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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 /** \defgroup PtlRPC Portal RPC and networking module.
34 * PortalRPC is the layer used by rest of lustre code to achieve network
35 * communications: establish connections with corresponding export and import
36 * states, listen for a service, send and receive RPCs.
37 * PortalRPC also includes base recovery framework: packet resending and
38 * replaying, reconnections, pinger.
40 * PortalRPC utilizes LNet as its transport layer.
53 #include <linux/kobject.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 4
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 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
377 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
381 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
382 * However, we need to allocate a much larger buffer for it because LNet
383 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
384 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
385 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
386 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
387 * utilization is very low.
389 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
390 * reused until all requests fit in it have been processed and released,
391 * which means one long blocked request can prevent the rqbd be reused.
392 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
393 * from LNet with unused 65 KB, buffer utilization will be about 59%.
394 * Please check LU-2432 for details.
396 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
400 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
401 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
402 * extra bytes to each request buffer to improve buffer utilization rate.
404 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
407 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
408 #define FLD_MAXREQSIZE (160)
410 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
411 #define FLD_MAXREPSIZE (152)
412 #define FLD_BUFSIZE (1 << 12)
415 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
417 #define SEQ_MAXREQSIZE (160)
419 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
420 #define SEQ_MAXREPSIZE (152)
421 #define SEQ_BUFSIZE (1 << 12)
423 /** MGS threads must be >= 3, see bug 22458 comment #28 */
424 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
425 #define MGS_NTHRS_MAX 32
428 #define MGS_BUFSIZE (8 * 1024)
429 #define MGS_MAXREQSIZE (7 * 1024)
430 #define MGS_MAXREPSIZE (9 * 1024)
433 * OSS threads constants:
435 * Given 8 as factor and 64 as base threads number
438 * On 8-core server configured to 2 partitions, we will have
439 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
442 * On 32-core machine configured to 4 partitions, we will have
443 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
444 * will be 112 * 4 = 448.
447 * On 64-core machine configured to 4 partitions, we will have
448 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
449 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
450 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
451 * for each partition.
453 * So we can see that with these constants, threads number wil be at the
454 * similar level of old versions, unless the server has many cores.
456 /* depress threads factor for VM with small memory size */
457 #define OSS_THR_FACTOR min_t(int, 8, \
458 NUM_CACHEPAGES >> (28 - PAGE_SHIFT))
459 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
460 #define OSS_NTHRS_BASE 64
462 /* threads for handling "create" request */
463 #define OSS_CR_THR_FACTOR 1
464 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
465 #define OSS_CR_NTHRS_BASE 8
466 #define OSS_CR_NTHRS_MAX 64
469 * OST_IO_MAXREQSIZE ~=
470 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
471 * DT_MAX_BRW_PAGES * niobuf_remote
473 * - single object with 16 pages is 512 bytes
474 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
475 * - Must be a multiple of 1024
477 #define _OST_MAXREQSIZE_BASE ((unsigned long)(sizeof(struct lustre_msg) + \
478 sizeof(struct ptlrpc_body) + \
479 sizeof(struct obdo) + \
480 sizeof(struct obd_ioobj) + \
481 sizeof(struct niobuf_remote)))
482 #define _OST_MAXREQSIZE_SUM ((unsigned long)(_OST_MAXREQSIZE_BASE + \
483 sizeof(struct niobuf_remote) * \
484 (DT_MAX_BRW_PAGES - 1)))
486 * FIEMAP request can be 4K+ for now
488 #define OST_MAXREQSIZE (16UL * 1024UL)
489 #define OST_IO_MAXREQSIZE max(OST_MAXREQSIZE, \
490 ((_OST_MAXREQSIZE_SUM - 1) | \
492 /* Safe estimate of free space in standard RPC, provides upper limit for # of
493 * bytes of i/o to pack in RPC (skipping bulk transfer). */
494 #define OST_SHORT_IO_SPACE (OST_IO_MAXREQSIZE - _OST_MAXREQSIZE_BASE)
496 /* Actual size used for short i/o buffer. Calculation means this:
497 * At least one page (for large PAGE_SIZE), or 16 KiB, but not more
498 * than the available space aligned to a page boundary. */
499 #define OBD_MAX_SHORT_IO_BYTES min(max(PAGE_SIZE, 16UL * 1024UL), \
500 OST_SHORT_IO_SPACE & PAGE_MASK)
502 #define OST_MAXREPSIZE (9 * 1024)
503 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
506 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
507 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 32 * 1024)
509 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
510 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
512 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
515 /* Macro to hide a typecast. */
516 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
518 struct ptlrpc_replay_async_args {
524 * Structure to single define portal connection.
526 struct ptlrpc_connection {
527 /** linkage for connections hash table */
528 struct hlist_node c_hash;
529 /** Our own lnet nid for this connection */
531 /** Remote side nid for this connection */
532 struct lnet_process_id c_peer;
533 /** UUID of the other side */
534 struct obd_uuid c_remote_uuid;
535 /** reference counter for this connection */
539 /** Client definition for PortalRPC */
540 struct ptlrpc_client {
541 /** What lnet portal does this client send messages to by default */
542 __u32 cli_request_portal;
543 /** What portal do we expect replies on */
544 __u32 cli_reply_portal;
545 /** Name of the client */
549 /** state flags of requests */
550 /* XXX only ones left are those used by the bulk descs as well! */
551 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
552 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
554 #define REQ_MAX_ACK_LOCKS 8
556 union ptlrpc_async_args {
558 * Scratchpad for passing args to completion interpreter. Users
559 * cast to the struct of their choosing, and CLASSERT that this is
560 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
561 * a pointer to it here. The pointer_arg ensures this struct is at
562 * least big enough for that.
564 void *pointer_arg[11];
568 struct ptlrpc_request_set;
569 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
572 * Definition of request set structure.
573 * Request set is a list of requests (not necessary to the same target) that
574 * once populated with RPCs could be sent in parallel.
575 * There are two kinds of request sets. General purpose and with dedicated
576 * serving thread. Example of the latter is ptlrpcd set.
577 * For general purpose sets once request set started sending it is impossible
578 * to add new requests to such set.
579 * Provides a way to call "completion callbacks" when all requests in the set
582 struct ptlrpc_request_set {
583 atomic_t set_refcount;
584 /** number of in queue requests */
585 atomic_t set_new_count;
586 /** number of uncompleted requests */
587 atomic_t set_remaining;
588 /** wait queue to wait on for request events */
589 wait_queue_head_t set_waitq;
590 /** List of requests in the set */
591 struct list_head set_requests;
593 * Lock for \a set_new_requests manipulations
594 * locked so that any old caller can communicate requests to
595 * the set holder who can then fold them into the lock-free set
597 spinlock_t set_new_req_lock;
598 /** List of new yet unsent requests. Only used with ptlrpcd now. */
599 struct list_head set_new_requests;
601 /** rq_status of requests that have been freed already */
603 /** Additional fields used by the flow control extension */
604 /** Maximum number of RPCs in flight */
605 int set_max_inflight;
606 /** Callback function used to generate RPCs */
607 set_producer_func set_producer;
608 /** opaq argument passed to the producer callback */
609 void *set_producer_arg;
610 unsigned int set_allow_intr:1;
613 struct ptlrpc_bulk_desc;
614 struct ptlrpc_service_part;
615 struct ptlrpc_service;
618 * ptlrpc callback & work item stuff
620 struct ptlrpc_cb_id {
621 void (*cbid_fn)(struct lnet_event *ev); /* specific callback fn */
622 void *cbid_arg; /* additional arg */
625 /** Maximum number of locks to fit into reply state */
626 #define RS_MAX_LOCKS 8
630 * Structure to define reply state on the server
631 * Reply state holds various reply message information. Also for "difficult"
632 * replies (rep-ack case) we store the state after sending reply and wait
633 * for the client to acknowledge the reception. In these cases locks could be
634 * added to the state for replay/failover consistency guarantees.
636 struct ptlrpc_reply_state {
637 /** Callback description */
638 struct ptlrpc_cb_id rs_cb_id;
639 /** Linkage for list of all reply states in a system */
640 struct list_head rs_list;
641 /** Linkage for list of all reply states on same export */
642 struct list_head rs_exp_list;
643 /** Linkage for list of all reply states for same obd */
644 struct list_head rs_obd_list;
646 struct list_head rs_debug_list;
648 /** A spinlock to protect the reply state flags */
650 /** Reply state flags */
651 unsigned long rs_difficult:1; /* ACK/commit stuff */
652 unsigned long rs_no_ack:1; /* no ACK, even for
653 difficult requests */
654 unsigned long rs_scheduled:1; /* being handled? */
655 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
656 unsigned long rs_handled:1; /* been handled yet? */
657 unsigned long rs_on_net:1; /* reply_out_callback pending? */
658 unsigned long rs_prealloc:1; /* rs from prealloc list */
659 unsigned long rs_committed:1;/* the transaction was committed
660 and the rs was dispatched
661 by ptlrpc_commit_replies */
662 unsigned long rs_convert_lock:1; /* need to convert saved
663 * locks to COS mode */
664 atomic_t rs_refcount; /* number of users */
665 /** Number of locks awaiting client ACK */
668 /** Size of the state */
672 /** Transaction number */
676 struct obd_export *rs_export;
677 struct ptlrpc_service_part *rs_svcpt;
678 /** Lnet metadata handle for the reply */
679 struct lnet_handle_md rs_md_h;
681 /** Context for the sevice thread */
682 struct ptlrpc_svc_ctx *rs_svc_ctx;
683 /** Reply buffer (actually sent to the client), encoded if needed */
684 struct lustre_msg *rs_repbuf; /* wrapper */
685 /** Size of the reply buffer */
686 int rs_repbuf_len; /* wrapper buf length */
687 /** Size of the reply message */
688 int rs_repdata_len; /* wrapper msg length */
690 * Actual reply message. Its content is encrupted (if needed) to
691 * produce reply buffer for actual sending. In simple case
692 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
694 struct lustre_msg *rs_msg; /* reply message */
696 /** Handles of locks awaiting client reply ACK */
697 struct lustre_handle rs_locks[RS_MAX_LOCKS];
698 /** Lock modes of locks in \a rs_locks */
699 enum ldlm_mode rs_modes[RS_MAX_LOCKS];
702 struct ptlrpc_thread;
706 RQ_PHASE_NEW = 0xebc0de00,
707 RQ_PHASE_RPC = 0xebc0de01,
708 RQ_PHASE_BULK = 0xebc0de02,
709 RQ_PHASE_INTERPRET = 0xebc0de03,
710 RQ_PHASE_COMPLETE = 0xebc0de04,
711 RQ_PHASE_UNREG_RPC = 0xebc0de05,
712 RQ_PHASE_UNREG_BULK = 0xebc0de06,
713 RQ_PHASE_UNDEFINED = 0xebc0de07
716 /** Type of request interpreter call-back */
717 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
718 struct ptlrpc_request *req,
720 /** Type of request resend call-back */
721 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
725 * Definition of request pool structure.
726 * The pool is used to store empty preallocated requests for the case
727 * when we would actually need to send something without performing
728 * any allocations (to avoid e.g. OOM).
730 struct ptlrpc_request_pool {
731 /** Locks the list */
733 /** list of ptlrpc_request structs */
734 struct list_head prp_req_list;
735 /** Maximum message size that would fit into a rquest from this pool */
737 /** Function to allocate more requests for this pool */
738 int (*prp_populate)(struct ptlrpc_request_pool *, int);
746 #include <lustre_nrs.h>
749 * Basic request prioritization operations structure.
750 * The whole idea is centered around locks and RPCs that might affect locks.
751 * When a lock is contended we try to give priority to RPCs that might lead
752 * to fastest release of that lock.
753 * Currently only implemented for OSTs only in a way that makes all
754 * IO and truncate RPCs that are coming from a locked region where a lock is
755 * contended a priority over other requests.
757 struct ptlrpc_hpreq_ops {
759 * Check if the lock handle of the given lock is the same as
760 * taken from the request.
762 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
764 * Check if the request is a high priority one.
766 int (*hpreq_check)(struct ptlrpc_request *);
768 * Called after the request has been handled.
770 void (*hpreq_fini)(struct ptlrpc_request *);
773 struct ptlrpc_cli_req {
774 /** For bulk requests on client only: bulk descriptor */
775 struct ptlrpc_bulk_desc *cr_bulk;
776 /** optional time limit for send attempts */
777 time64_t cr_delay_limit;
778 /** time request was first queued */
779 time64_t cr_queued_time;
780 /** request sent in nanoseconds */
782 /** time for request really sent out */
783 time64_t cr_sent_out;
784 /** when req reply unlink must finish. */
785 time64_t cr_reply_deadline;
786 /** when req bulk unlink must finish. */
787 time64_t cr_bulk_deadline;
788 /** when req unlink must finish. */
789 time64_t cr_req_deadline;
790 /** Portal to which this request would be sent */
792 /** Portal where to wait for reply and where reply would be sent */
794 /** request resending number */
795 unsigned int cr_resend_nr;
796 /** What was import generation when this request was sent */
798 enum lustre_imp_state cr_send_state;
799 /** Per-request waitq introduced by bug 21938 for recovery waiting */
800 wait_queue_head_t cr_set_waitq;
801 /** Link item for request set lists */
802 struct list_head cr_set_chain;
803 /** link to waited ctx */
804 struct list_head cr_ctx_chain;
806 /** client's half ctx */
807 struct ptlrpc_cli_ctx *cr_cli_ctx;
808 /** Link back to the request set */
809 struct ptlrpc_request_set *cr_set;
810 /** outgoing request MD handle */
811 struct lnet_handle_md cr_req_md_h;
812 /** request-out callback parameter */
813 struct ptlrpc_cb_id cr_req_cbid;
814 /** incoming reply MD handle */
815 struct lnet_handle_md cr_reply_md_h;
816 wait_queue_head_t cr_reply_waitq;
817 /** reply callback parameter */
818 struct ptlrpc_cb_id cr_reply_cbid;
819 /** Async completion handler, called when reply is received */
820 ptlrpc_interpterer_t cr_reply_interp;
821 /** Resend handler, called when request is resend to update RPC data */
822 ptlrpc_resend_cb_t cr_resend_cb;
823 /** Async completion context */
824 union ptlrpc_async_args cr_async_args;
825 /** Opaq data for replay and commit callbacks. */
827 /** Link to the imp->imp_unreplied_list */
828 struct list_head cr_unreplied_list;
830 * Commit callback, called when request is committed and about to be
833 void (*cr_commit_cb)(struct ptlrpc_request *);
834 /** Replay callback, called after request is replayed at recovery */
835 void (*cr_replay_cb)(struct ptlrpc_request *);
838 /** client request member alias */
839 /* NB: these alias should NOT be used by any new code, instead they should
840 * be removed step by step to avoid potential abuse */
841 #define rq_bulk rq_cli.cr_bulk
842 #define rq_delay_limit rq_cli.cr_delay_limit
843 #define rq_queued_time rq_cli.cr_queued_time
844 #define rq_sent_ns rq_cli.cr_sent_ns
845 #define rq_real_sent rq_cli.cr_sent_out
846 #define rq_reply_deadline rq_cli.cr_reply_deadline
847 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
848 #define rq_req_deadline rq_cli.cr_req_deadline
849 #define rq_nr_resend rq_cli.cr_resend_nr
850 #define rq_request_portal rq_cli.cr_req_ptl
851 #define rq_reply_portal rq_cli.cr_rep_ptl
852 #define rq_import_generation rq_cli.cr_imp_gen
853 #define rq_send_state rq_cli.cr_send_state
854 #define rq_set_chain rq_cli.cr_set_chain
855 #define rq_ctx_chain rq_cli.cr_ctx_chain
856 #define rq_set rq_cli.cr_set
857 #define rq_set_waitq rq_cli.cr_set_waitq
858 #define rq_cli_ctx rq_cli.cr_cli_ctx
859 #define rq_req_md_h rq_cli.cr_req_md_h
860 #define rq_req_cbid rq_cli.cr_req_cbid
861 #define rq_reply_md_h rq_cli.cr_reply_md_h
862 #define rq_reply_waitq rq_cli.cr_reply_waitq
863 #define rq_reply_cbid rq_cli.cr_reply_cbid
864 #define rq_interpret_reply rq_cli.cr_reply_interp
865 #define rq_resend_cb rq_cli.cr_resend_cb
866 #define rq_async_args rq_cli.cr_async_args
867 #define rq_cb_data rq_cli.cr_cb_data
868 #define rq_unreplied_list rq_cli.cr_unreplied_list
869 #define rq_commit_cb rq_cli.cr_commit_cb
870 #define rq_replay_cb rq_cli.cr_replay_cb
872 struct ptlrpc_srv_req {
873 /** initial thread servicing this request */
874 struct ptlrpc_thread *sr_svc_thread;
876 * Server side list of incoming unserved requests sorted by arrival
877 * time. Traversed from time to time to notice about to expire
878 * requests and sent back "early replies" to clients to let them
879 * know server is alive and well, just very busy to service their
882 struct list_head sr_timed_list;
883 /** server-side per-export list */
884 struct list_head sr_exp_list;
885 /** server-side history, used for debuging purposes. */
886 struct list_head sr_hist_list;
887 /** history sequence # */
889 /** the index of service's srv_at_array into which request is linked */
893 /** authed uid mapped to */
894 uid_t sr_auth_mapped_uid;
895 /** RPC is generated from what part of Lustre */
896 enum lustre_sec_part sr_sp_from;
897 /** request session context */
898 struct lu_context sr_ses;
902 /** stub for NRS request */
903 struct ptlrpc_nrs_request sr_nrq;
905 /** request arrival time */
906 struct timespec64 sr_arrival_time;
907 /** server's half ctx */
908 struct ptlrpc_svc_ctx *sr_svc_ctx;
909 /** (server side), pointed directly into req buffer */
910 struct ptlrpc_user_desc *sr_user_desc;
911 /** separated reply state, may be vmalloc'd */
912 struct ptlrpc_reply_state *sr_reply_state;
913 /** server-side hp handlers */
914 struct ptlrpc_hpreq_ops *sr_ops;
915 /** incoming request buffer */
916 struct ptlrpc_request_buffer_desc *sr_rqbd;
919 /** server request member alias */
920 /* NB: these alias should NOT be used by any new code, instead they should
921 * be removed step by step to avoid potential abuse */
922 #define rq_svc_thread rq_srv.sr_svc_thread
923 #define rq_timed_list rq_srv.sr_timed_list
924 #define rq_exp_list rq_srv.sr_exp_list
925 #define rq_history_list rq_srv.sr_hist_list
926 #define rq_history_seq rq_srv.sr_hist_seq
927 #define rq_at_index rq_srv.sr_at_index
928 #define rq_auth_uid rq_srv.sr_auth_uid
929 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
930 #define rq_sp_from rq_srv.sr_sp_from
931 #define rq_session rq_srv.sr_ses
932 #define rq_nrq rq_srv.sr_nrq
933 #define rq_arrival_time rq_srv.sr_arrival_time
934 #define rq_reply_state rq_srv.sr_reply_state
935 #define rq_svc_ctx rq_srv.sr_svc_ctx
936 #define rq_user_desc rq_srv.sr_user_desc
937 #define rq_ops rq_srv.sr_ops
938 #define rq_rqbd rq_srv.sr_rqbd
941 * Represents remote procedure call.
943 * This is a staple structure used by everybody wanting to send a request
946 struct ptlrpc_request {
947 /* Request type: one of PTL_RPC_MSG_* */
949 /** Result of request processing */
952 * Linkage item through which this request is included into
953 * sending/delayed lists on client and into rqbd list on server
955 struct list_head rq_list;
956 /** Lock to protect request flags and some other important bits, like
960 spinlock_t rq_early_free_lock;
961 /** client-side flags are serialized by rq_lock @{ */
962 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
963 rq_timedout:1, rq_resend:1, rq_restart:1,
965 * when ->rq_replay is set, request is kept by the client even
966 * after server commits corresponding transaction. This is
967 * used for operations that require sequence of multiple
968 * requests to be replayed. The only example currently is file
969 * open/close. When last request in such a sequence is
970 * committed, ->rq_replay is cleared on all requests in the
974 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
975 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
977 rq_req_unlinked:1, /* unlinked request buffer from lnet */
978 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
979 rq_memalloc:1, /* req originated from "kswapd" */
981 rq_reply_truncated:1,
982 /** whether the "rq_set" is a valid one */
985 /** do not resend request on -EINPROGRESS */
986 rq_no_retry_einprogress:1,
987 /* allow the req to be sent if the import is in recovery
990 /* bulk request, sent to server, but uncommitted */
992 rq_early_free_repbuf:1, /* free reply buffer in advance */
996 /** server-side flags @{ */
998 rq_hp:1, /**< high priority RPC */
999 rq_at_linked:1, /**< link into service's srv_at_array */
1000 rq_packed_final:1; /**< packed final reply */
1003 /** one of RQ_PHASE_* */
1004 enum rq_phase rq_phase;
1005 /** one of RQ_PHASE_* to be used next */
1006 enum rq_phase rq_next_phase;
1008 * client-side refcount for SENT race, server-side refcounf
1009 * for multiple replies
1011 atomic_t rq_refcount;
1014 * !rq_truncate : # reply bytes actually received,
1015 * rq_truncate : required repbuf_len for resend
1017 int rq_nob_received;
1018 /** Request length */
1022 /** Pool if request is from preallocated list */
1023 struct ptlrpc_request_pool *rq_pool;
1024 /** Request message - what client sent */
1025 struct lustre_msg *rq_reqmsg;
1026 /** Reply message - server response */
1027 struct lustre_msg *rq_repmsg;
1028 /** Transaction number */
1032 /** bulk match bits */
1035 * List item to for replay list. Not yet committed requests get linked
1037 * Also see \a rq_replay comment above.
1038 * It's also link chain on obd_export::exp_req_replay_queue
1040 struct list_head rq_replay_list;
1041 /** non-shared members for client & server request*/
1043 struct ptlrpc_cli_req rq_cli;
1044 struct ptlrpc_srv_req rq_srv;
1047 * security and encryption data
1049 /** description of flavors for client & server */
1050 struct sptlrpc_flavor rq_flvr;
1053 * SELinux policy info at the time of the request
1054 * sepol string format is:
1055 * <mode>:<policy name>:<policy version>:<policy hash>
1057 char rq_sepol[LUSTRE_NODEMAP_SEPOL_LENGTH + 1];
1059 /* client/server security flags */
1061 rq_ctx_init:1, /* context initiation */
1062 rq_ctx_fini:1, /* context destroy */
1063 rq_bulk_read:1, /* request bulk read */
1064 rq_bulk_write:1, /* request bulk write */
1065 /* server authentication flags */
1066 rq_auth_gss:1, /* authenticated by gss */
1067 rq_auth_usr_root:1, /* authed as root */
1068 rq_auth_usr_mdt:1, /* authed as mdt */
1069 rq_auth_usr_ost:1, /* authed as ost */
1070 /* security tfm flags */
1073 /* doesn't expect reply FIXME */
1075 rq_pill_init:1, /* pill initialized */
1076 rq_srv_req:1; /* server request */
1079 /** various buffer pointers */
1080 struct lustre_msg *rq_reqbuf; /**< req wrapper, vmalloc*/
1081 char *rq_repbuf; /**< rep buffer, vmalloc */
1082 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
1083 /** only in priv mode */
1084 struct lustre_msg *rq_clrbuf;
1085 int rq_reqbuf_len; /* req wrapper buf len */
1086 int rq_reqdata_len; /* req wrapper msg len */
1087 int rq_repbuf_len; /* rep buffer len */
1088 int rq_repdata_len; /* rep wrapper msg len */
1089 int rq_clrbuf_len; /* only in priv mode */
1090 int rq_clrdata_len; /* only in priv mode */
1092 /** early replies go to offset 0, regular replies go after that */
1093 unsigned int rq_reply_off;
1096 /** Fields that help to see if request and reply were swabbed or not */
1097 __u32 rq_req_swab_mask;
1098 __u32 rq_rep_swab_mask;
1100 /** how many early replies (for stats) */
1102 /** Server-side, export on which request was received */
1103 struct obd_export *rq_export;
1104 /** import where request is being sent */
1105 struct obd_import *rq_import;
1108 /** Peer description (the other side) */
1109 struct lnet_process_id rq_peer;
1110 /** Descriptor for the NID from which the peer sent the request. */
1111 struct lnet_process_id rq_source;
1113 * service time estimate (secs)
1114 * If the request is not served by this time, it is marked as timed out.
1115 * Do not change to time64_t since this is transmitted over the wire.
1119 * when request/reply sent (secs), or time when request should be sent
1122 /** when request must finish. */
1123 time64_t rq_deadline;
1124 /** request format description */
1125 struct req_capsule rq_pill;
1129 * Call completion handler for rpc if any, return it's status or original
1130 * rc if there was no handler defined for this request.
1132 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1133 struct ptlrpc_request *req, int rc)
1135 if (req->rq_interpret_reply != NULL) {
1136 req->rq_status = req->rq_interpret_reply(env, req,
1137 &req->rq_async_args,
1139 return req->rq_status;
1148 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
1149 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
1150 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1151 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1152 struct ptlrpc_nrs_pol_info *info);
1155 * Can the request be moved from the regular NRS head to the high-priority NRS
1156 * head (of the same PTLRPC service partition), if any?
1158 * For a reliable result, this should be checked under svcpt->scp_req lock.
1160 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1162 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1165 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1166 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1167 * to make sure it has not been scheduled yet (analogous to previous
1168 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1170 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1175 * Returns true if request buffer at offset \a index was already swabbed
1177 static inline bool lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
1179 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1180 return req->rq_req_swab_mask & (1 << index);
1184 * Returns true if request reply buffer at offset \a index was already swabbed
1186 static inline bool lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
1188 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1189 return req->rq_rep_swab_mask & (1 << index);
1193 * Returns true if request needs to be swabbed into local cpu byteorder
1195 static inline bool ptlrpc_req_need_swab(struct ptlrpc_request *req)
1197 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1201 * Returns true if request reply needs to be swabbed into local cpu byteorder
1203 static inline bool ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1205 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1209 * Mark request buffer at offset \a index that it was already swabbed
1211 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
1214 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1215 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1216 req->rq_req_swab_mask |= 1 << index;
1220 * Mark request reply buffer at offset \a index that it was already swabbed
1222 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
1225 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1226 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1227 req->rq_rep_swab_mask |= 1 << index;
1231 * Convert numerical request phase value \a phase into text string description
1233 static inline const char *
1234 ptlrpc_phase2str(enum rq_phase phase)
1243 case RQ_PHASE_INTERPRET:
1245 case RQ_PHASE_COMPLETE:
1247 case RQ_PHASE_UNREG_RPC:
1249 case RQ_PHASE_UNREG_BULK:
1257 * Convert numerical request phase of the request \a req into text stringi
1260 static inline const char *
1261 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1263 return ptlrpc_phase2str(req->rq_phase);
1267 * Debugging functions and helpers to print request structure into debug log
1270 /* Spare the preprocessor, spoil the bugs. */
1271 #define FLAG(field, str) (field ? str : "")
1273 /** Convert bit flags into a string */
1274 #define DEBUG_REQ_FLAGS(req) \
1275 ptlrpc_rqphase2str(req), \
1276 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1277 FLAG(req->rq_err, "E"), FLAG(req->rq_net_err, "e"), \
1278 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1279 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1280 FLAG(req->rq_no_resend, "N"), \
1281 FLAG(req->rq_waiting, "W"), \
1282 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1283 FLAG(req->rq_committed, "M")
1285 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s%s"
1287 void _debug_req(struct ptlrpc_request *req,
1288 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1289 __attribute__ ((format (printf, 3, 4)));
1292 * Helper that decides if we need to print request accordig to current debug
1295 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1297 CFS_CHECK_STACK(msgdata, mask, cdls); \
1299 if (((mask) & D_CANTMASK) != 0 || \
1300 ((libcfs_debug & (mask)) != 0 && \
1301 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1302 _debug_req((req), msgdata, fmt, ##a); \
1306 * This is the debug print function you need to use to print request sturucture
1307 * content into lustre debug log.
1308 * for most callers (level is a constant) this is resolved at compile time */
1309 #define DEBUG_REQ(level, req, fmt, args...) \
1311 if ((level) & (D_ERROR | D_WARNING)) { \
1312 static struct cfs_debug_limit_state cdls; \
1313 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1314 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1316 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1317 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1323 * Structure that defines a single page of a bulk transfer
1325 struct ptlrpc_bulk_page {
1326 /** Linkage to list of pages in a bulk */
1327 struct list_head bp_link;
1329 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1332 /** offset within a page */
1334 /** The page itself */
1335 struct page *bp_page;
1338 enum ptlrpc_bulk_op_type {
1339 PTLRPC_BULK_OP_ACTIVE = 0x00000001,
1340 PTLRPC_BULK_OP_PASSIVE = 0x00000002,
1341 PTLRPC_BULK_OP_PUT = 0x00000004,
1342 PTLRPC_BULK_OP_GET = 0x00000008,
1343 PTLRPC_BULK_BUF_KVEC = 0x00000010,
1344 PTLRPC_BULK_BUF_KIOV = 0x00000020,
1345 PTLRPC_BULK_GET_SOURCE = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_GET,
1346 PTLRPC_BULK_PUT_SINK = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_PUT,
1347 PTLRPC_BULK_GET_SINK = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_GET,
1348 PTLRPC_BULK_PUT_SOURCE = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_PUT,
1351 static inline bool ptlrpc_is_bulk_op_get(enum ptlrpc_bulk_op_type type)
1353 return (type & PTLRPC_BULK_OP_GET) == PTLRPC_BULK_OP_GET;
1356 static inline bool ptlrpc_is_bulk_get_source(enum ptlrpc_bulk_op_type type)
1358 return (type & PTLRPC_BULK_GET_SOURCE) == PTLRPC_BULK_GET_SOURCE;
1361 static inline bool ptlrpc_is_bulk_put_sink(enum ptlrpc_bulk_op_type type)
1363 return (type & PTLRPC_BULK_PUT_SINK) == PTLRPC_BULK_PUT_SINK;
1366 static inline bool ptlrpc_is_bulk_get_sink(enum ptlrpc_bulk_op_type type)
1368 return (type & PTLRPC_BULK_GET_SINK) == PTLRPC_BULK_GET_SINK;
1371 static inline bool ptlrpc_is_bulk_put_source(enum ptlrpc_bulk_op_type type)
1373 return (type & PTLRPC_BULK_PUT_SOURCE) == PTLRPC_BULK_PUT_SOURCE;
1376 static inline bool ptlrpc_is_bulk_desc_kvec(enum ptlrpc_bulk_op_type type)
1378 return ((type & PTLRPC_BULK_BUF_KVEC) | (type & PTLRPC_BULK_BUF_KIOV))
1379 == PTLRPC_BULK_BUF_KVEC;
1382 static inline bool ptlrpc_is_bulk_desc_kiov(enum ptlrpc_bulk_op_type type)
1384 return ((type & PTLRPC_BULK_BUF_KVEC) | (type & PTLRPC_BULK_BUF_KIOV))
1385 == PTLRPC_BULK_BUF_KIOV;
1388 static inline bool ptlrpc_is_bulk_op_active(enum ptlrpc_bulk_op_type type)
1390 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1391 (type & PTLRPC_BULK_OP_PASSIVE))
1392 == PTLRPC_BULK_OP_ACTIVE;
1395 static inline bool ptlrpc_is_bulk_op_passive(enum ptlrpc_bulk_op_type type)
1397 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1398 (type & PTLRPC_BULK_OP_PASSIVE))
1399 == PTLRPC_BULK_OP_PASSIVE;
1402 struct ptlrpc_bulk_frag_ops {
1404 * Add a page \a page to the bulk descriptor \a desc
1405 * Data to transfer in the page starts at offset \a pageoffset and
1406 * amount of data to transfer from the page is \a len
1408 void (*add_kiov_frag)(struct ptlrpc_bulk_desc *desc,
1409 struct page *page, int pageoffset, int len);
1412 * Add a \a fragment to the bulk descriptor \a desc.
1413 * Data to transfer in the fragment is pointed to by \a frag
1414 * The size of the fragment is \a len
1416 int (*add_iov_frag)(struct ptlrpc_bulk_desc *desc, void *frag, int len);
1419 * Uninitialize and free bulk descriptor \a desc.
1420 * Works on bulk descriptors both from server and client side.
1422 void (*release_frags)(struct ptlrpc_bulk_desc *desc);
1425 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops;
1426 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops;
1427 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops;
1430 * Definition of bulk descriptor.
1431 * Bulks are special "Two phase" RPCs where initial request message
1432 * is sent first and it is followed bt a transfer (o receiving) of a large
1433 * amount of data to be settled into pages referenced from the bulk descriptors.
1434 * Bulks transfers (the actual data following the small requests) are done
1435 * on separate LNet portals.
1436 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1437 * Another user is readpage for MDT.
1439 struct ptlrpc_bulk_desc {
1440 /** completed with failure */
1441 unsigned long bd_failure:1;
1443 unsigned long bd_registered:1;
1444 /** For serialization with callback */
1446 /** Import generation when request for this bulk was sent */
1447 int bd_import_generation;
1448 /** {put,get}{source,sink}{kvec,kiov} */
1449 enum ptlrpc_bulk_op_type bd_type;
1450 /** LNet portal for this bulk */
1452 /** Server side - export this bulk created for */
1453 struct obd_export *bd_export;
1454 /** Client side - import this bulk was sent on */
1455 struct obd_import *bd_import;
1456 /** Back pointer to the request */
1457 struct ptlrpc_request *bd_req;
1458 struct ptlrpc_bulk_frag_ops *bd_frag_ops;
1459 wait_queue_head_t bd_waitq; /* server side only WQ */
1460 int bd_iov_count; /* # entries in bd_iov */
1461 int bd_max_iov; /* allocated size of bd_iov */
1462 int bd_nob; /* # bytes covered */
1463 int bd_nob_transferred; /* # bytes GOT/PUT */
1465 __u64 bd_last_mbits;
1467 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1468 lnet_nid_t bd_sender; /* stash event::sender */
1469 int bd_md_count; /* # valid entries in bd_mds */
1470 int bd_md_max_brw; /* max entries in bd_mds */
1471 /** array of associated MDs */
1472 struct lnet_handle_md bd_mds[PTLRPC_BULK_OPS_COUNT];
1477 * encrypt iov, size is either 0 or bd_iov_count.
1479 lnet_kiov_t *bd_enc_vec;
1480 lnet_kiov_t *bd_vec;
1484 struct kvec *bd_enc_kvec;
1485 struct kvec *bd_kvec;
1491 #define GET_KIOV(desc) ((desc)->bd_u.bd_kiov.bd_vec)
1492 #define BD_GET_KIOV(desc, i) ((desc)->bd_u.bd_kiov.bd_vec[i])
1493 #define GET_ENC_KIOV(desc) ((desc)->bd_u.bd_kiov.bd_enc_vec)
1494 #define BD_GET_ENC_KIOV(desc, i) ((desc)->bd_u.bd_kiov.bd_enc_vec[i])
1495 #define GET_KVEC(desc) ((desc)->bd_u.bd_kvec.bd_kvec)
1496 #define BD_GET_KVEC(desc, i) ((desc)->bd_u.bd_kvec.bd_kvec[i])
1497 #define GET_ENC_KVEC(desc) ((desc)->bd_u.bd_kvec.bd_enc_kvec)
1498 #define BD_GET_ENC_KVEC(desc, i) ((desc)->bd_u.bd_kvec.bd_enc_kvec[i])
1502 SVC_STOPPED = 1 << 0,
1503 SVC_STOPPING = 1 << 1,
1504 SVC_STARTING = 1 << 2,
1505 SVC_RUNNING = 1 << 3,
1507 SVC_SIGNAL = 1 << 5,
1510 #define PTLRPC_THR_NAME_LEN 32
1512 * Definition of server service thread structure
1514 struct ptlrpc_thread {
1516 * List of active threads in svc->srv_threads
1518 struct list_head t_link;
1520 * thread-private data (preallocated vmalloc'd memory)
1525 * service thread index, from ptlrpc_start_threads
1531 struct task_struct *t_task;
1535 * put watchdog in the structure per thread b=14840
1537 struct delayed_work t_watchdog;
1539 * the svc this thread belonged to b=18582
1541 struct ptlrpc_service_part *t_svcpt;
1542 wait_queue_head_t t_ctl_waitq;
1543 struct lu_env *t_env;
1544 char t_name[PTLRPC_THR_NAME_LEN];
1547 static inline int thread_is_init(struct ptlrpc_thread *thread)
1549 return thread->t_flags == 0;
1552 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1554 return !!(thread->t_flags & SVC_STOPPED);
1557 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1559 return !!(thread->t_flags & SVC_STOPPING);
1562 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1564 return !!(thread->t_flags & SVC_STARTING);
1567 static inline int thread_is_running(struct ptlrpc_thread *thread)
1569 return !!(thread->t_flags & SVC_RUNNING);
1572 static inline int thread_is_event(struct ptlrpc_thread *thread)
1574 return !!(thread->t_flags & SVC_EVENT);
1577 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1579 return !!(thread->t_flags & SVC_SIGNAL);
1582 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1584 thread->t_flags &= ~flags;
1587 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1589 thread->t_flags = flags;
1592 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1594 thread->t_flags |= flags;
1597 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1600 if (thread->t_flags & flags) {
1601 thread->t_flags &= ~flags;
1608 * Request buffer descriptor structure.
1609 * This is a structure that contains one posted request buffer for service.
1610 * Once data land into a buffer, event callback creates actual request and
1611 * notifies wakes one of the service threads to process new incoming request.
1612 * More than one request can fit into the buffer.
1614 struct ptlrpc_request_buffer_desc {
1615 /** Link item for rqbds on a service */
1616 struct list_head rqbd_list;
1617 /** History of requests for this buffer */
1618 struct list_head rqbd_reqs;
1619 /** Back pointer to service for which this buffer is registered */
1620 struct ptlrpc_service_part *rqbd_svcpt;
1621 /** LNet descriptor */
1622 struct lnet_handle_md rqbd_md_h;
1624 /** The buffer itself */
1626 struct ptlrpc_cb_id rqbd_cbid;
1628 * This "embedded" request structure is only used for the
1629 * last request to fit into the buffer
1631 struct ptlrpc_request rqbd_req;
1634 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1636 struct ptlrpc_service_ops {
1638 * if non-NULL called during thread creation (ptlrpc_start_thread())
1639 * to initialize service specific per-thread state.
1641 int (*so_thr_init)(struct ptlrpc_thread *thr);
1643 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1644 * destruct state created by ->srv_init().
1646 void (*so_thr_done)(struct ptlrpc_thread *thr);
1648 * Handler function for incoming requests for this service
1650 int (*so_req_handler)(struct ptlrpc_request *req);
1652 * function to determine priority of the request, it's called
1653 * on every new request
1655 int (*so_hpreq_handler)(struct ptlrpc_request *);
1657 * service-specific print fn
1659 void (*so_req_printer)(void *, struct ptlrpc_request *);
1662 #ifndef __cfs_cacheline_aligned
1663 /* NB: put it here for reducing patche dependence */
1664 # define __cfs_cacheline_aligned
1668 * How many high priority requests to serve before serving one normal
1671 #define PTLRPC_SVC_HP_RATIO 10
1674 * Definition of PortalRPC service.
1675 * The service is listening on a particular portal (like tcp port)
1676 * and perform actions for a specific server like IO service for OST
1677 * or general metadata service for MDS.
1679 struct ptlrpc_service {
1680 /** serialize /proc operations */
1681 spinlock_t srv_lock;
1682 /** most often accessed fields */
1683 /** chain thru all services */
1684 struct list_head srv_list;
1685 /** service operations table */
1686 struct ptlrpc_service_ops srv_ops;
1687 /** only statically allocated strings here; we don't clean them */
1689 /** only statically allocated strings here; we don't clean them */
1690 char *srv_thread_name;
1691 /** service thread list */
1692 struct list_head srv_threads;
1693 /** threads # should be created for each partition on initializing */
1694 int srv_nthrs_cpt_init;
1695 /** limit of threads number for each partition */
1696 int srv_nthrs_cpt_limit;
1697 /** Root of debugfs dir tree for this service */
1698 struct dentry *srv_debugfs_entry;
1699 /** Pointer to statistic data for this service */
1700 struct lprocfs_stats *srv_stats;
1701 /** # hp per lp reqs to handle */
1702 int srv_hpreq_ratio;
1703 /** biggest request to receive */
1704 int srv_max_req_size;
1705 /** biggest reply to send */
1706 int srv_max_reply_size;
1707 /** size of individual buffers */
1709 /** # buffers to allocate in 1 group */
1710 int srv_nbuf_per_group;
1711 /** Local portal on which to receive requests */
1712 __u32 srv_req_portal;
1713 /** Portal on the client to send replies to */
1714 __u32 srv_rep_portal;
1716 * Tags for lu_context associated with this thread, see struct
1720 /** soft watchdog timeout multiplier */
1721 int srv_watchdog_factor;
1722 /** under unregister_service */
1723 unsigned srv_is_stopping:1;
1724 /** Whether or not to restrict service threads to CPUs in this CPT */
1725 unsigned srv_cpt_bind:1;
1727 /** max # request buffers */
1729 /** max # request buffers in history per partition */
1730 int srv_hist_nrqbds_cpt_max;
1731 /** number of CPTs this service associated with */
1733 /** CPTs array this service associated with */
1735 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1737 /** CPT table this service is running over */
1738 struct cfs_cpt_table *srv_cptable;
1741 struct kobject srv_kobj;
1742 struct completion srv_kobj_unregister;
1744 * partition data for ptlrpc service
1746 struct ptlrpc_service_part *srv_parts[0];
1750 * Definition of PortalRPC service partition data.
1751 * Although a service only has one instance of it right now, but we
1752 * will have multiple instances very soon (instance per CPT).
1754 * it has four locks:
1756 * serialize operations on rqbd and requests waiting for preprocess
1758 * serialize operations active requests sent to this portal
1760 * serialize adaptive timeout stuff
1762 * serialize operations on RS list (reply states)
1764 * We don't have any use-case to take two or more locks at the same time
1765 * for now, so there is no lock order issue.
1767 struct ptlrpc_service_part {
1768 /** back reference to owner */
1769 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1770 /* CPT id, reserved */
1772 /** always increasing number */
1774 /** # of starting threads */
1775 int scp_nthrs_starting;
1776 /** # of stopping threads, reserved for shrinking threads */
1777 int scp_nthrs_stopping;
1778 /** # running threads */
1779 int scp_nthrs_running;
1780 /** service threads list */
1781 struct list_head scp_threads;
1784 * serialize the following fields, used for protecting
1785 * rqbd list and incoming requests waiting for preprocess,
1786 * threads starting & stopping are also protected by this lock.
1788 spinlock_t scp_lock __cfs_cacheline_aligned;
1789 /** userland serialization */
1790 struct mutex scp_mutex;
1791 /** total # req buffer descs allocated */
1792 int scp_nrqbds_total;
1793 /** # posted request buffers for receiving */
1794 int scp_nrqbds_posted;
1795 /** in progress of allocating rqbd */
1796 int scp_rqbd_allocating;
1797 /** # incoming reqs */
1798 int scp_nreqs_incoming;
1799 /** request buffers to be reposted */
1800 struct list_head scp_rqbd_idle;
1801 /** req buffers receiving */
1802 struct list_head scp_rqbd_posted;
1803 /** incoming reqs */
1804 struct list_head scp_req_incoming;
1805 /** timeout before re-posting reqs, in jiffies */
1806 long scp_rqbd_timeout;
1808 * all threads sleep on this. This wait-queue is signalled when new
1809 * incoming request arrives and when difficult reply has to be handled.
1811 wait_queue_head_t scp_waitq;
1813 /** request history */
1814 struct list_head scp_hist_reqs;
1815 /** request buffer history */
1816 struct list_head scp_hist_rqbds;
1817 /** # request buffers in history */
1818 int scp_hist_nrqbds;
1819 /** sequence number for request */
1821 /** highest seq culled from history */
1822 __u64 scp_hist_seq_culled;
1825 * serialize the following fields, used for processing requests
1826 * sent to this portal
1828 spinlock_t scp_req_lock __cfs_cacheline_aligned;
1829 /** # reqs in either of the NRS heads below */
1830 /** # reqs being served */
1831 int scp_nreqs_active;
1832 /** # HPreqs being served */
1833 int scp_nhreqs_active;
1834 /** # hp requests handled */
1837 /** NRS head for regular requests */
1838 struct ptlrpc_nrs scp_nrs_reg;
1839 /** NRS head for HP requests; this is only valid for services that can
1840 * handle HP requests */
1841 struct ptlrpc_nrs *scp_nrs_hp;
1846 * serialize the following fields, used for changes on
1849 spinlock_t scp_at_lock __cfs_cacheline_aligned;
1850 /** estimated rpc service time */
1851 struct adaptive_timeout scp_at_estimate;
1852 /** reqs waiting for replies */
1853 struct ptlrpc_at_array scp_at_array;
1854 /** early reply timer */
1855 struct timer_list scp_at_timer;
1857 ktime_t scp_at_checktime;
1858 /** check early replies */
1859 unsigned scp_at_check;
1863 * serialize the following fields, used for processing
1864 * replies for this portal
1866 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1867 /** all the active replies */
1868 struct list_head scp_rep_active;
1869 /** List of free reply_states */
1870 struct list_head scp_rep_idle;
1871 /** waitq to run, when adding stuff to srv_free_rs_list */
1872 wait_queue_head_t scp_rep_waitq;
1873 /** # 'difficult' replies */
1874 atomic_t scp_nreps_difficult;
1877 #define ptlrpc_service_for_each_part(part, i, svc) \
1879 i < (svc)->srv_ncpts && \
1880 (svc)->srv_parts != NULL && \
1881 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1884 * Declaration of ptlrpcd control structure
1886 struct ptlrpcd_ctl {
1888 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1890 unsigned long pc_flags;
1892 * Thread lock protecting structure fields.
1898 struct completion pc_starting;
1902 struct completion pc_finishing;
1904 * Thread requests set.
1906 struct ptlrpc_request_set *pc_set;
1908 * Thread name used in kthread_run()
1912 * CPT the thread is bound on.
1916 * Index of ptlrpcd thread in the array.
1920 * Pointer to the array of partners' ptlrpcd_ctl structure.
1922 struct ptlrpcd_ctl **pc_partners;
1924 * Number of the ptlrpcd's partners.
1928 * Record the partner index to be processed next.
1932 * Error code if the thread failed to fully start.
1937 /* Bits for pc_flags */
1938 enum ptlrpcd_ctl_flags {
1940 * Ptlrpc thread start flag.
1942 LIOD_START = 1 << 0,
1944 * Ptlrpc thread stop flag.
1948 * Ptlrpc thread force flag (only stop force so far).
1949 * This will cause aborting any inflight rpcs handled
1950 * by thread if LIOD_STOP is specified.
1952 LIOD_FORCE = 1 << 2,
1954 * This is a recovery ptlrpc thread.
1956 LIOD_RECOVERY = 1 << 3,
1963 * Service compatibility function; the policy is compatible with all services.
1965 * \param[in] svc The service the policy is attempting to register with.
1966 * \param[in] desc The policy descriptor
1968 * \retval true The policy is compatible with the service
1970 * \see ptlrpc_nrs_pol_desc::pd_compat()
1972 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
1973 const struct ptlrpc_nrs_pol_desc *desc)
1979 * Service compatibility function; the policy is compatible with only a specific
1980 * service which is identified by its human-readable name at
1981 * ptlrpc_service::srv_name.
1983 * \param[in] svc The service the policy is attempting to register with.
1984 * \param[in] desc The policy descriptor
1986 * \retval false The policy is not compatible with the service
1987 * \retval true The policy is compatible with the service
1989 * \see ptlrpc_nrs_pol_desc::pd_compat()
1991 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
1992 const struct ptlrpc_nrs_pol_desc *desc)
1994 LASSERT(desc->pd_compat_svc_name != NULL);
1995 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2000 /* ptlrpc/events.c */
2001 extern struct lnet_handle_eq ptlrpc_eq_h;
2002 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2003 struct lnet_process_id *peer, lnet_nid_t *self);
2005 * These callbacks are invoked by LNet when something happened to
2009 extern void request_out_callback(struct lnet_event *ev);
2010 extern void reply_in_callback(struct lnet_event *ev);
2011 extern void client_bulk_callback(struct lnet_event *ev);
2012 extern void request_in_callback(struct lnet_event *ev);
2013 extern void reply_out_callback(struct lnet_event *ev);
2014 #ifdef HAVE_SERVER_SUPPORT
2015 extern void server_bulk_callback(struct lnet_event *ev);
2019 /* ptlrpc/connection.c */
2020 struct ptlrpc_connection *ptlrpc_connection_get(struct lnet_process_id peer,
2022 struct obd_uuid *uuid);
2023 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2024 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2025 int ptlrpc_connection_init(void);
2026 void ptlrpc_connection_fini(void);
2027 extern lnet_pid_t ptl_get_pid(void);
2030 * Check if the peer connection is on the local node. We need to use GFP_NOFS
2031 * for requests from a local client to avoid recursing into the filesystem
2032 * as we might end up waiting on a page sent in the request we're serving.
2034 * Use __GFP_HIGHMEM so that the pages can use all of the available memory
2035 * on 32-bit machines. Use more aggressive GFP_HIGHUSER flags from non-local
2036 * clients to be able to generate more memory pressure on the OSS and allow
2037 * inactive pages to be reclaimed, since it doesn't have any other processes
2038 * or allocations that generate memory reclaim pressure.
2040 * See b=17576 (bdf50dc9) and b=19529 (3dcf18d3) for details.
2042 static inline bool ptlrpc_connection_is_local(struct ptlrpc_connection *conn)
2047 if (conn->c_peer.nid == conn->c_self)
2050 RETURN(LNetIsPeerLocal(conn->c_peer.nid));
2053 /* ptlrpc/niobuf.c */
2055 * Actual interfacing with LNet to put/get/register/unregister stuff
2058 #ifdef HAVE_SERVER_SUPPORT
2059 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2060 unsigned nfrags, unsigned max_brw,
2063 const struct ptlrpc_bulk_frag_ops
2065 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2066 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2068 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2072 LASSERT(desc != NULL);
2074 spin_lock(&desc->bd_lock);
2075 rc = desc->bd_md_count;
2076 spin_unlock(&desc->bd_lock);
2081 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2082 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2084 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2086 struct ptlrpc_bulk_desc *desc;
2089 LASSERT(req != NULL);
2090 desc = req->rq_bulk;
2095 if (req->rq_bulk_deadline > ktime_get_real_seconds())
2099 spin_lock(&desc->bd_lock);
2100 rc = desc->bd_md_count;
2101 spin_unlock(&desc->bd_lock);
2105 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2106 #define PTLRPC_REPLY_EARLY 0x02
2107 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2108 int ptlrpc_reply(struct ptlrpc_request *req);
2109 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2110 int ptlrpc_error(struct ptlrpc_request *req);
2111 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2112 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2113 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2116 /* ptlrpc/client.c */
2118 * Client-side portals API. Everything to send requests, receive replies,
2119 * request queues, request management, etc.
2122 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2124 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2125 struct ptlrpc_client *);
2126 void ptlrpc_cleanup_client(struct obd_import *imp);
2127 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
2128 lnet_nid_t nid4refnet);
2130 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2131 int ptlrpc_replay_req(struct ptlrpc_request *req);
2132 void ptlrpc_restart_req(struct ptlrpc_request *req);
2133 void ptlrpc_abort_inflight(struct obd_import *imp);
2134 void ptlrpc_cleanup_imp(struct obd_import *imp);
2135 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2137 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2138 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2140 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2141 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *);
2142 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2143 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2144 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2146 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2147 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2149 struct ptlrpc_request_pool *
2150 ptlrpc_init_rq_pool(int, int,
2151 int (*populate_pool)(struct ptlrpc_request_pool *, int));
2153 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2154 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2155 const struct req_format *format);
2156 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2157 struct ptlrpc_request_pool *,
2158 const struct req_format *format);
2159 void ptlrpc_request_free(struct ptlrpc_request *request);
2160 int ptlrpc_request_pack(struct ptlrpc_request *request,
2161 __u32 version, int opcode);
2162 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2163 const struct req_format *format,
2164 __u32 version, int opcode);
2165 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2166 __u32 version, int opcode, char **bufs,
2167 struct ptlrpc_cli_ctx *ctx);
2168 void ptlrpc_req_finished(struct ptlrpc_request *request);
2169 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2170 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2171 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2172 unsigned nfrags, unsigned max_brw,
2175 const struct ptlrpc_bulk_frag_ops
2178 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
2179 void *frag, int len);
2180 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2181 struct page *page, int pageoffset, int len,
2183 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2184 struct page *page, int pageoffset,
2187 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2190 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2191 struct page *page, int pageoffset,
2194 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2197 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
2199 static inline void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
2203 for (i = 0; i < desc->bd_iov_count ; i++)
2204 put_page(BD_GET_KIOV(desc, i).kiov_page);
2207 static inline void ptlrpc_release_bulk_noop(struct ptlrpc_bulk_desc *desc)
2211 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2212 struct obd_import *imp);
2213 __u64 ptlrpc_next_xid(void);
2214 __u64 ptlrpc_sample_next_xid(void);
2215 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2217 /* Set of routines to run a function in ptlrpcd context */
2218 void *ptlrpcd_alloc_work(struct obd_import *imp,
2219 int (*cb)(const struct lu_env *, void *), void *data);
2220 void ptlrpcd_destroy_work(void *handler);
2221 int ptlrpcd_queue_work(void *handler);
2224 struct ptlrpc_service_buf_conf {
2225 /* nbufs is buffers # to allocate when growing the pool */
2226 unsigned int bc_nbufs;
2227 /* buffer size to post */
2228 unsigned int bc_buf_size;
2229 /* portal to listed for requests on */
2230 unsigned int bc_req_portal;
2231 /* portal of where to send replies to */
2232 unsigned int bc_rep_portal;
2233 /* maximum request size to be accepted for this service */
2234 unsigned int bc_req_max_size;
2235 /* maximum reply size this service can ever send */
2236 unsigned int bc_rep_max_size;
2239 struct ptlrpc_service_thr_conf {
2240 /* threadname should be 8 characters or less - 6 will be added on */
2242 /* threads increasing factor for each CPU */
2243 unsigned int tc_thr_factor;
2244 /* service threads # to start on each partition while initializing */
2245 unsigned int tc_nthrs_init;
2247 * low water of threads # upper-limit on each partition while running,
2248 * service availability may be impacted if threads number is lower
2249 * than this value. It can be ZERO if the service doesn't require
2250 * CPU affinity or there is only one partition.
2252 unsigned int tc_nthrs_base;
2253 /* "soft" limit for total threads number */
2254 unsigned int tc_nthrs_max;
2255 /* user specified threads number, it will be validated due to
2256 * other members of this structure. */
2257 unsigned int tc_nthrs_user;
2258 /* bind service threads to only CPUs in their associated CPT */
2259 unsigned int tc_cpu_bind;
2260 /* Tags for lu_context associated with service thread */
2264 struct ptlrpc_service_cpt_conf {
2265 struct cfs_cpt_table *cc_cptable;
2266 /* string pattern to describe CPTs for a service */
2268 /* whether or not to have per-CPT service partitions */
2272 struct ptlrpc_service_conf {
2275 /* soft watchdog timeout multiplifier to print stuck service traces */
2276 unsigned int psc_watchdog_factor;
2277 /* buffer information */
2278 struct ptlrpc_service_buf_conf psc_buf;
2279 /* thread information */
2280 struct ptlrpc_service_thr_conf psc_thr;
2281 /* CPU partition information */
2282 struct ptlrpc_service_cpt_conf psc_cpt;
2283 /* function table */
2284 struct ptlrpc_service_ops psc_ops;
2287 /* ptlrpc/service.c */
2289 * Server-side services API. Register/unregister service, request state
2290 * management, service thread management
2294 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
2295 int mode, bool no_ack, bool convert_lock);
2296 void ptlrpc_commit_replies(struct obd_export *exp);
2297 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2298 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2299 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2300 struct ptlrpc_service *ptlrpc_register_service(
2301 struct ptlrpc_service_conf *conf,
2302 struct kset *parent,
2303 struct dentry *debugfs_entry);
2304 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2306 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2307 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2308 int ptlrpc_service_health_check(struct ptlrpc_service *);
2309 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2310 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2311 struct obd_export *export);
2312 void ptlrpc_update_export_timer(struct obd_export *exp,
2313 time64_t extra_delay);
2315 int ptlrpc_hr_init(void);
2316 void ptlrpc_hr_fini(void);
2320 /* ptlrpc/import.c */
2325 int ptlrpc_connect_import(struct obd_import *imp);
2326 int ptlrpc_init_import(struct obd_import *imp);
2327 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2328 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp);
2329 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2330 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2332 void ptlrpc_import_enter_resend(struct obd_import *imp);
2333 /* ptlrpc/pack_generic.c */
2334 int ptlrpc_reconnect_import(struct obd_import *imp);
2338 * ptlrpc msg buffer and swab interface
2342 bool ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2344 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2346 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2347 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2349 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2350 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2352 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2353 __u32 *lens, char **bufs);
2354 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2356 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2357 __u32 *lens, char **bufs, int flags);
2358 #define LPRFL_EARLY_REPLY 1
2359 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2360 char **bufs, int flags);
2361 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2362 unsigned int newlen, int move_data);
2363 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2364 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2365 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
2366 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2367 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
2368 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
2369 __u32 lustre_msg_early_size(void);
2370 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
2371 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
2372 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
2373 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
2374 __u32 lustre_msg_bufcount(struct lustre_msg *m);
2375 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
2376 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2377 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2378 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2379 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
2380 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
2381 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
2382 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2383 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
2384 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2385 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2386 enum lustre_msg_version lustre_msg_get_version(struct lustre_msg *msg);
2387 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
2388 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2389 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2390 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
2391 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2392 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2393 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2394 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2395 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2396 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2397 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2398 int lustre_msg_get_status(struct lustre_msg *msg);
2399 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2400 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2401 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2402 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2403 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2404 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2405 __u64 lustre_msg_get_mbits(struct lustre_msg *msg);
2406 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2407 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2408 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2409 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2410 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2411 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
2412 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2413 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2414 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2415 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2416 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2417 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2418 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2419 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2420 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2421 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2422 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2423 void lustre_msg_set_mbits(struct lustre_msg *msg, __u64 mbits);
2426 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2427 unsigned int newlen, int move_data)
2429 LASSERT(req->rq_reply_state);
2430 LASSERT(req->rq_repmsg);
2431 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2435 #ifdef LUSTRE_TRANSLATE_ERRNOS
2437 static inline int ptlrpc_status_hton(int h)
2440 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2441 * ELDLM_LOCK_ABORTED, etc.
2444 return -lustre_errno_hton(-h);
2449 static inline int ptlrpc_status_ntoh(int n)
2452 * See the comment in ptlrpc_status_hton().
2455 return -lustre_errno_ntoh(-n);
2462 #define ptlrpc_status_hton(h) (h)
2463 #define ptlrpc_status_ntoh(n) (n)
2468 /** Change request phase of \a req to \a new_phase */
2470 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2472 if (req->rq_phase == new_phase)
2475 if (new_phase == RQ_PHASE_UNREG_RPC ||
2476 new_phase == RQ_PHASE_UNREG_BULK) {
2477 /* No embedded unregistering phases */
2478 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2479 req->rq_phase == RQ_PHASE_UNREG_BULK)
2482 req->rq_next_phase = req->rq_phase;
2484 atomic_inc(&req->rq_import->imp_unregistering);
2487 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2488 req->rq_phase == RQ_PHASE_UNREG_BULK) {
2490 atomic_dec(&req->rq_import->imp_unregistering);
2493 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2494 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2496 req->rq_phase = new_phase;
2500 * Returns true if request \a req got early reply and hard deadline is not met
2503 ptlrpc_client_early(struct ptlrpc_request *req)
2505 return req->rq_early;
2509 * Returns true if we got real reply from server for this request
2512 ptlrpc_client_replied(struct ptlrpc_request *req)
2514 if (req->rq_reply_deadline > ktime_get_real_seconds())
2516 return req->rq_replied;
2519 /** Returns true if request \a req is in process of receiving server reply */
2521 ptlrpc_client_recv(struct ptlrpc_request *req)
2523 if (req->rq_reply_deadline > ktime_get_real_seconds())
2525 return req->rq_receiving_reply;
2529 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2533 spin_lock(&req->rq_lock);
2534 if (req->rq_reply_deadline > ktime_get_real_seconds()) {
2535 spin_unlock(&req->rq_lock);
2538 if (req->rq_req_deadline > ktime_get_real_seconds()) {
2539 spin_unlock(&req->rq_lock);
2543 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
2544 req->rq_receiving_reply;
2545 spin_unlock(&req->rq_lock);
2550 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2553 if (req->rq_set == NULL)
2554 wake_up(&req->rq_reply_waitq);
2556 wake_up(&req->rq_set->set_waitq);
2560 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2562 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2563 atomic_inc(&rs->rs_refcount);
2567 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2569 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2570 if (atomic_dec_and_test(&rs->rs_refcount))
2571 lustre_free_reply_state(rs);
2574 /* Should only be called once per req */
2575 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2577 if (req->rq_reply_state == NULL)
2578 return; /* shouldn't occur */
2579 ptlrpc_rs_decref(req->rq_reply_state);
2580 req->rq_reply_state = NULL;
2581 req->rq_repmsg = NULL;
2584 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2586 return lustre_msg_get_magic(req->rq_reqmsg);
2589 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2591 switch (req->rq_reqmsg->lm_magic) {
2592 case LUSTRE_MSG_MAGIC_V2:
2593 return req->rq_reqmsg->lm_repsize;
2595 LASSERTF(0, "incorrect message magic: %08x\n",
2596 req->rq_reqmsg->lm_magic);
2601 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2603 if (req->rq_delay_limit != 0 &&
2604 req->rq_queued_time + req->rq_delay_limit < ktime_get_seconds())
2609 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2611 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2612 spin_lock(&req->rq_lock);
2613 req->rq_no_resend = 1;
2614 spin_unlock(&req->rq_lock);
2616 return req->rq_no_resend;
2620 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2622 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2624 return svcpt->scp_service->srv_watchdog_factor *
2625 max_t(int, at, obd_timeout);
2628 static inline struct ptlrpc_service *
2629 ptlrpc_req2svc(struct ptlrpc_request *req)
2631 LASSERT(req->rq_rqbd != NULL);
2632 return req->rq_rqbd->rqbd_svcpt->scp_service;
2635 /* ldlm/ldlm_lib.c */
2637 * Target client logic
2640 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2641 int client_obd_cleanup(struct obd_device *obddev);
2642 int client_connect_import(const struct lu_env *env,
2643 struct obd_export **exp, struct obd_device *obd,
2644 struct obd_uuid *cluuid, struct obd_connect_data *,
2646 int client_disconnect_export(struct obd_export *exp);
2647 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2649 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2650 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2651 struct obd_uuid *uuid);
2652 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2653 void client_destroy_import(struct obd_import *imp);
2656 #ifdef HAVE_SERVER_SUPPORT
2657 int server_disconnect_export(struct obd_export *exp);
2660 /* ptlrpc/pinger.c */
2662 * Pinger API (client side only)
2665 enum timeout_event {
2668 struct timeout_item;
2669 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2670 int ptlrpc_pinger_add_import(struct obd_import *imp);
2671 int ptlrpc_pinger_del_import(struct obd_import *imp);
2672 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
2673 timeout_cb_t cb, void *data,
2674 struct list_head *obd_list);
2675 int ptlrpc_del_timeout_client(struct list_head *obd_list,
2676 enum timeout_event event);
2677 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2678 int ptlrpc_obd_ping(struct obd_device *obd);
2679 void ping_evictor_start(void);
2680 void ping_evictor_stop(void);
2681 void ptlrpc_pinger_ir_up(void);
2682 void ptlrpc_pinger_ir_down(void);
2684 int ptlrpc_pinger_suppress_pings(void);
2686 /* ptlrpc/ptlrpcd.c */
2687 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2688 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2689 void ptlrpcd_wake(struct ptlrpc_request *req);
2690 void ptlrpcd_add_req(struct ptlrpc_request *req);
2691 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2692 int ptlrpcd_addref(void);
2693 void ptlrpcd_decref(void);
2695 /* ptlrpc/lproc_ptlrpc.c */
2697 * procfs output related functions
2700 const char* ll_opcode2str(__u32 opcode);
2701 const int ll_str2opcode(const char *ops);
2702 #ifdef CONFIG_PROC_FS
2703 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2704 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2705 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2707 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2708 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2709 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2713 /* ptlrpc/llog_server.c */
2714 int llog_origin_handle_open(struct ptlrpc_request *req);
2715 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2716 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2717 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2719 /* ptlrpc/llog_client.c */
2720 extern struct llog_operations llog_client_ops;