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, 2013, 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/lnet.h>
60 #include <lustre/lustre_idl.h>
61 #include <lustre_ha.h>
62 #include <lustre_sec.h>
63 #include <lustre_import.h>
64 #include <lprocfs_status.h>
65 #include <lu_object.h>
66 #include <lustre_req_layout.h>
67 #include <obd_support.h>
68 #include <lustre_ver.h>
70 /* MD flags we _always_ use */
71 #define PTLRPC_MD_OPTIONS 0
74 * Max # of bulk operations in one request.
75 * In order for the client and server to properly negotiate the maximum
76 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
77 * value. The client is free to limit the actual RPC size for any bulk
78 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
79 #define PTLRPC_BULK_OPS_BITS 2
80 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
82 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
83 * should not be used on the server at all. Otherwise, it imposes a
84 * protocol limitation on the maximum RPC size that can be used by any
85 * RPC sent to that server in the future. Instead, the server should
86 * use the negotiated per-client ocd_brw_size to determine the bulk
88 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
91 * Define maxima for bulk I/O.
93 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
94 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
95 * currently supported maximum between peers at connect via ocd_brw_size.
97 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
98 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
99 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
101 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
102 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
103 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
104 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
105 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
106 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
108 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
109 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
110 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
112 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
113 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
115 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
116 # error "PTLRPC_MAX_BRW_SIZE too big"
118 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
119 # error "PTLRPC_MAX_BRW_PAGES too big"
122 #define PTLRPC_NTHRS_INIT 2
127 * Constants determine how memory is used to buffer incoming service requests.
129 * ?_NBUFS # buffers to allocate when growing the pool
130 * ?_BUFSIZE # bytes in a single request buffer
131 * ?_MAXREQSIZE # maximum request service will receive
133 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
134 * of ?_NBUFS is added to the pool.
136 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
137 * considered full when less than ?_MAXREQSIZE is left in them.
142 * Constants determine how threads are created for ptlrpc service.
144 * ?_NTHRS_INIT # threads to create for each service partition on
145 * initializing. If it's non-affinity service and
146 * there is only one partition, it's the overall #
147 * threads for the service while initializing.
148 * ?_NTHRS_BASE # threads should be created at least for each
149 * ptlrpc partition to keep the service healthy.
150 * It's the low-water mark of threads upper-limit
151 * for each partition.
152 * ?_THR_FACTOR # threads can be added on threads upper-limit for
153 * each CPU core. This factor is only for reference,
154 * we might decrease value of factor if number of cores
155 * per CPT is above a limit.
156 * ?_NTHRS_MAX # overall threads can be created for a service,
157 * it's a soft limit because if service is running
158 * on machine with hundreds of cores and tens of
159 * CPU partitions, we need to guarantee each partition
160 * has ?_NTHRS_BASE threads, which means total threads
161 * will be ?_NTHRS_BASE * number_of_cpts which can
162 * exceed ?_NTHRS_MAX.
166 * #define MDS_NTHRS_INIT 2
167 * #define MDS_NTHRS_BASE 64
168 * #define MDS_NTHRS_FACTOR 8
169 * #define MDS_NTHRS_MAX 1024
172 * ---------------------------------------------------------------------
173 * Server(A) has 16 cores, user configured it to 4 partitions so each
174 * partition has 4 cores, then actual number of service threads on each
176 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
178 * Total number of threads for the service is:
179 * 96 * partitions(4) = 384
182 * ---------------------------------------------------------------------
183 * Server(B) has 32 cores, user configured it to 4 partitions so each
184 * partition has 8 cores, then actual number of service threads on each
186 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
188 * Total number of threads for the service is:
189 * 128 * partitions(4) = 512
192 * ---------------------------------------------------------------------
193 * Server(B) has 96 cores, user configured it to 8 partitions so each
194 * partition has 12 cores, then actual number of service threads on each
196 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
198 * Total number of threads for the service is:
199 * 160 * partitions(8) = 1280
201 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
202 * as upper limit of threads number for each partition:
203 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
206 * ---------------------------------------------------------------------
207 * Server(C) have a thousand of cores and user configured it to 32 partitions
208 * MDS_NTHRS_BASE(64) * 32 = 2048
210 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
211 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
212 * to keep service healthy, so total number of threads will just be 2048.
214 * NB: we don't suggest to choose server with that many cores because backend
215 * filesystem itself, buffer cache, or underlying network stack might
216 * have some SMP scalability issues at that large scale.
218 * If user already has a fat machine with hundreds or thousands of cores,
219 * there are two choices for configuration:
220 * a) create CPU table from subset of all CPUs and run Lustre on
222 * b) bind service threads on a few partitions, see modparameters of
223 * MDS and OSS for details
225 * NB: these calculations (and examples below) are simplified to help
226 * understanding, the real implementation is a little more complex,
227 * please see ptlrpc_server_nthreads_check() for details.
232 * LDLM threads constants:
234 * Given 8 as factor and 24 as base threads number
237 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
240 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
241 * threads for each partition and total threads number will be 112.
244 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
245 * threads for each partition to keep service healthy, so total threads
246 * number should be 24 * 8 = 192.
248 * So with these constants, threads number will be at the similar level
249 * of old versions, unless target machine has over a hundred cores
251 #define LDLM_THR_FACTOR 8
252 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
253 #define LDLM_NTHRS_BASE 24
254 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
256 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
257 #define LDLM_CLIENT_NBUFS 1
258 #define LDLM_SERVER_NBUFS 64
259 #define LDLM_BUFSIZE (8 * 1024)
260 #define LDLM_MAXREQSIZE (5 * 1024)
261 #define LDLM_MAXREPSIZE (1024)
264 * MDS threads constants:
266 * Please see examples in "Thread Constants", MDS threads number will be at
267 * the comparable level of old versions, unless the server has many cores.
269 #ifndef MDS_MAX_THREADS
270 #define MDS_MAX_THREADS 1024
271 #define MDS_MAX_OTHR_THREADS 256
273 #else /* MDS_MAX_THREADS */
274 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
275 #undef MDS_MAX_THREADS
276 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
278 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
281 /* default service */
282 #define MDS_THR_FACTOR 8
283 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
284 #define MDS_NTHRS_MAX MDS_MAX_THREADS
285 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
287 /* read-page service */
288 #define MDS_RDPG_THR_FACTOR 4
289 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
290 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
291 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
293 /* these should be removed when we remove setattr service in the future */
294 #define MDS_SETA_THR_FACTOR 4
295 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
296 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
297 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
299 /* non-affinity threads */
300 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
301 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
306 * Assume file name length = FNAME_MAX = 256 (true for ext3).
307 * path name length = PATH_MAX = 4096
308 * LOV MD size max = EA_MAX = 24 * 2000
309 * (NB: 24 is size of lov_ost_data)
310 * LOV LOGCOOKIE size max = 32 * 2000
311 * (NB: 32 is size of llog_cookie)
312 * symlink: FNAME_MAX + PATH_MAX <- largest
313 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
314 * rename: FNAME_MAX + FNAME_MAX
315 * open: FNAME_MAX + EA_MAX
317 * MDS_MAXREQSIZE ~= 4736 bytes =
318 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
319 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
321 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
322 * except in the open case where there are a large number of OSTs in a LOV.
324 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
325 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
328 * MDS incoming request with LOV EA
329 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
331 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
332 362 + LOV_MAX_STRIPE_COUNT * 24)
334 * MDS outgoing reply with LOV EA
336 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
337 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
339 * but 2.4 or later MDS will never send reply with llog_cookie to any
340 * version client. This macro is defined for server side reply buffer size.
342 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
345 * This is the size of a maximum REINT_SETXATTR request:
347 * lustre_msg 56 (32 + 4 x 5 + 4)
349 * mdt_rec_setxattr 136
351 * name 256 (XATTR_NAME_MAX)
352 * value 65536 (XATTR_SIZE_MAX)
354 #define MDS_EA_MAXREQSIZE 66288
357 * These are the maximum request and reply sizes (rounded up to 1 KB
358 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
360 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
361 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
362 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
365 * The update request includes all of updates from the create, which might
366 * include linkea (4K maxim), together with other updates, we set it to 9K:
367 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
369 #define OUT_MAXREQSIZE (9 * 1024)
370 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
372 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
373 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
377 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
378 * However, we need to allocate a much larger buffer for it because LNet
379 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
380 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
381 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
382 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
383 * utilization is very low.
385 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
386 * reused until all requests fit in it have been processed and released,
387 * which means one long blocked request can prevent the rqbd be reused.
388 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
389 * from LNet with unused 65 KB, buffer utilization will be about 59%.
390 * Please check LU-2432 for details.
392 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
396 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
397 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
398 * extra bytes to each request buffer to improve buffer utilization rate.
400 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
403 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
404 #define FLD_MAXREQSIZE (160)
406 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
407 #define FLD_MAXREPSIZE (152)
408 #define FLD_BUFSIZE (1 << 12)
411 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
413 #define SEQ_MAXREQSIZE (160)
415 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
416 #define SEQ_MAXREPSIZE (152)
417 #define SEQ_BUFSIZE (1 << 12)
419 /** MGS threads must be >= 3, see bug 22458 comment #28 */
420 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
421 #define MGS_NTHRS_MAX 32
424 #define MGS_BUFSIZE (8 * 1024)
425 #define MGS_MAXREQSIZE (7 * 1024)
426 #define MGS_MAXREPSIZE (9 * 1024)
429 * OSS threads constants:
431 * Given 8 as factor and 64 as base threads number
434 * On 8-core server configured to 2 partitions, we will have
435 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
438 * On 32-core machine configured to 4 partitions, we will have
439 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
440 * will be 112 * 4 = 448.
443 * On 64-core machine configured to 4 partitions, we will have
444 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
445 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
446 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
447 * for each partition.
449 * So we can see that with these constants, threads number wil be at the
450 * similar level of old versions, unless the server has many cores.
452 /* depress threads factor for VM with small memory size */
453 #define OSS_THR_FACTOR min_t(int, 8, \
454 NUM_CACHEPAGES >> (28 - PAGE_CACHE_SHIFT))
455 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
456 #define OSS_NTHRS_BASE 64
457 #define OSS_NTHRS_MAX 512
459 /* threads for handling "create" request */
460 #define OSS_CR_THR_FACTOR 1
461 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
462 #define OSS_CR_NTHRS_BASE 8
463 #define OSS_CR_NTHRS_MAX 64
466 * OST_IO_MAXREQSIZE ~=
467 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
468 * DT_MAX_BRW_PAGES * niobuf_remote
470 * - single object with 16 pages is 512 bytes
471 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
472 * - Must be a multiple of 1024
473 * - actual size is about 18K
475 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
476 sizeof(struct ptlrpc_body) + \
477 sizeof(struct obdo) + \
478 sizeof(struct obd_ioobj) + \
479 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
481 * FIEMAP request can be 4K+ for now
483 #define OST_MAXREQSIZE (16 * 1024)
484 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
485 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
487 #define OST_MAXREPSIZE (9 * 1024)
488 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
491 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
492 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
494 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
495 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
497 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
499 /* Macro to hide a typecast. */
500 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
502 struct ptlrpc_replay_async_args {
508 * Structure to single define portal connection.
510 struct ptlrpc_connection {
511 /** linkage for connections hash table */
512 struct hlist_node c_hash;
513 /** Our own lnet nid for this connection */
515 /** Remote side nid for this connection */
516 lnet_process_id_t c_peer;
517 /** UUID of the other side */
518 struct obd_uuid c_remote_uuid;
519 /** reference counter for this connection */
523 /** Client definition for PortalRPC */
524 struct ptlrpc_client {
525 /** What lnet portal does this client send messages to by default */
526 __u32 cli_request_portal;
527 /** What portal do we expect replies on */
528 __u32 cli_reply_portal;
529 /** Name of the client */
533 /** state flags of requests */
534 /* XXX only ones left are those used by the bulk descs as well! */
535 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
536 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
538 #define REQ_MAX_ACK_LOCKS 8
540 union ptlrpc_async_args {
542 * Scratchpad for passing args to completion interpreter. Users
543 * cast to the struct of their choosing, and CLASSERT that this is
544 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
545 * a pointer to it here. The pointer_arg ensures this struct is at
546 * least big enough for that.
548 void *pointer_arg[11];
552 struct ptlrpc_request_set;
553 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
554 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
557 * Definition of request set structure.
558 * Request set is a list of requests (not necessary to the same target) that
559 * once populated with RPCs could be sent in parallel.
560 * There are two kinds of request sets. General purpose and with dedicated
561 * serving thread. Example of the latter is ptlrpcd set.
562 * For general purpose sets once request set started sending it is impossible
563 * to add new requests to such set.
564 * Provides a way to call "completion callbacks" when all requests in the set
567 struct ptlrpc_request_set {
568 atomic_t set_refcount;
569 /** number of in queue requests */
570 atomic_t set_new_count;
571 /** number of uncompleted requests */
572 atomic_t set_remaining;
573 /** wait queue to wait on for request events */
574 wait_queue_head_t set_waitq;
575 wait_queue_head_t *set_wakeup_ptr;
576 /** List of requests in the set */
577 struct list_head set_requests;
579 * List of completion callbacks to be called when the set is completed
580 * This is only used if \a set_interpret is NULL.
581 * Links struct ptlrpc_set_cbdata.
583 struct list_head set_cblist;
584 /** Completion callback, if only one. */
585 set_interpreter_func set_interpret;
586 /** opaq argument passed to completion \a set_interpret callback. */
589 * Lock for \a set_new_requests manipulations
590 * locked so that any old caller can communicate requests to
591 * the set holder who can then fold them into the lock-free set
593 spinlock_t set_new_req_lock;
594 /** List of new yet unsent requests. Only used with ptlrpcd now. */
595 struct list_head set_new_requests;
597 /** rq_status of requests that have been freed already */
599 /** Additional fields used by the flow control extension */
600 /** Maximum number of RPCs in flight */
601 int set_max_inflight;
602 /** Callback function used to generate RPCs */
603 set_producer_func set_producer;
604 /** opaq argument passed to the producer callback */
605 void *set_producer_arg;
609 * Description of a single ptrlrpc_set callback
611 struct ptlrpc_set_cbdata {
612 /** List linkage item */
613 struct list_head psc_item;
614 /** Pointer to interpreting function */
615 set_interpreter_func psc_interpret;
616 /** Opaq argument to pass to the callback */
620 struct ptlrpc_bulk_desc;
621 struct ptlrpc_service_part;
622 struct ptlrpc_service;
625 * ptlrpc callback & work item stuff
627 struct ptlrpc_cb_id {
628 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
629 void *cbid_arg; /* additional arg */
632 /** Maximum number of locks to fit into reply state */
633 #define RS_MAX_LOCKS 8
637 * Structure to define reply state on the server
638 * Reply state holds various reply message information. Also for "difficult"
639 * replies (rep-ack case) we store the state after sending reply and wait
640 * for the client to acknowledge the reception. In these cases locks could be
641 * added to the state for replay/failover consistency guarantees.
643 struct ptlrpc_reply_state {
644 /** Callback description */
645 struct ptlrpc_cb_id rs_cb_id;
646 /** Linkage for list of all reply states in a system */
647 struct list_head rs_list;
648 /** Linkage for list of all reply states on same export */
649 struct list_head rs_exp_list;
650 /** Linkage for list of all reply states for same obd */
651 struct list_head rs_obd_list;
653 struct list_head rs_debug_list;
655 /** A spinlock to protect the reply state flags */
657 /** Reply state flags */
658 unsigned long rs_difficult:1; /* ACK/commit stuff */
659 unsigned long rs_no_ack:1; /* no ACK, even for
660 difficult requests */
661 unsigned long rs_scheduled:1; /* being handled? */
662 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
663 unsigned long rs_handled:1; /* been handled yet? */
664 unsigned long rs_on_net:1; /* reply_out_callback pending? */
665 unsigned long rs_prealloc:1; /* rs from prealloc list */
666 unsigned long rs_committed:1;/* the transaction was committed
667 and the rs was dispatched
668 by ptlrpc_commit_replies */
669 /** Size of the state */
673 /** Transaction number */
677 struct obd_export *rs_export;
678 struct ptlrpc_service_part *rs_svcpt;
679 /** Lnet metadata handle for the reply */
680 lnet_handle_md_t rs_md_h;
681 atomic_t rs_refcount;
683 /** Context for the sevice thread */
684 struct ptlrpc_svc_ctx *rs_svc_ctx;
685 /** Reply buffer (actually sent to the client), encoded if needed */
686 struct lustre_msg *rs_repbuf; /* wrapper */
687 /** Size of the reply buffer */
688 int rs_repbuf_len; /* wrapper buf length */
689 /** Size of the reply message */
690 int rs_repdata_len; /* wrapper msg length */
692 * Actual reply message. Its content is encrupted (if needed) to
693 * produce reply buffer for actual sending. In simple case
694 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
696 struct lustre_msg *rs_msg; /* reply message */
698 /** Number of locks awaiting client ACK */
700 /** Handles of locks awaiting client reply ACK */
701 struct lustre_handle rs_locks[RS_MAX_LOCKS];
702 /** Lock modes of locks in \a rs_locks */
703 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
706 struct ptlrpc_thread;
710 RQ_PHASE_NEW = 0xebc0de00,
711 RQ_PHASE_RPC = 0xebc0de01,
712 RQ_PHASE_BULK = 0xebc0de02,
713 RQ_PHASE_INTERPRET = 0xebc0de03,
714 RQ_PHASE_COMPLETE = 0xebc0de04,
715 RQ_PHASE_UNREGISTERING = 0xebc0de05,
716 RQ_PHASE_UNDEFINED = 0xebc0de06
719 /** Type of request interpreter call-back */
720 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
721 struct ptlrpc_request *req,
723 /** Type of request resend call-back */
724 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
728 * Definition of request pool structure.
729 * The pool is used to store empty preallocated requests for the case
730 * when we would actually need to send something without performing
731 * any allocations (to avoid e.g. OOM).
733 struct ptlrpc_request_pool {
734 /** Locks the list */
736 /** list of ptlrpc_request structs */
737 struct list_head prp_req_list;
738 /** Maximum message size that would fit into a rquest from this pool */
740 /** Function to allocate more requests for this pool */
741 void (*prp_populate)(struct ptlrpc_request_pool *, int);
750 * \defgroup nrs Network Request Scheduler
753 struct ptlrpc_nrs_policy;
754 struct ptlrpc_nrs_resource;
755 struct ptlrpc_nrs_request;
758 * NRS control operations.
760 * These are common for all policies.
762 enum ptlrpc_nrs_ctl {
764 * Not a valid opcode.
766 PTLRPC_NRS_CTL_INVALID,
768 * Activate the policy.
770 PTLRPC_NRS_CTL_START,
772 * Reserved for multiple primary policies, which may be a possibility
777 * Policies can start using opcodes from this value and onwards for
778 * their own purposes; the assigned value itself is arbitrary.
780 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
784 * ORR policy operations
787 NRS_CTL_ORR_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
788 NRS_CTL_ORR_WR_QUANTUM,
789 NRS_CTL_ORR_RD_OFF_TYPE,
790 NRS_CTL_ORR_WR_OFF_TYPE,
791 NRS_CTL_ORR_RD_SUPP_REQ,
792 NRS_CTL_ORR_WR_SUPP_REQ,
796 * NRS policy operations.
798 * These determine the behaviour of a policy, and are called in response to
801 struct ptlrpc_nrs_pol_ops {
803 * Called during policy registration; this operation is optional.
805 * \param[in,out] policy The policy being initialized
807 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
809 * Called during policy unregistration; this operation is optional.
811 * \param[in,out] policy The policy being unregistered/finalized
813 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
815 * Called when activating a policy via lprocfs; policies allocate and
816 * initialize their resources here; this operation is optional.
818 * \param[in,out] policy The policy being started
819 * \param[in,out] arg A generic char buffer
821 * \see nrs_policy_start_locked()
823 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy,
826 * Called when deactivating a policy via lprocfs; policies deallocate
827 * their resources here; this operation is optional
829 * \param[in,out] policy The policy being stopped
831 * \see nrs_policy_stop0()
833 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
835 * Used for policy-specific operations; i.e. not generic ones like
836 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
837 * to an ioctl; this operation is optional.
839 * \param[in,out] policy The policy carrying out operation \a opc
840 * \param[in] opc The command operation being carried out
841 * \param[in,out] arg An generic buffer for communication between the
842 * user and the control operation
847 * \see ptlrpc_nrs_policy_control()
849 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
850 enum ptlrpc_nrs_ctl opc, void *arg);
853 * Called when obtaining references to the resources of the resource
854 * hierarchy for a request that has arrived for handling at the PTLRPC
855 * service. Policies should return -ve for requests they do not wish
856 * to handle. This operation is mandatory.
858 * \param[in,out] policy The policy we're getting resources for.
859 * \param[in,out] nrq The request we are getting resources for.
860 * \param[in] parent The parent resource of the resource being
861 * requested; set to NULL if none.
862 * \param[out] resp The resource is to be returned here; the
863 * fallback policy in an NRS head should
864 * \e always return a non-NULL pointer value.
865 * \param[in] moving_req When set, signifies that this is an attempt
866 * to obtain resources for a request being moved
867 * to the high-priority NRS head by
868 * ldlm_lock_reorder_req().
869 * This implies two things:
870 * 1. We are under obd_export::exp_rpc_lock and
871 * so should not sleep.
872 * 2. We should not perform non-idempotent or can
873 * skip performing idempotent operations that
874 * were carried out when resources were first
875 * taken for the request when it was initialized
876 * in ptlrpc_nrs_req_initialize().
878 * \retval 0, +ve The level of the returned resource in the resource
879 * hierarchy; currently only 0 (for a non-leaf resource)
880 * and 1 (for a leaf resource) are supported by the
884 * \see ptlrpc_nrs_req_initialize()
885 * \see ptlrpc_nrs_hpreq_add_nolock()
886 * \see ptlrpc_nrs_req_hp_move()
888 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
889 struct ptlrpc_nrs_request *nrq,
890 const struct ptlrpc_nrs_resource *parent,
891 struct ptlrpc_nrs_resource **resp,
894 * Called when releasing references taken for resources in the resource
895 * hierarchy for the request; this operation is optional.
897 * \param[in,out] policy The policy the resource belongs to
898 * \param[in] res The resource to be freed
900 * \see ptlrpc_nrs_req_finalize()
901 * \see ptlrpc_nrs_hpreq_add_nolock()
902 * \see ptlrpc_nrs_req_hp_move()
904 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
905 const struct ptlrpc_nrs_resource *res);
908 * Obtains a request for handling from the policy, and optionally
909 * removes the request from the policy; this operation is mandatory.
911 * \param[in,out] policy The policy to poll
912 * \param[in] peek When set, signifies that we just want to
913 * examine the request, and not handle it, so the
914 * request is not removed from the policy.
915 * \param[in] force When set, it will force a policy to return a
916 * request if it has one queued.
918 * \retval NULL No request available for handling
919 * \retval valid-pointer The request polled for handling
921 * \see ptlrpc_nrs_req_get_nolock()
923 struct ptlrpc_nrs_request *
924 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
927 * Called when attempting to add a request to a policy for later
928 * handling; this operation is mandatory.
930 * \param[in,out] policy The policy on which to enqueue \a nrq
931 * \param[in,out] nrq The request to enqueue
936 * \see ptlrpc_nrs_req_add_nolock()
938 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
939 struct ptlrpc_nrs_request *nrq);
941 * Removes a request from the policy's set of pending requests. Normally
942 * called after a request has been polled successfully from the policy
943 * for handling; this operation is mandatory.
945 * \param[in,out] policy The policy the request \a nrq belongs to
946 * \param[in,out] nrq The request to dequeue
948 * \see ptlrpc_nrs_req_del_nolock()
950 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
951 struct ptlrpc_nrs_request *nrq);
953 * Called after the request being carried out. Could be used for
954 * job/resource control; this operation is optional.
956 * \param[in,out] policy The policy which is stopping to handle request
958 * \param[in,out] nrq The request
960 * \pre assert_spin_locked(&svcpt->scp_req_lock)
962 * \see ptlrpc_nrs_req_stop_nolock()
964 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
965 struct ptlrpc_nrs_request *nrq);
967 * Registers the policy's lprocfs interface with a PTLRPC service.
969 * \param[in] svc The service
974 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
976 * Unegisters the policy's lprocfs interface with a PTLRPC service.
978 * In cases of failed policy registration in
979 * \e ptlrpc_nrs_policy_register(), this function may be called for a
980 * service which has not registered the policy successfully, so
981 * implementations of this method should make sure their operations are
982 * safe in such cases.
984 * \param[in] svc The service
986 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
992 enum nrs_policy_flags {
994 * Fallback policy, use this flag only on a single supported policy per
995 * service. The flag cannot be used on policies that use
996 * \e PTLRPC_NRS_FL_REG_EXTERN
998 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
1000 * Start policy immediately after registering.
1002 PTLRPC_NRS_FL_REG_START = (1 << 1),
1004 * This is a policy registering from a module different to the one NRS
1005 * core ships in (currently ptlrpc).
1007 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1013 * Denotes whether an NRS instance is for handling normal or high-priority
1014 * RPCs, or whether an operation pertains to one or both of the NRS instances
1017 enum ptlrpc_nrs_queue_type {
1018 PTLRPC_NRS_QUEUE_REG = (1 << 0),
1019 PTLRPC_NRS_QUEUE_HP = (1 << 1),
1020 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
1026 * A PTLRPC service has at least one NRS head instance for handling normal
1027 * priority RPCs, and may optionally have a second NRS head instance for
1028 * handling high-priority RPCs. Each NRS head maintains a list of available
1029 * policies, of which one and only one policy is acting as the fallback policy,
1030 * and optionally a different policy may be acting as the primary policy. For
1031 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1032 * enqueue the RPC using the primary policy (if any). The fallback policy is
1033 * used in the following cases:
1034 * - when there was no primary policy in the
1035 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1037 * - when the primary policy that was at the
1038 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1039 * RPC was initialized, denoted it did not wish, or for some other reason was
1040 * not able to handle the request, by returning a non-valid NRS resource
1042 * - when the primary policy that was at the
1043 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1044 * RPC was initialized, fails later during the request enqueueing stage.
1046 * \see nrs_resource_get_safe()
1047 * \see nrs_request_enqueue()
1050 spinlock_t nrs_lock;
1051 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1053 * List of registered policies
1055 struct list_head nrs_policy_list;
1057 * List of policies with queued requests. Policies that have any
1058 * outstanding requests are queued here, and this list is queried
1059 * in a round-robin manner from NRS core when obtaining a request
1060 * for handling. This ensures that requests from policies that at some
1061 * point transition away from the
1062 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1064 struct list_head nrs_policy_queued;
1066 * Service partition for this NRS head
1068 struct ptlrpc_service_part *nrs_svcpt;
1070 * Primary policy, which is the preferred policy for handling RPCs
1072 struct ptlrpc_nrs_policy *nrs_policy_primary;
1074 * Fallback policy, which is the backup policy for handling RPCs
1076 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1078 * This NRS head handles either HP or regular requests
1080 enum ptlrpc_nrs_queue_type nrs_queue_type;
1082 * # queued requests from all policies in this NRS head
1084 unsigned long nrs_req_queued;
1086 * # scheduled requests from all policies in this NRS head
1088 unsigned long nrs_req_started;
1090 * # policies on this NRS
1092 unsigned nrs_num_pols;
1094 * This NRS head is in progress of starting a policy
1096 unsigned nrs_policy_starting:1;
1098 * In progress of shutting down the whole NRS head; used during
1101 unsigned nrs_stopping:1;
1103 * NRS policy is throttling reqeust
1105 unsigned nrs_throttling:1;
1108 #define NRS_POL_NAME_MAX 16
1110 struct ptlrpc_nrs_pol_desc;
1113 * Service compatibility predicate; this determines whether a policy is adequate
1114 * for handling RPCs of a particular PTLRPC service.
1116 * XXX:This should give the same result during policy registration and
1117 * unregistration, and for all partitions of a service; so the result should not
1118 * depend on temporal service or other properties, that may influence the
1121 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
1122 const struct ptlrpc_nrs_pol_desc *desc);
1124 struct ptlrpc_nrs_pol_conf {
1126 * Human-readable policy name
1128 char nc_name[NRS_POL_NAME_MAX];
1130 * NRS operations for this policy
1132 const struct ptlrpc_nrs_pol_ops *nc_ops;
1134 * Service compatibility predicate
1136 nrs_pol_desc_compat_t nc_compat;
1138 * Set for policies that support a single ptlrpc service, i.e. ones that
1139 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
1140 * depicts the name of the single service that such policies are
1143 const char *nc_compat_svc_name;
1145 * Owner module for this policy descriptor; policies registering from a
1146 * different module to the one the NRS framework is held within
1147 * (currently ptlrpc), should set this field to THIS_MODULE.
1149 struct module *nc_owner;
1151 * Policy registration flags; a bitmast of \e nrs_policy_flags
1157 * NRS policy registering descriptor
1159 * Is used to hold a description of a policy that can be passed to NRS core in
1160 * order to register the policy with NRS heads in different PTLRPC services.
1162 struct ptlrpc_nrs_pol_desc {
1164 * Human-readable policy name
1166 char pd_name[NRS_POL_NAME_MAX];
1168 * Link into nrs_core::nrs_policies
1170 struct list_head pd_list;
1172 * NRS operations for this policy
1174 const struct ptlrpc_nrs_pol_ops *pd_ops;
1176 * Service compatibility predicate
1178 nrs_pol_desc_compat_t pd_compat;
1180 * Set for policies that are compatible with only one PTLRPC service.
1182 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
1184 const char *pd_compat_svc_name;
1186 * Owner module for this policy descriptor.
1188 * We need to hold a reference to the module whenever we might make use
1189 * of any of the module's contents, i.e.
1190 * - If one or more instances of the policy are at a state where they
1191 * might be handling a request, i.e.
1192 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1193 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
1194 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
1195 * is taken on the module when
1196 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
1197 * becomes 0, so that we hold only one reference to the module maximum
1200 * We do not need to hold a reference to the module, even though we
1201 * might use code and data from the module, in the following cases:
1202 * - During external policy registration, because this should happen in
1203 * the module's init() function, in which case the module is safe from
1204 * removal because a reference is being held on the module by the
1205 * kernel, and iirc kmod (and I guess module-init-tools also) will
1206 * serialize any racing processes properly anyway.
1207 * - During external policy unregistration, because this should happen
1208 * in a module's exit() function, and any attempts to start a policy
1209 * instance would need to take a reference on the module, and this is
1210 * not possible once we have reached the point where the exit()
1211 * handler is called.
1212 * - During service registration and unregistration, as service setup
1213 * and cleanup, and policy registration, unregistration and policy
1214 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
1215 * as long as users adhere to the convention of registering policies
1216 * in init() and unregistering them in module exit() functions, there
1217 * should not be a race between these operations.
1218 * - During any policy-specific lprocfs operations, because a reference
1219 * is held by the kernel on a proc entry that has been entered by a
1220 * syscall, so as long as proc entries are removed during unregistration time,
1221 * then unregistration and lprocfs operations will be properly
1224 struct module *pd_owner;
1226 * Bitmask of \e nrs_policy_flags
1230 * # of references on this descriptor
1238 * Policies transition from one state to the other during their lifetime
1240 enum ptlrpc_nrs_pol_state {
1242 * Not a valid policy state.
1244 NRS_POL_STATE_INVALID,
1246 * Policies are at this state either at the start of their life, or
1247 * transition here when the user selects a different policy to act
1248 * as the primary one.
1250 NRS_POL_STATE_STOPPED,
1252 * Policy is progress of stopping
1254 NRS_POL_STATE_STOPPING,
1256 * Policy is in progress of starting
1258 NRS_POL_STATE_STARTING,
1260 * A policy is in this state in two cases:
1261 * - it is the fallback policy, which is always in this state.
1262 * - it has been activated by the user; i.e. it is the primary policy,
1264 NRS_POL_STATE_STARTED,
1268 * NRS policy information
1270 * Used for obtaining information for the status of a policy via lprocfs
1272 struct ptlrpc_nrs_pol_info {
1276 char pi_name[NRS_POL_NAME_MAX];
1278 * Current policy state
1280 enum ptlrpc_nrs_pol_state pi_state;
1282 * # RPCs enqueued for later dispatching by the policy
1286 * # RPCs started for dispatch by the policy
1288 long pi_req_started;
1290 * Is this a fallback policy?
1292 unsigned pi_fallback:1;
1298 * There is one instance of this for each policy in each NRS head of each
1299 * PTLRPC service partition.
1301 struct ptlrpc_nrs_policy {
1303 * Linkage into the NRS head's list of policies,
1304 * ptlrpc_nrs:nrs_policy_list
1306 struct list_head pol_list;
1308 * Linkage into the NRS head's list of policies with enqueued
1309 * requests ptlrpc_nrs:nrs_policy_queued
1311 struct list_head pol_list_queued;
1313 * Current state of this policy
1315 enum ptlrpc_nrs_pol_state pol_state;
1317 * Bitmask of nrs_policy_flags
1321 * # RPCs enqueued for later dispatching by the policy
1323 long pol_req_queued;
1325 * # RPCs started for dispatch by the policy
1327 long pol_req_started;
1329 * Usage Reference count taken on the policy instance
1333 * The NRS head this policy has been created at
1335 struct ptlrpc_nrs *pol_nrs;
1337 * Private policy data; varies by policy type
1341 * Policy descriptor for this policy instance.
1343 struct ptlrpc_nrs_pol_desc *pol_desc;
1349 * Resources are embedded into two types of NRS entities:
1350 * - Inside NRS policies, in the policy's private data in
1351 * ptlrpc_nrs_policy::pol_private
1352 * - In objects that act as prime-level scheduling entities in different NRS
1353 * policies; e.g. on a policy that performs round robin or similar order
1354 * scheduling across client NIDs, there would be one NRS resource per unique
1355 * client NID. On a policy which performs round robin scheduling across
1356 * backend filesystem objects, there would be one resource associated with
1357 * each of the backend filesystem objects partaking in the scheduling
1358 * performed by the policy.
1360 * NRS resources share a parent-child relationship, in which resources embedded
1361 * in policy instances are the parent entities, with all scheduling entities
1362 * a policy schedules across being the children, thus forming a simple resource
1363 * hierarchy. This hierarchy may be extended with one or more levels in the
1364 * future if the ability to have more than one primary policy is added.
1366 * Upon request initialization, references to the then active NRS policies are
1367 * taken and used to later handle the dispatching of the request with one of
1370 * \see nrs_resource_get_safe()
1371 * \see ptlrpc_nrs_req_add()
1373 struct ptlrpc_nrs_resource {
1375 * This NRS resource's parent; is NULL for resources embedded in NRS
1376 * policy instances; i.e. those are top-level ones.
1378 struct ptlrpc_nrs_resource *res_parent;
1380 * The policy associated with this resource.
1382 struct ptlrpc_nrs_policy *res_policy;
1395 * This policy is a logical wrapper around previous, non-NRS functionality.
1396 * It dispatches RPCs in the same order as they arrive from the network. This
1397 * policy is currently used as the fallback policy, and the only enabled policy
1398 * on all NRS heads of all PTLRPC service partitions.
1403 * Private data structure for the FIFO policy
1405 struct nrs_fifo_head {
1407 * Resource object for policy instance.
1409 struct ptlrpc_nrs_resource fh_res;
1411 * List of queued requests.
1413 struct list_head fh_list;
1415 * For debugging purposes.
1420 struct nrs_fifo_req {
1421 struct list_head fr_list;
1430 * CRR-N, Client Round Robin over NIDs
1435 * private data structure for CRR-N NRS
1437 struct nrs_crrn_net {
1438 struct ptlrpc_nrs_resource cn_res;
1439 cfs_binheap_t *cn_binheap;
1440 cfs_hash_t *cn_cli_hash;
1442 * Used when a new scheduling round commences, in order to synchronize
1443 * all clients with the new round number.
1447 * Determines the relevant ordering amongst request batches within a
1452 * Round Robin quantum; the maximum number of RPCs that each request
1453 * batch for each client can have in a scheduling round.
1459 * Object representing a client in CRR-N, as identified by its NID
1461 struct nrs_crrn_client {
1462 struct ptlrpc_nrs_resource cc_res;
1463 struct hlist_node cc_hnode;
1466 * The round number against which this client is currently scheduling
1471 * The sequence number used for requests scheduled by this client during
1472 * the current round number.
1477 * Round Robin quantum; the maximum number of RPCs the client is allowed
1478 * to schedule in a single batch of each round.
1482 * # of pending requests for this client, on all existing rounds
1488 * CRR-N NRS request definition
1490 struct nrs_crrn_req {
1492 * Round number for this request; shared with all other requests in the
1497 * Sequence number for this request; shared with all other requests in
1504 * CRR-N policy operations.
1508 * Read the RR quantum size of a CRR-N policy.
1510 NRS_CTL_CRRN_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
1512 * Write the RR quantum size of a CRR-N policy.
1514 NRS_CTL_CRRN_WR_QUANTUM,
1522 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
1527 * Lower and upper byte offsets of a brw RPC
1529 struct nrs_orr_req_range {
1535 * RPC types supported by the ORR/TRR policies
1538 NOS_OST_READ = (1 << 0),
1539 NOS_OST_WRITE = (1 << 1),
1540 NOS_OST_RW = (NOS_OST_READ | NOS_OST_WRITE),
1542 * Default value for policies.
1544 NOS_DFLT = NOS_OST_READ
1548 * As unique keys for grouping RPCs together, we use the object's OST FID for
1549 * the ORR policy, and the OST index for the TRR policy.
1551 * XXX: We waste some space for TRR policy instances by using a union, but it
1552 * allows to consolidate some of the code between ORR and TRR, and these
1553 * policies will probably eventually merge into one anyway.
1555 struct nrs_orr_key {
1557 /** object FID for ORR */
1558 struct lu_fid ok_fid;
1559 /** OST index for TRR */
1565 * The largest base string for unique hash/slab object names is
1566 * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT
1567 * id number, so this _should_ be more than enough for the maximum number of
1568 * CPTs on any system. If it does happen that this statement is incorrect,
1569 * nrs_orr_genobjname() will inevitably yield a non-unique name and cause
1570 * kmem_cache_create() to complain (on Linux), so the erroneous situation
1571 * will hopefully not go unnoticed.
1573 #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3)
1576 * private data structure for ORR and TRR NRS
1578 struct nrs_orr_data {
1579 struct ptlrpc_nrs_resource od_res;
1580 cfs_binheap_t *od_binheap;
1581 cfs_hash_t *od_obj_hash;
1582 struct kmem_cache *od_cache;
1584 * Used when a new scheduling round commences, in order to synchronize
1585 * all object or OST batches with the new round number.
1589 * Determines the relevant ordering amongst request batches within a
1594 * RPC types that are currently supported.
1596 enum nrs_orr_supp od_supp;
1598 * Round Robin quantum; the maxium number of RPCs that each request
1599 * batch for each object or OST can have in a scheduling round.
1603 * Whether to use physical disk offsets or logical file offsets.
1607 * XXX: We need to provide a persistently allocated string to hold
1608 * unique object names for this policy, since in currently supported
1609 * versions of Linux by Lustre, kmem_cache_create() just sets a pointer
1610 * to the name string provided. kstrdup() is used in the version of
1611 * kmeme_cache_create() in current Linux mainline, so we may be able to
1612 * remove this in the future.
1614 char od_objname[NRS_ORR_OBJ_NAME_MAX];
1618 * Represents a backend-fs object or OST in the ORR and TRR policies
1621 struct nrs_orr_object {
1622 struct ptlrpc_nrs_resource oo_res;
1623 struct hlist_node oo_hnode;
1625 * The round number against which requests are being scheduled for this
1630 * The sequence number used for requests scheduled for this object or
1631 * OST during the current round number.
1635 * The key of the object or OST for which this structure instance is
1638 struct nrs_orr_key oo_key;
1641 * Round Robin quantum; the maximum number of RPCs that are allowed to
1642 * be scheduled for the object or OST in a single batch of each round.
1646 * # of pending requests for this object or OST, on all existing rounds
1652 * ORR/TRR NRS request definition
1654 struct nrs_orr_req {
1656 * The offset range this request covers
1658 struct nrs_orr_req_range or_range;
1660 * Round number for this request; shared with all other requests in the
1665 * Sequence number for this request; shared with all other requests in
1670 * For debugging purposes.
1672 struct nrs_orr_key or_key;
1674 * An ORR policy instance has filled in request information while
1675 * enqueueing the request on the service partition's regular NRS head.
1677 unsigned int or_orr_set:1;
1679 * A TRR policy instance has filled in request information while
1680 * enqueueing the request on the service partition's regular NRS head.
1682 unsigned int or_trr_set:1;
1684 * Request offset ranges have been filled in with logical offset
1687 unsigned int or_logical_set:1;
1689 * Request offset ranges have been filled in with physical offset
1692 unsigned int or_physical_set:1;
1697 #include <lustre_nrs_tbf.h>
1702 * Instances of this object exist embedded within ptlrpc_request; the main
1703 * purpose of this object is to hold references to the request's resources
1704 * for the lifetime of the request, and to hold properties that policies use
1705 * use for determining the request's scheduling priority.
1707 struct ptlrpc_nrs_request {
1709 * The request's resource hierarchy.
1711 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1713 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1714 * policy that was used to enqueue the request.
1716 * \see nrs_request_enqueue()
1718 unsigned nr_res_idx;
1719 unsigned nr_initialized:1;
1720 unsigned nr_enqueued:1;
1721 unsigned nr_started:1;
1722 unsigned nr_finalized:1;
1723 cfs_binheap_node_t nr_node;
1726 * Policy-specific fields, used for determining a request's scheduling
1727 * priority, and other supporting functionality.
1731 * Fields for the FIFO policy
1733 struct nrs_fifo_req fifo;
1735 * CRR-N request defintion
1737 struct nrs_crrn_req crr;
1738 /** ORR and TRR share the same request definition */
1739 struct nrs_orr_req orr;
1741 * TBF request definition
1743 struct nrs_tbf_req tbf;
1746 * Externally-registering policies may want to use this to allocate
1747 * their own request properties.
1755 * Basic request prioritization operations structure.
1756 * The whole idea is centered around locks and RPCs that might affect locks.
1757 * When a lock is contended we try to give priority to RPCs that might lead
1758 * to fastest release of that lock.
1759 * Currently only implemented for OSTs only in a way that makes all
1760 * IO and truncate RPCs that are coming from a locked region where a lock is
1761 * contended a priority over other requests.
1763 struct ptlrpc_hpreq_ops {
1765 * Check if the lock handle of the given lock is the same as
1766 * taken from the request.
1768 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1770 * Check if the request is a high priority one.
1772 int (*hpreq_check)(struct ptlrpc_request *);
1774 * Called after the request has been handled.
1776 void (*hpreq_fini)(struct ptlrpc_request *);
1779 struct ptlrpc_cli_req {
1780 /** For bulk requests on client only: bulk descriptor */
1781 struct ptlrpc_bulk_desc *cr_bulk;
1782 /** optional time limit for send attempts */
1783 cfs_duration_t cr_delay_limit;
1784 /** time request was first queued */
1785 cfs_time_t cr_queued_time;
1786 /** request sent timeval */
1787 struct timeval cr_sent_tv;
1788 /** time for request really sent out */
1790 /** when req reply unlink must finish. */
1791 time_t cr_reply_deadline;
1792 /** when req bulk unlink must finish. */
1793 time_t cr_bulk_deadline;
1794 /** Portal to which this request would be sent */
1796 /** Portal where to wait for reply and where reply would be sent */
1798 /** request resending number */
1799 unsigned int cr_resend_nr;
1800 /** What was import generation when this request was sent */
1802 enum lustre_imp_state cr_send_state;
1803 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1804 wait_queue_head_t cr_set_waitq;
1805 /** Link item for request set lists */
1806 struct list_head cr_set_chain;
1807 /** link to waited ctx */
1808 struct list_head cr_ctx_chain;
1810 /** client's half ctx */
1811 struct ptlrpc_cli_ctx *cr_cli_ctx;
1812 /** Link back to the request set */
1813 struct ptlrpc_request_set *cr_set;
1814 /** outgoing request MD handle */
1815 lnet_handle_md_t cr_req_md_h;
1816 /** request-out callback parameter */
1817 struct ptlrpc_cb_id cr_req_cbid;
1818 /** incoming reply MD handle */
1819 lnet_handle_md_t cr_reply_md_h;
1820 wait_queue_head_t cr_reply_waitq;
1821 /** reply callback parameter */
1822 struct ptlrpc_cb_id cr_reply_cbid;
1823 /** Async completion handler, called when reply is received */
1824 ptlrpc_interpterer_t cr_reply_interp;
1825 /** Resend handler, called when request is resend to update RPC data */
1826 ptlrpc_resend_cb_t cr_resend_cb;
1827 /** Async completion context */
1828 union ptlrpc_async_args cr_async_args;
1829 /** Opaq data for replay and commit callbacks. */
1832 * Commit callback, called when request is committed and about to be
1835 void (*cr_commit_cb)(struct ptlrpc_request *);
1836 /** Replay callback, called after request is replayed at recovery */
1837 void (*cr_replay_cb)(struct ptlrpc_request *);
1840 /** client request member alias */
1841 /* NB: these alias should NOT be used by any new code, instead they should
1842 * be removed step by step to avoid potential abuse */
1843 #define rq_bulk rq_cli.cr_bulk
1844 #define rq_delay_limit rq_cli.cr_delay_limit
1845 #define rq_queued_time rq_cli.cr_queued_time
1846 #define rq_sent_tv rq_cli.cr_sent_tv
1847 #define rq_real_sent rq_cli.cr_sent_out
1848 #define rq_reply_deadline rq_cli.cr_reply_deadline
1849 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
1850 #define rq_nr_resend rq_cli.cr_resend_nr
1851 #define rq_request_portal rq_cli.cr_req_ptl
1852 #define rq_reply_portal rq_cli.cr_rep_ptl
1853 #define rq_import_generation rq_cli.cr_imp_gen
1854 #define rq_send_state rq_cli.cr_send_state
1855 #define rq_set_chain rq_cli.cr_set_chain
1856 #define rq_ctx_chain rq_cli.cr_ctx_chain
1857 #define rq_set rq_cli.cr_set
1858 #define rq_set_waitq rq_cli.cr_set_waitq
1859 #define rq_cli_ctx rq_cli.cr_cli_ctx
1860 #define rq_req_md_h rq_cli.cr_req_md_h
1861 #define rq_req_cbid rq_cli.cr_req_cbid
1862 #define rq_reply_md_h rq_cli.cr_reply_md_h
1863 #define rq_reply_waitq rq_cli.cr_reply_waitq
1864 #define rq_reply_cbid rq_cli.cr_reply_cbid
1865 #define rq_interpret_reply rq_cli.cr_reply_interp
1866 #define rq_resend_cb rq_cli.cr_resend_cb
1867 #define rq_async_args rq_cli.cr_async_args
1868 #define rq_cb_data rq_cli.cr_cb_data
1869 #define rq_commit_cb rq_cli.cr_commit_cb
1870 #define rq_replay_cb rq_cli.cr_replay_cb
1872 struct ptlrpc_srv_req {
1873 /** initial thread servicing this request */
1874 struct ptlrpc_thread *sr_svc_thread;
1876 * Server side list of incoming unserved requests sorted by arrival
1877 * time. Traversed from time to time to notice about to expire
1878 * requests and sent back "early replies" to clients to let them
1879 * know server is alive and well, just very busy to service their
1882 struct list_head sr_timed_list;
1883 /** server-side per-export list */
1884 struct list_head sr_exp_list;
1885 /** server-side history, used for debuging purposes. */
1886 struct list_head sr_hist_list;
1887 /** history sequence # */
1889 /** the index of service's srv_at_array into which request is linked */
1893 /** authed uid mapped to */
1894 uid_t sr_auth_mapped_uid;
1895 /** RPC is generated from what part of Lustre */
1896 enum lustre_sec_part sr_sp_from;
1897 /** request session context */
1898 struct lu_context sr_ses;
1902 /** stub for NRS request */
1903 struct ptlrpc_nrs_request sr_nrq;
1905 /** request arrival time */
1906 struct timeval sr_arrival_time;
1907 /** server's half ctx */
1908 struct ptlrpc_svc_ctx *sr_svc_ctx;
1909 /** (server side), pointed directly into req buffer */
1910 struct ptlrpc_user_desc *sr_user_desc;
1911 /** separated reply state */
1912 struct ptlrpc_reply_state *sr_reply_state;
1913 /** server-side hp handlers */
1914 struct ptlrpc_hpreq_ops *sr_ops;
1915 /** incoming request buffer */
1916 struct ptlrpc_request_buffer_desc *sr_rqbd;
1919 /** server request member alias */
1920 /* NB: these alias should NOT be used by any new code, instead they should
1921 * be removed step by step to avoid potential abuse */
1922 #define rq_svc_thread rq_srv.sr_svc_thread
1923 #define rq_timed_list rq_srv.sr_timed_list
1924 #define rq_exp_list rq_srv.sr_exp_list
1925 #define rq_history_list rq_srv.sr_hist_list
1926 #define rq_history_seq rq_srv.sr_hist_seq
1927 #define rq_at_index rq_srv.sr_at_index
1928 #define rq_auth_uid rq_srv.sr_auth_uid
1929 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
1930 #define rq_sp_from rq_srv.sr_sp_from
1931 #define rq_session rq_srv.sr_ses
1932 #define rq_nrq rq_srv.sr_nrq
1933 #define rq_arrival_time rq_srv.sr_arrival_time
1934 #define rq_reply_state rq_srv.sr_reply_state
1935 #define rq_svc_ctx rq_srv.sr_svc_ctx
1936 #define rq_user_desc rq_srv.sr_user_desc
1937 #define rq_ops rq_srv.sr_ops
1938 #define rq_rqbd rq_srv.sr_rqbd
1941 * Represents remote procedure call.
1943 * This is a staple structure used by everybody wanting to send a request
1946 struct ptlrpc_request {
1947 /* Request type: one of PTL_RPC_MSG_* */
1949 /** Result of request processing */
1952 * Linkage item through which this request is included into
1953 * sending/delayed lists on client and into rqbd list on server
1955 struct list_head rq_list;
1956 /** Lock to protect request flags and some other important bits, like
1960 /** client-side flags are serialized by rq_lock @{ */
1961 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1962 rq_timedout:1, rq_resend:1, rq_restart:1,
1964 * when ->rq_replay is set, request is kept by the client even
1965 * after server commits corresponding transaction. This is
1966 * used for operations that require sequence of multiple
1967 * requests to be replayed. The only example currently is file
1968 * open/close. When last request in such a sequence is
1969 * committed, ->rq_replay is cleared on all requests in the
1973 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1974 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1976 rq_req_unlinked:1, /* unlinked request buffer from lnet */
1977 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
1978 rq_memalloc:1, /* req originated from "kswapd" */
1980 rq_reply_truncated:1,
1981 /** whether the "rq_set" is a valid one */
1983 rq_generation_set:1,
1984 /** do not resend request on -EINPROGRESS */
1985 rq_no_retry_einprogress:1,
1986 /* allow the req to be sent if the import is in recovery
1989 /* bulk request, sent to server, but uncommitted */
1993 /** server-side flags @{ */
1995 rq_hp:1, /**< high priority RPC */
1996 rq_at_linked:1, /**< link into service's srv_at_array */
1997 rq_packed_final:1; /**< packed final reply */
2000 /** one of RQ_PHASE_* */
2001 enum rq_phase rq_phase;
2002 /** one of RQ_PHASE_* to be used next */
2003 enum rq_phase rq_next_phase;
2005 * client-side refcount for SENT race, server-side refcounf
2006 * for multiple replies
2008 atomic_t rq_refcount;
2011 * !rq_truncate : # reply bytes actually received,
2012 * rq_truncate : required repbuf_len for resend
2014 int rq_nob_received;
2015 /** Request length */
2019 /** Pool if request is from preallocated list */
2020 struct ptlrpc_request_pool *rq_pool;
2021 /** Request message - what client sent */
2022 struct lustre_msg *rq_reqmsg;
2023 /** Reply message - server response */
2024 struct lustre_msg *rq_repmsg;
2025 /** Transaction number */
2030 * List item to for replay list. Not yet commited requests get linked
2032 * Also see \a rq_replay comment above.
2033 * It's also link chain on obd_export::exp_req_replay_queue
2035 struct list_head rq_replay_list;
2036 /** non-shared members for client & server request*/
2038 struct ptlrpc_cli_req rq_cli;
2039 struct ptlrpc_srv_req rq_srv;
2042 * security and encryption data
2044 /** description of flavors for client & server */
2045 struct sptlrpc_flavor rq_flvr;
2047 /* client/server security flags */
2049 rq_ctx_init:1, /* context initiation */
2050 rq_ctx_fini:1, /* context destroy */
2051 rq_bulk_read:1, /* request bulk read */
2052 rq_bulk_write:1, /* request bulk write */
2053 /* server authentication flags */
2054 rq_auth_gss:1, /* authenticated by gss */
2055 rq_auth_remote:1, /* authed as remote user */
2056 rq_auth_usr_root:1, /* authed as root */
2057 rq_auth_usr_mdt:1, /* authed as mdt */
2058 rq_auth_usr_ost:1, /* authed as ost */
2059 /* security tfm flags */
2062 /* doesn't expect reply FIXME */
2064 rq_pill_init:1, /* pill initialized */
2065 rq_srv_req:1; /* server request */
2068 /** various buffer pointers */
2069 struct lustre_msg *rq_reqbuf; /**< req wrapper */
2070 char *rq_repbuf; /**< rep buffer */
2071 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
2072 /** only in priv mode */
2073 struct lustre_msg *rq_clrbuf;
2074 int rq_reqbuf_len; /* req wrapper buf len */
2075 int rq_reqdata_len; /* req wrapper msg len */
2076 int rq_repbuf_len; /* rep buffer len */
2077 int rq_repdata_len; /* rep wrapper msg len */
2078 int rq_clrbuf_len; /* only in priv mode */
2079 int rq_clrdata_len; /* only in priv mode */
2081 /** early replies go to offset 0, regular replies go after that */
2082 unsigned int rq_reply_off;
2086 /** Fields that help to see if request and reply were swabbed or not */
2087 __u32 rq_req_swab_mask;
2088 __u32 rq_rep_swab_mask;
2090 /** how many early replies (for stats) */
2092 /** Server-side, export on which request was received */
2093 struct obd_export *rq_export;
2094 /** import where request is being sent */
2095 struct obd_import *rq_import;
2098 /** Peer description (the other side) */
2099 lnet_process_id_t rq_peer;
2101 * service time estimate (secs)
2102 * If the request is not served by this time, it is marked as timed out.
2106 * when request/reply sent (secs), or time when request should be sent
2109 /** when request must finish. */
2111 /** request format description */
2112 struct req_capsule rq_pill;
2116 * Call completion handler for rpc if any, return it's status or original
2117 * rc if there was no handler defined for this request.
2119 static inline int ptlrpc_req_interpret(const struct lu_env *env,
2120 struct ptlrpc_request *req, int rc)
2122 if (req->rq_interpret_reply != NULL) {
2123 req->rq_status = req->rq_interpret_reply(env, req,
2124 &req->rq_async_args,
2126 return req->rq_status;
2134 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
2135 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
2136 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
2137 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
2138 struct ptlrpc_nrs_pol_info *info);
2141 * Can the request be moved from the regular NRS head to the high-priority NRS
2142 * head (of the same PTLRPC service partition), if any?
2144 * For a reliable result, this should be checked under svcpt->scp_req lock.
2146 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
2148 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
2151 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
2152 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
2153 * to make sure it has not been scheduled yet (analogous to previous
2154 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
2156 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
2161 * Returns 1 if request buffer at offset \a index was already swabbed
2163 static inline int lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
2165 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2166 return req->rq_req_swab_mask & (1 << index);
2170 * Returns 1 if request reply buffer at offset \a index was already swabbed
2172 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
2174 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2175 return req->rq_rep_swab_mask & (1 << index);
2179 * Returns 1 if request needs to be swabbed into local cpu byteorder
2181 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
2183 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2187 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
2189 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
2191 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2195 * Mark request buffer at offset \a index that it was already swabbed
2197 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
2200 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2201 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
2202 req->rq_req_swab_mask |= 1 << index;
2206 * Mark request reply buffer at offset \a index that it was already swabbed
2208 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
2211 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2212 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
2213 req->rq_rep_swab_mask |= 1 << index;
2217 * Convert numerical request phase value \a phase into text string description
2219 static inline const char *
2220 ptlrpc_phase2str(enum rq_phase phase)
2229 case RQ_PHASE_INTERPRET:
2231 case RQ_PHASE_COMPLETE:
2233 case RQ_PHASE_UNREGISTERING:
2234 return "Unregistering";
2241 * Convert numerical request phase of the request \a req into text stringi
2244 static inline const char *
2245 ptlrpc_rqphase2str(struct ptlrpc_request *req)
2247 return ptlrpc_phase2str(req->rq_phase);
2251 * Debugging functions and helpers to print request structure into debug log
2254 /* Spare the preprocessor, spoil the bugs. */
2255 #define FLAG(field, str) (field ? str : "")
2257 /** Convert bit flags into a string */
2258 #define DEBUG_REQ_FLAGS(req) \
2259 ptlrpc_rqphase2str(req), \
2260 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
2261 FLAG(req->rq_err, "E"), \
2262 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
2263 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
2264 FLAG(req->rq_no_resend, "N"), \
2265 FLAG(req->rq_waiting, "W"), \
2266 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
2267 FLAG(req->rq_committed, "M")
2269 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
2271 void _debug_req(struct ptlrpc_request *req,
2272 struct libcfs_debug_msg_data *data, const char *fmt, ...)
2273 __attribute__ ((format (printf, 3, 4)));
2276 * Helper that decides if we need to print request accordig to current debug
2279 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
2281 CFS_CHECK_STACK(msgdata, mask, cdls); \
2283 if (((mask) & D_CANTMASK) != 0 || \
2284 ((libcfs_debug & (mask)) != 0 && \
2285 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
2286 _debug_req((req), msgdata, fmt, ##a); \
2290 * This is the debug print function you need to use to print request sturucture
2291 * content into lustre debug log.
2292 * for most callers (level is a constant) this is resolved at compile time */
2293 #define DEBUG_REQ(level, req, fmt, args...) \
2295 if ((level) & (D_ERROR | D_WARNING)) { \
2296 static cfs_debug_limit_state_t cdls; \
2297 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
2298 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
2300 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
2301 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
2307 * Structure that defines a single page of a bulk transfer
2309 struct ptlrpc_bulk_page {
2310 /** Linkage to list of pages in a bulk */
2311 struct list_head bp_link;
2313 * Number of bytes in a page to transfer starting from \a bp_pageoffset
2316 /** offset within a page */
2318 /** The page itself */
2319 struct page *bp_page;
2322 #define BULK_GET_SOURCE 0
2323 #define BULK_PUT_SINK 1
2324 #define BULK_GET_SINK 2
2325 #define BULK_PUT_SOURCE 3
2328 * Definition of bulk descriptor.
2329 * Bulks are special "Two phase" RPCs where initial request message
2330 * is sent first and it is followed bt a transfer (o receiving) of a large
2331 * amount of data to be settled into pages referenced from the bulk descriptors.
2332 * Bulks transfers (the actual data following the small requests) are done
2333 * on separate LNet portals.
2334 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
2335 * Another user is readpage for MDT.
2337 struct ptlrpc_bulk_desc {
2338 /** completed with failure */
2339 unsigned long bd_failure:1;
2340 /** {put,get}{source,sink} */
2341 unsigned long bd_type:2;
2343 unsigned long bd_registered:1;
2344 /** For serialization with callback */
2346 /** Import generation when request for this bulk was sent */
2347 int bd_import_generation;
2348 /** LNet portal for this bulk */
2350 /** Server side - export this bulk created for */
2351 struct obd_export *bd_export;
2352 /** Client side - import this bulk was sent on */
2353 struct obd_import *bd_import;
2354 /** Back pointer to the request */
2355 struct ptlrpc_request *bd_req;
2356 wait_queue_head_t bd_waitq; /* server side only WQ */
2357 int bd_iov_count; /* # entries in bd_iov */
2358 int bd_max_iov; /* allocated size of bd_iov */
2359 int bd_nob; /* # bytes covered */
2360 int bd_nob_transferred; /* # bytes GOT/PUT */
2364 struct ptlrpc_cb_id bd_cbid; /* network callback info */
2365 lnet_nid_t bd_sender; /* stash event::sender */
2366 int bd_md_count; /* # valid entries in bd_mds */
2367 int bd_md_max_brw; /* max entries in bd_mds */
2368 /** array of associated MDs */
2369 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
2372 * encrypt iov, size is either 0 or bd_iov_count.
2374 lnet_kiov_t *bd_enc_iov;
2376 lnet_kiov_t bd_iov[0];
2380 SVC_STOPPED = 1 << 0,
2381 SVC_STOPPING = 1 << 1,
2382 SVC_STARTING = 1 << 2,
2383 SVC_RUNNING = 1 << 3,
2385 SVC_SIGNAL = 1 << 5,
2388 #define PTLRPC_THR_NAME_LEN 32
2390 * Definition of server service thread structure
2392 struct ptlrpc_thread {
2394 * List of active threads in svc->srv_threads
2396 struct list_head t_link;
2398 * thread-private data (preallocated memory)
2403 * service thread index, from ptlrpc_start_threads
2407 * service thread pid
2411 * put watchdog in the structure per thread b=14840
2413 struct lc_watchdog *t_watchdog;
2415 * the svc this thread belonged to b=18582
2417 struct ptlrpc_service_part *t_svcpt;
2418 wait_queue_head_t t_ctl_waitq;
2419 struct lu_env *t_env;
2420 char t_name[PTLRPC_THR_NAME_LEN];
2423 static inline int thread_is_init(struct ptlrpc_thread *thread)
2425 return thread->t_flags == 0;
2428 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2430 return !!(thread->t_flags & SVC_STOPPED);
2433 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2435 return !!(thread->t_flags & SVC_STOPPING);
2438 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2440 return !!(thread->t_flags & SVC_STARTING);
2443 static inline int thread_is_running(struct ptlrpc_thread *thread)
2445 return !!(thread->t_flags & SVC_RUNNING);
2448 static inline int thread_is_event(struct ptlrpc_thread *thread)
2450 return !!(thread->t_flags & SVC_EVENT);
2453 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2455 return !!(thread->t_flags & SVC_SIGNAL);
2458 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2460 thread->t_flags &= ~flags;
2463 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2465 thread->t_flags = flags;
2468 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2470 thread->t_flags |= flags;
2473 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2476 if (thread->t_flags & flags) {
2477 thread->t_flags &= ~flags;
2484 * Request buffer descriptor structure.
2485 * This is a structure that contains one posted request buffer for service.
2486 * Once data land into a buffer, event callback creates actual request and
2487 * notifies wakes one of the service threads to process new incoming request.
2488 * More than one request can fit into the buffer.
2490 struct ptlrpc_request_buffer_desc {
2491 /** Link item for rqbds on a service */
2492 struct list_head rqbd_list;
2493 /** History of requests for this buffer */
2494 struct list_head rqbd_reqs;
2495 /** Back pointer to service for which this buffer is registered */
2496 struct ptlrpc_service_part *rqbd_svcpt;
2497 /** LNet descriptor */
2498 lnet_handle_md_t rqbd_md_h;
2500 /** The buffer itself */
2502 struct ptlrpc_cb_id rqbd_cbid;
2504 * This "embedded" request structure is only used for the
2505 * last request to fit into the buffer
2507 struct ptlrpc_request rqbd_req;
2510 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2512 struct ptlrpc_service_ops {
2514 * if non-NULL called during thread creation (ptlrpc_start_thread())
2515 * to initialize service specific per-thread state.
2517 int (*so_thr_init)(struct ptlrpc_thread *thr);
2519 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2520 * destruct state created by ->srv_init().
2522 void (*so_thr_done)(struct ptlrpc_thread *thr);
2524 * Handler function for incoming requests for this service
2526 int (*so_req_handler)(struct ptlrpc_request *req);
2528 * function to determine priority of the request, it's called
2529 * on every new request
2531 int (*so_hpreq_handler)(struct ptlrpc_request *);
2533 * service-specific print fn
2535 void (*so_req_printer)(void *, struct ptlrpc_request *);
2538 #ifndef __cfs_cacheline_aligned
2539 /* NB: put it here for reducing patche dependence */
2540 # define __cfs_cacheline_aligned
2544 * How many high priority requests to serve before serving one normal
2547 #define PTLRPC_SVC_HP_RATIO 10
2550 * Definition of PortalRPC service.
2551 * The service is listening on a particular portal (like tcp port)
2552 * and perform actions for a specific server like IO service for OST
2553 * or general metadata service for MDS.
2555 struct ptlrpc_service {
2556 /** serialize /proc operations */
2557 spinlock_t srv_lock;
2558 /** most often accessed fields */
2559 /** chain thru all services */
2560 struct list_head srv_list;
2561 /** service operations table */
2562 struct ptlrpc_service_ops srv_ops;
2563 /** only statically allocated strings here; we don't clean them */
2565 /** only statically allocated strings here; we don't clean them */
2566 char *srv_thread_name;
2567 /** service thread list */
2568 struct list_head srv_threads;
2569 /** threads # should be created for each partition on initializing */
2570 int srv_nthrs_cpt_init;
2571 /** limit of threads number for each partition */
2572 int srv_nthrs_cpt_limit;
2573 /** Root of /proc dir tree for this service */
2574 struct proc_dir_entry *srv_procroot;
2575 /** Pointer to statistic data for this service */
2576 struct lprocfs_stats *srv_stats;
2577 /** # hp per lp reqs to handle */
2578 int srv_hpreq_ratio;
2579 /** biggest request to receive */
2580 int srv_max_req_size;
2581 /** biggest reply to send */
2582 int srv_max_reply_size;
2583 /** size of individual buffers */
2585 /** # buffers to allocate in 1 group */
2586 int srv_nbuf_per_group;
2587 /** Local portal on which to receive requests */
2588 __u32 srv_req_portal;
2589 /** Portal on the client to send replies to */
2590 __u32 srv_rep_portal;
2592 * Tags for lu_context associated with this thread, see struct
2596 /** soft watchdog timeout multiplier */
2597 int srv_watchdog_factor;
2598 /** under unregister_service */
2599 unsigned srv_is_stopping:1;
2601 /** max # request buffers in history per partition */
2602 int srv_hist_nrqbds_cpt_max;
2603 /** number of CPTs this service bound on */
2605 /** CPTs array this service bound on */
2607 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2609 /** CPT table this service is running over */
2610 struct cfs_cpt_table *srv_cptable;
2612 * partition data for ptlrpc service
2614 struct ptlrpc_service_part *srv_parts[0];
2618 * Definition of PortalRPC service partition data.
2619 * Although a service only has one instance of it right now, but we
2620 * will have multiple instances very soon (instance per CPT).
2622 * it has four locks:
2624 * serialize operations on rqbd and requests waiting for preprocess
2626 * serialize operations active requests sent to this portal
2628 * serialize adaptive timeout stuff
2630 * serialize operations on RS list (reply states)
2632 * We don't have any use-case to take two or more locks at the same time
2633 * for now, so there is no lock order issue.
2635 struct ptlrpc_service_part {
2636 /** back reference to owner */
2637 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2638 /* CPT id, reserved */
2640 /** always increasing number */
2642 /** # of starting threads */
2643 int scp_nthrs_starting;
2644 /** # of stopping threads, reserved for shrinking threads */
2645 int scp_nthrs_stopping;
2646 /** # running threads */
2647 int scp_nthrs_running;
2648 /** service threads list */
2649 struct list_head scp_threads;
2652 * serialize the following fields, used for protecting
2653 * rqbd list and incoming requests waiting for preprocess,
2654 * threads starting & stopping are also protected by this lock.
2656 spinlock_t scp_lock __cfs_cacheline_aligned;
2657 /** total # req buffer descs allocated */
2658 int scp_nrqbds_total;
2659 /** # posted request buffers for receiving */
2660 int scp_nrqbds_posted;
2661 /** in progress of allocating rqbd */
2662 int scp_rqbd_allocating;
2663 /** # incoming reqs */
2664 int scp_nreqs_incoming;
2665 /** request buffers to be reposted */
2666 struct list_head scp_rqbd_idle;
2667 /** req buffers receiving */
2668 struct list_head scp_rqbd_posted;
2669 /** incoming reqs */
2670 struct list_head scp_req_incoming;
2671 /** timeout before re-posting reqs, in tick */
2672 cfs_duration_t scp_rqbd_timeout;
2674 * all threads sleep on this. This wait-queue is signalled when new
2675 * incoming request arrives and when difficult reply has to be handled.
2677 wait_queue_head_t scp_waitq;
2679 /** request history */
2680 struct list_head scp_hist_reqs;
2681 /** request buffer history */
2682 struct list_head scp_hist_rqbds;
2683 /** # request buffers in history */
2684 int scp_hist_nrqbds;
2685 /** sequence number for request */
2687 /** highest seq culled from history */
2688 __u64 scp_hist_seq_culled;
2691 * serialize the following fields, used for processing requests
2692 * sent to this portal
2694 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2695 /** # reqs in either of the NRS heads below */
2696 /** # reqs being served */
2697 int scp_nreqs_active;
2698 /** # HPreqs being served */
2699 int scp_nhreqs_active;
2700 /** # hp requests handled */
2703 /** NRS head for regular requests */
2704 struct ptlrpc_nrs scp_nrs_reg;
2705 /** NRS head for HP requests; this is only valid for services that can
2706 * handle HP requests */
2707 struct ptlrpc_nrs *scp_nrs_hp;
2712 * serialize the following fields, used for changes on
2715 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2716 /** estimated rpc service time */
2717 struct adaptive_timeout scp_at_estimate;
2718 /** reqs waiting for replies */
2719 struct ptlrpc_at_array scp_at_array;
2720 /** early reply timer */
2721 struct timer_list scp_at_timer;
2723 cfs_time_t scp_at_checktime;
2724 /** check early replies */
2725 unsigned scp_at_check;
2729 * serialize the following fields, used for processing
2730 * replies for this portal
2732 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2733 /** all the active replies */
2734 struct list_head scp_rep_active;
2735 /** List of free reply_states */
2736 struct list_head scp_rep_idle;
2737 /** waitq to run, when adding stuff to srv_free_rs_list */
2738 wait_queue_head_t scp_rep_waitq;
2739 /** # 'difficult' replies */
2740 atomic_t scp_nreps_difficult;
2743 #define ptlrpc_service_for_each_part(part, i, svc) \
2745 i < (svc)->srv_ncpts && \
2746 (svc)->srv_parts != NULL && \
2747 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2750 * Declaration of ptlrpcd control structure
2752 struct ptlrpcd_ctl {
2754 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2756 unsigned long pc_flags;
2758 * Thread lock protecting structure fields.
2764 struct completion pc_starting;
2768 struct completion pc_finishing;
2770 * Thread requests set.
2772 struct ptlrpc_request_set *pc_set;
2774 * Thread name used in kthread_run()
2778 * Environment for request interpreters to run in.
2780 struct lu_env pc_env;
2782 * Index of ptlrpcd thread in the array.
2786 * Number of the ptlrpcd's partners.
2790 * Pointer to the array of partners' ptlrpcd_ctl structure.
2792 struct ptlrpcd_ctl **pc_partners;
2794 * Record the partner index to be processed next.
2799 /* Bits for pc_flags */
2800 enum ptlrpcd_ctl_flags {
2802 * Ptlrpc thread start flag.
2804 LIOD_START = 1 << 0,
2806 * Ptlrpc thread stop flag.
2810 * Ptlrpc thread force flag (only stop force so far).
2811 * This will cause aborting any inflight rpcs handled
2812 * by thread if LIOD_STOP is specified.
2814 LIOD_FORCE = 1 << 2,
2816 * This is a recovery ptlrpc thread.
2818 LIOD_RECOVERY = 1 << 3,
2820 * The ptlrpcd is bound to some CPU core.
2829 * Service compatibility function; the policy is compatible with all services.
2831 * \param[in] svc The service the policy is attempting to register with.
2832 * \param[in] desc The policy descriptor
2834 * \retval true The policy is compatible with the service
2836 * \see ptlrpc_nrs_pol_desc::pd_compat()
2838 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2839 const struct ptlrpc_nrs_pol_desc *desc)
2845 * Service compatibility function; the policy is compatible with only a specific
2846 * service which is identified by its human-readable name at
2847 * ptlrpc_service::srv_name.
2849 * \param[in] svc The service the policy is attempting to register with.
2850 * \param[in] desc The policy descriptor
2852 * \retval false The policy is not compatible with the service
2853 * \retval true The policy is compatible with the service
2855 * \see ptlrpc_nrs_pol_desc::pd_compat()
2857 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2858 const struct ptlrpc_nrs_pol_desc *desc)
2860 LASSERT(desc->pd_compat_svc_name != NULL);
2861 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2866 /* ptlrpc/events.c */
2867 extern lnet_handle_eq_t ptlrpc_eq_h;
2868 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2869 lnet_process_id_t *peer, lnet_nid_t *self);
2871 * These callbacks are invoked by LNet when something happened to
2875 extern void request_out_callback(lnet_event_t *ev);
2876 extern void reply_in_callback(lnet_event_t *ev);
2877 extern void client_bulk_callback(lnet_event_t *ev);
2878 extern void request_in_callback(lnet_event_t *ev);
2879 extern void reply_out_callback(lnet_event_t *ev);
2880 #ifdef HAVE_SERVER_SUPPORT
2881 extern void server_bulk_callback(lnet_event_t *ev);
2885 /* ptlrpc/connection.c */
2886 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2888 struct obd_uuid *uuid);
2889 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2890 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2891 int ptlrpc_connection_init(void);
2892 void ptlrpc_connection_fini(void);
2893 extern lnet_pid_t ptl_get_pid(void);
2895 /* ptlrpc/niobuf.c */
2897 * Actual interfacing with LNet to put/get/register/unregister stuff
2900 #ifdef HAVE_SERVER_SUPPORT
2901 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2902 unsigned npages, unsigned max_brw,
2903 unsigned type, unsigned portal);
2904 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2905 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2907 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2911 LASSERT(desc != NULL);
2913 spin_lock(&desc->bd_lock);
2914 rc = desc->bd_md_count;
2915 spin_unlock(&desc->bd_lock);
2920 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2921 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2923 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2925 struct ptlrpc_bulk_desc *desc;
2928 LASSERT(req != NULL);
2929 desc = req->rq_bulk;
2931 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2932 req->rq_bulk_deadline > cfs_time_current_sec())
2938 spin_lock(&desc->bd_lock);
2939 rc = desc->bd_md_count;
2940 spin_unlock(&desc->bd_lock);
2944 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2945 #define PTLRPC_REPLY_EARLY 0x02
2946 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2947 int ptlrpc_reply(struct ptlrpc_request *req);
2948 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2949 int ptlrpc_error(struct ptlrpc_request *req);
2950 void ptlrpc_resend_req(struct ptlrpc_request *request);
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 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2972 void ptlrpc_restart_req(struct ptlrpc_request *req);
2973 void ptlrpc_abort_inflight(struct obd_import *imp);
2974 void ptlrpc_cleanup_imp(struct obd_import *imp);
2975 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2977 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2978 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2980 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2981 set_interpreter_func fn, void *data);
2982 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2983 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2984 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2985 int ptlrpc_expired_set(void *data);
2986 void ptlrpc_interrupted_set(void *data);
2987 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2988 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2989 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2990 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2991 struct ptlrpc_request *req);
2993 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2994 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2996 struct ptlrpc_request_pool *
2997 ptlrpc_init_rq_pool(int, int,
2998 void (*populate_pool)(struct ptlrpc_request_pool *, int));
3000 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
3001 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
3002 const struct req_format *format);
3003 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
3004 struct ptlrpc_request_pool *,
3005 const struct req_format *format);
3006 void ptlrpc_request_free(struct ptlrpc_request *request);
3007 int ptlrpc_request_pack(struct ptlrpc_request *request,
3008 __u32 version, int opcode);
3009 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
3010 const struct req_format *format,
3011 __u32 version, int opcode);
3012 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
3013 __u32 version, int opcode, char **bufs,
3014 struct ptlrpc_cli_ctx *ctx);
3015 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
3016 int opcode, int count, __u32 *lengths,
3018 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
3019 __u32 version, int opcode,
3020 int count, __u32 *lengths, char **bufs,
3021 struct ptlrpc_request_pool *pool);
3022 void ptlrpc_req_finished(struct ptlrpc_request *request);
3023 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
3024 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
3025 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
3026 unsigned npages, unsigned max_brw,
3027 unsigned type, unsigned portal);
3028 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
3029 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
3031 __ptlrpc_free_bulk(bulk, 1);
3033 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
3035 __ptlrpc_free_bulk(bulk, 0);
3037 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
3038 struct page *page, int pageoffset, int len, int);
3039 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
3040 struct page *page, int pageoffset,
3043 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
3046 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
3047 struct page *page, int pageoffset,
3050 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
3053 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3054 struct obd_import *imp);
3055 __u64 ptlrpc_next_xid(void);
3056 __u64 ptlrpc_sample_next_xid(void);
3057 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
3059 /* Set of routines to run a function in ptlrpcd context */
3060 void *ptlrpcd_alloc_work(struct obd_import *imp,
3061 int (*cb)(const struct lu_env *, void *), void *data);
3062 void ptlrpcd_destroy_work(void *handler);
3063 int ptlrpcd_queue_work(void *handler);
3066 struct ptlrpc_service_buf_conf {
3067 /* nbufs is buffers # to allocate when growing the pool */
3068 unsigned int bc_nbufs;
3069 /* buffer size to post */
3070 unsigned int bc_buf_size;
3071 /* portal to listed for requests on */
3072 unsigned int bc_req_portal;
3073 /* portal of where to send replies to */
3074 unsigned int bc_rep_portal;
3075 /* maximum request size to be accepted for this service */
3076 unsigned int bc_req_max_size;
3077 /* maximum reply size this service can ever send */
3078 unsigned int bc_rep_max_size;
3081 struct ptlrpc_service_thr_conf {
3082 /* threadname should be 8 characters or less - 6 will be added on */
3084 /* threads increasing factor for each CPU */
3085 unsigned int tc_thr_factor;
3086 /* service threads # to start on each partition while initializing */
3087 unsigned int tc_nthrs_init;
3089 * low water of threads # upper-limit on each partition while running,
3090 * service availability may be impacted if threads number is lower
3091 * than this value. It can be ZERO if the service doesn't require
3092 * CPU affinity or there is only one partition.
3094 unsigned int tc_nthrs_base;
3095 /* "soft" limit for total threads number */
3096 unsigned int tc_nthrs_max;
3097 /* user specified threads number, it will be validated due to
3098 * other members of this structure. */
3099 unsigned int tc_nthrs_user;
3100 /* set NUMA node affinity for service threads */
3101 unsigned int tc_cpu_affinity;
3102 /* Tags for lu_context associated with service thread */
3106 struct ptlrpc_service_cpt_conf {
3107 struct cfs_cpt_table *cc_cptable;
3108 /* string pattern to describe CPTs for a service */
3112 struct ptlrpc_service_conf {
3115 /* soft watchdog timeout multiplifier to print stuck service traces */
3116 unsigned int psc_watchdog_factor;
3117 /* buffer information */
3118 struct ptlrpc_service_buf_conf psc_buf;
3119 /* thread information */
3120 struct ptlrpc_service_thr_conf psc_thr;
3121 /* CPU partition information */
3122 struct ptlrpc_service_cpt_conf psc_cpt;
3123 /* function table */
3124 struct ptlrpc_service_ops psc_ops;
3127 /* ptlrpc/service.c */
3129 * Server-side services API. Register/unregister service, request state
3130 * management, service thread management
3134 void ptlrpc_save_lock(struct ptlrpc_request *req,
3135 struct lustre_handle *lock, int mode, int no_ack);
3136 void ptlrpc_commit_replies(struct obd_export *exp);
3137 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
3138 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
3139 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
3140 struct ptlrpc_service *ptlrpc_register_service(
3141 struct ptlrpc_service_conf *conf,
3142 struct proc_dir_entry *proc_entry);
3143 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
3145 int ptlrpc_start_threads(struct ptlrpc_service *svc);
3146 int ptlrpc_unregister_service(struct ptlrpc_service *service);
3147 int liblustre_check_services(void *arg);
3148 void ptlrpc_daemonize(char *name);
3149 int ptlrpc_service_health_check(struct ptlrpc_service *);
3150 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
3151 void ptlrpc_request_change_export(struct ptlrpc_request *req,
3152 struct obd_export *export);
3153 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay);
3155 int ptlrpc_hr_init(void);
3156 void ptlrpc_hr_fini(void);
3160 /* ptlrpc/import.c */
3165 int ptlrpc_connect_import(struct obd_import *imp);
3166 int ptlrpc_init_import(struct obd_import *imp);
3167 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
3168 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
3169 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
3172 /* ptlrpc/pack_generic.c */
3173 int ptlrpc_reconnect_import(struct obd_import *imp);
3177 * ptlrpc msg buffer and swab interface
3181 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
3183 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
3185 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
3186 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
3188 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
3189 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
3191 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
3192 __u32 *lens, char **bufs);
3193 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
3195 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
3196 __u32 *lens, char **bufs, int flags);
3197 #define LPRFL_EARLY_REPLY 1
3198 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
3199 char **bufs, int flags);
3200 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
3201 unsigned int newlen, int move_data);
3202 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
3203 int __lustre_unpack_msg(struct lustre_msg *m, int len);
3204 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
3205 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
3206 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
3207 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
3208 __u32 lustre_msg_early_size(void);
3209 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
3210 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
3211 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
3212 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
3213 __u32 lustre_msg_bufcount(struct lustre_msg *m);
3214 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
3215 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
3216 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
3217 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
3218 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
3219 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
3220 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
3221 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
3222 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
3223 void lustre_msg_set_op_flags(struct lustre_msg *msg, __u32 flags);
3224 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
3225 __u32 lustre_msg_get_type(struct lustre_msg *msg);
3226 __u32 lustre_msg_get_version(struct lustre_msg *msg);
3227 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
3228 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
3229 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
3230 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
3231 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
3232 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
3233 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
3234 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
3235 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
3236 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
3237 int lustre_msg_get_status(struct lustre_msg *msg);
3238 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
3239 int lustre_msg_is_v1(struct lustre_msg *msg);
3240 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
3241 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
3242 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
3243 char *lustre_msg_get_jobid(struct lustre_msg *msg);
3244 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
3245 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 53, 0)
3246 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
3248 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
3250 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
3251 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
3252 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
3253 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
3254 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
3255 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
3256 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
3257 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
3258 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
3259 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
3260 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
3261 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
3262 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
3263 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
3264 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
3267 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
3268 unsigned int newlen, int move_data)
3270 LASSERT(req->rq_reply_state);
3271 LASSERT(req->rq_repmsg);
3272 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
3276 #ifdef LUSTRE_TRANSLATE_ERRNOS
3278 static inline int ptlrpc_status_hton(int h)
3281 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
3282 * ELDLM_LOCK_ABORTED, etc.
3285 return -lustre_errno_hton(-h);
3290 static inline int ptlrpc_status_ntoh(int n)
3293 * See the comment in ptlrpc_status_hton().
3296 return -lustre_errno_ntoh(-n);
3303 #define ptlrpc_status_hton(h) (h)
3304 #define ptlrpc_status_ntoh(n) (n)
3309 /** Change request phase of \a req to \a new_phase */
3311 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
3313 if (req->rq_phase == new_phase)
3316 if (new_phase == RQ_PHASE_UNREGISTERING) {
3317 req->rq_next_phase = req->rq_phase;
3319 atomic_inc(&req->rq_import->imp_unregistering);
3322 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
3324 atomic_dec(&req->rq_import->imp_unregistering);
3327 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
3328 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
3330 req->rq_phase = new_phase;
3334 * Returns true if request \a req got early reply and hard deadline is not met
3337 ptlrpc_client_early(struct ptlrpc_request *req)
3339 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3340 req->rq_reply_deadline > cfs_time_current_sec())
3342 return req->rq_early;
3346 * Returns true if we got real reply from server for this request
3349 ptlrpc_client_replied(struct ptlrpc_request *req)
3351 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3352 req->rq_reply_deadline > cfs_time_current_sec())
3354 return req->rq_replied;
3357 /** Returns true if request \a req is in process of receiving server reply */
3359 ptlrpc_client_recv(struct ptlrpc_request *req)
3361 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3362 req->rq_reply_deadline > cfs_time_current_sec())
3364 return req->rq_receiving_reply;
3368 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
3372 spin_lock(&req->rq_lock);
3373 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3374 req->rq_reply_deadline > cfs_time_current_sec()) {
3375 spin_unlock(&req->rq_lock);
3378 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
3379 req->rq_receiving_reply;
3380 spin_unlock(&req->rq_lock);
3385 ptlrpc_client_wake_req(struct ptlrpc_request *req)
3387 if (req->rq_set == NULL)
3388 wake_up(&req->rq_reply_waitq);
3390 wake_up(&req->rq_set->set_waitq);
3394 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
3396 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3397 atomic_inc(&rs->rs_refcount);
3401 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
3403 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3404 if (atomic_dec_and_test(&rs->rs_refcount))
3405 lustre_free_reply_state(rs);
3408 /* Should only be called once per req */
3409 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
3411 if (req->rq_reply_state == NULL)
3412 return; /* shouldn't occur */
3413 ptlrpc_rs_decref(req->rq_reply_state);
3414 req->rq_reply_state = NULL;
3415 req->rq_repmsg = NULL;
3418 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3420 return lustre_msg_get_magic(req->rq_reqmsg);
3423 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3425 switch (req->rq_reqmsg->lm_magic) {
3426 case LUSTRE_MSG_MAGIC_V2:
3427 return req->rq_reqmsg->lm_repsize;
3429 LASSERTF(0, "incorrect message magic: %08x\n",
3430 req->rq_reqmsg->lm_magic);
3435 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3437 if (req->rq_delay_limit != 0 &&
3438 cfs_time_before(cfs_time_add(req->rq_queued_time,
3439 cfs_time_seconds(req->rq_delay_limit)),
3440 cfs_time_current())) {
3446 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3448 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3449 spin_lock(&req->rq_lock);
3450 req->rq_no_resend = 1;
3451 spin_unlock(&req->rq_lock);
3453 return req->rq_no_resend;
3457 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3459 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3461 return svcpt->scp_service->srv_watchdog_factor *
3462 max_t(int, at, obd_timeout);
3465 static inline struct ptlrpc_service *
3466 ptlrpc_req2svc(struct ptlrpc_request *req)
3468 LASSERT(req->rq_rqbd != NULL);
3469 return req->rq_rqbd->rqbd_svcpt->scp_service;
3472 /* ldlm/ldlm_lib.c */
3474 * Target client logic
3477 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3478 int client_obd_cleanup(struct obd_device *obddev);
3479 int client_connect_import(const struct lu_env *env,
3480 struct obd_export **exp, struct obd_device *obd,
3481 struct obd_uuid *cluuid, struct obd_connect_data *,
3483 int client_disconnect_export(struct obd_export *exp);
3484 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3486 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3487 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3488 struct obd_uuid *uuid);
3489 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3490 void client_destroy_import(struct obd_import *imp);
3493 #ifdef HAVE_SERVER_SUPPORT
3494 int server_disconnect_export(struct obd_export *exp);
3497 /* ptlrpc/pinger.c */
3499 * Pinger API (client side only)
3502 enum timeout_event {
3505 struct timeout_item;
3506 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3507 int ptlrpc_pinger_add_import(struct obd_import *imp);
3508 int ptlrpc_pinger_del_import(struct obd_import *imp);
3509 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3510 timeout_cb_t cb, void *data,
3511 struct list_head *obd_list);
3512 int ptlrpc_del_timeout_client(struct list_head *obd_list,
3513 enum timeout_event event);
3514 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3515 int ptlrpc_obd_ping(struct obd_device *obd);
3516 void ping_evictor_start(void);
3517 void ping_evictor_stop(void);
3518 void ptlrpc_pinger_ir_up(void);
3519 void ptlrpc_pinger_ir_down(void);
3521 int ptlrpc_pinger_suppress_pings(void);
3523 /* ptlrpc daemon bind policy */
3525 /* all ptlrpcd threads are free mode */
3526 PDB_POLICY_NONE = 1,
3527 /* all ptlrpcd threads are bound mode */
3528 PDB_POLICY_FULL = 2,
3529 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3530 PDB_POLICY_PAIR = 3,
3531 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3532 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3533 * If kernel supports NUMA, pthrpcd threads are binded and
3534 * grouped by NUMA node */
3535 PDB_POLICY_NEIGHBOR = 4,
3538 /* ptlrpc daemon load policy
3539 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3540 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3541 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3542 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3543 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3545 /* on the same CPU core as the caller */
3546 PDL_POLICY_SAME = 1,
3547 /* within the same CPU partition, but not the same core as the caller */
3548 PDL_POLICY_LOCAL = 2,
3549 /* round-robin on all CPU cores, but not the same core as the caller */
3550 PDL_POLICY_ROUND = 3,
3551 /* the specified CPU core is preferred, but not enforced */
3552 PDL_POLICY_PREFERRED = 4,
3555 /* ptlrpc/ptlrpcd.c */
3556 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3557 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3558 void ptlrpcd_wake(struct ptlrpc_request *req);
3559 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3560 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3561 int ptlrpcd_addref(void);
3562 void ptlrpcd_decref(void);
3564 /* ptlrpc/lproc_ptlrpc.c */
3566 * procfs output related functions
3569 const char* ll_opcode2str(__u32 opcode);
3571 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3572 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3573 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3575 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3576 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3577 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3581 /* ptlrpc/llog_server.c */
3582 int llog_origin_handle_open(struct ptlrpc_request *req);
3583 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3584 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3585 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3586 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3587 int llog_origin_handle_close(struct ptlrpc_request *req);
3589 /* ptlrpc/llog_client.c */
3590 extern struct llog_operations llog_client_ops;