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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 /** \defgroup PtlRPC Portal RPC and networking module.
33 * PortalRPC is the layer used by rest of lustre code to achieve network
34 * communications: establish connections with corresponding export and import
35 * states, listen for a service, send and receive RPCs.
36 * PortalRPC also includes base recovery framework: packet resending and
37 * replaying, reconnections, pinger.
39 * PortalRPC utilizes LNet as its transport layer.
52 #include <linux/kobject.h>
53 #include <linux/rhashtable.h>
54 #include <linux/uio.h>
55 #include <libcfs/libcfs.h>
57 #include <lnet/lib-types.h>
58 #include <uapi/linux/lnet/nidstr.h>
59 #include <uapi/linux/lustre/lustre_idl.h>
60 #include <lustre_ha.h>
61 #include <lustre_sec.h>
62 #include <lustre_import.h>
63 #include <lprocfs_status.h>
64 #include <lu_object.h>
65 #include <lustre_req_layout.h>
66 #include <obd_support.h>
67 #include <uapi/linux/lustre/lustre_ver.h>
69 /* MD flags we _always_ use */
70 #define PTLRPC_MD_OPTIONS 0
73 * log2 max # of bulk operations in one request: 2=4MB/RPC, 5=32MB/RPC, ...
74 * In order for the client and server to properly negotiate the maximum
75 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
76 * value. The client is free to limit the actual RPC size for any bulk
77 * transfer via cl_max_pages_per_rpc to some non-power-of-two value.
78 * NOTE: This is limited to 16 (=64GB RPCs) by IOOBJ_MAX_BRW_BITS. */
79 #define PTLRPC_BULK_OPS_BITS 6
80 #if PTLRPC_BULK_OPS_BITS > 16
81 #error "More than 65536 BRW RPCs not allowed by IOOBJ_MAX_BRW_BITS."
83 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
85 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
86 * should not be used on the server at all. Otherwise, it imposes a
87 * protocol limitation on the maximum RPC size that can be used by any
88 * RPC sent to that server in the future. Instead, the server should
89 * use the negotiated per-client ocd_brw_size to determine the bulk
91 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
94 * Define maxima for bulk I/O.
96 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
97 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
98 * currently supported maximum between peers at connect via ocd_brw_size.
100 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
101 #define PTLRPC_MAX_BRW_SIZE (1U << PTLRPC_MAX_BRW_BITS)
102 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_SHIFT)
104 #define ONE_MB_BRW_SIZE (1U << LNET_MTU_BITS)
105 #define MD_MAX_BRW_SIZE (1U << LNET_MTU_BITS)
106 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_SHIFT)
107 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
108 #define DT_DEF_BRW_SIZE (4 * ONE_MB_BRW_SIZE)
109 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_SHIFT)
110 #define OFD_MAX_BRW_SIZE (1U << LNET_MTU_BITS)
112 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
113 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
114 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
116 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_SIZE))
117 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_SIZE"
119 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
120 # error "PTLRPC_MAX_BRW_SIZE too big"
122 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
123 # error "PTLRPC_MAX_BRW_PAGES too big"
126 #define PTLRPC_NTHRS_INIT 2
131 * Constants determine how memory is used to buffer incoming service requests.
133 * ?_NBUFS # buffers to allocate when growing the pool
134 * ?_BUFSIZE # bytes in a single request buffer
135 * ?_MAXREQSIZE # maximum request service will receive
137 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
138 * of ?_NBUFS is added to the pool.
140 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
141 * considered full when less than ?_MAXREQSIZE is left in them.
146 * Constants determine how threads are created for ptlrpc service.
148 * ?_NTHRS_INIT # threads to create for each service partition on
149 * initializing. If it's non-affinity service and
150 * there is only one partition, it's the overall #
151 * threads for the service while initializing.
152 * ?_NTHRS_BASE # threads should be created at least for each
153 * ptlrpc partition to keep the service healthy.
154 * It's the low-water mark of threads upper-limit
155 * for each partition.
156 * ?_THR_FACTOR # threads can be added on threads upper-limit for
157 * each CPU core. This factor is only for reference,
158 * we might decrease value of factor if number of cores
159 * per CPT is above a limit.
160 * ?_NTHRS_MAX # overall threads can be created for a service,
161 * it's a soft limit because if service is running
162 * on machine with hundreds of cores and tens of
163 * CPU partitions, we need to guarantee each partition
164 * has ?_NTHRS_BASE threads, which means total threads
165 * will be ?_NTHRS_BASE * number_of_cpts which can
166 * exceed ?_NTHRS_MAX.
170 * #define MDS_NTHRS_INIT 2
171 * #define MDS_NTHRS_BASE 64
172 * #define MDS_NTHRS_FACTOR 8
173 * #define MDS_NTHRS_MAX 1024
176 * ---------------------------------------------------------------------
177 * Server(A) has 16 cores, user configured it to 4 partitions so each
178 * partition has 4 cores, then actual number of service threads on each
180 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
182 * Total number of threads for the service is:
183 * 96 * partitions(4) = 384
186 * ---------------------------------------------------------------------
187 * Server(B) has 32 cores, user configured it to 4 partitions so each
188 * partition has 8 cores, then actual number of service threads on each
190 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
192 * Total number of threads for the service is:
193 * 128 * partitions(4) = 512
196 * ---------------------------------------------------------------------
197 * Server(B) has 96 cores, user configured it to 8 partitions so each
198 * partition has 12 cores, then actual number of service threads on each
200 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
202 * Total number of threads for the service is:
203 * 160 * partitions(8) = 1280
205 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
206 * as upper limit of threads number for each partition:
207 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
210 * ---------------------------------------------------------------------
211 * Server(C) have a thousand of cores and user configured it to 32 partitions
212 * MDS_NTHRS_BASE(64) * 32 = 2048
214 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
215 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
216 * to keep service healthy, so total number of threads will just be 2048.
218 * NB: we don't suggest to choose server with that many cores because backend
219 * filesystem itself, buffer cache, or underlying network stack might
220 * have some SMP scalability issues at that large scale.
222 * If user already has a fat machine with hundreds or thousands of cores,
223 * there are two choices for configuration:
224 * a) create CPU table from subset of all CPUs and run Lustre on
226 * b) bind service threads on a few partitions, see modparameters of
227 * MDS and OSS for details
229 * NB: these calculations (and examples below) are simplified to help
230 * understanding, the real implementation is a little more complex,
231 * please see ptlrpc_server_nthreads_check() for details.
236 * LDLM threads constants:
238 * Given 8 as factor and 24 as base threads number
241 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
244 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
245 * threads for each partition and total threads number will be 112.
248 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
249 * threads for each partition to keep service healthy, so total threads
250 * number should be 24 * 8 = 192.
252 * So with these constants, threads number will be at the similar level
253 * of old versions, unless target machine has over a hundred cores
255 #define LDLM_THR_FACTOR 8
256 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
257 #define LDLM_NTHRS_BASE 24
258 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
260 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
261 #define LDLM_CLIENT_NBUFS 1
262 #define LDLM_SERVER_NBUFS 64
263 #define LDLM_BUFSIZE (8 * 1024)
264 #define LDLM_MAXREQSIZE (5 * 1024)
265 #define LDLM_MAXREPSIZE (1024)
268 * MDS threads constants:
270 * Please see examples in "Thread Constants", MDS threads number will be at
271 * the comparable level of old versions, unless the server has many cores.
273 #ifndef MDS_MAX_THREADS
274 #define MDS_MAX_THREADS 1024
275 #define MDS_MAX_OTHR_THREADS 256
277 #else /* MDS_MAX_THREADS */
278 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
279 #undef MDS_MAX_THREADS
280 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
282 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
285 /* default service */
286 #define MDS_THR_FACTOR 8
287 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
288 #define MDS_NTHRS_MAX MDS_MAX_THREADS
289 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
291 /* read-page service */
292 #define MDS_RDPG_THR_FACTOR 4
293 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
294 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
295 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
297 /* these should be removed when we remove setattr service in the future */
298 #define MDS_SETA_THR_FACTOR 4
299 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
300 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
301 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
303 /* non-affinity threads */
304 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
305 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
310 * Assume file name length = FNAME_MAX = 256 (true for ext3).
311 * path name length = PATH_MAX = 4096
312 * LOV MD size max = EA_MAX = 24 * 2000
313 * (NB: 24 is size of lov_ost_data)
314 * LOV LOGCOOKIE size max = 32 * 2000
315 * (NB: 32 is size of llog_cookie)
316 * symlink: FNAME_MAX + PATH_MAX <- largest
317 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
318 * rename: FNAME_MAX + FNAME_MAX
319 * open: FNAME_MAX + EA_MAX
321 * MDS_MAXREQSIZE ~= 4736 bytes =
322 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
323 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
325 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
326 * except in the open case where there are a large number of OSTs in a LOV.
328 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
329 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
332 * MDS incoming request with LOV EA
333 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
335 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
336 362 + LOV_MAX_STRIPE_COUNT * 24)
338 * MDS outgoing reply with LOV EA
340 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
341 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
343 * but 2.4 or later MDS will never send reply with llog_cookie to any
344 * version client. This macro is defined for server side reply buffer size.
346 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
349 * This is the size of a maximum REINT_SETXATTR request:
351 * lustre_msg 56 (32 + 4 x 5 + 4)
353 * mdt_rec_setxattr 136
355 * name 256 (XATTR_NAME_MAX)
356 * value 65536 (XATTR_SIZE_MAX)
358 #define MDS_EA_MAXREQSIZE 66288
361 * These are the maximum request and reply sizes (rounded up to 1 KB
362 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
364 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
365 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
366 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
369 * The update request includes all of updates from the create, which might
370 * include linkea (4K maxim), together with other updates, we set it to 1000K:
371 * lustre_msg + ptlrpc_body + OUT_UPDATE_BUFFER_SIZE_MAX
373 #define OUT_MAXREQSIZE (1000 * 1024)
374 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
376 #define BUT_MAXREQSIZE OUT_MAXREQSIZE
377 #define BUT_MAXREPSIZE BUT_MAXREQSIZE
379 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
380 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
384 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
385 * However, we need to allocate a much larger buffer for it because LNet
386 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
387 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
388 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
389 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
390 * utilization is very low.
392 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
393 * reused until all requests fit in it have been processed and released,
394 * which means one long blocked request can prevent the rqbd be reused.
395 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
396 * from LNet with unused 65 KB, buffer utilization will be about 59%.
397 * Please check LU-2432 for details.
399 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
403 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
404 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
405 * extra bytes to each request buffer to improve buffer utilization rate.
407 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
410 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
411 #define FLD_MAXREQSIZE (160)
413 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
414 #define FLD_MAXREPSIZE (152)
415 #define FLD_BUFSIZE (1 << 12)
418 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
420 #define SEQ_MAXREQSIZE (160)
422 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
423 #define SEQ_MAXREPSIZE (152)
424 #define SEQ_BUFSIZE (1 << 12)
426 /** MGS threads must be >= 3, see bug 22458 comment #28 */
427 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
428 #define MGS_NTHRS_MAX 32
431 #define MGS_BUFSIZE (8 * 1024)
432 #define MGS_MAXREQSIZE (7 * 1024)
433 #define MGS_MAXREPSIZE (9 * 1024)
436 * OSS threads constants:
438 * Given 8 as factor and 64 as base threads number
441 * On 8-core server configured to 2 partitions, we will have
442 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
445 * On 32-core machine configured to 4 partitions, we will have
446 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
447 * will be 112 * 4 = 448.
450 * On 64-core machine configured to 4 partitions, we will have
451 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
452 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
453 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
454 * for each partition.
456 * So we can see that with these constants, threads number wil be at the
457 * similar level of old versions, unless the server has many cores.
459 /* depress threads factor for VM with small memory size */
460 #define OSS_THR_FACTOR min_t(int, 8, \
461 NUM_CACHEPAGES >> (28 - PAGE_SHIFT))
462 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
463 #define OSS_NTHRS_BASE 64
465 /* threads for handling "create" request */
466 #define OSS_CR_THR_FACTOR 1
467 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
468 #define OSS_CR_NTHRS_BASE 8
469 #define OSS_CR_NTHRS_MAX 64
472 * OST_IO_MAXREQSIZE ~=
473 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
474 * DT_MAX_BRW_PAGES * niobuf_remote
476 * - single object with 16 pages is 512 bytes
477 * - OST_IO_MAXREQSIZE must be at least 1 niobuf per page of data
478 * - Must be a multiple of 1024
479 * - should allow a reasonably large SHORT_IO_BYTES size (64KB)
481 #define _OST_MAXREQSIZE_BASE ((unsigned long)(sizeof(struct lustre_msg) + \
482 /* lm_buflens */ sizeof(__u32) * 4 + \
483 sizeof(struct ptlrpc_body) + \
484 sizeof(struct obdo) + \
485 sizeof(struct obd_ioobj) + \
486 sizeof(struct niobuf_remote)))
487 #define _OST_MAXREQSIZE_SUM ((unsigned long)(_OST_MAXREQSIZE_BASE + \
488 sizeof(struct niobuf_remote) * \
491 * FIEMAP request can be 4K+ for now
493 #define OST_MAXREQSIZE (16UL * 1024UL)
494 #define OST_IO_MAXREQSIZE max(OST_MAXREQSIZE, \
495 ((_OST_MAXREQSIZE_SUM - 1) | \
497 /* Safe estimate of free space in standard RPC, provides upper limit for # of
498 * bytes of i/o to pack in RPC (skipping bulk transfer). */
499 #define OST_MAX_SHORT_IO_BYTES ((OST_IO_MAXREQSIZE - _OST_MAXREQSIZE_BASE) & \
502 /* Actual size used for short i/o buffer. Calculation means this:
503 * At least one page (for large PAGE_SIZE), or 16 KiB, but not more
504 * than the available space aligned to a page boundary. */
505 #define OBD_DEF_SHORT_IO_BYTES min(max(PAGE_SIZE, 16UL * 1024UL), \
506 OST_MAX_SHORT_IO_BYTES)
508 #define OST_MAXREPSIZE (9 * 1024)
509 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
512 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
513 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 32 * 1024)
515 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
516 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
518 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
520 /* Macro to hide a typecast and BUILD_BUG. */
521 #define ptlrpc_req_async_args(_var, req) ({ \
522 BUILD_BUG_ON(sizeof(*_var) > sizeof(req->rq_async_args)); \
523 (typeof(_var))&req->rq_async_args; \
526 struct ptlrpc_replay_async_args {
532 * Structure to single define portal connection.
534 struct ptlrpc_connection {
535 /** linkage for connections hash table */
536 struct rhash_head c_hash;
537 /** Our own lnet nid for this connection */
538 struct lnet_nid c_self;
539 /** Remote side nid for this connection */
540 struct lnet_processid c_peer;
541 /** UUID of the other side */
542 struct obd_uuid c_remote_uuid;
543 /** reference counter for this connection */
547 /** Client definition for PortalRPC */
548 struct ptlrpc_client {
549 /** What lnet portal does this client send messages to by default */
550 __u32 cli_request_portal;
551 /** What portal do we expect replies on */
552 __u32 cli_reply_portal;
553 /** Name of the client */
554 const char *cli_name;
557 /** state flags of requests */
558 /* XXX only ones left are those used by the bulk descs as well! */
559 #define PTL_RPC_FL_INTR BIT(0) /* reply wait was interrupted by user */
560 #define PTL_RPC_FL_TIMEOUT BIT(7) /* request timed out waiting for reply */
562 #define REQ_MAX_ACK_LOCKS 8
564 union ptlrpc_async_args {
566 * Scratchpad for passing args to completion interpreter. Users
567 * cast to the struct of their choosing, and BUILD_BUG_ON that this is
568 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
569 * a pointer to it here. The pointer_arg ensures this struct is at
570 * least big enough for that.
572 void *pointer_arg[11];
576 struct ptlrpc_request_set;
577 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
580 * Definition of request set structure.
581 * Request set is a list of requests (not necessary to the same target) that
582 * once populated with RPCs could be sent in parallel.
583 * There are two kinds of request sets. General purpose and with dedicated
584 * serving thread. Example of the latter is ptlrpcd set.
585 * For general purpose sets once request set started sending it is impossible
586 * to add new requests to such set.
587 * Provides a way to call "completion callbacks" when all requests in the set
590 struct ptlrpc_request_set {
591 atomic_t set_refcount;
592 /** number of in queue requests */
593 atomic_t set_new_count;
594 /** number of uncompleted requests */
595 atomic_t set_remaining;
596 /** wait queue to wait on for request events */
597 wait_queue_head_t set_waitq;
598 /** List of requests in the set */
599 struct list_head set_requests;
601 * Lock for \a set_new_requests manipulations
602 * locked so that any old caller can communicate requests to
603 * the set holder who can then fold them into the lock-free set
605 spinlock_t set_new_req_lock;
606 /** List of new yet unsent requests. Only used with ptlrpcd now. */
607 struct list_head set_new_requests;
609 /** rq_status of requests that have been freed already */
611 /** Additional fields used by the flow control extension */
612 /** Maximum number of RPCs in flight */
613 int set_max_inflight;
614 /** Callback function used to generate RPCs */
615 set_producer_func set_producer;
616 /** opaq argument passed to the producer callback */
617 void *set_producer_arg;
618 unsigned int set_allow_intr:1;
621 struct ptlrpc_bulk_desc;
622 struct ptlrpc_service_part;
623 struct ptlrpc_service;
626 * ptlrpc callback & work item stuff
628 struct ptlrpc_cb_id {
629 void (*cbid_fn)(struct lnet_event *ev); /* specific callback fn */
630 void *cbid_arg; /* additional arg */
633 /** Maximum number of locks to fit into reply state */
634 #define RS_MAX_LOCKS 8
638 * Structure to define reply state on the server
639 * Reply state holds various reply message information. Also for "difficult"
640 * replies (rep-ack case) we store the state after sending reply and wait
641 * for the client to acknowledge the reception. In these cases locks could be
642 * added to the state for replay/failover consistency guarantees.
644 struct ptlrpc_reply_state {
645 /** Callback description */
646 struct ptlrpc_cb_id rs_cb_id;
647 /** Linkage for list of all reply states in a system */
648 struct list_head rs_list;
649 /** Linkage for list of all reply states on same export */
650 struct list_head rs_exp_list;
651 /** Linkage for list of all reply states for same obd */
652 struct list_head rs_obd_list;
654 struct list_head rs_debug_list;
656 /** A spinlock to protect the reply state flags */
658 /** Reply state flags */
659 unsigned long rs_difficult:1; /* ACK/commit stuff */
660 unsigned long rs_no_ack:1; /* no ACK, even for
661 difficult requests */
662 unsigned long rs_scheduled:1; /* being handled? */
663 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
664 unsigned long rs_handled:1; /* been handled yet? */
665 unsigned long rs_sent:1; /* Got LNET_EVENT_SEND? */
666 unsigned long rs_unlinked:1; /* Reply MD unlinked? */
667 unsigned long rs_prealloc:1; /* rs from prealloc list */
668 unsigned long rs_committed:1;/* the transaction was committed
669 and the rs was dispatched
670 by ptlrpc_commit_replies */
671 unsigned long rs_convert_lock:1; /* need to convert saved
672 * locks to COS mode */
673 atomic_t rs_refcount; /* number of users */
674 /** Number of locks awaiting client ACK */
677 /** Size of the state */
681 /** Transaction number */
685 struct obd_export *rs_export;
686 struct ptlrpc_service_part *rs_svcpt;
687 /** Lnet metadata handle for the reply */
688 struct lnet_handle_md rs_md_h;
690 /** Context for the sevice thread */
691 struct ptlrpc_svc_ctx *rs_svc_ctx;
692 /** Reply buffer (actually sent to the client), encoded if needed */
693 struct lustre_msg *rs_repbuf; /* wrapper */
694 /** Size of the reply buffer */
695 int rs_repbuf_len; /* wrapper buf length */
696 /** Size of the reply message */
697 int rs_repdata_len; /* wrapper msg length */
699 * Actual reply message. Its content is encrupted (if needed) to
700 * produce reply buffer for actual sending. In simple case
701 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
703 struct lustre_msg *rs_msg; /* reply message */
705 /** Handles of locks awaiting client reply ACK */
706 struct lustre_handle rs_locks[RS_MAX_LOCKS];
707 /** Lock modes of locks in \a rs_locks */
708 enum ldlm_mode rs_modes[RS_MAX_LOCKS];
711 struct ptlrpc_thread;
715 RQ_PHASE_NEW = 0xebc0de00,
716 RQ_PHASE_RPC = 0xebc0de01,
717 RQ_PHASE_BULK = 0xebc0de02,
718 RQ_PHASE_INTERPRET = 0xebc0de03,
719 RQ_PHASE_COMPLETE = 0xebc0de04,
720 RQ_PHASE_UNREG_RPC = 0xebc0de05,
721 RQ_PHASE_UNREG_BULK = 0xebc0de06,
722 RQ_PHASE_UNDEFINED = 0xebc0de07
725 /** Type of request interpreter call-back */
726 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
727 struct ptlrpc_request *req,
729 /** Type of request resend call-back */
730 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
734 * Definition of request pool structure.
735 * The pool is used to store empty preallocated requests for the case
736 * when we would actually need to send something without performing
737 * any allocations (to avoid e.g. OOM).
739 struct ptlrpc_request_pool {
740 /** Locks the list */
742 /** list of ptlrpc_request structs */
743 struct list_head prp_req_list;
744 /** Maximum message size that would fit into a rquest from this pool */
746 /** Function to allocate more requests for this pool */
747 int (*prp_populate)(struct ptlrpc_request_pool *, int);
755 #include <lustre_nrs.h>
758 * Basic request prioritization operations structure.
759 * The whole idea is centered around locks and RPCs that might affect locks.
760 * When a lock is contended we try to give priority to RPCs that might lead
761 * to fastest release of that lock.
762 * Currently only implemented for OSTs only in a way that makes all
763 * IO and truncate RPCs that are coming from a locked region where a lock is
764 * contended a priority over other requests.
766 struct ptlrpc_hpreq_ops {
768 * Check if the lock handle of the given lock is the same as
769 * taken from the request.
771 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
773 * Check if the request is a high priority one.
775 int (*hpreq_check)(struct ptlrpc_request *);
777 * Called after the request has been handled.
779 void (*hpreq_fini)(struct ptlrpc_request *);
782 struct ptlrpc_cli_req {
783 /** For bulk requests on client only: bulk descriptor */
784 struct ptlrpc_bulk_desc *cr_bulk;
785 /** optional time limit for send attempts. This is a timeout
786 * not a timestamp so timeout_t (s32) is used instead of time64_t
788 timeout_t cr_delay_limit;
789 /** time request was first queued */
790 time64_t cr_queued_time;
791 /** request sent in nanoseconds */
793 /** time for request really sent out */
794 time64_t cr_sent_out;
795 /** when req reply unlink must finish. */
796 time64_t cr_reply_deadline;
797 /** when req bulk unlink must finish. */
798 time64_t cr_bulk_deadline;
799 /** when req unlink must finish. */
800 time64_t cr_req_deadline;
801 /** Portal to which this request would be sent */
803 /** Portal where to wait for reply and where reply would be sent */
805 /** request resending number */
806 unsigned int cr_resend_nr;
807 /** What was import generation when this request was sent */
809 enum lustre_imp_state cr_send_state;
810 /** Per-request waitq introduced by bug 21938 for recovery waiting */
811 wait_queue_head_t cr_set_waitq;
812 /** Link item for request set lists */
813 struct list_head cr_set_chain;
814 /** link to waited ctx */
815 struct list_head cr_ctx_chain;
817 /** client's half ctx */
818 struct ptlrpc_cli_ctx *cr_cli_ctx;
819 /** Link back to the request set */
820 struct ptlrpc_request_set *cr_set;
821 /** outgoing request MD handle */
822 struct lnet_handle_md cr_req_md_h;
823 /** request-out callback parameter */
824 struct ptlrpc_cb_id cr_req_cbid;
825 /** incoming reply MD handle */
826 struct lnet_handle_md cr_reply_md_h;
827 wait_queue_head_t cr_reply_waitq;
828 /** reply callback parameter */
829 struct ptlrpc_cb_id cr_reply_cbid;
830 /** Async completion handler, called when reply is received */
831 ptlrpc_interpterer_t cr_reply_interp;
832 /** Resend handler, called when request is resend to update RPC data */
833 ptlrpc_resend_cb_t cr_resend_cb;
834 /** Async completion context */
835 union ptlrpc_async_args cr_async_args;
836 /** Opaq data for replay and commit callbacks. */
838 /** Link to the imp->imp_unreplied_list */
839 struct list_head cr_unreplied_list;
841 * Commit callback, called when request is committed and about to be
844 void (*cr_commit_cb)(struct ptlrpc_request *);
845 /** Replay callback, called after request is replayed at recovery */
846 void (*cr_replay_cb)(struct ptlrpc_request *);
849 /** client request member alias */
850 /* NB: these alias should NOT be used by any new code, instead they should
851 * be removed step by step to avoid potential abuse */
852 #define rq_bulk rq_cli.cr_bulk
853 #define rq_delay_limit rq_cli.cr_delay_limit
854 #define rq_queued_time rq_cli.cr_queued_time
855 #define rq_sent_ns rq_cli.cr_sent_ns
856 #define rq_real_sent rq_cli.cr_sent_out
857 #define rq_reply_deadline rq_cli.cr_reply_deadline
858 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
859 #define rq_req_deadline rq_cli.cr_req_deadline
860 #define rq_nr_resend rq_cli.cr_resend_nr
861 #define rq_request_portal rq_cli.cr_req_ptl
862 #define rq_reply_portal rq_cli.cr_rep_ptl
863 #define rq_import_generation rq_cli.cr_imp_gen
864 #define rq_send_state rq_cli.cr_send_state
865 #define rq_set_chain rq_cli.cr_set_chain
866 #define rq_ctx_chain rq_cli.cr_ctx_chain
867 #define rq_set rq_cli.cr_set
868 #define rq_set_waitq rq_cli.cr_set_waitq
869 #define rq_cli_ctx rq_cli.cr_cli_ctx
870 #define rq_req_md_h rq_cli.cr_req_md_h
871 #define rq_req_cbid rq_cli.cr_req_cbid
872 #define rq_reply_md_h rq_cli.cr_reply_md_h
873 #define rq_reply_waitq rq_cli.cr_reply_waitq
874 #define rq_reply_cbid rq_cli.cr_reply_cbid
875 #define rq_interpret_reply rq_cli.cr_reply_interp
876 #define rq_resend_cb rq_cli.cr_resend_cb
877 #define rq_async_args rq_cli.cr_async_args
878 #define rq_cb_data rq_cli.cr_cb_data
879 #define rq_unreplied_list rq_cli.cr_unreplied_list
880 #define rq_commit_cb rq_cli.cr_commit_cb
881 #define rq_replay_cb rq_cli.cr_replay_cb
883 struct ptlrpc_srv_req {
884 /** initial thread servicing this request */
885 struct ptlrpc_thread *sr_svc_thread;
887 * Server side list of incoming unserved requests sorted by arrival
888 * time. Traversed from time to time to notice about to expire
889 * requests and sent back "early replies" to clients to let them
890 * know server is alive and well, just very busy to service their
893 struct list_head sr_timed_list;
894 /** server-side per-export list */
895 struct list_head sr_exp_list;
896 /** server-side history, used for debuging purposes. */
897 struct list_head sr_hist_list;
898 /** history sequence # */
900 /** the index of service's srv_at_array into which request is linked */
904 /** authed uid mapped to */
905 uid_t sr_auth_mapped_uid;
906 /** RPC is generated from what part of Lustre */
907 enum lustre_sec_part sr_sp_from;
908 /** request session context */
909 struct lu_context sr_ses;
913 /** stub for NRS request */
914 struct ptlrpc_nrs_request sr_nrq;
916 /** request arrival time */
917 struct timespec64 sr_arrival_time;
918 /** server's half ctx */
919 struct ptlrpc_svc_ctx *sr_svc_ctx;
920 /** (server side), pointed directly into req buffer */
921 struct ptlrpc_user_desc *sr_user_desc;
922 /** separated reply state, may be vmalloc'd */
923 struct ptlrpc_reply_state *sr_reply_state;
924 /** server-side hp handlers */
925 struct ptlrpc_hpreq_ops *sr_ops;
926 /** incoming request buffer */
927 struct ptlrpc_request_buffer_desc *sr_rqbd;
930 /** server request member alias */
931 /* NB: these alias should NOT be used by any new code, instead they should
932 * be removed step by step to avoid potential abuse */
933 #define rq_svc_thread rq_srv.sr_svc_thread
934 #define rq_timed_list rq_srv.sr_timed_list
935 #define rq_exp_list rq_srv.sr_exp_list
936 #define rq_history_list rq_srv.sr_hist_list
937 #define rq_history_seq rq_srv.sr_hist_seq
938 #define rq_at_index rq_srv.sr_at_index
939 #define rq_auth_uid rq_srv.sr_auth_uid
940 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
941 #define rq_sp_from rq_srv.sr_sp_from
942 #define rq_session rq_srv.sr_ses
943 #define rq_nrq rq_srv.sr_nrq
944 #define rq_arrival_time rq_srv.sr_arrival_time
945 #define rq_reply_state rq_srv.sr_reply_state
946 #define rq_svc_ctx rq_srv.sr_svc_ctx
947 #define rq_user_desc rq_srv.sr_user_desc
948 #define rq_ops rq_srv.sr_ops
949 #define rq_rqbd rq_srv.sr_rqbd
950 #define rq_reqmsg rq_pill.rc_reqmsg
951 #define rq_repmsg rq_pill.rc_repmsg
952 #define rq_req_swab_mask rq_pill.rc_req_swab_mask
953 #define rq_rep_swab_mask rq_pill.rc_rep_swab_mask
956 * Represents remote procedure call.
958 * This is a staple structure used by everybody wanting to send a request
961 struct ptlrpc_request {
962 /* Request type: one of PTL_RPC_MSG_* */
964 /** Result of request processing */
967 * Linkage item through which this request is included into
968 * sending/delayed lists on client and into rqbd list on server
970 struct list_head rq_list;
971 /** Lock to protect request flags and some other important bits, like
975 spinlock_t rq_early_free_lock;
976 /** client-side flags are serialized by rq_lock @{ */
977 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
978 rq_timedout:1, rq_resend:1, rq_restart:1,
980 * when ->rq_replay is set, request is kept by the client even
981 * after server commits corresponding transaction. This is
982 * used for operations that require sequence of multiple
983 * requests to be replayed. The only example currently is file
984 * open/close. When last request in such a sequence is
985 * committed, ->rq_replay is cleared on all requests in the
989 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
990 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
992 rq_req_unlinked:1, /* unlinked request buffer from lnet */
993 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
994 rq_memalloc:1, /* req originated from "kswapd" */
996 rq_reply_truncated:1,
997 /** whether the "rq_set" is a valid one */
1000 /** do not resend request on -EINPROGRESS */
1001 rq_no_retry_einprogress:1,
1002 /* allow the req to be sent if the import is in recovery
1005 /* bulk request, sent to server, but uncommitted */
1007 rq_early_free_repbuf:1, /* free reply buffer in advance */
1011 /** server-side flags @{ */
1013 rq_hp:1, /**< high priority RPC */
1014 rq_at_linked:1, /**< link into service's srv_at_array */
1015 rq_packed_final:1, /**< packed final reply */
1016 rq_obsolete:1; /* aborted by a signal on a client */
1019 /** one of RQ_PHASE_* */
1020 enum rq_phase rq_phase;
1021 /** one of RQ_PHASE_* to be used next */
1022 enum rq_phase rq_next_phase;
1024 * client-side refcount for SENT race, server-side refcounf
1025 * for multiple replies
1027 atomic_t rq_refcount;
1030 * !rq_truncate : # reply bytes actually received,
1031 * rq_truncate : required repbuf_len for resend
1033 int rq_nob_received;
1034 /** Request length */
1038 /** Pool if request is from preallocated list */
1039 struct ptlrpc_request_pool *rq_pool;
1040 /** Transaction number */
1044 /** bulk match bits */
1046 /** reply match bits */
1049 * List item to for replay list. Not yet committed requests get linked
1051 * Also see \a rq_replay comment above.
1052 * It's also link chain on obd_export::exp_req_replay_queue
1054 struct list_head rq_replay_list;
1055 /** non-shared members for client & server request*/
1057 struct ptlrpc_cli_req rq_cli;
1058 struct ptlrpc_srv_req rq_srv;
1061 * security and encryption data
1063 /** description of flavors for client & server */
1064 struct sptlrpc_flavor rq_flvr;
1067 * SELinux policy info at the time of the request
1068 * sepol string format is:
1069 * <mode>:<policy name>:<policy version>:<policy hash>
1071 char rq_sepol[LUSTRE_NODEMAP_SEPOL_LENGTH + 1];
1073 /* client/server security flags */
1075 rq_ctx_init:1, /* context initiation */
1076 rq_ctx_fini:1, /* context destroy */
1077 rq_bulk_read:1, /* request bulk read */
1078 rq_bulk_write:1, /* request bulk write */
1079 /* server authentication flags */
1080 rq_auth_gss:1, /* authenticated by gss */
1081 rq_auth_usr_root:1, /* authed as root */
1082 rq_auth_usr_mdt:1, /* authed as mdt */
1083 rq_auth_usr_ost:1, /* authed as ost */
1084 /* security tfm flags */
1087 /* doesn't expect reply FIXME */
1089 rq_pill_init:1, /* pill initialized */
1090 rq_srv_req:1; /* server request */
1093 /** various buffer pointers */
1094 struct lustre_msg *rq_reqbuf; /**< req wrapper, vmalloc*/
1095 char *rq_repbuf; /**< rep buffer, vmalloc */
1096 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
1097 /** only in priv mode */
1098 struct lustre_msg *rq_clrbuf;
1099 int rq_reqbuf_len; /* req wrapper buf len */
1100 int rq_reqdata_len; /* req wrapper msg len */
1101 int rq_repbuf_len; /* rep buffer len */
1102 int rq_repdata_len; /* rep wrapper msg len */
1103 int rq_clrbuf_len; /* only in priv mode */
1104 int rq_clrdata_len; /* only in priv mode */
1106 /** early replies go to offset 0, regular replies go after that */
1107 unsigned int rq_reply_off;
1110 /** how many early replies (for stats) */
1112 /** Server-side, export on which request was received */
1113 struct obd_export *rq_export;
1114 /** import where request is being sent */
1115 struct obd_import *rq_import;
1117 struct lnet_nid rq_self;
1118 /** Peer description (the other side) */
1119 struct lnet_processid rq_peer;
1120 /** Descriptor for the NID from which the peer sent the request. */
1121 struct lnet_processid rq_source;
1123 * service time estimate (secs)
1124 * If the request is not served by this time, it is marked as timed out.
1125 * Do not change to time64_t since this is transmitted over the wire.
1127 * The linux kernel handles timestamps with time64_t and timeouts
1128 * are normally done with jiffies. Lustre shares the rq_timeout between
1129 * nodes. Since jiffies can vary from node to node Lustre instead
1130 * will express the timeout value in seconds. To avoid confusion with
1131 * timestamps (time64_t) and jiffy timeouts (long) Lustre timeouts
1132 * are expressed in s32 (timeout_t). Also what is transmitted over
1133 * the wire is 32 bits.
1135 timeout_t rq_timeout;
1137 * when request/reply sent (secs), or time when request should be sent
1140 /** when request must finish. */
1141 time64_t rq_deadline;
1142 /** request format description */
1143 struct req_capsule rq_pill;
1147 * Call completion handler for rpc if any, return it's status or original
1148 * rc if there was no handler defined for this request.
1150 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1151 struct ptlrpc_request *req, int rc)
1153 if (req->rq_interpret_reply != NULL) {
1154 req->rq_status = req->rq_interpret_reply(env, req,
1155 &req->rq_async_args,
1157 return req->rq_status;
1166 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1169 * Can the request be moved from the regular NRS head to the high-priority NRS
1170 * head (of the same PTLRPC service partition), if any?
1172 * For a reliable result, this should be checked under svcpt->scp_req lock.
1174 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1176 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1179 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1180 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1181 * to make sure it has not been scheduled yet (analogous to previous
1182 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1184 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1188 static inline bool req_capsule_ptlreq(struct req_capsule *pill)
1190 struct ptlrpc_request *req = pill->rc_req;
1192 return req != NULL && pill == &req->rq_pill;
1195 static inline bool req_capsule_subreq(struct req_capsule *pill)
1197 struct ptlrpc_request *req = pill->rc_req;
1199 return req == NULL || pill != &req->rq_pill;
1203 * Returns true if request needs to be swabbed into local cpu byteorder
1205 static inline bool req_capsule_req_need_swab(struct req_capsule *pill)
1207 struct ptlrpc_request *req = pill->rc_req;
1209 return req && req_capsule_req_swabbed(&req->rq_pill,
1210 MSG_PTLRPC_HEADER_OFF);
1214 * Returns true if request reply needs to be swabbed into local cpu byteorder
1216 static inline bool req_capsule_rep_need_swab(struct req_capsule *pill)
1218 struct ptlrpc_request *req = pill->rc_req;
1220 return req && req_capsule_rep_swabbed(&req->rq_pill,
1221 MSG_PTLRPC_HEADER_OFF);
1225 * Convert numerical request phase value \a phase into text string description
1227 static inline const char *
1228 ptlrpc_phase2str(enum rq_phase phase)
1237 case RQ_PHASE_INTERPRET:
1239 case RQ_PHASE_COMPLETE:
1241 case RQ_PHASE_UNREG_RPC:
1243 case RQ_PHASE_UNREG_BULK:
1251 * Convert numerical request phase of the request \a req into text stringi
1254 static inline const char *
1255 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1257 return ptlrpc_phase2str(req->rq_phase);
1261 * Debugging functions and helpers to print request structure into debug log
1264 /* Spare the preprocessor, spoil the bugs. */
1265 #define FLAG(field, str) (field ? str : "")
1267 /** Convert bit flags into a string */
1268 #define DEBUG_REQ_FLAGS(req) \
1269 ptlrpc_rqphase2str(req), \
1270 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1271 FLAG(req->rq_err, "E"), FLAG(req->rq_net_err, "e"), \
1272 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1273 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1274 FLAG(req->rq_no_resend, "N"), FLAG(req->rq_no_reply, "n"), \
1275 FLAG(req->rq_waiting, "W"), \
1276 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1277 FLAG(req->rq_committed, "M"), \
1278 FLAG(req->rq_req_unlinked, "Q"), \
1279 FLAG(req->rq_reply_unlinked, "U"), \
1280 FLAG(req->rq_receiving_reply, "r")
1282 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s"
1284 void _debug_req(struct ptlrpc_request *req,
1285 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1286 __attribute__ ((format (printf, 3, 4)));
1289 * Helper that decides if we need to print request accordig to current debug
1292 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1294 CFS_CHECK_STACK(msgdata, mask, cdls); \
1296 if (((mask) & D_CANTMASK) != 0 || \
1297 ((libcfs_debug & (mask)) != 0 && \
1298 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1299 _debug_req((req), msgdata, fmt, ##a); \
1303 * This is the debug print function you need to use to print request sturucture
1304 * content into lustre debug log.
1305 * for most callers (level is a constant) this is resolved at compile time */
1306 #define DEBUG_REQ(level, req, fmt, args...) \
1308 if ((level) & (D_ERROR | D_WARNING)) { \
1309 static struct cfs_debug_limit_state cdls; \
1310 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1311 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1313 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1314 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1319 enum ptlrpc_bulk_op_type {
1320 PTLRPC_BULK_OP_ACTIVE = 0x00000001,
1321 PTLRPC_BULK_OP_PASSIVE = 0x00000002,
1322 PTLRPC_BULK_OP_PUT = 0x00000004,
1323 PTLRPC_BULK_OP_GET = 0x00000008,
1324 PTLRPC_BULK_GET_SOURCE = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_GET,
1325 PTLRPC_BULK_PUT_SINK = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_PUT,
1326 PTLRPC_BULK_GET_SINK = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_GET,
1327 PTLRPC_BULK_PUT_SOURCE = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_PUT,
1330 static inline bool ptlrpc_is_bulk_op_get(enum ptlrpc_bulk_op_type type)
1332 return (type & PTLRPC_BULK_OP_GET) == PTLRPC_BULK_OP_GET;
1335 static inline bool ptlrpc_is_bulk_get_source(enum ptlrpc_bulk_op_type type)
1337 return (type & PTLRPC_BULK_GET_SOURCE) == PTLRPC_BULK_GET_SOURCE;
1340 static inline bool ptlrpc_is_bulk_put_sink(enum ptlrpc_bulk_op_type type)
1342 return (type & PTLRPC_BULK_PUT_SINK) == PTLRPC_BULK_PUT_SINK;
1345 static inline bool ptlrpc_is_bulk_get_sink(enum ptlrpc_bulk_op_type type)
1347 return (type & PTLRPC_BULK_GET_SINK) == PTLRPC_BULK_GET_SINK;
1350 static inline bool ptlrpc_is_bulk_put_source(enum ptlrpc_bulk_op_type type)
1352 return (type & PTLRPC_BULK_PUT_SOURCE) == PTLRPC_BULK_PUT_SOURCE;
1355 static inline bool ptlrpc_is_bulk_op_active(enum ptlrpc_bulk_op_type type)
1357 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1358 (type & PTLRPC_BULK_OP_PASSIVE))
1359 == PTLRPC_BULK_OP_ACTIVE;
1362 static inline bool ptlrpc_is_bulk_op_passive(enum ptlrpc_bulk_op_type type)
1364 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1365 (type & PTLRPC_BULK_OP_PASSIVE))
1366 == PTLRPC_BULK_OP_PASSIVE;
1369 struct ptlrpc_bulk_frag_ops {
1371 * Add a page \a page to the bulk descriptor \a desc
1372 * Data to transfer in the page starts at offset \a pageoffset and
1373 * amount of data to transfer from the page is \a len
1375 void (*add_kiov_frag)(struct ptlrpc_bulk_desc *desc,
1376 struct page *page, int pageoffset, int len);
1379 * Add a \a fragment to the bulk descriptor \a desc.
1380 * Data to transfer in the fragment is pointed to by \a frag
1381 * The size of the fragment is \a len
1383 int (*add_iov_frag)(struct ptlrpc_bulk_desc *desc, void *frag, int len);
1386 * Uninitialize and free bulk descriptor \a desc.
1387 * Works on bulk descriptors both from server and client side.
1389 void (*release_frags)(struct ptlrpc_bulk_desc *desc);
1392 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops;
1393 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops;
1396 * Definition of bulk descriptor.
1397 * Bulks are special "Two phase" RPCs where initial request message
1398 * is sent first and it is followed bt a transfer (o receiving) of a large
1399 * amount of data to be settled into pages referenced from the bulk descriptors.
1400 * Bulks transfers (the actual data following the small requests) are done
1401 * on separate LNet portals.
1402 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1403 * Another user is readpage for MDT.
1405 struct ptlrpc_bulk_desc {
1406 unsigned int bd_refs; /* number MD's assigned including zero-sends */
1407 /** completed with failure */
1408 unsigned long bd_failure:1;
1410 unsigned long bd_registered:1,
1411 /* bulk request is RDMA transfer, use page->host as real address */
1413 /** For serialization with callback */
1415 /** {put,get}{source,sink}{kvec,kiov} */
1416 enum ptlrpc_bulk_op_type bd_type;
1417 /** LNet portal for this bulk */
1419 /** Server side - export this bulk created for */
1420 struct obd_export *bd_export;
1421 /** Client side - import this bulk was sent on */
1422 struct obd_import *bd_import;
1423 /** Back pointer to the request */
1424 struct ptlrpc_request *bd_req;
1425 const struct ptlrpc_bulk_frag_ops *bd_frag_ops;
1426 wait_queue_head_t bd_waitq; /* server side only WQ */
1427 int bd_iov_count; /* # entries in bd_iov */
1428 int bd_max_iov; /* allocated size of bd_iov */
1429 int bd_nob; /* # bytes covered */
1430 int bd_nob_transferred; /* # bytes GOT/PUT */
1431 unsigned int bd_nob_last; /* # bytes in last MD */
1433 __u64 bd_last_mbits;
1435 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1436 struct lnet_nid bd_sender; /* stash event::sender */
1437 int bd_md_count; /* # valid entries in bd_mds */
1438 int bd_md_max_brw; /* max entries in bd_mds */
1440 /** array of offsets for each MD */
1441 unsigned int bd_mds_off[PTLRPC_BULK_OPS_COUNT];
1442 /** array of associated MDs */
1443 struct lnet_handle_md bd_mds[PTLRPC_BULK_OPS_COUNT];
1445 /* encrypted iov, size is either 0 or bd_iov_count. */
1446 struct bio_vec *bd_enc_vec;
1447 struct bio_vec *bd_vec;
1452 SVC_STOPPED = BIT(0),
1453 SVC_STOPPING = BIT(1),
1454 SVC_STARTING = BIT(2),
1455 SVC_RUNNING = BIT(3),
1458 #define PTLRPC_THR_NAME_LEN 32
1460 * Definition of server service thread structure
1462 struct ptlrpc_thread {
1464 * List of active threads in svcpt->scp_threads
1466 struct list_head t_link;
1468 * thread-private data (preallocated vmalloc'd memory)
1473 * service thread index, from ptlrpc_start_threads
1479 struct task_struct *t_task;
1483 * put watchdog in the structure per thread b=14840
1485 struct delayed_work t_watchdog;
1487 * the svc this thread belonged to b=18582
1489 struct ptlrpc_service_part *t_svcpt;
1490 wait_queue_head_t t_ctl_waitq;
1491 struct lu_env *t_env;
1492 char t_name[PTLRPC_THR_NAME_LEN];
1495 static inline int thread_is_init(struct ptlrpc_thread *thread)
1497 return thread->t_flags == 0;
1500 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1502 return !!(thread->t_flags & SVC_STOPPED);
1505 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1507 return !!(thread->t_flags & SVC_STOPPING);
1510 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1512 return !!(thread->t_flags & SVC_STARTING);
1515 static inline int thread_is_running(struct ptlrpc_thread *thread)
1517 return !!(thread->t_flags & SVC_RUNNING);
1520 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1522 thread->t_flags &= ~flags;
1525 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1527 thread->t_flags = flags;
1530 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1532 thread->t_flags |= flags;
1535 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1538 if (thread->t_flags & flags) {
1539 thread->t_flags &= ~flags;
1546 * Request buffer descriptor structure.
1547 * This is a structure that contains one posted request buffer for service.
1548 * Once data land into a buffer, event callback creates actual request and
1549 * notifies wakes one of the service threads to process new incoming request.
1550 * More than one request can fit into the buffer.
1552 struct ptlrpc_request_buffer_desc {
1553 /** Link item for rqbds on a service */
1554 struct list_head rqbd_list;
1555 /** History of requests for this buffer */
1556 struct list_head rqbd_reqs;
1557 /** Back pointer to service for which this buffer is registered */
1558 struct ptlrpc_service_part *rqbd_svcpt;
1559 /** LNet descriptor */
1560 struct lnet_handle_md rqbd_md_h;
1562 /** The buffer itself */
1564 struct ptlrpc_cb_id rqbd_cbid;
1566 * This "embedded" request structure is only used for the
1567 * last request to fit into the buffer
1569 struct ptlrpc_request rqbd_req;
1572 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1574 struct ptlrpc_service_ops {
1576 * if non-NULL called during thread creation (ptlrpc_start_thread())
1577 * to initialize service specific per-thread state.
1579 int (*so_thr_init)(struct ptlrpc_thread *thr);
1581 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1582 * destruct state created by ->srv_init().
1584 void (*so_thr_done)(struct ptlrpc_thread *thr);
1586 * Handler function for incoming requests for this service
1588 int (*so_req_handler)(struct ptlrpc_request *req);
1590 * function to determine priority of the request, it's called
1591 * on every new request
1593 int (*so_hpreq_handler)(struct ptlrpc_request *);
1595 * service-specific print fn
1597 void (*so_req_printer)(void *, struct ptlrpc_request *);
1600 #ifndef __cfs_cacheline_aligned
1601 /* NB: put it here for reducing patche dependence */
1602 # define __cfs_cacheline_aligned
1606 * How many high priority requests to serve before serving one normal
1609 #define PTLRPC_SVC_HP_RATIO 10
1612 * Definition of PortalRPC service.
1613 * The service is listening on a particular portal (like tcp port)
1614 * and perform actions for a specific server like IO service for OST
1615 * or general metadata service for MDS.
1617 struct ptlrpc_service {
1618 /** serialize /proc operations */
1619 spinlock_t srv_lock;
1620 /** most often accessed fields */
1621 /** chain thru all services */
1622 struct list_head srv_list;
1623 /** service operations table */
1624 struct ptlrpc_service_ops srv_ops;
1625 /** only statically allocated strings here; we don't clean them */
1627 /** only statically allocated strings here; we don't clean them */
1628 char *srv_thread_name;
1629 /** threads # should be created for each partition on initializing */
1630 int srv_nthrs_cpt_init;
1631 /** limit of threads number for each partition */
1632 int srv_nthrs_cpt_limit;
1633 /** Root of debugfs dir tree for this service */
1634 struct dentry *srv_debugfs_entry;
1635 /** Pointer to statistic data for this service */
1636 struct lprocfs_stats *srv_stats;
1637 /** # hp per lp reqs to handle */
1638 int srv_hpreq_ratio;
1639 /** biggest request to receive */
1640 int srv_max_req_size;
1641 /** biggest reply to send */
1642 int srv_max_reply_size;
1643 /** size of individual buffers */
1645 /** # buffers to allocate in 1 group */
1646 int srv_nbuf_per_group;
1647 /** Local portal on which to receive requests */
1648 __u32 srv_req_portal;
1649 /** Portal on the client to send replies to */
1650 __u32 srv_rep_portal;
1652 * Tags for lu_context associated with this thread, see struct
1656 /** soft watchdog timeout multiplier */
1657 int srv_watchdog_factor;
1658 /** under unregister_service */
1659 unsigned srv_is_stopping:1;
1660 /** Whether or not to restrict service threads to CPUs in this CPT */
1661 unsigned srv_cpt_bind:1;
1663 /** max # request buffers */
1665 /** max # request buffers in history per partition */
1666 int srv_hist_nrqbds_cpt_max;
1667 /** number of CPTs this service associated with */
1669 /** CPTs array this service associated with */
1671 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1673 /** CPT table this service is running over */
1674 struct cfs_cpt_table *srv_cptable;
1677 struct kobject srv_kobj;
1678 struct completion srv_kobj_unregister;
1680 * partition data for ptlrpc service
1682 struct ptlrpc_service_part *srv_parts[0];
1686 * Definition of PortalRPC service partition data.
1687 * Although a service only has one instance of it right now, but we
1688 * will have multiple instances very soon (instance per CPT).
1690 * it has four locks:
1692 * serialize operations on rqbd and requests waiting for preprocess
1694 * serialize operations active requests sent to this portal
1696 * serialize adaptive timeout stuff
1698 * serialize operations on RS list (reply states)
1700 * We don't have any use-case to take two or more locks at the same time
1701 * for now, so there is no lock order issue.
1703 struct ptlrpc_service_part {
1704 /** back reference to owner */
1705 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1706 /* CPT id, reserved */
1708 /** always increasing number */
1710 /** # of starting threads */
1711 int scp_nthrs_starting;
1712 /** # running threads */
1713 int scp_nthrs_running;
1714 /** service threads list */
1715 struct list_head scp_threads;
1718 * serialize the following fields, used for protecting
1719 * rqbd list and incoming requests waiting for preprocess,
1720 * threads starting & stopping are also protected by this lock.
1722 spinlock_t scp_lock __cfs_cacheline_aligned;
1723 /** userland serialization */
1724 struct mutex scp_mutex;
1725 /** total # req buffer descs allocated */
1726 int scp_nrqbds_total;
1727 /** # posted request buffers for receiving */
1728 int scp_nrqbds_posted;
1729 /** in progress of allocating rqbd */
1730 int scp_rqbd_allocating;
1731 /** # incoming reqs */
1732 int scp_nreqs_incoming;
1733 /** request buffers to be reposted */
1734 struct list_head scp_rqbd_idle;
1735 /** req buffers receiving */
1736 struct list_head scp_rqbd_posted;
1737 /** incoming reqs */
1738 struct list_head scp_req_incoming;
1739 /** timeout before re-posting reqs, in jiffies */
1740 long scp_rqbd_timeout;
1742 * all threads sleep on this. This wait-queue is signalled when new
1743 * incoming request arrives and when difficult reply has to be handled.
1745 wait_queue_head_t scp_waitq;
1747 /** request history */
1748 struct list_head scp_hist_reqs;
1749 /** request buffer history */
1750 struct list_head scp_hist_rqbds;
1751 /** # request buffers in history */
1752 int scp_hist_nrqbds;
1753 /** sequence number for request */
1755 /** highest seq culled from history */
1756 __u64 scp_hist_seq_culled;
1759 * serialize the following fields, used for processing requests
1760 * sent to this portal
1762 spinlock_t scp_req_lock __cfs_cacheline_aligned;
1763 /** # reqs in either of the NRS heads below */
1764 /** # reqs being served */
1765 int scp_nreqs_active;
1766 /** # HPreqs being served */
1767 int scp_nhreqs_active;
1768 /** # hp requests handled */
1771 /** NRS head for regular requests */
1772 struct ptlrpc_nrs scp_nrs_reg;
1773 /** NRS head for HP requests; this is only valid for services that can
1774 * handle HP requests */
1775 struct ptlrpc_nrs *scp_nrs_hp;
1780 * serialize the following fields, used for changes on
1783 spinlock_t scp_at_lock __cfs_cacheline_aligned;
1784 /** estimated rpc service time */
1785 struct adaptive_timeout scp_at_estimate;
1786 /** reqs waiting for replies */
1787 struct ptlrpc_at_array scp_at_array;
1788 /** early reply timer */
1789 struct timer_list scp_at_timer;
1791 ktime_t scp_at_checktime;
1792 /** check early replies */
1793 unsigned scp_at_check;
1797 * serialize the following fields, used for processing
1798 * replies for this portal
1800 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1801 /** all the active replies */
1802 struct list_head scp_rep_active;
1803 /** List of free reply_states */
1804 struct list_head scp_rep_idle;
1805 /** waitq to run, when adding stuff to srv_free_rs_list */
1806 wait_queue_head_t scp_rep_waitq;
1807 /** # 'difficult' replies */
1808 atomic_t scp_nreps_difficult;
1811 #define ptlrpc_service_for_each_part(part, i, svc) \
1813 i < (svc)->srv_ncpts && \
1814 (svc)->srv_parts != NULL && \
1815 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1818 * Declaration of ptlrpcd control structure
1820 struct ptlrpcd_ctl {
1822 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1824 unsigned long pc_flags;
1826 * Thread lock protecting structure fields.
1832 struct completion pc_starting;
1836 struct completion pc_finishing;
1838 * Thread requests set.
1840 struct ptlrpc_request_set *pc_set;
1842 * Thread name used in kthread_run()
1846 * CPT the thread is bound on.
1850 * Index of ptlrpcd thread in the array.
1854 * Pointer to the array of partners' ptlrpcd_ctl structure.
1856 struct ptlrpcd_ctl **pc_partners;
1858 * Number of the ptlrpcd's partners.
1862 * Record the partner index to be processed next.
1866 * Error code if the thread failed to fully start.
1871 /* Bits for pc_flags */
1872 enum ptlrpcd_ctl_flags {
1874 * Ptlrpc thread start flag.
1876 LIOD_START = BIT(0),
1878 * Ptlrpc thread stop flag.
1882 * Ptlrpc thread force flag (only stop force so far).
1883 * This will cause aborting any inflight rpcs handled
1884 * by thread if LIOD_STOP is specified.
1886 LIOD_FORCE = BIT(2),
1888 * This is a recovery ptlrpc thread.
1890 LIOD_RECOVERY = BIT(3),
1897 * Service compatibility function; the policy is compatible with all services.
1899 * \param[in] svc The service the policy is attempting to register with.
1900 * \param[in] desc The policy descriptor
1902 * \retval true The policy is compatible with the service
1904 * \see ptlrpc_nrs_pol_desc::pd_compat()
1906 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
1907 const struct ptlrpc_nrs_pol_desc *desc)
1913 * Service compatibility function; the policy is compatible with only a specific
1914 * service which is identified by its human-readable name at
1915 * ptlrpc_service::srv_name.
1917 * \param[in] svc The service the policy is attempting to register with.
1918 * \param[in] desc The policy descriptor
1920 * \retval false The policy is not compatible with the service
1921 * \retval true The policy is compatible with the service
1923 * \see ptlrpc_nrs_pol_desc::pd_compat()
1925 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
1926 const struct ptlrpc_nrs_pol_desc *desc)
1928 LASSERT(desc->pd_compat_svc_name != NULL);
1929 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
1934 /* ptlrpc/events.c */
1935 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1936 struct lnet_processid *peer,
1937 struct lnet_nid *self,
1940 * These callbacks are invoked by LNet when something happened to
1944 extern void request_out_callback(struct lnet_event *ev);
1945 extern void reply_in_callback(struct lnet_event *ev);
1946 extern void client_bulk_callback(struct lnet_event *ev);
1947 extern void request_in_callback(struct lnet_event *ev);
1948 extern void reply_out_callback(struct lnet_event *ev);
1949 #ifdef HAVE_SERVER_SUPPORT
1950 extern void server_bulk_callback(struct lnet_event *ev);
1954 /* ptlrpc/connection.c */
1955 struct ptlrpc_connection *ptlrpc_connection_get(struct lnet_processid *peer,
1956 struct lnet_nid *self,
1957 struct obd_uuid *uuid);
1959 static inline void ptlrpc_connection_put(struct ptlrpc_connection *conn)
1964 LASSERT(atomic_read(&conn->c_refcount) > 0);
1967 * We do not remove connection from hashtable and
1968 * do not free it even if last caller released ref,
1969 * as we want to have it cached for the case it is
1972 * Deallocating it and later creating new connection
1973 * again would be wastful. This way we also avoid
1974 * expensive locking to protect things from get/put
1975 * race when found cached connection is freed by
1976 * ptlrpc_connection_put().
1978 * It will be freed later in module unload time,
1979 * when ptlrpc_connection_fini()->lh_exit->conn_exit()
1982 atomic_dec(&conn->c_refcount);
1984 CDEBUG(D_INFO, "PUT conn=%p refcount %d to %s\n",
1985 conn, atomic_read(&conn->c_refcount),
1986 libcfs_nidstr(&conn->c_peer.nid));
1989 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1990 int ptlrpc_connection_init(void);
1991 void ptlrpc_connection_fini(void);
1992 extern lnet_pid_t ptl_get_pid(void);
1995 * Check if the peer connection is on the local node. We need to use GFP_NOFS
1996 * for requests from a local client to avoid recursing into the filesystem
1997 * as we might end up waiting on a page sent in the request we're serving.
1999 * Use __GFP_HIGHMEM so that the pages can use all of the available memory
2000 * on 32-bit machines. Use more aggressive GFP_HIGHUSER flags from non-local
2001 * clients to be able to generate more memory pressure on the OSS and allow
2002 * inactive pages to be reclaimed, since it doesn't have any other processes
2003 * or allocations that generate memory reclaim pressure.
2005 * See b=17576 (bdf50dc9) and b=19529 (3dcf18d3) for details.
2007 static inline bool ptlrpc_connection_is_local(struct ptlrpc_connection *conn)
2012 if (nid_same(&conn->c_peer.nid, &conn->c_self))
2015 RETURN(LNetIsPeerLocal(&conn->c_peer.nid));
2018 /* ptlrpc/niobuf.c */
2020 * Actual interfacing with LNet to put/get/register/unregister stuff
2023 #ifdef HAVE_SERVER_SUPPORT
2024 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2025 unsigned nfrags, unsigned max_brw,
2028 const struct ptlrpc_bulk_frag_ops
2030 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2031 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2033 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2037 LASSERT(desc != NULL);
2039 spin_lock(&desc->bd_lock);
2041 spin_unlock(&desc->bd_lock);
2046 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2047 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2049 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2051 struct ptlrpc_bulk_desc *desc;
2054 LASSERT(req != NULL);
2055 desc = req->rq_bulk;
2060 if (req->rq_bulk_deadline > ktime_get_real_seconds())
2064 spin_lock(&desc->bd_lock);
2066 spin_unlock(&desc->bd_lock);
2070 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2071 #define PTLRPC_REPLY_EARLY 0x02
2072 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2073 int ptlrpc_reply(struct ptlrpc_request *req);
2074 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2075 int ptlrpc_error(struct ptlrpc_request *req);
2076 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2077 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2078 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2081 /* ptlrpc/client.c */
2083 * Client-side portals API. Everything to send requests, receive replies,
2084 * request queues, request management, etc.
2087 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2089 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
2090 struct ptlrpc_client *);
2091 void ptlrpc_cleanup_client(struct obd_import *imp);
2092 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
2095 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2096 int ptlrpc_replay_req(struct ptlrpc_request *req);
2097 void ptlrpc_restart_req(struct ptlrpc_request *req);
2098 void ptlrpc_abort_inflight(struct obd_import *imp);
2099 void ptlrpc_cleanup_imp(struct obd_import *imp);
2100 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2102 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2103 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2105 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2106 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *);
2107 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2108 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2109 #define PTLRPCD_SET ((struct ptlrpc_request_set *)1)
2111 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2112 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2114 struct ptlrpc_request_pool *
2115 ptlrpc_init_rq_pool(int, int,
2116 int (*populate_pool)(struct ptlrpc_request_pool *, int));
2118 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2119 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2120 const struct req_format *format);
2121 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2122 struct ptlrpc_request_pool *,
2123 const struct req_format *format);
2124 void ptlrpc_request_free(struct ptlrpc_request *request);
2125 int ptlrpc_request_pack(struct ptlrpc_request *request,
2126 __u32 version, int opcode);
2127 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2128 const struct req_format *format,
2129 __u32 version, int opcode);
2130 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2131 __u32 version, int opcode, char **bufs,
2132 struct ptlrpc_cli_ctx *ctx);
2133 void ptlrpc_req_finished(struct ptlrpc_request *request);
2134 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2135 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2136 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2137 unsigned nfrags, unsigned max_brw,
2140 const struct ptlrpc_bulk_frag_ops
2143 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2144 struct page *page, int pageoffset, int len,
2147 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
2149 static inline void ptlrpc_release_bulk_noop(struct ptlrpc_bulk_desc *desc)
2153 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2154 struct obd_import *imp);
2155 __u64 ptlrpc_next_xid(void);
2156 __u64 ptlrpc_sample_next_xid(void);
2157 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2158 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req);
2159 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req);
2161 /* Set of routines to run a function in ptlrpcd context */
2162 void *ptlrpcd_alloc_work(struct obd_import *imp,
2163 int (*cb)(const struct lu_env *, void *), void *data);
2164 void ptlrpcd_destroy_work(void *handler);
2165 int ptlrpcd_queue_work(void *handler);
2168 struct ptlrpc_service_buf_conf {
2169 /* nbufs is buffers # to allocate when growing the pool */
2170 unsigned int bc_nbufs;
2171 /* buffer size to post */
2172 unsigned int bc_buf_size;
2173 /* portal to listed for requests on */
2174 unsigned int bc_req_portal;
2175 /* portal of where to send replies to */
2176 unsigned int bc_rep_portal;
2177 /* maximum request size to be accepted for this service */
2178 unsigned int bc_req_max_size;
2179 /* maximum reply size this service can ever send */
2180 unsigned int bc_rep_max_size;
2183 struct ptlrpc_service_thr_conf {
2184 /* threadname should be 8 characters or less - 6 will be added on */
2186 /* threads increasing factor for each CPU */
2187 unsigned int tc_thr_factor;
2188 /* service threads # to start on each partition while initializing */
2189 unsigned int tc_nthrs_init;
2191 * low water of threads # upper-limit on each partition while running,
2192 * service availability may be impacted if threads number is lower
2193 * than this value. It can be ZERO if the service doesn't require
2194 * CPU affinity or there is only one partition.
2196 unsigned int tc_nthrs_base;
2197 /* "soft" limit for total threads number */
2198 unsigned int tc_nthrs_max;
2199 /* user specified threads number, it will be validated due to
2200 * other members of this structure. */
2201 unsigned int tc_nthrs_user;
2202 /* bind service threads to only CPUs in their associated CPT */
2203 unsigned int tc_cpu_bind;
2204 /* Tags for lu_context associated with service thread */
2208 struct ptlrpc_service_cpt_conf {
2209 struct cfs_cpt_table *cc_cptable;
2210 /* string pattern to describe CPTs for a service */
2212 /* whether or not to have per-CPT service partitions */
2216 struct ptlrpc_service_conf {
2219 /* soft watchdog timeout multiplifier to print stuck service traces */
2220 unsigned int psc_watchdog_factor;
2221 /* buffer information */
2222 struct ptlrpc_service_buf_conf psc_buf;
2223 /* thread information */
2224 struct ptlrpc_service_thr_conf psc_thr;
2225 /* CPU partition information */
2226 struct ptlrpc_service_cpt_conf psc_cpt;
2227 /* function table */
2228 struct ptlrpc_service_ops psc_ops;
2231 /* ptlrpc/service.c */
2233 * Server-side services API. Register/unregister service, request state
2234 * management, service thread management
2238 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
2239 int mode, bool no_ack, bool convert_lock);
2240 void ptlrpc_commit_replies(struct obd_export *exp);
2241 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2242 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2243 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2244 struct ptlrpc_service *ptlrpc_register_service(
2245 struct ptlrpc_service_conf *conf,
2246 struct kset *parent,
2247 struct dentry *debugfs_entry);
2249 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2250 int ptlrpc_service_health_check(struct ptlrpc_service *);
2251 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2252 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2253 struct obd_export *export);
2254 void ptlrpc_update_export_timer(struct obd_export *exp,
2255 time64_t extra_delay);
2257 int ptlrpc_hr_init(void);
2258 void ptlrpc_hr_fini(void);
2260 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout);
2261 void ptlrpc_watchdog_disable(struct delayed_work *work);
2262 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout);
2266 /* ptlrpc/import.c */
2271 int ptlrpc_connect_import(struct obd_import *imp);
2272 int ptlrpc_connect_import_locked(struct obd_import *imp);
2273 int ptlrpc_init_import(struct obd_import *imp);
2274 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2275 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp);
2276 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2277 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2279 void ptlrpc_import_enter_resend(struct obd_import *imp);
2280 /* ptlrpc/pack_generic.c */
2281 int ptlrpc_reconnect_import(struct obd_import *imp);
2285 * ptlrpc msg buffer and swab interface
2289 #define PTLRPC_MAX_BUFCOUNT \
2290 (sizeof(((struct ptlrpc_request *)0)->rq_req_swab_mask) * 8)
2291 #define MD_MAX_BUFLEN (MDS_REG_MAXREQSIZE > OUT_MAXREQSIZE ? \
2292 MDS_REG_MAXREQSIZE : OUT_MAXREQSIZE)
2293 #define PTLRPC_MAX_BUFLEN (OST_IO_MAXREQSIZE > MD_MAX_BUFLEN ? \
2294 OST_IO_MAXREQSIZE : MD_MAX_BUFLEN)
2295 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2296 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2298 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2299 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2301 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2302 __u32 *lens, char **bufs);
2303 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2305 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2306 __u32 *lens, char **bufs, int flags);
2307 #define LPRFL_EARLY_REPLY 1
2308 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2309 char **bufs, int flags);
2310 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2311 unsigned int newlen, int move_data);
2312 int lustre_grow_msg(struct lustre_msg *msg, int segment, unsigned int newlen);
2313 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2314 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2315 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
2316 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2317 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
2318 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
2319 extern __u32 lustre_msg_early_size;
2320 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
2321 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
2322 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
2323 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
2324 __u32 lustre_msg_bufcount(struct lustre_msg *m);
2325 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
2326 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2327 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2328 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2329 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
2330 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
2331 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
2332 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2333 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
2334 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2335 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2336 enum lustre_msg_version lustre_msg_get_version(struct lustre_msg *msg);
2337 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
2338 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2339 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2340 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
2341 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2342 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2343 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2344 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2345 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2346 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2347 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2348 int lustre_msg_get_status(struct lustre_msg *msg);
2349 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2350 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2351 timeout_t lustre_msg_get_timeout(struct lustre_msg *msg);
2352 timeout_t lustre_msg_get_service_timeout(struct lustre_msg *msg);
2353 int lustre_msg_get_uid_gid(struct lustre_msg *msg, __u32 *uid, __u32 *gid);
2354 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2355 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2356 __u64 lustre_msg_get_mbits(struct lustre_msg *msg);
2357 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, __u32 buf);
2358 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2359 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2360 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2361 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2362 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
2363 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2364 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2365 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2366 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2367 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2368 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2369 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2370 void lustre_msg_set_timeout(struct lustre_msg *msg, timeout_t timeout);
2371 void lustre_msg_set_service_timeout(struct lustre_msg *msg,
2372 timeout_t service_timeout);
2373 void lustre_msg_set_uid_gid(struct lustre_msg *msg, __u32 *uid, __u32 *gid);
2374 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2375 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2376 void lustre_msg_set_mbits(struct lustre_msg *msg, __u64 mbits);
2379 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2380 unsigned int newlen, int move_data)
2382 LASSERT(req->rq_reply_state);
2383 LASSERT(req->rq_repmsg);
2384 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2388 #ifdef LUSTRE_TRANSLATE_ERRNOS
2390 static inline int ptlrpc_status_hton(int h)
2393 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2394 * ELDLM_LOCK_ABORTED, etc.
2397 return -lustre_errno_hton(-h);
2402 static inline int ptlrpc_status_ntoh(int n)
2405 * See the comment in ptlrpc_status_hton().
2408 return -lustre_errno_ntoh(-n);
2415 #define ptlrpc_status_hton(h) (h)
2416 #define ptlrpc_status_ntoh(n) (n)
2421 /** Change request phase of \a req to \a new_phase */
2423 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2425 if (req->rq_phase == new_phase)
2428 if (new_phase == RQ_PHASE_UNREG_RPC ||
2429 new_phase == RQ_PHASE_UNREG_BULK) {
2430 /* No embedded unregistering phases */
2431 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2432 req->rq_phase == RQ_PHASE_UNREG_BULK)
2435 req->rq_next_phase = req->rq_phase;
2437 atomic_inc(&req->rq_import->imp_unregistering);
2440 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2441 req->rq_phase == RQ_PHASE_UNREG_BULK) {
2443 atomic_dec(&req->rq_import->imp_unregistering);
2446 DEBUG_REQ(D_INFO, req, "move request phase from %s to %s",
2447 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2449 req->rq_phase = new_phase;
2453 * Returns true if request \a req got early reply and hard deadline is not met
2456 ptlrpc_client_early(struct ptlrpc_request *req)
2458 return req->rq_early;
2462 * Returns true if we got real reply from server for this request
2465 ptlrpc_client_replied(struct ptlrpc_request *req)
2467 if (req->rq_reply_deadline > ktime_get_real_seconds())
2469 return req->rq_replied;
2472 /** Returns true if request \a req is in process of receiving server reply */
2474 ptlrpc_client_recv(struct ptlrpc_request *req)
2476 if (req->rq_reply_deadline > ktime_get_real_seconds())
2478 return req->rq_receiving_reply;
2481 #define ptlrpc_cli_wait_unlink(req) __ptlrpc_cli_wait_unlink(req, NULL)
2484 __ptlrpc_cli_wait_unlink(struct ptlrpc_request *req, bool *discard)
2488 spin_lock(&req->rq_lock);
2489 if (req->rq_reply_deadline > ktime_get_real_seconds()) {
2490 spin_unlock(&req->rq_lock);
2493 if (req->rq_req_deadline > ktime_get_real_seconds()) {
2494 spin_unlock(&req->rq_lock);
2500 if (req->rq_reply_unlinked && req->rq_req_unlinked == 0) {
2502 spin_unlock(&req->rq_lock);
2503 return 1; /* Should call again after LNetMDUnlink */
2507 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
2508 req->rq_receiving_reply;
2509 spin_unlock(&req->rq_lock);
2514 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2517 if (req->rq_set == NULL)
2518 wake_up(&req->rq_reply_waitq);
2520 wake_up(&req->rq_set->set_waitq);
2524 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2526 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2527 atomic_inc(&rs->rs_refcount);
2531 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2533 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2534 if (atomic_dec_and_test(&rs->rs_refcount))
2535 lustre_free_reply_state(rs);
2538 /* Should only be called once per req */
2539 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2541 if (req->rq_reply_state == NULL)
2542 return; /* shouldn't occur */
2544 /* req_repmsg equals rq_reply_state->rs_msg,
2545 * so set it to NULL before rq_reply_state is possibly freed
2547 spin_lock(&req->rq_early_free_lock);
2548 req->rq_repmsg = NULL;
2549 spin_unlock(&req->rq_early_free_lock);
2551 ptlrpc_rs_decref(req->rq_reply_state);
2552 req->rq_reply_state = NULL;
2555 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2557 return lustre_msg_get_magic(req->rq_reqmsg);
2560 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2562 switch (req->rq_reqmsg->lm_magic) {
2563 case LUSTRE_MSG_MAGIC_V2:
2564 return req->rq_reqmsg->lm_repsize;
2566 LASSERTF(0, "incorrect message magic: %08x\n",
2567 req->rq_reqmsg->lm_magic);
2572 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2574 if (req->rq_delay_limit != 0 &&
2575 req->rq_queued_time + req->rq_delay_limit < ktime_get_seconds())
2580 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2582 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2583 spin_lock(&req->rq_lock);
2584 req->rq_no_resend = 1;
2585 spin_unlock(&req->rq_lock);
2587 return req->rq_no_resend;
2591 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2593 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2595 return svcpt->scp_service->srv_watchdog_factor *
2596 max_t(int, at, obd_timeout);
2599 static inline struct ptlrpc_service *
2600 ptlrpc_req2svc(struct ptlrpc_request *req)
2602 LASSERT(req->rq_rqbd != NULL);
2603 return req->rq_rqbd->rqbd_svcpt->scp_service;
2606 /* ldlm/ldlm_lib.c */
2608 * Target client logic
2611 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg);
2612 int client_obd_cleanup(struct obd_device *obd);
2613 int client_connect_import(const struct lu_env *env,
2614 struct obd_export **exp, struct obd_device *obd,
2615 struct obd_uuid *cluuid, struct obd_connect_data *,
2617 int client_disconnect_export(struct obd_export *exp);
2618 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2620 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2621 lnet_nid_t prim_nid, int priority);
2622 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
2623 int nid_count, struct obd_uuid *uuid);
2624 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2625 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2626 void client_destroy_import(struct obd_import *imp);
2629 #ifdef HAVE_SERVER_SUPPORT
2630 int server_disconnect_export(struct obd_export *exp);
2633 /* ptlrpc/pinger.c */
2635 * Pinger API (client side only)
2638 enum timeout_event {
2641 struct timeout_item;
2642 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2643 int ptlrpc_pinger_add_import(struct obd_import *imp);
2644 int ptlrpc_pinger_del_import(struct obd_import *imp);
2645 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2646 int ptlrpc_obd_ping(struct obd_device *obd);
2647 void ping_evictor_start(void);
2648 void ping_evictor_stop(void);
2649 void ptlrpc_pinger_ir_up(void);
2650 void ptlrpc_pinger_ir_down(void);
2652 int ptlrpc_pinger_suppress_pings(void);
2654 /* ptlrpc/ptlrpcd.c */
2655 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2656 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2657 void ptlrpcd_wake(struct ptlrpc_request *req);
2658 void ptlrpcd_add_req(struct ptlrpc_request *req);
2659 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2660 int ptlrpcd_addref(void);
2661 void ptlrpcd_decref(void);
2663 /* ptlrpc/lproc_ptlrpc.c */
2665 * procfs output related functions
2668 const char* ll_opcode2str(__u32 opcode);
2669 const int ll_str2opcode(const char *ops);
2670 #ifdef CONFIG_PROC_FS
2671 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2672 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2673 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2675 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2676 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2677 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2681 /* ptlrpc/llog_server.c */
2682 int llog_origin_handle_open(struct ptlrpc_request *req);
2683 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2684 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2685 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2687 /* ptlrpc/llog_client.c */
2688 extern const struct llog_operations llog_client_ops;