
94 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
95 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
96 * currently supported maximum between peers at connect via ocd_brw_size.
98 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
99 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
100 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
102 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
103 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
104 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
105 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
106 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
107 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
109 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
110 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
111 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
113 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
114 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
116 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
117 # error "PTLRPC_MAX_BRW_SIZE too big"
119 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
120 # error "PTLRPC_MAX_BRW_PAGES too big"
123 #define PTLRPC_NTHRS_INIT 2
128 * Constants determine how memory is used to buffer incoming service requests.
130 * ?_NBUFS # buffers to allocate when growing the pool
131 * ?_BUFSIZE # bytes in a single request buffer
132 * ?_MAXREQSIZE # maximum request service will receive
134 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
135 * of ?_NBUFS is added to the pool.
137 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
138 * considered full when less than ?_MAXREQSIZE is left in them.
143 * Constants determine how threads are created for ptlrpc service.
145 * ?_NTHRS_INIT # threads to create for each service partition on
146 * initializing. If it's non-affinity service and
147 * there is only one partition, it's the overall #
148 * threads for the service while initializing.
149 * ?_NTHRS_BASE # threads should be created at least for each
150 * ptlrpc partition to keep the service healthy.
151 * It's the low-water mark of threads upper-limit
152 * for each partition.
153 * ?_THR_FACTOR # threads can be added on threads upper-limit for
154 * each CPU core. This factor is only for reference,
155 * we might decrease value of factor if number of cores
156 * per CPT is above a limit.
157 * ?_NTHRS_MAX # overall threads can be created for a service,
158 * it's a soft limit because if service is running
159 * on machine with hundreds of cores and tens of
160 * CPU partitions, we need to guarantee each partition
161 * has ?_NTHRS_BASE threads, which means total threads
162 * will be ?_NTHRS_BASE * number_of_cpts which can
163 * exceed ?_NTHRS_MAX.
167 * #define MDS_NTHRS_INIT 2
168 * #define MDS_NTHRS_BASE 64
169 * #define MDS_NTHRS_FACTOR 8
170 * #define MDS_NTHRS_MAX 1024
173 * ---------------------------------------------------------------------
174 * Server(A) has 16 cores, user configured it to 4 partitions so each
175 * partition has 4 cores, then actual number of service threads on each
177 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
179 * Total number of threads for the service is:
180 * 96 * partitions(4) = 384
183 * ---------------------------------------------------------------------
184 * Server(B) has 32 cores, user configured it to 4 partitions so each
185 * partition has 8 cores, then actual number of service threads on each
187 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
189 * Total number of threads for the service is:
190 * 128 * partitions(4) = 512
193 * ---------------------------------------------------------------------
194 * Server(B) has 96 cores, user configured it to 8 partitions so each
195 * partition has 12 cores, then actual number of service threads on each
197 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
199 * Total number of threads for the service is:
200 * 160 * partitions(8) = 1280
202 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
203 * as upper limit of threads number for each partition:
204 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
207 * ---------------------------------------------------------------------
208 * Server(C) have a thousand of cores and user configured it to 32 partitions
209 * MDS_NTHRS_BASE(64) * 32 = 2048
211 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
212 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
213 * to keep service healthy, so total number of threads will just be 2048.
215 * NB: we don't suggest to choose server with that many cores because backend
216 * filesystem itself, buffer cache, or underlying network stack might
217 * have some SMP scalability issues at that large scale.
219 * If user already has a fat machine with hundreds or thousands of cores,
220 * there are two choices for configuration:
221 * a) create CPU table from subset of all CPUs and run Lustre on
223 * b) bind service threads on a few partitions, see modparameters of
224 * MDS and OSS for details
226 * NB: these calculations (and examples below) are simplified to help
227 * understanding, the real implementation is a little more complex,
228 * please see ptlrpc_server_nthreads_check() for details.
233 * LDLM threads constants:
235 * Given 8 as factor and 24 as base threads number
238 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
241 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
242 * threads for each partition and total threads number will be 112.
245 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
246 * threads for each partition to keep service healthy, so total threads
247 * number should be 24 * 8 = 192.
249 * So with these constants, threads number will be at the similar level
250 * of old versions, unless target machine has over a hundred cores
252 #define LDLM_THR_FACTOR 8
253 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
254 #define LDLM_NTHRS_BASE 24
255 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
257 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
258 #define LDLM_CLIENT_NBUFS 1
259 #define LDLM_SERVER_NBUFS 64
260 #define LDLM_BUFSIZE (8 * 1024)
261 #define LDLM_MAXREQSIZE (5 * 1024)
262 #define LDLM_MAXREPSIZE (1024)
265 * MDS threads constants:
267 * Please see examples in "Thread Constants", MDS threads number will be at
268 * the comparable level of old versions, unless the server has many cores.
270 #ifndef MDS_MAX_THREADS
271 #define MDS_MAX_THREADS 1024
272 #define MDS_MAX_OTHR_THREADS 256
274 #else /* MDS_MAX_THREADS */
275 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
276 #undef MDS_MAX_THREADS
277 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
279 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
282 /* default service */
283 #define MDS_THR_FACTOR 8
284 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
285 #define MDS_NTHRS_MAX MDS_MAX_THREADS
286 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
288 /* read-page service */
289 #define MDS_RDPG_THR_FACTOR 4
290 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
291 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
292 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
294 /* these should be removed when we remove setattr service in the future */
295 #define MDS_SETA_THR_FACTOR 4
296 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
297 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
298 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
300 /* non-affinity threads */
301 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
302 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
307 * Assume file name length = FNAME_MAX = 256 (true for ext3).
308 * path name length = PATH_MAX = 4096
309 * LOV MD size max = EA_MAX = 24 * 2000
310 * (NB: 24 is size of lov_ost_data)
311 * LOV LOGCOOKIE size max = 32 * 2000
312 * (NB: 32 is size of llog_cookie)
313 * symlink: FNAME_MAX + PATH_MAX <- largest
314 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
315 * rename: FNAME_MAX + FNAME_MAX
316 * open: FNAME_MAX + EA_MAX
318 * MDS_MAXREQSIZE ~= 4736 bytes =
319 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
320 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
322 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
323 * except in the open case where there are a large number of OSTs in a LOV.
325 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
326 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
329 * MDS incoming request with LOV EA
330 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
332 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
333 362 + LOV_MAX_STRIPE_COUNT * 24)
335 * MDS outgoing reply with LOV EA
337 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
338 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
340 * but 2.4 or later MDS will never send reply with llog_cookie to any
341 * version client. This macro is defined for server side reply buffer size.
343 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
346 * This is the size of a maximum REINT_SETXATTR request:
348 * lustre_msg 56 (32 + 4 x 5 + 4)
350 * mdt_rec_setxattr 136
352 * name 256 (XATTR_NAME_MAX)
353 * value 65536 (XATTR_SIZE_MAX)
355 #define MDS_EA_MAXREQSIZE 66288
358 * These are the maximum request and reply sizes (rounded up to 1 KB
359 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
361 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
362 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
363 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
366 * The update request includes all of updates from the create, which might
367 * include linkea (4K maxim), together with other updates, we set it to 9K:
368 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
370 #define OUT_MAXREQSIZE (9 * 1024)
371 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
373 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
374 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
378 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
379 * However, we need to allocate a much larger buffer for it because LNet
380 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
381 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
382 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
383 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
384 * utilization is very low.
386 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
387 * reused until all requests fit in it have been processed and released,
388 * which means one long blocked request can prevent the rqbd be reused.
389 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
390 * from LNet with unused 65 KB, buffer utilization will be about 59%.
391 * Please check LU-2432 for details.
393 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
397 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
398 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
399 * extra bytes to each request buffer to improve buffer utilization rate.
401 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
404 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
405 #define FLD_MAXREQSIZE (160)
407 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
408 #define FLD_MAXREPSIZE (152)
409 #define FLD_BUFSIZE (1 << 12)
412 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
414 #define SEQ_MAXREQSIZE (160)
416 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
417 #define SEQ_MAXREPSIZE (152)
418 #define SEQ_BUFSIZE (1 << 12)
420 /** MGS threads must be >= 3, see bug 22458 comment #28 */
421 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
422 #define MGS_NTHRS_MAX 32
425 #define MGS_BUFSIZE (8 * 1024)
426 #define MGS_MAXREQSIZE (7 * 1024)
427 #define MGS_MAXREPSIZE (9 * 1024)
430 * OSS threads constants:
432 * Given 8 as factor and 64 as base threads number
435 * On 8-core server configured to 2 partitions, we will have
436 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
439 * On 32-core machine configured to 4 partitions, we will have
440 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
441 * will be 112 * 4 = 448.
444 * On 64-core machine configured to 4 partitions, we will have
445 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
446 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
447 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
448 * for each partition.
450 * So we can see that with these constants, threads number wil be at the
451 * similar level of old versions, unless the server has many cores.
453 /* depress threads factor for VM with small memory size */
454 #define OSS_THR_FACTOR min_t(int, 8, \
455 NUM_CACHEPAGES >> (28 - PAGE_CACHE_SHIFT))
456 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
457 #define OSS_NTHRS_BASE 64
458 #define OSS_NTHRS_MAX 512
460 /* threads for handling "create" request */
461 #define OSS_CR_THR_FACTOR 1
462 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
463 #define OSS_CR_NTHRS_BASE 8
464 #define OSS_CR_NTHRS_MAX 64
467 * OST_IO_MAXREQSIZE ~=
468 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
469 * DT_MAX_BRW_PAGES * niobuf_remote
471 * - single object with 16 pages is 512 bytes
472 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
473 * - Must be a multiple of 1024
474 * - actual size is about 18K
476 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
477 sizeof(struct ptlrpc_body) + \
478 sizeof(struct obdo) + \
479 sizeof(struct obd_ioobj) + \
480 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
482 * FIEMAP request can be 4K+ for now
484 #define OST_MAXREQSIZE (16 * 1024)
485 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
486 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
488 #define OST_MAXREPSIZE (9 * 1024)
489 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
492 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
493 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
495 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
496 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
498 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
500 /* Macro to hide a typecast. */
501 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
503 struct ptlrpc_replay_async_args {
509 * Structure to single define portal connection.
511 struct ptlrpc_connection {
512 /** linkage for connections hash table */
513 struct hlist_node c_hash;
514 /** Our own lnet nid for this connection */
516 /** Remote side nid for this connection */
517 lnet_process_id_t c_peer;
518 /** UUID of the other side */
519 struct obd_uuid c_remote_uuid;
520 /** reference counter for this connection */
524 /** Client definition for PortalRPC */
525 struct ptlrpc_client {
526 /** What lnet portal does this client send messages to by default */
527 __u32 cli_request_portal;
528 /** What portal do we expect replies on */
529 __u32 cli_reply_portal;
530 /** Name of the client */
534 /** state flags of requests */
535 /* XXX only ones left are those used by the bulk descs as well! */
536 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
537 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
539 #define REQ_MAX_ACK_LOCKS 8
541 union ptlrpc_async_args {
543 * Scratchpad for passing args to completion interpreter. Users
544 * cast to the struct of their choosing, and CLASSERT that this is
545 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
546 * a pointer to it here. The pointer_arg ensures this struct is at
547 * least big enough for that.
549 void *pointer_arg[11];
553 struct ptlrpc_request_set;
554 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
555 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
558 * Definition of request set structure.
559 * Request set is a list of requests (not necessary to the same target) that
560 * once populated with RPCs could be sent in parallel.
561 * There are two kinds of request sets. General purpose and with dedicated
562 * serving thread. Example of the latter is ptlrpcd set.
563 * For general purpose sets once request set started sending it is impossible
564 * to add new requests to such set.
565 * Provides a way to call "completion callbacks" when all requests in the set
568 struct ptlrpc_request_set {
569 atomic_t set_refcount;
570 /** number of in queue requests */
571 atomic_t set_new_count;
572 /** number of uncompleted requests */
573 atomic_t set_remaining;
574 /** wait queue to wait on for request events */
575 wait_queue_head_t set_waitq;
576 wait_queue_head_t *set_wakeup_ptr;
577 /** List of requests in the set */
578 struct list_head set_requests;
580 * List of completion callbacks to be called when the set is completed
581 * This is only used if \a set_interpret is NULL.
582 * Links struct ptlrpc_set_cbdata.
584 struct list_head set_cblist;
585 /** Completion callback, if only one. */
586 set_interpreter_func set_interpret;
587 /** opaq argument passed to completion \a set_interpret callback. */
590 * Lock for \a set_new_requests manipulations
591 * locked so that any old caller can communicate requests to
592 * the set holder who can then fold them into the lock-free set
594 spinlock_t set_new_req_lock;
595 /** List of new yet unsent requests. Only used with ptlrpcd now. */
596 struct list_head set_new_requests;
598 /** rq_status of requests that have been freed already */
600 /** Additional fields used by the flow control extension */
601 /** Maximum number of RPCs in flight */
602 int set_max_inflight;
603 /** Callback function used to generate RPCs */
604 set_producer_func set_producer;
605 /** opaq argument passed to the producer callback */
606 void *set_producer_arg;
610 * Description of a single ptrlrpc_set callback
612 struct ptlrpc_set_cbdata {
613 /** List linkage item */
614 struct list_head psc_item;
615 /** Pointer to interpreting function */
616 set_interpreter_func psc_interpret;
617 /** Opaq argument to pass to the callback */
621 struct ptlrpc_bulk_desc;
622 struct ptlrpc_service_part;
623 struct ptlrpc_service;
626 * ptlrpc callback & work item stuff
628 struct ptlrpc_cb_id {
629 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
630 void *cbid_arg; /* additional arg */
633 /** Maximum number of locks to fit into reply state */
634 #define RS_MAX_LOCKS 8
638 * Structure to define reply state on the server
639 * Reply state holds various reply message information. Also for "difficult"
640 * replies (rep-ack case) we store the state after sending reply and wait
641 * for the client to acknowledge the reception. In these cases locks could be
642 * added to the state for replay/failover consistency guarantees.
644 struct ptlrpc_reply_state {
645 /** Callback description */
646 struct ptlrpc_cb_id rs_cb_id;
647 /** Linkage for list of all reply states in a system */
648 struct list_head rs_list;
649 /** Linkage for list of all reply states on same export */
650 struct list_head rs_exp_list;
651 /** Linkage for list of all reply states for same obd */
652 struct list_head rs_obd_list;
654 struct list_head rs_debug_list;
656 /** A spinlock to protect the reply state flags */
658 /** Reply state flags */
659 unsigned long rs_difficult:1; /* ACK/commit stuff */
660 unsigned long rs_no_ack:1; /* no ACK, even for
661 difficult requests */
662 unsigned long rs_scheduled:1; /* being handled? */
663 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
664 unsigned long rs_handled:1; /* been handled yet? */
665 unsigned long rs_on_net:1; /* reply_out_callback pending? */
666 unsigned long rs_prealloc:1; /* rs from prealloc list */
667 unsigned long rs_committed:1;/* the transaction was committed
668 and the rs was dispatched
669 by ptlrpc_commit_replies */
670 /** Size of the state */
674 /** Transaction number */
678 struct obd_export *rs_export;
679 struct ptlrpc_service_part *rs_svcpt;
680 /** Lnet metadata handle for the reply */
681 lnet_handle_md_t rs_md_h;
682 atomic_t rs_refcount;
684 /** Context for the sevice thread */
685 struct ptlrpc_svc_ctx *rs_svc_ctx;
686 /** Reply buffer (actually sent to the client), encoded if needed */
687 struct lustre_msg *rs_repbuf; /* wrapper */
688 /** Size of the reply buffer */
689 int rs_repbuf_len; /* wrapper buf length */
690 /** Size of the reply message */
691 int rs_repdata_len; /* wrapper msg length */
693 * Actual reply message. Its content is encrupted (if needed) to
694 * produce reply buffer for actual sending. In simple case
695 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
697 struct lustre_msg *rs_msg; /* reply message */
699 /** Number of locks awaiting client ACK */
701 /** Handles of locks awaiting client reply ACK */
702 struct lustre_handle rs_locks[RS_MAX_LOCKS];
703 /** Lock modes of locks in \a rs_locks */
704 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
707 struct ptlrpc_thread;
711 RQ_PHASE_NEW = 0xebc0de00,
712 RQ_PHASE_RPC = 0xebc0de01,
713 RQ_PHASE_BULK = 0xebc0de02,
714 RQ_PHASE_INTERPRET = 0xebc0de03,
715 RQ_PHASE_COMPLETE = 0xebc0de04,
716 RQ_PHASE_UNREGISTERING = 0xebc0de05,
717 RQ_PHASE_UNDEFINED = 0xebc0de06
720 /** Type of request interpreter call-back */
721 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
722 struct ptlrpc_request *req,
724 /** Type of request resend call-back */
725 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
729 * Definition of request pool structure.
730 * The pool is used to store empty preallocated requests for the case
731 * when we would actually need to send something without performing
732 * any allocations (to avoid e.g. OOM).
734 struct ptlrpc_request_pool {
735 /** Locks the list */
737 /** list of ptlrpc_request structs */
738 struct list_head prp_req_list;
739 /** Maximum message size that would fit into a rquest from this pool */
741 /** Function to allocate more requests for this pool */
742 void (*prp_populate)(struct ptlrpc_request_pool *, int);
751 * \defgroup nrs Network Request Scheduler
754 struct ptlrpc_nrs_policy;
755 struct ptlrpc_nrs_resource;
756 struct ptlrpc_nrs_request;
759 * NRS control operations.
761 * These are common for all policies.
763 enum ptlrpc_nrs_ctl {
765 * Not a valid opcode.
767 PTLRPC_NRS_CTL_INVALID,
769 * Activate the policy.
771 PTLRPC_NRS_CTL_START,
773 * Reserved for multiple primary policies, which may be a possibility
778 * Policies can start using opcodes from this value and onwards for
779 * their own purposes; the assigned value itself is arbitrary.
781 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
785 * ORR policy operations
788 NRS_CTL_ORR_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
789 NRS_CTL_ORR_WR_QUANTUM,
790 NRS_CTL_ORR_RD_OFF_TYPE,
791 NRS_CTL_ORR_WR_OFF_TYPE,
792 NRS_CTL_ORR_RD_SUPP_REQ,
793 NRS_CTL_ORR_WR_SUPP_REQ,
797 * NRS policy operations.
799 * These determine the behaviour of a policy, and are called in response to
802 struct ptlrpc_nrs_pol_ops {
804 * Called during policy registration; this operation is optional.
806 * \param[in,out] policy The policy being initialized
808 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
810 * Called during policy unregistration; this operation is optional.
812 * \param[in,out] policy The policy being unregistered/finalized
814 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
816 * Called when activating a policy via lprocfs; policies allocate and
817 * initialize their resources here; this operation is optional.
819 * \param[in,out] policy The policy being started
820 * \param[in,out] arg A generic char buffer
822 * \see nrs_policy_start_locked()
824 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy,
827 * Called when deactivating a policy via lprocfs; policies deallocate
828 * their resources here; this operation is optional
830 * \param[in,out] policy The policy being stopped
832 * \see nrs_policy_stop0()
834 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
836 * Used for policy-specific operations; i.e. not generic ones like
837 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
838 * to an ioctl; this operation is optional.
840 * \param[in,out] policy The policy carrying out operation \a opc
841 * \param[in] opc The command operation being carried out
842 * \param[in,out] arg An generic buffer for communication between the
843 * user and the control operation
848 * \see ptlrpc_nrs_policy_control()
850 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
851 enum ptlrpc_nrs_ctl opc, void *arg);
854 * Called when obtaining references to the resources of the resource
855 * hierarchy for a request that has arrived for handling at the PTLRPC
856 * service. Policies should return -ve for requests they do not wish
857 * to handle. This operation is mandatory.
859 * \param[in,out] policy The policy we're getting resources for.
860 * \param[in,out] nrq The request we are getting resources for.
861 * \param[in] parent The parent resource of the resource being
862 * requested; set to NULL if none.
863 * \param[out] resp The resource is to be returned here; the
864 * fallback policy in an NRS head should
865 * \e always return a non-NULL pointer value.
866 * \param[in] moving_req When set, signifies that this is an attempt
867 * to obtain resources for a request being moved
868 * to the high-priority NRS head by
869 * ldlm_lock_reorder_req().
870 * This implies two things:
871 * 1. We are under obd_export::exp_rpc_lock and
872 * so should not sleep.
873 * 2. We should not perform non-idempotent or can
874 * skip performing idempotent operations that
875 * were carried out when resources were first
876 * taken for the request when it was initialized
877 * in ptlrpc_nrs_req_initialize().
879 * \retval 0, +ve The level of the returned resource in the resource
880 * hierarchy; currently only 0 (for a non-leaf resource)
881 * and 1 (for a leaf resource) are supported by the
885 * \see ptlrpc_nrs_req_initialize()
886 * \see ptlrpc_nrs_hpreq_add_nolock()
887 * \see ptlrpc_nrs_req_hp_move()
889 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
890 struct ptlrpc_nrs_request *nrq,
891 const struct ptlrpc_nrs_resource *parent,
892 struct ptlrpc_nrs_resource **resp,
895 * Called when releasing references taken for resources in the resource
896 * hierarchy for the request; this operation is optional.
898 * \param[in,out] policy The policy the resource belongs to
899 * \param[in] res The resource to be freed
901 * \see ptlrpc_nrs_req_finalize()
902 * \see ptlrpc_nrs_hpreq_add_nolock()
903 * \see ptlrpc_nrs_req_hp_move()
905 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
906 const struct ptlrpc_nrs_resource *res);
909 * Obtains a request for handling from the policy, and optionally
910 * removes the request from the policy; this operation is mandatory.
912 * \param[in,out] policy The policy to poll
913 * \param[in] peek When set, signifies that we just want to
914 * examine the request, and not handle it, so the
915 * request is not removed from the policy.
916 * \param[in] force When set, it will force a policy to return a
917 * request if it has one queued.
919 * \retval NULL No request available for handling
920 * \retval valid-pointer The request polled for handling
922 * \see ptlrpc_nrs_req_get_nolock()
924 struct ptlrpc_nrs_request *
925 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
928 * Called when attempting to add a request to a policy for later
929 * handling; this operation is mandatory.
931 * \param[in,out] policy The policy on which to enqueue \a nrq
932 * \param[in,out] nrq The request to enqueue
937 * \see ptlrpc_nrs_req_add_nolock()
939 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
940 struct ptlrpc_nrs_request *nrq);
942 * Removes a request from the policy's set of pending requests. Normally
943 * called after a request has been polled successfully from the policy
944 * for handling; this operation is mandatory.
946 * \param[in,out] policy The policy the request \a nrq belongs to
947 * \param[in,out] nrq The request to dequeue
949 * \see ptlrpc_nrs_req_del_nolock()
951 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
952 struct ptlrpc_nrs_request *nrq);
954 * Called after the request being carried out. Could be used for
955 * job/resource control; this operation is optional.
957 * \param[in,out] policy The policy which is stopping to handle request
959 * \param[in,out] nrq The request
961 * \pre assert_spin_locked(&svcpt->scp_req_lock)
963 * \see ptlrpc_nrs_req_stop_nolock()
965 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
966 struct ptlrpc_nrs_request *nrq);
968 * Registers the policy's lprocfs interface with a PTLRPC service.
970 * \param[in] svc The service
975 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
977 * Unegisters the policy's lprocfs interface with a PTLRPC service.
979 * In cases of failed policy registration in
980 * \e ptlrpc_nrs_policy_register(), this function may be called for a
981 * service which has not registered the policy successfully, so
982 * implementations of this method should make sure their operations are
983 * safe in such cases.
985 * \param[in] svc The service
987 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
993 enum nrs_policy_flags {
995 * Fallback policy, use this flag only on a single supported policy per
996 * service. The flag cannot be used on policies that use
997 * \e PTLRPC_NRS_FL_REG_EXTERN
999 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
1001 * Start policy immediately after registering.
1003 PTLRPC_NRS_FL_REG_START = (1 << 1),
1005 * This is a policy registering from a module different to the one NRS
1006 * core ships in (currently ptlrpc).
1008 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1014 * Denotes whether an NRS instance is for handling normal or high-priority
1015 * RPCs, or whether an operation pertains to one or both of the NRS instances
1018 enum ptlrpc_nrs_queue_type {
1019 PTLRPC_NRS_QUEUE_REG = (1 << 0),
1020 PTLRPC_NRS_QUEUE_HP = (1 << 1),
1021 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
1027 * A PTLRPC service has at least one NRS head instance for handling normal
1028 * priority RPCs, and may optionally have a second NRS head instance for
1029 * handling high-priority RPCs. Each NRS head maintains a list of available
1030 * policies, of which one and only one policy is acting as the fallback policy,
1031 * and optionally a different policy may be acting as the primary policy. For
1032 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1033 * enqueue the RPC using the primary policy (if any). The fallback policy is
1034 * used in the following cases:
1035 * - when there was no primary policy in the
1036 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1038 * - when the primary policy that was at the
1039 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1040 * RPC was initialized, denoted it did not wish, or for some other reason was
1041 * not able to handle the request, by returning a non-valid NRS resource
1043 * - when the primary policy that was at the
1044 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1045 * RPC was initialized, fails later during the request enqueueing stage.
1047 * \see nrs_resource_get_safe()
1048 * \see nrs_request_enqueue()
1051 spinlock_t nrs_lock;
1052 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1054 * List of registered policies
1056 struct list_head nrs_policy_list;
1058 * List of policies with queued requests. Policies that have any
1059 * outstanding requests are queued here, and this list is queried
1060 * in a round-robin manner from NRS core when obtaining a request
1061 * for handling. This ensures that requests from policies that at some
1062 * point transition away from the
1063 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1065 struct list_head nrs_policy_queued;
1067 * Service partition for this NRS head
1069 struct ptlrpc_service_part *nrs_svcpt;
1071 * Primary policy, which is the preferred policy for handling RPCs
1073 struct ptlrpc_nrs_policy *nrs_policy_primary;
1075 * Fallback policy, which is the backup policy for handling RPCs
1077 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1079 * This NRS head handles either HP or regular requests
1081 enum ptlrpc_nrs_queue_type nrs_queue_type;
1083 * # queued requests from all policies in this NRS head
1085 unsigned long nrs_req_queued;
1087 * # scheduled requests from all policies in this NRS head
1089 unsigned long nrs_req_started;
1091 * # policies on this NRS
1093 unsigned nrs_num_pols;
1095 * This NRS head is in progress of starting a policy
1097 unsigned nrs_policy_starting:1;
1099 * In progress of shutting down the whole NRS head; used during
1102 unsigned nrs_stopping:1;
1104 * NRS policy is throttling reqeust
1106 unsigned nrs_throttling:1;
1109 #define NRS_POL_NAME_MAX 16
1111 struct ptlrpc_nrs_pol_desc;
1114 * Service compatibility predicate; this determines whether a policy is adequate
1115 * for handling RPCs of a particular PTLRPC service.
1117 * XXX:This should give the same result during policy registration and
1118 * unregistration, and for all partitions of a service; so the result should not
1119 * depend on temporal service or other properties, that may influence the
1122 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
1123 const struct ptlrpc_nrs_pol_desc *desc);
1125 struct ptlrpc_nrs_pol_conf {
1127 * Human-readable policy name
1129 char nc_name[NRS_POL_NAME_MAX];
1131 * NRS operations for this policy
1133 const struct ptlrpc_nrs_pol_ops *nc_ops;
1135 * Service compatibility predicate
1137 nrs_pol_desc_compat_t nc_compat;
1139 * Set for policies that support a single ptlrpc service, i.e. ones that
1140 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
1141 * depicts the name of the single service that such policies are
1144 const char *nc_compat_svc_name;
1146 * Owner module for this policy descriptor; policies registering from a
1147 * different module to the one the NRS framework is held within
1148 * (currently ptlrpc), should set this field to THIS_MODULE.
1150 struct module *nc_owner;
1152 * Policy registration flags; a bitmast of \e nrs_policy_flags
1158 * NRS policy registering descriptor
1160 * Is used to hold a description of a policy that can be passed to NRS core in
1161 * order to register the policy with NRS heads in different PTLRPC services.
1163 struct ptlrpc_nrs_pol_desc {
1165 * Human-readable policy name
1167 char pd_name[NRS_POL_NAME_MAX];
1169 * Link into nrs_core::nrs_policies
1171 struct list_head pd_list;
1173 * NRS operations for this policy
1175 const struct ptlrpc_nrs_pol_ops *pd_ops;
1177 * Service compatibility predicate
1179 nrs_pol_desc_compat_t pd_compat;
1181 * Set for policies that are compatible with only one PTLRPC service.
1183 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
1185 const char *pd_compat_svc_name;
1187 * Owner module for this policy descriptor.
1189 * We need to hold a reference to the module whenever we might make use
1190 * of any of the module's contents, i.e.
1191 * - If one or more instances of the policy are at a state where they
1192 * might be handling a request, i.e.
1193 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1194 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
1195 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
1196 * is taken on the module when
1197 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
1198 * becomes 0, so that we hold only one reference to the module maximum
1201 * We do not need to hold a reference to the module, even though we
1202 * might use code and data from the module, in the following cases:
1203 * - During external policy registration, because this should happen in
1204 * the module's init() function, in which case the module is safe from
1205 * removal because a reference is being held on the module by the
1206 * kernel, and iirc kmod (and I guess module-init-tools also) will
1207 * serialize any racing processes properly anyway.
1208 * - During external policy unregistration, because this should happen
1209 * in a module's exit() function, and any attempts to start a policy
1210 * instance would need to take a reference on the module, and this is
1211 * not possible once we have reached the point where the exit()
1212 * handler is called.
1213 * - During service registration and unregistration, as service setup
1214 * and cleanup, and policy registration, unregistration and policy
1215 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
1216 * as long as users adhere to the convention of registering policies
1217 * in init() and unregistering them in module exit() functions, there
1218 * should not be a race between these operations.
1219 * - During any policy-specific lprocfs operations, because a reference
1220 * is held by the kernel on a proc entry that has been entered by a
1221 * syscall, so as long as proc entries are removed during unregistration time,
1222 * then unregistration and lprocfs operations will be properly
1225 struct module *pd_owner;
1227 * Bitmask of \e nrs_policy_flags
1231 * # of references on this descriptor
1239 * Policies transition from one state to the other during their lifetime
1241 enum ptlrpc_nrs_pol_state {
1243 * Not a valid policy state.
1245 NRS_POL_STATE_INVALID,
1247 * Policies are at this state either at the start of their life, or
1248 * transition here when the user selects a different policy to act
1249 * as the primary one.
1251 NRS_POL_STATE_STOPPED,
1253 * Policy is progress of stopping
1255 NRS_POL_STATE_STOPPING,
1257 * Policy is in progress of starting
1259 NRS_POL_STATE_STARTING,
1261 * A policy is in this state in two cases:
1262 * - it is the fallback policy, which is always in this state.
1263 * - it has been activated by the user; i.e. it is the primary policy,
1265 NRS_POL_STATE_STARTED,
1269 * NRS policy information
1271 * Used for obtaining information for the status of a policy via lprocfs
1273 struct ptlrpc_nrs_pol_info {
1277 char pi_name[NRS_POL_NAME_MAX];
1279 * Current policy state
1281 enum ptlrpc_nrs_pol_state pi_state;
1283 * # RPCs enqueued for later dispatching by the policy
1287 * # RPCs started for dispatch by the policy
1289 long pi_req_started;
1291 * Is this a fallback policy?
1293 unsigned pi_fallback:1;
1299 * There is one instance of this for each policy in each NRS head of each
1300 * PTLRPC service partition.
1302 struct ptlrpc_nrs_policy {
1304 * Linkage into the NRS head's list of policies,
1305 * ptlrpc_nrs:nrs_policy_list
1307 struct list_head pol_list;
1309 * Linkage into the NRS head's list of policies with enqueued
1310 * requests ptlrpc_nrs:nrs_policy_queued
1312 struct list_head pol_list_queued;
1314 * Current state of this policy
1316 enum ptlrpc_nrs_pol_state pol_state;
1318 * Bitmask of nrs_policy_flags
1322 * # RPCs enqueued for later dispatching by the policy
1324 long pol_req_queued;
1326 * # RPCs started for dispatch by the policy
1328 long pol_req_started;
1330 * Usage Reference count taken on the policy instance
1334 * The NRS head this policy has been created at
1336 struct ptlrpc_nrs *pol_nrs;
1338 * Private policy data; varies by policy type
1342 * Policy descriptor for this policy instance.
1344 struct ptlrpc_nrs_pol_desc *pol_desc;
1350 * Resources are embedded into two types of NRS entities:
1351 * - Inside NRS policies, in the policy's private data in
1352 * ptlrpc_nrs_policy::pol_private
1353 * - In objects that act as prime-level scheduling entities in different NRS
1354 * policies; e.g. on a policy that performs round robin or similar order
1355 * scheduling across client NIDs, there would be one NRS resource per unique
1356 * client NID. On a policy which performs round robin scheduling across
1357 * backend filesystem objects, there would be one resource associated with
1358 * each of the backend filesystem objects partaking in the scheduling
1359 * performed by the policy.
1361 * NRS resources share a parent-child relationship, in which resources embedded
1362 * in policy instances are the parent entities, with all scheduling entities
1363 * a policy schedules across being the children, thus forming a simple resource
1364 * hierarchy. This hierarchy may be extended with one or more levels in the
1365 * future if the ability to have more than one primary policy is added.
1367 * Upon request initialization, references to the then active NRS policies are
1368 * taken and used to later handle the dispatching of the request with one of
1371 * \see nrs_resource_get_safe()
1372 * \see ptlrpc_nrs_req_add()
1374 struct ptlrpc_nrs_resource {
1376 * This NRS resource's parent; is NULL for resources embedded in NRS
1377 * policy instances; i.e. those are top-level ones.
1379 struct ptlrpc_nrs_resource *res_parent;
1381 * The policy associated with this resource.
1383 struct ptlrpc_nrs_policy *res_policy;
1396 * This policy is a logical wrapper around previous, non-NRS functionality.
1397 * It dispatches RPCs in the same order as they arrive from the network. This
1398 * policy is currently used as the fallback policy, and the only enabled policy
1399 * on all NRS heads of all PTLRPC service partitions.
1404 * Private data structure for the FIFO policy
1406 struct nrs_fifo_head {
1408 * Resource object for policy instance.
1410 struct ptlrpc_nrs_resource fh_res;
1412 * List of queued requests.
1414 struct list_head fh_list;
1416 * For debugging purposes.
1421 struct nrs_fifo_req {
1422 struct list_head fr_list;
1431 * CRR-N, Client Round Robin over NIDs
1436 * private data structure for CRR-N NRS
1438 struct nrs_crrn_net {
1439 struct ptlrpc_nrs_resource cn_res;
1440 cfs_binheap_t *cn_binheap;
1441 cfs_hash_t *cn_cli_hash;
1443 * Used when a new scheduling round commences, in order to synchronize
1444 * all clients with the new round number.
1448 * Determines the relevant ordering amongst request batches within a
1453 * Round Robin quantum; the maximum number of RPCs that each request
1454 * batch for each client can have in a scheduling round.
1460 * Object representing a client in CRR-N, as identified by its NID
1462 struct nrs_crrn_client {
1463 struct ptlrpc_nrs_resource cc_res;
1464 struct hlist_node cc_hnode;
1467 * The round number against which this client is currently scheduling
1472 * The sequence number used for requests scheduled by this client during
1473 * the current round number.
1478 * Round Robin quantum; the maximum number of RPCs the client is allowed
1479 * to schedule in a single batch of each round.
1483 * # of pending requests for this client, on all existing rounds
1489 * CRR-N NRS request definition
1491 struct nrs_crrn_req {
1493 * Round number for this request; shared with all other requests in the
1498 * Sequence number for this request; shared with all other requests in
1505 * CRR-N policy operations.
1509 * Read the RR quantum size of a CRR-N policy.
1511 NRS_CTL_CRRN_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
1513 * Write the RR quantum size of a CRR-N policy.
1515 NRS_CTL_CRRN_WR_QUANTUM,
1523 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
1528 * Lower and upper byte offsets of a brw RPC
1530 struct nrs_orr_req_range {
1536 * RPC types supported by the ORR/TRR policies
1539 NOS_OST_READ = (1 << 0),
1540 NOS_OST_WRITE = (1 << 1),
1541 NOS_OST_RW = (NOS_OST_READ | NOS_OST_WRITE),
1543 * Default value for policies.
1545 NOS_DFLT = NOS_OST_READ
1549 * As unique keys for grouping RPCs together, we use the object's OST FID for
1550 * the ORR policy, and the OST index for the TRR policy.
1552 * XXX: We waste some space for TRR policy instances by using a union, but it
1553 * allows to consolidate some of the code between ORR and TRR, and these
1554 * policies will probably eventually merge into one anyway.
1556 struct nrs_orr_key {
1558 /** object FID for ORR */
1559 struct lu_fid ok_fid;
1560 /** OST index for TRR */
1566 * The largest base string for unique hash/slab object names is
1567 * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT
1568 * id number, so this _should_ be more than enough for the maximum number of
1569 * CPTs on any system. If it does happen that this statement is incorrect,
1570 * nrs_orr_genobjname() will inevitably yield a non-unique name and cause
1571 * kmem_cache_create() to complain (on Linux), so the erroneous situation
1572 * will hopefully not go unnoticed.
1574 #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3)
1577 * private data structure for ORR and TRR NRS
1579 struct nrs_orr_data {
1580 struct ptlrpc_nrs_resource od_res;
1581 cfs_binheap_t *od_binheap;
1582 cfs_hash_t *od_obj_hash;
1583 struct kmem_cache *od_cache;
1585 * Used when a new scheduling round commences, in order to synchronize
1586 * all object or OST batches with the new round number.
1590 * Determines the relevant ordering amongst request batches within a
1595 * RPC types that are currently supported.
1597 enum nrs_orr_supp od_supp;
1599 * Round Robin quantum; the maxium number of RPCs that each request
1600 * batch for each object or OST can have in a scheduling round.
1604 * Whether to use physical disk offsets or logical file offsets.
1608 * XXX: We need to provide a persistently allocated string to hold
1609 * unique object names for this policy, since in currently supported
1610 * versions of Linux by Lustre, kmem_cache_create() just sets a pointer
1611 * to the name string provided. kstrdup() is used in the version of
1612 * kmeme_cache_create() in current Linux mainline, so we may be able to
1613 * remove this in the future.
1615 char od_objname[NRS_ORR_OBJ_NAME_MAX];
1619 * Represents a backend-fs object or OST in the ORR and TRR policies
1622 struct nrs_orr_object {
1623 struct ptlrpc_nrs_resource oo_res;
1624 struct hlist_node oo_hnode;
1626 * The round number against which requests are being scheduled for this
1631 * The sequence number used for requests scheduled for this object or
1632 * OST during the current round number.
1636 * The key of the object or OST for which this structure instance is
1639 struct nrs_orr_key oo_key;
1642 * Round Robin quantum; the maximum number of RPCs that are allowed to
1643 * be scheduled for the object or OST in a single batch of each round.
1647 * # of pending requests for this object or OST, on all existing rounds
1653 * ORR/TRR NRS request definition
1655 struct nrs_orr_req {
1657 * The offset range this request covers
1659 struct nrs_orr_req_range or_range;
1661 * Round number for this request; shared with all other requests in the
1666 * Sequence number for this request; shared with all other requests in
1671 * For debugging purposes.
1673 struct nrs_orr_key or_key;
1675 * An ORR policy instance has filled in request information while
1676 * enqueueing the request on the service partition's regular NRS head.
1678 unsigned int or_orr_set:1;
1680 * A TRR policy instance has filled in request information while
1681 * enqueueing the request on the service partition's regular NRS head.
1683 unsigned int or_trr_set:1;
1685 * Request offset ranges have been filled in with logical offset
1688 unsigned int or_logical_set:1;
1690 * Request offset ranges have been filled in with physical offset
1693 unsigned int or_physical_set:1;
1698 #include <lustre_nrs_tbf.h>
1703 * Instances of this object exist embedded within ptlrpc_request; the main
1704 * purpose of this object is to hold references to the request's resources
1705 * for the lifetime of the request, and to hold properties that policies use
1706 * use for determining the request's scheduling priority.
1708 struct ptlrpc_nrs_request {
1710 * The request's resource hierarchy.
1712 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1714 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1715 * policy that was used to enqueue the request.
1717 * \see nrs_request_enqueue()
1719 unsigned nr_res_idx;
1720 unsigned nr_initialized:1;
1721 unsigned nr_enqueued:1;
1722 unsigned nr_started:1;
1723 unsigned nr_finalized:1;
1724 cfs_binheap_node_t nr_node;
1727 * Policy-specific fields, used for determining a request's scheduling
1728 * priority, and other supporting functionality.
1732 * Fields for the FIFO policy
1734 struct nrs_fifo_req fifo;
1736 * CRR-N request defintion
1738 struct nrs_crrn_req crr;
1739 /** ORR and TRR share the same request definition */
1740 struct nrs_orr_req orr;
1742 * TBF request definition
1744 struct nrs_tbf_req tbf;
1747 * Externally-registering policies may want to use this to allocate
1748 * their own request properties.
1756 * Basic request prioritization operations structure.
1757 * The whole idea is centered around locks and RPCs that might affect locks.
1758 * When a lock is contended we try to give priority to RPCs that might lead
1759 * to fastest release of that lock.
1760 * Currently only implemented for OSTs only in a way that makes all
1761 * IO and truncate RPCs that are coming from a locked region where a lock is
1762 * contended a priority over other requests.
1764 struct ptlrpc_hpreq_ops {
1766 * Check if the lock handle of the given lock is the same as
1767 * taken from the request.
1769 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1771 * Check if the request is a high priority one.
1773 int (*hpreq_check)(struct ptlrpc_request *);
1775 * Called after the request has been handled.
1777 void (*hpreq_fini)(struct ptlrpc_request *);
1780 struct ptlrpc_cli_req {
1781 /** For bulk requests on client only: bulk descriptor */
1782 struct ptlrpc_bulk_desc *cr_bulk;
1783 /** optional time limit for send attempts */
1784 cfs_duration_t cr_delay_limit;
1785 /** time request was first queued */
1786 cfs_time_t cr_queued_time;
1787 /** request sent timeval */
1788 struct timeval cr_sent_tv;
1789 /** time for request really sent out */
1791 /** when req reply unlink must finish. */
1792 time_t cr_reply_deadline;
1793 /** when req bulk unlink must finish. */
1794 time_t cr_bulk_deadline;
1795 /** Portal to which this request would be sent */
1797 /** Portal where to wait for reply and where reply would be sent */
1799 /** request resending number */
1800 unsigned int cr_resend_nr;
1801 /** What was import generation when this request was sent */
1803 enum lustre_imp_state cr_send_state;
1804 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1805 wait_queue_head_t cr_set_waitq;
1806 /** Link item for request set lists */
1807 struct list_head cr_set_chain;
1808 /** link to waited ctx */
1809 struct list_head cr_ctx_chain;
1811 /** client's half ctx */
1812 struct ptlrpc_cli_ctx *cr_cli_ctx;
1813 /** Link back to the request set */
1814 struct ptlrpc_request_set *cr_set;
1815 /** outgoing request MD handle */
1816 lnet_handle_md_t cr_req_md_h;
1817 /** request-out callback parameter */
1818 struct ptlrpc_cb_id cr_req_cbid;
1819 /** incoming reply MD handle */
1820 lnet_handle_md_t cr_reply_md_h;
1821 wait_queue_head_t cr_reply_waitq;
1822 /** reply callback parameter */
1823 struct ptlrpc_cb_id cr_reply_cbid;
1824 /** Async completion handler, called when reply is received */
1825 ptlrpc_interpterer_t cr_reply_interp;
1826 /** Resend handler, called when request is resend to update RPC data */
1827 ptlrpc_resend_cb_t cr_resend_cb;
1828 /** Async completion context */
1829 union ptlrpc_async_args cr_async_args;
1830 /** Opaq data for replay and commit callbacks. */
1833 * Commit callback, called when request is committed and about to be
1836 void (*cr_commit_cb)(struct ptlrpc_request *);
1837 /** Replay callback, called after request is replayed at recovery */
1838 void (*cr_replay_cb)(struct ptlrpc_request *);
1841 /** client request member alias */
1842 /* NB: these alias should NOT be used by any new code, instead they should
1843 * be removed step by step to avoid potential abuse */
1844 #define rq_bulk rq_cli.cr_bulk
1845 #define rq_delay_limit rq_cli.cr_delay_limit
1846 #define rq_queued_time rq_cli.cr_queued_time
1847 #define rq_sent_tv rq_cli.cr_sent_tv
1848 #define rq_real_sent rq_cli.cr_sent_out
1849 #define rq_reply_deadline rq_cli.cr_reply_deadline
1850 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
1851 #define rq_nr_resend rq_cli.cr_resend_nr
1852 #define rq_request_portal rq_cli.cr_req_ptl
1853 #define rq_reply_portal rq_cli.cr_rep_ptl
1854 #define rq_import_generation rq_cli.cr_imp_gen
1855 #define rq_send_state rq_cli.cr_send_state
1856 #define rq_set_chain rq_cli.cr_set_chain
1857 #define rq_ctx_chain rq_cli.cr_ctx_chain
1858 #define rq_set rq_cli.cr_set
1859 #define rq_set_waitq rq_cli.cr_set_waitq
1860 #define rq_cli_ctx rq_cli.cr_cli_ctx
1861 #define rq_req_md_h rq_cli.cr_req_md_h
1862 #define rq_req_cbid rq_cli.cr_req_cbid
1863 #define rq_reply_md_h rq_cli.cr_reply_md_h
1864 #define rq_reply_waitq rq_cli.cr_reply_waitq
1865 #define rq_reply_cbid rq_cli.cr_reply_cbid
1866 #define rq_interpret_reply rq_cli.cr_reply_interp
1867 #define rq_resend_cb rq_cli.cr_resend_cb
1868 #define rq_async_args rq_cli.cr_async_args
1869 #define rq_cb_data rq_cli.cr_cb_data
1870 #define rq_commit_cb rq_cli.cr_commit_cb
1871 #define rq_replay_cb rq_cli.cr_replay_cb
1873 struct ptlrpc_srv_req {
1874 /** initial thread servicing this request */
1875 struct ptlrpc_thread *sr_svc_thread;
1877 * Server side list of incoming unserved requests sorted by arrival
1878 * time. Traversed from time to time to notice about to expire
1879 * requests and sent back "early replies" to clients to let them
1880 * know server is alive and well, just very busy to service their
1883 struct list_head sr_timed_list;
1884 /** server-side per-export list */
1885 struct list_head sr_exp_list;
1886 /** server-side history, used for debuging purposes. */
1887 struct list_head sr_hist_list;
1888 /** history sequence # */
1890 /** the index of service's srv_at_array into which request is linked */
1894 /** authed uid mapped to */
1895 uid_t sr_auth_mapped_uid;
1896 /** RPC is generated from what part of Lustre */
1897 enum lustre_sec_part sr_sp_from;
1898 /** request session context */
1899 struct lu_context sr_ses;
1903 /** stub for NRS request */
1904 struct ptlrpc_nrs_request sr_nrq;
1906 /** request arrival time */
1907 struct timeval sr_arrival_time;
1908 /** server's half ctx */
1909 struct ptlrpc_svc_ctx *sr_svc_ctx;
1910 /** (server side), pointed directly into req buffer */
1911 struct ptlrpc_user_desc *sr_user_desc;
1912 /** separated reply state */
1913 struct ptlrpc_reply_state *sr_reply_state;
1914 /** server-side hp handlers */
1915 struct ptlrpc_hpreq_ops *sr_ops;
1916 /** incoming request buffer */
1917 struct ptlrpc_request_buffer_desc *sr_rqbd;
1920 /** server request member alias */
1921 /* NB: these alias should NOT be used by any new code, instead they should
1922 * be removed step by step to avoid potential abuse */
1923 #define rq_svc_thread rq_srv.sr_svc_thread
1924 #define rq_timed_list rq_srv.sr_timed_list
1925 #define rq_exp_list rq_srv.sr_exp_list
1926 #define rq_history_list rq_srv.sr_hist_list
1927 #define rq_history_seq rq_srv.sr_hist_seq
1928 #define rq_at_index rq_srv.sr_at_index
1929 #define rq_auth_uid rq_srv.sr_auth_uid
1930 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
1931 #define rq_sp_from rq_srv.sr_sp_from
1932 #define rq_session rq_srv.sr_ses
1933 #define rq_nrq rq_srv.sr_nrq
1934 #define rq_arrival_time rq_srv.sr_arrival_time
1935 #define rq_reply_state rq_srv.sr_reply_state
1936 #define rq_svc_ctx rq_srv.sr_svc_ctx
1937 #define rq_user_desc rq_srv.sr_user_desc
1938 #define rq_ops rq_srv.sr_ops
1939 #define rq_rqbd rq_srv.sr_rqbd
1942 * Represents remote procedure call.
1944 * This is a staple structure used by everybody wanting to send a request
1947 struct ptlrpc_request {
1948 /* Request type: one of PTL_RPC_MSG_* */
1950 /** Result of request processing */
1953 * Linkage item through which this request is included into
1954 * sending/delayed lists on client and into rqbd list on server
1956 struct list_head rq_list;
1957 /** Lock to protect request flags and some other important bits, like
1961 /** client-side flags are serialized by rq_lock @{ */
1962 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1963 rq_timedout:1, rq_resend:1, rq_restart:1,
1965 * when ->rq_replay is set, request is kept by the client even
1966 * after server commits corresponding transaction. This is
1967 * used for operations that require sequence of multiple
1968 * requests to be replayed. The only example currently is file
1969 * open/close. When last request in such a sequence is
1970 * committed, ->rq_replay is cleared on all requests in the
1974 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1975 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1977 rq_req_unlinked:1, /* unlinked request buffer from lnet */
1978 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
1979 rq_memalloc:1, /* req originated from "kswapd" */
1981 rq_reply_truncated:1,
1982 /** whether the "rq_set" is a valid one */
1984 rq_generation_set:1,
1985 /** do not resend request on -EINPROGRESS */
1986 rq_no_retry_einprogress:1,
1987 /* allow the req to be sent if the import is in recovery
1990 /* bulk request, sent to server, but uncommitted */
1994 /** server-side flags @{ */
1996 rq_hp:1, /**< high priority RPC */
1997 rq_at_linked:1, /**< link into service's srv_at_array */
1998 rq_packed_final:1; /**< packed final reply */
2001 /** one of RQ_PHASE_* */
2002 enum rq_phase rq_phase;
2003 /** one of RQ_PHASE_* to be used next */
2004 enum rq_phase rq_next_phase;
2006 * client-side refcount for SENT race, server-side refcounf
2007 * for multiple replies
2009 atomic_t rq_refcount;
2012 * !rq_truncate : # reply bytes actually received,
2013 * rq_truncate : required repbuf_len for resend
2015 int rq_nob_received;
2016 /** Request length */
2020 /** Pool if request is from preallocated list */
2021 struct ptlrpc_request_pool *rq_pool;
2022 /** Request message - what client sent */
2023 struct lustre_msg *rq_reqmsg;
2024 /** Reply message - server response */
2025 struct lustre_msg *rq_repmsg;
2026 /** Transaction number */
2031 * List item to for replay list. Not yet commited requests get linked
2033 * Also see \a rq_replay comment above.
2034 * It's also link chain on obd_export::exp_req_replay_queue
2036 struct list_head rq_replay_list;
2037 /** non-shared members for client & server request*/
2039 struct ptlrpc_cli_req rq_cli;
2040 struct ptlrpc_srv_req rq_srv;
2043 * security and encryption data
2045 /** description of flavors for client & server */
2046 struct sptlrpc_flavor rq_flvr;
2048 /* client/server security flags */
2050 rq_ctx_init:1, /* context initiation */
2051 rq_ctx_fini:1, /* context destroy */
2052 rq_bulk_read:1, /* request bulk read */
2053 rq_bulk_write:1, /* request bulk write */
2054 /* server authentication flags */
2055 rq_auth_gss:1, /* authenticated by gss */
2056 rq_auth_remote:1, /* authed as remote user */
2057 rq_auth_usr_root:1, /* authed as root */
2058 rq_auth_usr_mdt:1, /* authed as mdt */
2059 rq_auth_usr_ost:1, /* authed as ost */
2060 /* security tfm flags */
2063 /* doesn't expect reply FIXME */
2065 rq_pill_init:1, /* pill initialized */
2066 rq_srv_req:1; /* server request */
2069 /** various buffer pointers */
2070 struct lustre_msg *rq_reqbuf; /**< req wrapper */
2071 char *rq_repbuf; /**< rep buffer */
2072 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
2073 /** only in priv mode */
2074 struct lustre_msg *rq_clrbuf;
2075 int rq_reqbuf_len; /* req wrapper buf len */
2076 int rq_reqdata_len; /* req wrapper msg len */
2077 int rq_repbuf_len; /* rep buffer len */
2078 int rq_repdata_len; /* rep wrapper msg len */
2079 int rq_clrbuf_len; /* only in priv mode */
2080 int rq_clrdata_len; /* only in priv mode */
2082 /** early replies go to offset 0, regular replies go after that */
2083 unsigned int rq_reply_off;
2087 /** Fields that help to see if request and reply were swabbed or not */
2088 __u32 rq_req_swab_mask;
2089 __u32 rq_rep_swab_mask;
2091 /** how many early replies (for stats) */
2093 /** Server-side, export on which request was received */
2094 struct obd_export *rq_export;
2095 /** import where request is being sent */
2096 struct obd_import *rq_import;
2099 /** Peer description (the other side) */
2100 lnet_process_id_t rq_peer;
2102 * service time estimate (secs)
2103 * If the request is not served by this time, it is marked as timed out.
2107 * when request/reply sent (secs), or time when request should be sent
2110 /** when request must finish. */
2112 /** request format description */
2113 struct req_capsule rq_pill;
2117 * Call completion handler for rpc if any, return it's status or original
2118 * rc if there was no handler defined for this request.
2120 static inline int ptlrpc_req_interpret(const struct lu_env *env,
2121 struct ptlrpc_request *req, int rc)
2123 if (req->rq_interpret_reply != NULL) {
2124 req->rq_status = req->rq_interpret_reply(env, req,
2125 &req->rq_async_args,
2127 return req->rq_status;
2135 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
2136 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
2137 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
2138 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
2139 struct ptlrpc_nrs_pol_info *info);
2142 * Can the request be moved from the regular NRS head to the high-priority NRS
2143 * head (of the same PTLRPC service partition), if any?
2145 * For a reliable result, this should be checked under svcpt->scp_req lock.
2147 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
2149 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
2152 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
2153 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
2154 * to make sure it has not been scheduled yet (analogous to previous
2155 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
2157 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
2162 * Returns 1 if request buffer at offset \a index was already swabbed
2164 static inline int lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
2166 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2167 return req->rq_req_swab_mask & (1 << index);
2171 * Returns 1 if request reply buffer at offset \a index was already swabbed
2173 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
2175 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2176 return req->rq_rep_swab_mask & (1 << index);
2180 * Returns 1 if request needs to be swabbed into local cpu byteorder
2182 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
2184 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2188 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
2190 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
2192 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2196 * Mark request buffer at offset \a index that it was already swabbed
2198 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
2201 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2202 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
2203 req->rq_req_swab_mask |= 1 << index;
2207 * Mark request reply buffer at offset \a index that it was already swabbed
2209 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
2212 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2213 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
2214 req->rq_rep_swab_mask |= 1 << index;
2218 * Convert numerical request phase value \a phase into text string description
2220 static inline const char *
2221 ptlrpc_phase2str(enum rq_phase phase)
2230 case RQ_PHASE_INTERPRET:
2232 case RQ_PHASE_COMPLETE:
2234 case RQ_PHASE_UNREGISTERING:
2235 return "Unregistering";
2242 * Convert numerical request phase of the request \a req into text stringi
2245 static inline const char *
2246 ptlrpc_rqphase2str(struct ptlrpc_request *req)
2248 return ptlrpc_phase2str(req->rq_phase);
2252 * Debugging functions and helpers to print request structure into debug log
2255 /* Spare the preprocessor, spoil the bugs. */
2256 #define FLAG(field, str) (field ? str : "")
2258 /** Convert bit flags into a string */
2259 #define DEBUG_REQ_FLAGS(req) \
2260 ptlrpc_rqphase2str(req), \
2261 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
2262 FLAG(req->rq_err, "E"), \
2263 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
2264 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
2265 FLAG(req->rq_no_resend, "N"), \
2266 FLAG(req->rq_waiting, "W"), \
2267 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
2268 FLAG(req->rq_committed, "M")
2270 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
2272 void _debug_req(struct ptlrpc_request *req,
2273 struct libcfs_debug_msg_data *data, const char *fmt, ...)
2274 __attribute__ ((format (printf, 3, 4)));
2277 * Helper that decides if we need to print request accordig to current debug
2280 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
2282 CFS_CHECK_STACK(msgdata, mask, cdls); \
2284 if (((mask) & D_CANTMASK) != 0 || \
2285 ((libcfs_debug & (mask)) != 0 && \
2286 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
2287 _debug_req((req), msgdata, fmt, ##a); \
2291 * This is the debug print function you need to use to print request sturucture
2292 * content into lustre debug log.
2293 * for most callers (level is a constant) this is resolved at compile time */
2294 #define DEBUG_REQ(level, req, fmt, args...) \
2296 if ((level) & (D_ERROR | D_WARNING)) { \
2297 static cfs_debug_limit_state_t cdls; \
2298 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
2299 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
2301 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
2302 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
2308 * Structure that defines a single page of a bulk transfer
2310 struct ptlrpc_bulk_page {
2311 /** Linkage to list of pages in a bulk */
2312 struct list_head bp_link;
2314 * Number of bytes in a page to transfer starting from \a bp_pageoffset
2317 /** offset within a page */
2319 /** The page itself */
2320 struct page *bp_page;
2323 #define BULK_GET_SOURCE 0
2324 #define BULK_PUT_SINK 1
2325 #define BULK_GET_SINK 2
2326 #define BULK_PUT_SOURCE 3
2329 * Definition of bulk descriptor.
2330 * Bulks are special "Two phase" RPCs where initial request message
2331 * is sent first and it is followed bt a transfer (o receiving) of a large
2332 * amount of data to be settled into pages referenced from the bulk descriptors.
2333 * Bulks transfers (the actual data following the small requests) are done
2334 * on separate LNet portals.
2335 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
2336 * Another user is readpage for MDT.
2338 struct ptlrpc_bulk_desc {
2339 /** completed with failure */
2340 unsigned long bd_failure:1;
2341 /** {put,get}{source,sink} */
2342 unsigned long bd_type:2;
2344 unsigned long bd_registered:1;
2345 /** For serialization with callback */
2347 /** Import generation when request for this bulk was sent */
2348 int bd_import_generation;
2349 /** LNet portal for this bulk */
2351 /** Server side - export this bulk created for */
2352 struct obd_export *bd_export;
2353 /** Client side - import this bulk was sent on */
2354 struct obd_import *bd_import;
2355 /** Back pointer to the request */
2356 struct ptlrpc_request *bd_req;
2357 wait_queue_head_t bd_waitq; /* server side only WQ */
2358 int bd_iov_count; /* # entries in bd_iov */
2359 int bd_max_iov; /* allocated size of bd_iov */
2360 int bd_nob; /* # bytes covered */
2361 int bd_nob_transferred; /* # bytes GOT/PUT */
2365 struct ptlrpc_cb_id bd_cbid; /* network callback info */
2366 lnet_nid_t bd_sender; /* stash event::sender */
2367 int bd_md_count; /* # valid entries in bd_mds */
2368 int bd_md_max_brw; /* max entries in bd_mds */
2369 /** array of associated MDs */
2370 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
2373 * encrypt iov, size is either 0 or bd_iov_count.
2375 lnet_kiov_t *bd_enc_iov;
2377 lnet_kiov_t bd_iov[0];
2381 SVC_STOPPED = 1 << 0,
2382 SVC_STOPPING = 1 << 1,
2383 SVC_STARTING = 1 << 2,
2384 SVC_RUNNING = 1 << 3,
2386 SVC_SIGNAL = 1 << 5,
2389 #define PTLRPC_THR_NAME_LEN 32
2391 * Definition of server service thread structure
2393 struct ptlrpc_thread {
2395 * List of active threads in svc->srv_threads
2397 struct list_head t_link;
2399 * thread-private data (preallocated memory)
2404 * service thread index, from ptlrpc_start_threads
2408 * service thread pid
2412 * put watchdog in the structure per thread b=14840
2414 struct lc_watchdog *t_watchdog;
2416 * the svc this thread belonged to b=18582
2418 struct ptlrpc_service_part *t_svcpt;
2419 wait_queue_head_t t_ctl_waitq;
2420 struct lu_env *t_env;
2421 char t_name[PTLRPC_THR_NAME_LEN];
2424 static inline int thread_is_init(struct ptlrpc_thread *thread)
2426 return thread->t_flags == 0;
2429 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2431 return !!(thread->t_flags & SVC_STOPPED);
2434 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2436 return !!(thread->t_flags & SVC_STOPPING);
2439 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2441 return !!(thread->t_flags & SVC_STARTING);
2444 static inline int thread_is_running(struct ptlrpc_thread *thread)
2446 return !!(thread->t_flags & SVC_RUNNING);
2449 static inline int thread_is_event(struct ptlrpc_thread *thread)
2451 return !!(thread->t_flags & SVC_EVENT);
2454 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2456 return !!(thread->t_flags & SVC_SIGNAL);
2459 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2461 thread->t_flags &= ~flags;
2464 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2466 thread->t_flags = flags;
2469 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2471 thread->t_flags |= flags;
2474 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2477 if (thread->t_flags & flags) {
2478 thread->t_flags &= ~flags;
2485 * Request buffer descriptor structure.
2486 * This is a structure that contains one posted request buffer for service.
2487 * Once data land into a buffer, event callback creates actual request and
2488 * notifies wakes one of the service threads to process new incoming request.
2489 * More than one request can fit into the buffer.
2491 struct ptlrpc_request_buffer_desc {
2492 /** Link item for rqbds on a service */
2493 struct list_head rqbd_list;
2494 /** History of requests for this buffer */
2495 struct list_head rqbd_reqs;
2496 /** Back pointer to service for which this buffer is registered */
2497 struct ptlrpc_service_part *rqbd_svcpt;
2498 /** LNet descriptor */
2499 lnet_handle_md_t rqbd_md_h;
2501 /** The buffer itself */
2503 struct ptlrpc_cb_id rqbd_cbid;
2505 * This "embedded" request structure is only used for the
2506 * last request to fit into the buffer
2508 struct ptlrpc_request rqbd_req;
2511 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2513 struct ptlrpc_service_ops {
2515 * if non-NULL called during thread creation (ptlrpc_start_thread())
2516 * to initialize service specific per-thread state.
2518 int (*so_thr_init)(struct ptlrpc_thread *thr);
2520 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2521 * destruct state created by ->srv_init().
2523 void (*so_thr_done)(struct ptlrpc_thread *thr);
2525 * Handler function for incoming requests for this service
2527 int (*so_req_handler)(struct ptlrpc_request *req);
2529 * function to determine priority of the request, it's called
2530 * on every new request
2532 int (*so_hpreq_handler)(struct ptlrpc_request *);
2534 * service-specific print fn
2536 void (*so_req_printer)(void *, struct ptlrpc_request *);
2539 #ifndef __cfs_cacheline_aligned
2540 /* NB: put it here for reducing patche dependence */
2541 # define __cfs_cacheline_aligned
2545 * How many high priority requests to serve before serving one normal
2548 #define PTLRPC_SVC_HP_RATIO 10
2551 * Definition of PortalRPC service.
2552 * The service is listening on a particular portal (like tcp port)
2553 * and perform actions for a specific server like IO service for OST
2554 * or general metadata service for MDS.
2556 struct ptlrpc_service {
2557 /** serialize /proc operations */
2558 spinlock_t srv_lock;
2559 /** most often accessed fields */
2560 /** chain thru all services */
2561 struct list_head srv_list;
2562 /** service operations table */
2563 struct ptlrpc_service_ops srv_ops;
2564 /** only statically allocated strings here; we don't clean them */
2566 /** only statically allocated strings here; we don't clean them */
2567 char *srv_thread_name;
2568 /** service thread list */
2569 struct list_head srv_threads;
2570 /** threads # should be created for each partition on initializing */
2571 int srv_nthrs_cpt_init;
2572 /** limit of threads number for each partition */
2573 int srv_nthrs_cpt_limit;
2574 /** Root of /proc dir tree for this service */
2575 struct proc_dir_entry *srv_procroot;
2576 /** Pointer to statistic data for this service */
2577 struct lprocfs_stats *srv_stats;
2578 /** # hp per lp reqs to handle */
2579 int srv_hpreq_ratio;
2580 /** biggest request to receive */
2581 int srv_max_req_size;
2582 /** biggest reply to send */
2583 int srv_max_reply_size;
2584 /** size of individual buffers */
2586 /** # buffers to allocate in 1 group */
2587 int srv_nbuf_per_group;
2588 /** Local portal on which to receive requests */
2589 __u32 srv_req_portal;
2590 /** Portal on the client to send replies to */
2591 __u32 srv_rep_portal;
2593 * Tags for lu_context associated with this thread, see struct
2597 /** soft watchdog timeout multiplier */
2598 int srv_watchdog_factor;
2599 /** under unregister_service */
2600 unsigned srv_is_stopping:1;
2602 /** max # request buffers in history per partition */
2603 int srv_hist_nrqbds_cpt_max;
2604 /** number of CPTs this service bound on */
2606 /** CPTs array this service bound on */
2608 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2610 /** CPT table this service is running over */
2611 struct cfs_cpt_table *srv_cptable;
2613 * partition data for ptlrpc service
2615 struct ptlrpc_service_part *srv_parts[0];
2619 * Definition of PortalRPC service partition data.
2620 * Although a service only has one instance of it right now, but we
2621 * will have multiple instances very soon (instance per CPT).
2623 * it has four locks:
2625 * serialize operations on rqbd and requests waiting for preprocess
2627 * serialize operations active requests sent to this portal
2629 * serialize adaptive timeout stuff
2631 * serialize operations on RS list (reply states)
2633 * We don't have any use-case to take two or more locks at the same time
2634 * for now, so there is no lock order issue.
2636 struct ptlrpc_service_part {
2637 /** back reference to owner */
2638 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2639 /* CPT id, reserved */
2641 /** always increasing number */
2643 /** # of starting threads */
2644 int scp_nthrs_starting;
2645 /** # of stopping threads, reserved for shrinking threads */
2646 int scp_nthrs_stopping;
2647 /** # running threads */
2648 int scp_nthrs_running;
2649 /** service threads list */
2650 struct list_head scp_threads;
2653 * serialize the following fields, used for protecting
2654 * rqbd list and incoming requests waiting for preprocess,
2655 * threads starting & stopping are also protected by this lock.
2657 spinlock_t scp_lock __cfs_cacheline_aligned;
2658 /** total # req buffer descs allocated */
2659 int scp_nrqbds_total;
2660 /** # posted request buffers for receiving */
2661 int scp_nrqbds_posted;
2662 /** in progress of allocating rqbd */
2663 int scp_rqbd_allocating;
2664 /** # incoming reqs */
2665 int scp_nreqs_incoming;
2666 /** request buffers to be reposted */
2667 struct list_head scp_rqbd_idle;
2668 /** req buffers receiving */
2669 struct list_head scp_rqbd_posted;
2670 /** incoming reqs */
2671 struct list_head scp_req_incoming;
2672 /** timeout before re-posting reqs, in tick */
2673 cfs_duration_t scp_rqbd_timeout;
2675 * all threads sleep on this. This wait-queue is signalled when new
2676 * incoming request arrives and when difficult reply has to be handled.
2678 wait_queue_head_t scp_waitq;
2680 /** request history */
2681 struct list_head scp_hist_reqs;
2682 /** request buffer history */
2683 struct list_head scp_hist_rqbds;
2684 /** # request buffers in history */
2685 int scp_hist_nrqbds;
2686 /** sequence number for request */
2688 /** highest seq culled from history */
2689 __u64 scp_hist_seq_culled;
2692 * serialize the following fields, used for processing requests
2693 * sent to this portal
2695 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2696 /** # reqs in either of the NRS heads below */
2697 /** # reqs being served */
2698 int scp_nreqs_active;
2699 /** # HPreqs being served */
2700 int scp_nhreqs_active;
2701 /** # hp requests handled */
2704 /** NRS head for regular requests */
2705 struct ptlrpc_nrs scp_nrs_reg;
2706 /** NRS head for HP requests; this is only valid for services that can
2707 * handle HP requests */
2708 struct ptlrpc_nrs *scp_nrs_hp;
2713 * serialize the following fields, used for changes on
2716 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2717 /** estimated rpc service time */
2718 struct adaptive_timeout scp_at_estimate;
2719 /** reqs waiting for replies */
2720 struct ptlrpc_at_array scp_at_array;
2721 /** early reply timer */
2722 struct timer_list scp_at_timer;
2724 cfs_time_t scp_at_checktime;
2725 /** check early replies */
2726 unsigned scp_at_check;
2730 * serialize the following fields, used for processing
2731 * replies for this portal
2733 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2734 /** all the active replies */
2735 struct list_head scp_rep_active;
2736 /** List of free reply_states */
2737 struct list_head scp_rep_idle;
2738 /** waitq to run, when adding stuff to srv_free_rs_list */
2739 wait_queue_head_t scp_rep_waitq;
2740 /** # 'difficult' replies */
2741 atomic_t scp_nreps_difficult;
2744 #define ptlrpc_service_for_each_part(part, i, svc) \
2746 i < (svc)->srv_ncpts && \
2747 (svc)->srv_parts != NULL && \
2748 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2751 * Declaration of ptlrpcd control structure
2753 struct ptlrpcd_ctl {
2755 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2757 unsigned long pc_flags;
2759 * Thread lock protecting structure fields.
2765 struct completion pc_starting;
2769 struct completion pc_finishing;
2771 * Thread requests set.
2773 struct ptlrpc_request_set *pc_set;
2775 * Thread name used in kthread_run()
2779 * Environment for request interpreters to run in.
2781 struct lu_env pc_env;
2783 * Index of ptlrpcd thread in the array.
2787 * Number of the ptlrpcd's partners.
2791 * Pointer to the array of partners' ptlrpcd_ctl structure.
2793 struct ptlrpcd_ctl **pc_partners;
2795 * Record the partner index to be processed next.
2800 /* Bits for pc_flags */
2801 enum ptlrpcd_ctl_flags {
2803 * Ptlrpc thread start flag.
2805 LIOD_START = 1 << 0,
2807 * Ptlrpc thread stop flag.
2811 * Ptlrpc thread force flag (only stop force so far).
2812 * This will cause aborting any inflight rpcs handled
2813 * by thread if LIOD_STOP is specified.
2815 LIOD_FORCE = 1 << 2,
2817 * This is a recovery ptlrpc thread.
2819 LIOD_RECOVERY = 1 << 3,
2821 * The ptlrpcd is bound to some CPU core.
2830 * Service compatibility function; the policy is compatible with all services.
2832 * \param[in] svc The service the policy is attempting to register with.
2833 * \param[in] desc The policy descriptor
2835 * \retval true The policy is compatible with the service
2837 * \see ptlrpc_nrs_pol_desc::pd_compat()
2839 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2840 const struct ptlrpc_nrs_pol_desc *desc)
2846 * Service compatibility function; the policy is compatible with only a specific
2847 * service which is identified by its human-readable name at
2848 * ptlrpc_service::srv_name.
2850 * \param[in] svc The service the policy is attempting to register with.
2851 * \param[in] desc The policy descriptor
2853 * \retval false The policy is not compatible with the service
2854 * \retval true The policy is compatible with the service
2856 * \see ptlrpc_nrs_pol_desc::pd_compat()
2858 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2859 const struct ptlrpc_nrs_pol_desc *desc)
2861 LASSERT(desc->pd_compat_svc_name != NULL);
2862 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2867 /* ptlrpc/events.c */
2868 extern lnet_handle_eq_t ptlrpc_eq_h;
2869 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2870 lnet_process_id_t *peer, lnet_nid_t *self);
2872 * These callbacks are invoked by LNet when something happened to
2876 extern void request_out_callback(lnet_event_t *ev);
2877 extern void reply_in_callback(lnet_event_t *ev);
2878 extern void client_bulk_callback(lnet_event_t *ev);
2879 extern void request_in_callback(lnet_event_t *ev);
2880 extern void reply_out_callback(lnet_event_t *ev);
2881 #ifdef HAVE_SERVER_SUPPORT
2882 extern void server_bulk_callback(lnet_event_t *ev);
2886 /* ptlrpc/connection.c */
2887 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2889 struct obd_uuid *uuid);
2890 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2891 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2892 int ptlrpc_connection_init(void);
2893 void ptlrpc_connection_fini(void);
2894 extern lnet_pid_t ptl_get_pid(void);
2896 /* ptlrpc/niobuf.c */
2898 * Actual interfacing with LNet to put/get/register/unregister stuff
2901 #ifdef HAVE_SERVER_SUPPORT
2902 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2903 unsigned npages, unsigned max_brw,
2904 unsigned type, unsigned portal);
2905 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2906 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2908 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2912 LASSERT(desc != NULL);
2914 spin_lock(&desc->bd_lock);
2915 rc = desc->bd_md_count;
2916 spin_unlock(&desc->bd_lock);
2921 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2922 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2924 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2926 struct ptlrpc_bulk_desc *desc;
2929 LASSERT(req != NULL);
2930 desc = req->rq_bulk;
2932 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2933 req->rq_bulk_deadline > cfs_time_current_sec())
2939 spin_lock(&desc->bd_lock);
2940 rc = desc->bd_md_count;
2941 spin_unlock(&desc->bd_lock);
2945 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2946 #define PTLRPC_REPLY_EARLY 0x02
2947 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2948 int ptlrpc_reply(struct ptlrpc_request *req);
2949 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2950 int ptlrpc_error(struct ptlrpc_request *req);
2951 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2952 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2953 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2956 /* ptlrpc/client.c */
2958 * Client-side portals API. Everything to send requests, receive replies,
2959 * request queues, request management, etc.
2962 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2964 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2965 struct ptlrpc_client *);
2966 void ptlrpc_cleanup_client(struct obd_import *imp);
2967 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2969 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2970 int ptlrpc_replay_req(struct ptlrpc_request *req);
2971 void ptlrpc_restart_req(struct ptlrpc_request *req);
2972 void ptlrpc_abort_inflight(struct obd_import *imp);
2973 void ptlrpc_cleanup_imp(struct obd_import *imp);
2974 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2976 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2977 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2979 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2980 set_interpreter_func fn, void *data);
2981 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2982 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2983 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2984 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2985 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2987 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2988 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2990 struct ptlrpc_request_pool *
2991 ptlrpc_init_rq_pool(int, int,
2992 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2994 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2995 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2996 const struct req_format *format);
2997 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2998 struct ptlrpc_request_pool *,
2999 const struct req_format *format);
3000 void ptlrpc_request_free(struct ptlrpc_request *request);
3001 int ptlrpc_request_pack(struct ptlrpc_request *request,
3002 __u32 version, int opcode);
3003 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
3004 const struct req_format *format,
3005 __u32 version, int opcode);
3006 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
3007 __u32 version, int opcode, char **bufs,
3008 struct ptlrpc_cli_ctx *ctx);
3009 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
3010 int opcode, int count, __u32 *lengths,
3012 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
3013 __u32 version, int opcode,
3014 int count, __u32 *lengths, char **bufs,
3015 struct ptlrpc_request_pool *pool);
3016 void ptlrpc_req_finished(struct ptlrpc_request *request);
3017 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
3018 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
3019 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
3020 unsigned npages, unsigned max_brw,
3021 unsigned type, unsigned portal);
3022 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
3023 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
3025 __ptlrpc_free_bulk(bulk, 1);
3027 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
3029 __ptlrpc_free_bulk(bulk, 0);
3031 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
3032 struct page *page, int pageoffset, int len, int);
3033 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
3034 struct page *page, int pageoffset,
3037 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
3040 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
3041 struct page *page, int pageoffset,
3044 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
3047 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3048 struct obd_import *imp);
3049 __u64 ptlrpc_next_xid(void);
3050 __u64 ptlrpc_sample_next_xid(void);
3051 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
3053 /* Set of routines to run a function in ptlrpcd context */
3054 void *ptlrpcd_alloc_work(struct obd_import *imp,
3055 int (*cb)(const struct lu_env *, void *), void *data);
3056 void ptlrpcd_destroy_work(void *handler);
3057 int ptlrpcd_queue_work(void *handler);
3060 struct ptlrpc_service_buf_conf {
3061 /* nbufs is buffers # to allocate when growing the pool */
3062 unsigned int bc_nbufs;
3063 /* buffer size to post */
3064 unsigned int bc_buf_size;
3065 /* portal to listed for requests on */
3066 unsigned int bc_req_portal;
3067 /* portal of where to send replies to */
3068 unsigned int bc_rep_portal;
3069 /* maximum request size to be accepted for this service */
3070 unsigned int bc_req_max_size;
3071 /* maximum reply size this service can ever send */
3072 unsigned int bc_rep_max_size;
3075 struct ptlrpc_service_thr_conf {
3076 /* threadname should be 8 characters or less - 6 will be added on */
3078 /* threads increasing factor for each CPU */
3079 unsigned int tc_thr_factor;
3080 /* service threads # to start on each partition while initializing */
3081 unsigned int tc_nthrs_init;
3083 * low water of threads # upper-limit on each partition while running,
3084 * service availability may be impacted if threads number is lower
3085 * than this value. It can be ZERO if the service doesn't require
3086 * CPU affinity or there is only one partition.
3088 unsigned int tc_nthrs_base;
3089 /* "soft" limit for total threads number */
3090 unsigned int tc_nthrs_max;
3091 /* user specified threads number, it will be validated due to
3092 * other members of this structure. */
3093 unsigned int tc_nthrs_user;
3094 /* set NUMA node affinity for service threads */
3095 unsigned int tc_cpu_affinity;
3096 /* Tags for lu_context associated with service thread */
3100 struct ptlrpc_service_cpt_conf {
3101 struct cfs_cpt_table *cc_cptable;
3102 /* string pattern to describe CPTs for a service */
3106 struct ptlrpc_service_conf {
3109 /* soft watchdog timeout multiplifier to print stuck service traces */
3110 unsigned int psc_watchdog_factor;
3111 /* buffer information */
3112 struct ptlrpc_service_buf_conf psc_buf;
3113 /* thread information */
3114 struct ptlrpc_service_thr_conf psc_thr;
3115 /* CPU partition information */
3116 struct ptlrpc_service_cpt_conf psc_cpt;
3117 /* function table */
3118 struct ptlrpc_service_ops psc_ops;
3121 /* ptlrpc/service.c */
3123 * Server-side services API. Register/unregister service, request state
3124 * management, service thread management
3128 void ptlrpc_save_lock(struct ptlrpc_request *req,
3129 struct lustre_handle *lock, int mode, int no_ack);
3130 void ptlrpc_commit_replies(struct obd_export *exp);
3131 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
3132 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
3133 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
3134 struct ptlrpc_service *ptlrpc_register_service(
3135 struct ptlrpc_service_conf *conf,
3136 struct proc_dir_entry *proc_entry);
3137 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
3139 int ptlrpc_start_threads(struct ptlrpc_service *svc);
3140 int ptlrpc_unregister_service(struct ptlrpc_service *service);
3141 int liblustre_check_services(void *arg);
3142 void ptlrpc_daemonize(char *name);
3143 int ptlrpc_service_health_check(struct ptlrpc_service *);
3144 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
3145 void ptlrpc_request_change_export(struct ptlrpc_request *req,
3146 struct obd_export *export);
3147 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay);
3149 int ptlrpc_hr_init(void);
3150 void ptlrpc_hr_fini(void);
3154 /* ptlrpc/import.c */
3159 int ptlrpc_connect_import(struct obd_import *imp);
3160 int ptlrpc_init_import(struct obd_import *imp);
3161 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
3162 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
3163 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
3166 /* ptlrpc/pack_generic.c */
3167 int ptlrpc_reconnect_import(struct obd_import *imp);
3171 * ptlrpc msg buffer and swab interface
3175 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
3177 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
3179 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
3180 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
3182 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
3183 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
3185 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
3186 __u32 *lens, char **bufs);
3187 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
3189 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
3190 __u32 *lens, char **bufs, int flags);
3191 #define LPRFL_EARLY_REPLY 1
3192 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
3193 char **bufs, int flags);
3194 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
3195 unsigned int newlen, int move_data);
3196 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
3197 int __lustre_unpack_msg(struct lustre_msg *m, int len);
3198 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
3199 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
3200 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
3201 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
3202 __u32 lustre_msg_early_size(void);
3203 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
3204 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
3205 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
3206 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
3207 __u32 lustre_msg_bufcount(struct lustre_msg *m);
3208 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
3209 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
3210 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
3211 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
3212 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
3213 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
3214 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
3215 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
3216 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
3217 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
3218 __u32 lustre_msg_get_type(struct lustre_msg *msg);
3219 __u32 lustre_msg_get_version(struct lustre_msg *msg);
3220 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
3221 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
3222 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
3223 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
3224 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
3225 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
3226 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
3227 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
3228 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
3229 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
3230 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
3231 int lustre_msg_get_status(struct lustre_msg *msg);
3232 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
3233 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
3234 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
3235 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
3236 char *lustre_msg_get_jobid(struct lustre_msg *msg);
3237 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
3238 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 53, 0)
3239 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
3241 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
3243 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
3244 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
3245 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
3246 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
3247 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
3248 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
3249 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
3250 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
3251 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
3252 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
3253 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
3254 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
3255 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
3256 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
3257 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
3258 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
3261 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
3262 unsigned int newlen, int move_data)
3264 LASSERT(req->rq_reply_state);
3265 LASSERT(req->rq_repmsg);
3266 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
3270 #ifdef LUSTRE_TRANSLATE_ERRNOS
3272 static inline int ptlrpc_status_hton(int h)
3275 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
3276 * ELDLM_LOCK_ABORTED, etc.
3279 return -lustre_errno_hton(-h);
3284 static inline int ptlrpc_status_ntoh(int n)
3287 * See the comment in ptlrpc_status_hton().
3290 return -lustre_errno_ntoh(-n);
3297 #define ptlrpc_status_hton(h) (h)
3298 #define ptlrpc_status_ntoh(n) (n)
3303 /** Change request phase of \a req to \a new_phase */
3305 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
3307 if (req->rq_phase == new_phase)
3310 if (new_phase == RQ_PHASE_UNREGISTERING) {
3311 req->rq_next_phase = req->rq_phase;
3313 atomic_inc(&req->rq_import->imp_unregistering);
3316 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
3318 atomic_dec(&req->rq_import->imp_unregistering);
3321 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
3322 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
3324 req->rq_phase = new_phase;
3328 * Returns true if request \a req got early reply and hard deadline is not met
3331 ptlrpc_client_early(struct ptlrpc_request *req)
3333 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3334 req->rq_reply_deadline > cfs_time_current_sec())
3336 return req->rq_early;
3340 * Returns true if we got real reply from server for this request
3343 ptlrpc_client_replied(struct ptlrpc_request *req)
3345 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3346 req->rq_reply_deadline > cfs_time_current_sec())
3348 return req->rq_replied;
3351 /** Returns true if request \a req is in process of receiving server reply */
3353 ptlrpc_client_recv(struct ptlrpc_request *req)
3355 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3356 req->rq_reply_deadline > cfs_time_current_sec())
3358 return req->rq_receiving_reply;
3362 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
3366 spin_lock(&req->rq_lock);
3367 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3368 req->rq_reply_deadline > cfs_time_current_sec()) {
3369 spin_unlock(&req->rq_lock);
3372 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
3373 req->rq_receiving_reply;
3374 spin_unlock(&req->rq_lock);
3379 ptlrpc_client_wake_req(struct ptlrpc_request *req)
3381 if (req->rq_set == NULL)
3382 wake_up(&req->rq_reply_waitq);
3384 wake_up(&req->rq_set->set_waitq);
3388 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
3390 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3391 atomic_inc(&rs->rs_refcount);
3395 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
3397 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3398 if (atomic_dec_and_test(&rs->rs_refcount))
3399 lustre_free_reply_state(rs);
3402 /* Should only be called once per req */
3403 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
3405 if (req->rq_reply_state == NULL)
3406 return; /* shouldn't occur */
3407 ptlrpc_rs_decref(req->rq_reply_state);
3408 req->rq_reply_state = NULL;
3409 req->rq_repmsg = NULL;
3412 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3414 return lustre_msg_get_magic(req->rq_reqmsg);
3417 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3419 switch (req->rq_reqmsg->lm_magic) {
3420 case LUSTRE_MSG_MAGIC_V2:
3421 return req->rq_reqmsg->lm_repsize;
3423 LASSERTF(0, "incorrect message magic: %08x\n",
3424 req->rq_reqmsg->lm_magic);
3429 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3431 if (req->rq_delay_limit != 0 &&
3432 cfs_time_before(cfs_time_add(req->rq_queued_time,
3433 cfs_time_seconds(req->rq_delay_limit)),
3434 cfs_time_current())) {
3440 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3442 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3443 spin_lock(&req->rq_lock);
3444 req->rq_no_resend = 1;
3445 spin_unlock(&req->rq_lock);
3447 return req->rq_no_resend;
3451 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3453 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3455 return svcpt->scp_service->srv_watchdog_factor *
3456 max_t(int, at, obd_timeout);
3459 static inline struct ptlrpc_service *
3460 ptlrpc_req2svc(struct ptlrpc_request *req)
3462 LASSERT(req->rq_rqbd != NULL);
3463 return req->rq_rqbd->rqbd_svcpt->scp_service;
3466 /* ldlm/ldlm_lib.c */
3468 * Target client logic
3471 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3472 int client_obd_cleanup(struct obd_device *obddev);
3473 int client_connect_import(const struct lu_env *env,
3474 struct obd_export **exp, struct obd_device *obd,
3475 struct obd_uuid *cluuid, struct obd_connect_data *,
3477 int client_disconnect_export(struct obd_export *exp);
3478 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3480 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3481 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3482 struct obd_uuid *uuid);
3483 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3484 void client_destroy_import(struct obd_import *imp);
3487 #ifdef HAVE_SERVER_SUPPORT
3488 int server_disconnect_export(struct obd_export *exp);
3491 /* ptlrpc/pinger.c */
3493 * Pinger API (client side only)
3496 enum timeout_event {
3499 struct timeout_item;
3500 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3501 int ptlrpc_pinger_add_import(struct obd_import *imp);
3502 int ptlrpc_pinger_del_import(struct obd_import *imp);
3503 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3504 timeout_cb_t cb, void *data,
3505 struct list_head *obd_list);
3506 int ptlrpc_del_timeout_client(struct list_head *obd_list,
3507 enum timeout_event event);
3508 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3509 int ptlrpc_obd_ping(struct obd_device *obd);
3510 void ping_evictor_start(void);
3511 void ping_evictor_stop(void);
3512 void ptlrpc_pinger_ir_up(void);
3513 void ptlrpc_pinger_ir_down(void);
3515 int ptlrpc_pinger_suppress_pings(void);
3517 /* ptlrpc daemon bind policy */
3519 /* all ptlrpcd threads are free mode */
3520 PDB_POLICY_NONE = 1,
3521 /* all ptlrpcd threads are bound mode */
3522 PDB_POLICY_FULL = 2,
3523 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3524 PDB_POLICY_PAIR = 3,
3525 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3526 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3527 * If kernel supports NUMA, pthrpcd threads are binded and
3528 * grouped by NUMA node */
3529 PDB_POLICY_NEIGHBOR = 4,
3532 /* ptlrpc daemon load policy
3533 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3534 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3535 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3536 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3537 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3539 /* on the same CPU core as the caller */
3540 PDL_POLICY_SAME = 1,
3541 /* within the same CPU partition, but not the same core as the caller */
3542 PDL_POLICY_LOCAL = 2,
3543 /* round-robin on all CPU cores, but not the same core as the caller */
3544 PDL_POLICY_ROUND = 3,
3545 /* the specified CPU core is preferred, but not enforced */
3546 PDL_POLICY_PREFERRED = 4,
3549 /* ptlrpc/ptlrpcd.c */
3550 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3551 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3552 void ptlrpcd_wake(struct ptlrpc_request *req);
3553 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3554 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3555 int ptlrpcd_addref(void);
3556 void ptlrpcd_decref(void);
3558 /* ptlrpc/lproc_ptlrpc.c */
3560 * procfs output related functions
3563 const char* ll_opcode2str(__u32 opcode);
3564 #ifdef CONFIG_PROC_FS
3565 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3566 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3567 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3569 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3570 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3571 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3575 /* ptlrpc/llog_server.c */
3576 int llog_origin_handle_open(struct ptlrpc_request *req);
3577 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3578 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3579 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3580 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3581 int llog_origin_handle_close(struct ptlrpc_request *req);
3583 /* ptlrpc/llog_client.c */
3584 extern struct llog_operations llog_client_ops;