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 (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 (_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 (16 * 1024)
489 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
490 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
491 /* Safe estimate of free space in standard RPC, provides upper limit for # of
492 * bytes of i/o to pack in RPC (skipping bulk transfer). */
493 #define OST_SHORT_IO_SPACE (OST_IO_MAXREQSIZE - _OST_MAXREQSIZE_BASE)
495 /* Actual size used for short i/o buffer. Calculation means this:
496 * At least one page (for large PAGE_SIZE), or 16 KiB, but not more
497 * than the available space aligned to a page boundary. */
498 #define OBD_MAX_SHORT_IO_BYTES (min(max(PAGE_SIZE, 16UL * 1024UL), \
499 OST_SHORT_IO_SPACE & PAGE_MASK))
501 #define OST_MAXREPSIZE (9 * 1024)
502 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
505 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
506 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
508 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
509 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
511 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
514 /* Macro to hide a typecast. */
515 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
517 struct ptlrpc_replay_async_args {
523 * Structure to single define portal connection.
525 struct ptlrpc_connection {
526 /** linkage for connections hash table */
527 struct hlist_node c_hash;
528 /** Our own lnet nid for this connection */
530 /** Remote side nid for this connection */
531 struct lnet_process_id c_peer;
532 /** UUID of the other side */
533 struct obd_uuid c_remote_uuid;
534 /** reference counter for this connection */
538 /** Client definition for PortalRPC */
539 struct ptlrpc_client {
540 /** What lnet portal does this client send messages to by default */
541 __u32 cli_request_portal;
542 /** What portal do we expect replies on */
543 __u32 cli_reply_portal;
544 /** Name of the client */
548 /** state flags of requests */
549 /* XXX only ones left are those used by the bulk descs as well! */
550 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
551 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
553 #define REQ_MAX_ACK_LOCKS 8
555 union ptlrpc_async_args {
557 * Scratchpad for passing args to completion interpreter. Users
558 * cast to the struct of their choosing, and CLASSERT that this is
559 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
560 * a pointer to it here. The pointer_arg ensures this struct is at
561 * least big enough for that.
563 void *pointer_arg[11];
567 struct ptlrpc_request_set;
568 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
571 * Definition of request set structure.
572 * Request set is a list of requests (not necessary to the same target) that
573 * once populated with RPCs could be sent in parallel.
574 * There are two kinds of request sets. General purpose and with dedicated
575 * serving thread. Example of the latter is ptlrpcd set.
576 * For general purpose sets once request set started sending it is impossible
577 * to add new requests to such set.
578 * Provides a way to call "completion callbacks" when all requests in the set
581 struct ptlrpc_request_set {
582 atomic_t set_refcount;
583 /** number of in queue requests */
584 atomic_t set_new_count;
585 /** number of uncompleted requests */
586 atomic_t set_remaining;
587 /** wait queue to wait on for request events */
588 wait_queue_head_t set_waitq;
589 /** List of requests in the set */
590 struct list_head set_requests;
592 * Lock for \a set_new_requests manipulations
593 * locked so that any old caller can communicate requests to
594 * the set holder who can then fold them into the lock-free set
596 spinlock_t set_new_req_lock;
597 /** List of new yet unsent requests. Only used with ptlrpcd now. */
598 struct list_head set_new_requests;
600 /** rq_status of requests that have been freed already */
602 /** Additional fields used by the flow control extension */
603 /** Maximum number of RPCs in flight */
604 int set_max_inflight;
605 /** Callback function used to generate RPCs */
606 set_producer_func set_producer;
607 /** opaq argument passed to the producer callback */
608 void *set_producer_arg;
609 unsigned int set_allow_intr:1;
612 struct ptlrpc_bulk_desc;
613 struct ptlrpc_service_part;
614 struct ptlrpc_service;
617 * ptlrpc callback & work item stuff
619 struct ptlrpc_cb_id {
620 void (*cbid_fn)(struct lnet_event *ev); /* specific callback fn */
621 void *cbid_arg; /* additional arg */
624 /** Maximum number of locks to fit into reply state */
625 #define RS_MAX_LOCKS 8
629 * Structure to define reply state on the server
630 * Reply state holds various reply message information. Also for "difficult"
631 * replies (rep-ack case) we store the state after sending reply and wait
632 * for the client to acknowledge the reception. In these cases locks could be
633 * added to the state for replay/failover consistency guarantees.
635 struct ptlrpc_reply_state {
636 /** Callback description */
637 struct ptlrpc_cb_id rs_cb_id;
638 /** Linkage for list of all reply states in a system */
639 struct list_head rs_list;
640 /** Linkage for list of all reply states on same export */
641 struct list_head rs_exp_list;
642 /** Linkage for list of all reply states for same obd */
643 struct list_head rs_obd_list;
645 struct list_head rs_debug_list;
647 /** A spinlock to protect the reply state flags */
649 /** Reply state flags */
650 unsigned long rs_difficult:1; /* ACK/commit stuff */
651 unsigned long rs_no_ack:1; /* no ACK, even for
652 difficult requests */
653 unsigned long rs_scheduled:1; /* being handled? */
654 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
655 unsigned long rs_handled:1; /* been handled yet? */
656 unsigned long rs_on_net:1; /* reply_out_callback pending? */
657 unsigned long rs_prealloc:1; /* rs from prealloc list */
658 unsigned long rs_committed:1;/* the transaction was committed
659 and the rs was dispatched
660 by ptlrpc_commit_replies */
661 unsigned long rs_convert_lock:1; /* need to convert saved
662 * locks to COS mode */
663 atomic_t rs_refcount; /* number of users */
664 /** Number of locks awaiting client ACK */
667 /** Size of the state */
671 /** Transaction number */
675 struct obd_export *rs_export;
676 struct ptlrpc_service_part *rs_svcpt;
677 /** Lnet metadata handle for the reply */
678 struct lnet_handle_md rs_md_h;
680 /** Context for the sevice thread */
681 struct ptlrpc_svc_ctx *rs_svc_ctx;
682 /** Reply buffer (actually sent to the client), encoded if needed */
683 struct lustre_msg *rs_repbuf; /* wrapper */
684 /** Size of the reply buffer */
685 int rs_repbuf_len; /* wrapper buf length */
686 /** Size of the reply message */
687 int rs_repdata_len; /* wrapper msg length */
689 * Actual reply message. Its content is encrupted (if needed) to
690 * produce reply buffer for actual sending. In simple case
691 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
693 struct lustre_msg *rs_msg; /* reply message */
695 /** Handles of locks awaiting client reply ACK */
696 struct lustre_handle rs_locks[RS_MAX_LOCKS];
697 /** Lock modes of locks in \a rs_locks */
698 enum ldlm_mode rs_modes[RS_MAX_LOCKS];
701 struct ptlrpc_thread;
705 RQ_PHASE_NEW = 0xebc0de00,
706 RQ_PHASE_RPC = 0xebc0de01,
707 RQ_PHASE_BULK = 0xebc0de02,
708 RQ_PHASE_INTERPRET = 0xebc0de03,
709 RQ_PHASE_COMPLETE = 0xebc0de04,
710 RQ_PHASE_UNREG_RPC = 0xebc0de05,
711 RQ_PHASE_UNREG_BULK = 0xebc0de06,
712 RQ_PHASE_UNDEFINED = 0xebc0de07
715 /** Type of request interpreter call-back */
716 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
717 struct ptlrpc_request *req,
719 /** Type of request resend call-back */
720 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
724 * Definition of request pool structure.
725 * The pool is used to store empty preallocated requests for the case
726 * when we would actually need to send something without performing
727 * any allocations (to avoid e.g. OOM).
729 struct ptlrpc_request_pool {
730 /** Locks the list */
732 /** list of ptlrpc_request structs */
733 struct list_head prp_req_list;
734 /** Maximum message size that would fit into a rquest from this pool */
736 /** Function to allocate more requests for this pool */
737 int (*prp_populate)(struct ptlrpc_request_pool *, int);
745 #include <lustre_nrs.h>
748 * Basic request prioritization operations structure.
749 * The whole idea is centered around locks and RPCs that might affect locks.
750 * When a lock is contended we try to give priority to RPCs that might lead
751 * to fastest release of that lock.
752 * Currently only implemented for OSTs only in a way that makes all
753 * IO and truncate RPCs that are coming from a locked region where a lock is
754 * contended a priority over other requests.
756 struct ptlrpc_hpreq_ops {
758 * Check if the lock handle of the given lock is the same as
759 * taken from the request.
761 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
763 * Check if the request is a high priority one.
765 int (*hpreq_check)(struct ptlrpc_request *);
767 * Called after the request has been handled.
769 void (*hpreq_fini)(struct ptlrpc_request *);
772 struct ptlrpc_cli_req {
773 /** For bulk requests on client only: bulk descriptor */
774 struct ptlrpc_bulk_desc *cr_bulk;
775 /** optional time limit for send attempts */
776 time64_t cr_delay_limit;
777 /** time request was first queued */
778 time64_t cr_queued_time;
779 /** request sent in nanoseconds */
781 /** time for request really sent out */
782 time64_t cr_sent_out;
783 /** when req reply unlink must finish. */
784 time64_t cr_reply_deadline;
785 /** when req bulk unlink must finish. */
786 time64_t cr_bulk_deadline;
787 /** when req unlink must finish. */
788 time64_t cr_req_deadline;
789 /** Portal to which this request would be sent */
791 /** Portal where to wait for reply and where reply would be sent */
793 /** request resending number */
794 unsigned int cr_resend_nr;
795 /** What was import generation when this request was sent */
797 enum lustre_imp_state cr_send_state;
798 /** Per-request waitq introduced by bug 21938 for recovery waiting */
799 wait_queue_head_t cr_set_waitq;
800 /** Link item for request set lists */
801 struct list_head cr_set_chain;
802 /** link to waited ctx */
803 struct list_head cr_ctx_chain;
805 /** client's half ctx */
806 struct ptlrpc_cli_ctx *cr_cli_ctx;
807 /** Link back to the request set */
808 struct ptlrpc_request_set *cr_set;
809 /** outgoing request MD handle */
810 struct lnet_handle_md cr_req_md_h;
811 /** request-out callback parameter */
812 struct ptlrpc_cb_id cr_req_cbid;
813 /** incoming reply MD handle */
814 struct lnet_handle_md cr_reply_md_h;
815 wait_queue_head_t cr_reply_waitq;
816 /** reply callback parameter */
817 struct ptlrpc_cb_id cr_reply_cbid;
818 /** Async completion handler, called when reply is received */
819 ptlrpc_interpterer_t cr_reply_interp;
820 /** Resend handler, called when request is resend to update RPC data */
821 ptlrpc_resend_cb_t cr_resend_cb;
822 /** Async completion context */
823 union ptlrpc_async_args cr_async_args;
824 /** Opaq data for replay and commit callbacks. */
826 /** Link to the imp->imp_unreplied_list */
827 struct list_head cr_unreplied_list;
829 * Commit callback, called when request is committed and about to be
832 void (*cr_commit_cb)(struct ptlrpc_request *);
833 /** Replay callback, called after request is replayed at recovery */
834 void (*cr_replay_cb)(struct ptlrpc_request *);
837 /** client request member alias */
838 /* NB: these alias should NOT be used by any new code, instead they should
839 * be removed step by step to avoid potential abuse */
840 #define rq_bulk rq_cli.cr_bulk
841 #define rq_delay_limit rq_cli.cr_delay_limit
842 #define rq_queued_time rq_cli.cr_queued_time
843 #define rq_sent_ns rq_cli.cr_sent_ns
844 #define rq_real_sent rq_cli.cr_sent_out
845 #define rq_reply_deadline rq_cli.cr_reply_deadline
846 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
847 #define rq_req_deadline rq_cli.cr_req_deadline
848 #define rq_nr_resend rq_cli.cr_resend_nr
849 #define rq_request_portal rq_cli.cr_req_ptl
850 #define rq_reply_portal rq_cli.cr_rep_ptl
851 #define rq_import_generation rq_cli.cr_imp_gen
852 #define rq_send_state rq_cli.cr_send_state
853 #define rq_set_chain rq_cli.cr_set_chain
854 #define rq_ctx_chain rq_cli.cr_ctx_chain
855 #define rq_set rq_cli.cr_set
856 #define rq_set_waitq rq_cli.cr_set_waitq
857 #define rq_cli_ctx rq_cli.cr_cli_ctx
858 #define rq_req_md_h rq_cli.cr_req_md_h
859 #define rq_req_cbid rq_cli.cr_req_cbid
860 #define rq_reply_md_h rq_cli.cr_reply_md_h
861 #define rq_reply_waitq rq_cli.cr_reply_waitq
862 #define rq_reply_cbid rq_cli.cr_reply_cbid
863 #define rq_interpret_reply rq_cli.cr_reply_interp
864 #define rq_resend_cb rq_cli.cr_resend_cb
865 #define rq_async_args rq_cli.cr_async_args
866 #define rq_cb_data rq_cli.cr_cb_data
867 #define rq_unreplied_list rq_cli.cr_unreplied_list
868 #define rq_commit_cb rq_cli.cr_commit_cb
869 #define rq_replay_cb rq_cli.cr_replay_cb
871 struct ptlrpc_srv_req {
872 /** initial thread servicing this request */
873 struct ptlrpc_thread *sr_svc_thread;
875 * Server side list of incoming unserved requests sorted by arrival
876 * time. Traversed from time to time to notice about to expire
877 * requests and sent back "early replies" to clients to let them
878 * know server is alive and well, just very busy to service their
881 struct list_head sr_timed_list;
882 /** server-side per-export list */
883 struct list_head sr_exp_list;
884 /** server-side history, used for debuging purposes. */
885 struct list_head sr_hist_list;
886 /** history sequence # */
888 /** the index of service's srv_at_array into which request is linked */
892 /** authed uid mapped to */
893 uid_t sr_auth_mapped_uid;
894 /** RPC is generated from what part of Lustre */
895 enum lustre_sec_part sr_sp_from;
896 /** request session context */
897 struct lu_context sr_ses;
901 /** stub for NRS request */
902 struct ptlrpc_nrs_request sr_nrq;
904 /** request arrival time */
905 struct timespec64 sr_arrival_time;
906 /** server's half ctx */
907 struct ptlrpc_svc_ctx *sr_svc_ctx;
908 /** (server side), pointed directly into req buffer */
909 struct ptlrpc_user_desc *sr_user_desc;
910 /** separated reply state, may be vmalloc'd */
911 struct ptlrpc_reply_state *sr_reply_state;
912 /** server-side hp handlers */
913 struct ptlrpc_hpreq_ops *sr_ops;
914 /** incoming request buffer */
915 struct ptlrpc_request_buffer_desc *sr_rqbd;
918 /** server request member alias */
919 /* NB: these alias should NOT be used by any new code, instead they should
920 * be removed step by step to avoid potential abuse */
921 #define rq_svc_thread rq_srv.sr_svc_thread
922 #define rq_timed_list rq_srv.sr_timed_list
923 #define rq_exp_list rq_srv.sr_exp_list
924 #define rq_history_list rq_srv.sr_hist_list
925 #define rq_history_seq rq_srv.sr_hist_seq
926 #define rq_at_index rq_srv.sr_at_index
927 #define rq_auth_uid rq_srv.sr_auth_uid
928 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
929 #define rq_sp_from rq_srv.sr_sp_from
930 #define rq_session rq_srv.sr_ses
931 #define rq_nrq rq_srv.sr_nrq
932 #define rq_arrival_time rq_srv.sr_arrival_time
933 #define rq_reply_state rq_srv.sr_reply_state
934 #define rq_svc_ctx rq_srv.sr_svc_ctx
935 #define rq_user_desc rq_srv.sr_user_desc
936 #define rq_ops rq_srv.sr_ops
937 #define rq_rqbd rq_srv.sr_rqbd
940 * Represents remote procedure call.
942 * This is a staple structure used by everybody wanting to send a request
945 struct ptlrpc_request {
946 /* Request type: one of PTL_RPC_MSG_* */
948 /** Result of request processing */
951 * Linkage item through which this request is included into
952 * sending/delayed lists on client and into rqbd list on server
954 struct list_head rq_list;
955 /** Lock to protect request flags and some other important bits, like
959 spinlock_t rq_early_free_lock;
960 /** client-side flags are serialized by rq_lock @{ */
961 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
962 rq_timedout:1, rq_resend:1, rq_restart:1,
964 * when ->rq_replay is set, request is kept by the client even
965 * after server commits corresponding transaction. This is
966 * used for operations that require sequence of multiple
967 * requests to be replayed. The only example currently is file
968 * open/close. When last request in such a sequence is
969 * committed, ->rq_replay is cleared on all requests in the
973 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
974 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
976 rq_req_unlinked:1, /* unlinked request buffer from lnet */
977 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
978 rq_memalloc:1, /* req originated from "kswapd" */
980 rq_reply_truncated:1,
981 /** whether the "rq_set" is a valid one */
984 /** do not resend request on -EINPROGRESS */
985 rq_no_retry_einprogress:1,
986 /* allow the req to be sent if the import is in recovery
989 /* bulk request, sent to server, but uncommitted */
991 rq_early_free_repbuf:1, /* free reply buffer in advance */
995 /** server-side flags @{ */
997 rq_hp:1, /**< high priority RPC */
998 rq_at_linked:1, /**< link into service's srv_at_array */
999 rq_packed_final:1; /**< packed final reply */
1002 /** one of RQ_PHASE_* */
1003 enum rq_phase rq_phase;
1004 /** one of RQ_PHASE_* to be used next */
1005 enum rq_phase rq_next_phase;
1007 * client-side refcount for SENT race, server-side refcounf
1008 * for multiple replies
1010 atomic_t rq_refcount;
1013 * !rq_truncate : # reply bytes actually received,
1014 * rq_truncate : required repbuf_len for resend
1016 int rq_nob_received;
1017 /** Request length */
1021 /** Pool if request is from preallocated list */
1022 struct ptlrpc_request_pool *rq_pool;
1023 /** Request message - what client sent */
1024 struct lustre_msg *rq_reqmsg;
1025 /** Reply message - server response */
1026 struct lustre_msg *rq_repmsg;
1027 /** Transaction number */
1031 /** bulk match bits */
1034 * List item to for replay list. Not yet committed requests get linked
1036 * Also see \a rq_replay comment above.
1037 * It's also link chain on obd_export::exp_req_replay_queue
1039 struct list_head rq_replay_list;
1040 /** non-shared members for client & server request*/
1042 struct ptlrpc_cli_req rq_cli;
1043 struct ptlrpc_srv_req rq_srv;
1046 * security and encryption data
1048 /** description of flavors for client & server */
1049 struct sptlrpc_flavor rq_flvr;
1051 /* client/server security flags */
1053 rq_ctx_init:1, /* context initiation */
1054 rq_ctx_fini:1, /* context destroy */
1055 rq_bulk_read:1, /* request bulk read */
1056 rq_bulk_write:1, /* request bulk write */
1057 /* server authentication flags */
1058 rq_auth_gss:1, /* authenticated by gss */
1059 rq_auth_usr_root:1, /* authed as root */
1060 rq_auth_usr_mdt:1, /* authed as mdt */
1061 rq_auth_usr_ost:1, /* authed as ost */
1062 /* security tfm flags */
1065 /* doesn't expect reply FIXME */
1067 rq_pill_init:1, /* pill initialized */
1068 rq_srv_req:1; /* server request */
1071 /** various buffer pointers */
1072 struct lustre_msg *rq_reqbuf; /**< req wrapper, vmalloc*/
1073 char *rq_repbuf; /**< rep buffer, vmalloc */
1074 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
1075 /** only in priv mode */
1076 struct lustre_msg *rq_clrbuf;
1077 int rq_reqbuf_len; /* req wrapper buf len */
1078 int rq_reqdata_len; /* req wrapper msg len */
1079 int rq_repbuf_len; /* rep buffer len */
1080 int rq_repdata_len; /* rep wrapper msg len */
1081 int rq_clrbuf_len; /* only in priv mode */
1082 int rq_clrdata_len; /* only in priv mode */
1084 /** early replies go to offset 0, regular replies go after that */
1085 unsigned int rq_reply_off;
1088 /** Fields that help to see if request and reply were swabbed or not */
1089 __u32 rq_req_swab_mask;
1090 __u32 rq_rep_swab_mask;
1092 /** how many early replies (for stats) */
1094 /** Server-side, export on which request was received */
1095 struct obd_export *rq_export;
1096 /** import where request is being sent */
1097 struct obd_import *rq_import;
1100 /** Peer description (the other side) */
1101 struct lnet_process_id rq_peer;
1102 /** Descriptor for the NID from which the peer sent the request. */
1103 struct lnet_process_id rq_source;
1105 * service time estimate (secs)
1106 * If the request is not served by this time, it is marked as timed out.
1107 * Do not change to time64_t since this is transmitted over the wire.
1111 * when request/reply sent (secs), or time when request should be sent
1114 /** when request must finish. */
1115 time64_t rq_deadline;
1116 /** request format description */
1117 struct req_capsule rq_pill;
1121 * Call completion handler for rpc if any, return it's status or original
1122 * rc if there was no handler defined for this request.
1124 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1125 struct ptlrpc_request *req, int rc)
1127 if (req->rq_interpret_reply != NULL) {
1128 req->rq_status = req->rq_interpret_reply(env, req,
1129 &req->rq_async_args,
1131 return req->rq_status;
1140 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
1141 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
1142 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1143 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1144 struct ptlrpc_nrs_pol_info *info);
1147 * Can the request be moved from the regular NRS head to the high-priority NRS
1148 * head (of the same PTLRPC service partition), if any?
1150 * For a reliable result, this should be checked under svcpt->scp_req lock.
1152 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1154 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1157 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1158 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1159 * to make sure it has not been scheduled yet (analogous to previous
1160 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1162 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1167 * Returns true if request buffer at offset \a index was already swabbed
1169 static inline bool lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
1171 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1172 return req->rq_req_swab_mask & (1 << index);
1176 * Returns true if request reply buffer at offset \a index was already swabbed
1178 static inline bool lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
1180 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1181 return req->rq_rep_swab_mask & (1 << index);
1185 * Returns true if request needs to be swabbed into local cpu byteorder
1187 static inline bool ptlrpc_req_need_swab(struct ptlrpc_request *req)
1189 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1193 * Returns true if request reply needs to be swabbed into local cpu byteorder
1195 static inline bool ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1197 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1201 * Mark request buffer at offset \a index that it was already swabbed
1203 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
1206 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1207 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1208 req->rq_req_swab_mask |= 1 << index;
1212 * Mark request reply buffer at offset \a index that it was already swabbed
1214 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
1217 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1218 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1219 req->rq_rep_swab_mask |= 1 << index;
1223 * Convert numerical request phase value \a phase into text string description
1225 static inline const char *
1226 ptlrpc_phase2str(enum rq_phase phase)
1235 case RQ_PHASE_INTERPRET:
1237 case RQ_PHASE_COMPLETE:
1239 case RQ_PHASE_UNREG_RPC:
1241 case RQ_PHASE_UNREG_BULK:
1249 * Convert numerical request phase of the request \a req into text stringi
1252 static inline const char *
1253 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1255 return ptlrpc_phase2str(req->rq_phase);
1259 * Debugging functions and helpers to print request structure into debug log
1262 /* Spare the preprocessor, spoil the bugs. */
1263 #define FLAG(field, str) (field ? str : "")
1265 /** Convert bit flags into a string */
1266 #define DEBUG_REQ_FLAGS(req) \
1267 ptlrpc_rqphase2str(req), \
1268 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1269 FLAG(req->rq_err, "E"), FLAG(req->rq_net_err, "e"), \
1270 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1271 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1272 FLAG(req->rq_no_resend, "N"), \
1273 FLAG(req->rq_waiting, "W"), \
1274 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1275 FLAG(req->rq_committed, "M")
1277 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s%s"
1279 void _debug_req(struct ptlrpc_request *req,
1280 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1281 __attribute__ ((format (printf, 3, 4)));
1284 * Helper that decides if we need to print request accordig to current debug
1287 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1289 CFS_CHECK_STACK(msgdata, mask, cdls); \
1291 if (((mask) & D_CANTMASK) != 0 || \
1292 ((libcfs_debug & (mask)) != 0 && \
1293 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1294 _debug_req((req), msgdata, fmt, ##a); \
1298 * This is the debug print function you need to use to print request sturucture
1299 * content into lustre debug log.
1300 * for most callers (level is a constant) this is resolved at compile time */
1301 #define DEBUG_REQ(level, req, fmt, args...) \
1303 if ((level) & (D_ERROR | D_WARNING)) { \
1304 static struct cfs_debug_limit_state cdls; \
1305 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1306 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1308 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1309 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1315 * Structure that defines a single page of a bulk transfer
1317 struct ptlrpc_bulk_page {
1318 /** Linkage to list of pages in a bulk */
1319 struct list_head bp_link;
1321 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1324 /** offset within a page */
1326 /** The page itself */
1327 struct page *bp_page;
1330 enum ptlrpc_bulk_op_type {
1331 PTLRPC_BULK_OP_ACTIVE = 0x00000001,
1332 PTLRPC_BULK_OP_PASSIVE = 0x00000002,
1333 PTLRPC_BULK_OP_PUT = 0x00000004,
1334 PTLRPC_BULK_OP_GET = 0x00000008,
1335 PTLRPC_BULK_BUF_KVEC = 0x00000010,
1336 PTLRPC_BULK_BUF_KIOV = 0x00000020,
1337 PTLRPC_BULK_GET_SOURCE = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_GET,
1338 PTLRPC_BULK_PUT_SINK = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_PUT,
1339 PTLRPC_BULK_GET_SINK = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_GET,
1340 PTLRPC_BULK_PUT_SOURCE = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_PUT,
1343 static inline bool ptlrpc_is_bulk_op_get(enum ptlrpc_bulk_op_type type)
1345 return (type & PTLRPC_BULK_OP_GET) == PTLRPC_BULK_OP_GET;
1348 static inline bool ptlrpc_is_bulk_get_source(enum ptlrpc_bulk_op_type type)
1350 return (type & PTLRPC_BULK_GET_SOURCE) == PTLRPC_BULK_GET_SOURCE;
1353 static inline bool ptlrpc_is_bulk_put_sink(enum ptlrpc_bulk_op_type type)
1355 return (type & PTLRPC_BULK_PUT_SINK) == PTLRPC_BULK_PUT_SINK;
1358 static inline bool ptlrpc_is_bulk_get_sink(enum ptlrpc_bulk_op_type type)
1360 return (type & PTLRPC_BULK_GET_SINK) == PTLRPC_BULK_GET_SINK;
1363 static inline bool ptlrpc_is_bulk_put_source(enum ptlrpc_bulk_op_type type)
1365 return (type & PTLRPC_BULK_PUT_SOURCE) == PTLRPC_BULK_PUT_SOURCE;
1368 static inline bool ptlrpc_is_bulk_desc_kvec(enum ptlrpc_bulk_op_type type)
1370 return ((type & PTLRPC_BULK_BUF_KVEC) | (type & PTLRPC_BULK_BUF_KIOV))
1371 == PTLRPC_BULK_BUF_KVEC;
1374 static inline bool ptlrpc_is_bulk_desc_kiov(enum ptlrpc_bulk_op_type type)
1376 return ((type & PTLRPC_BULK_BUF_KVEC) | (type & PTLRPC_BULK_BUF_KIOV))
1377 == PTLRPC_BULK_BUF_KIOV;
1380 static inline bool ptlrpc_is_bulk_op_active(enum ptlrpc_bulk_op_type type)
1382 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1383 (type & PTLRPC_BULK_OP_PASSIVE))
1384 == PTLRPC_BULK_OP_ACTIVE;
1387 static inline bool ptlrpc_is_bulk_op_passive(enum ptlrpc_bulk_op_type type)
1389 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1390 (type & PTLRPC_BULK_OP_PASSIVE))
1391 == PTLRPC_BULK_OP_PASSIVE;
1394 struct ptlrpc_bulk_frag_ops {
1396 * Add a page \a page to the bulk descriptor \a desc
1397 * Data to transfer in the page starts at offset \a pageoffset and
1398 * amount of data to transfer from the page is \a len
1400 void (*add_kiov_frag)(struct ptlrpc_bulk_desc *desc,
1401 struct page *page, int pageoffset, int len);
1404 * Add a \a fragment to the bulk descriptor \a desc.
1405 * Data to transfer in the fragment is pointed to by \a frag
1406 * The size of the fragment is \a len
1408 int (*add_iov_frag)(struct ptlrpc_bulk_desc *desc, void *frag, int len);
1411 * Uninitialize and free bulk descriptor \a desc.
1412 * Works on bulk descriptors both from server and client side.
1414 void (*release_frags)(struct ptlrpc_bulk_desc *desc);
1417 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops;
1418 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops;
1419 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops;
1422 * Definition of bulk descriptor.
1423 * Bulks are special "Two phase" RPCs where initial request message
1424 * is sent first and it is followed bt a transfer (o receiving) of a large
1425 * amount of data to be settled into pages referenced from the bulk descriptors.
1426 * Bulks transfers (the actual data following the small requests) are done
1427 * on separate LNet portals.
1428 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1429 * Another user is readpage for MDT.
1431 struct ptlrpc_bulk_desc {
1432 /** completed with failure */
1433 unsigned long bd_failure:1;
1435 unsigned long bd_registered:1;
1436 /** For serialization with callback */
1438 /** Import generation when request for this bulk was sent */
1439 int bd_import_generation;
1440 /** {put,get}{source,sink}{kvec,kiov} */
1441 enum ptlrpc_bulk_op_type bd_type;
1442 /** LNet portal for this bulk */
1444 /** Server side - export this bulk created for */
1445 struct obd_export *bd_export;
1446 /** Client side - import this bulk was sent on */
1447 struct obd_import *bd_import;
1448 /** Back pointer to the request */
1449 struct ptlrpc_request *bd_req;
1450 struct ptlrpc_bulk_frag_ops *bd_frag_ops;
1451 wait_queue_head_t bd_waitq; /* server side only WQ */
1452 int bd_iov_count; /* # entries in bd_iov */
1453 int bd_max_iov; /* allocated size of bd_iov */
1454 int bd_nob; /* # bytes covered */
1455 int bd_nob_transferred; /* # bytes GOT/PUT */
1457 __u64 bd_last_mbits;
1459 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1460 lnet_nid_t bd_sender; /* stash event::sender */
1461 int bd_md_count; /* # valid entries in bd_mds */
1462 int bd_md_max_brw; /* max entries in bd_mds */
1463 /** array of associated MDs */
1464 struct lnet_handle_md bd_mds[PTLRPC_BULK_OPS_COUNT];
1469 * encrypt iov, size is either 0 or bd_iov_count.
1471 lnet_kiov_t *bd_enc_vec;
1472 lnet_kiov_t *bd_vec;
1476 struct kvec *bd_enc_kvec;
1477 struct kvec *bd_kvec;
1483 #define GET_KIOV(desc) ((desc)->bd_u.bd_kiov.bd_vec)
1484 #define BD_GET_KIOV(desc, i) ((desc)->bd_u.bd_kiov.bd_vec[i])
1485 #define GET_ENC_KIOV(desc) ((desc)->bd_u.bd_kiov.bd_enc_vec)
1486 #define BD_GET_ENC_KIOV(desc, i) ((desc)->bd_u.bd_kiov.bd_enc_vec[i])
1487 #define GET_KVEC(desc) ((desc)->bd_u.bd_kvec.bd_kvec)
1488 #define BD_GET_KVEC(desc, i) ((desc)->bd_u.bd_kvec.bd_kvec[i])
1489 #define GET_ENC_KVEC(desc) ((desc)->bd_u.bd_kvec.bd_enc_kvec)
1490 #define BD_GET_ENC_KVEC(desc, i) ((desc)->bd_u.bd_kvec.bd_enc_kvec[i])
1494 SVC_STOPPED = 1 << 0,
1495 SVC_STOPPING = 1 << 1,
1496 SVC_STARTING = 1 << 2,
1497 SVC_RUNNING = 1 << 3,
1499 SVC_SIGNAL = 1 << 5,
1502 #define PTLRPC_THR_NAME_LEN 32
1504 * Definition of server service thread structure
1506 struct ptlrpc_thread {
1508 * List of active threads in svc->srv_threads
1510 struct list_head t_link;
1512 * thread-private data (preallocated vmalloc'd memory)
1517 * service thread index, from ptlrpc_start_threads
1521 * service thread pid
1525 * put watchdog in the structure per thread b=14840
1527 struct lc_watchdog *t_watchdog;
1529 * the svc this thread belonged to b=18582
1531 struct ptlrpc_service_part *t_svcpt;
1532 wait_queue_head_t t_ctl_waitq;
1533 struct lu_env *t_env;
1534 char t_name[PTLRPC_THR_NAME_LEN];
1537 static inline int thread_is_init(struct ptlrpc_thread *thread)
1539 return thread->t_flags == 0;
1542 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1544 return !!(thread->t_flags & SVC_STOPPED);
1547 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1549 return !!(thread->t_flags & SVC_STOPPING);
1552 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1554 return !!(thread->t_flags & SVC_STARTING);
1557 static inline int thread_is_running(struct ptlrpc_thread *thread)
1559 return !!(thread->t_flags & SVC_RUNNING);
1562 static inline int thread_is_event(struct ptlrpc_thread *thread)
1564 return !!(thread->t_flags & SVC_EVENT);
1567 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1569 return !!(thread->t_flags & SVC_SIGNAL);
1572 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1574 thread->t_flags &= ~flags;
1577 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1579 thread->t_flags = flags;
1582 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1584 thread->t_flags |= flags;
1587 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1590 if (thread->t_flags & flags) {
1591 thread->t_flags &= ~flags;
1598 * Request buffer descriptor structure.
1599 * This is a structure that contains one posted request buffer for service.
1600 * Once data land into a buffer, event callback creates actual request and
1601 * notifies wakes one of the service threads to process new incoming request.
1602 * More than one request can fit into the buffer.
1604 struct ptlrpc_request_buffer_desc {
1605 /** Link item for rqbds on a service */
1606 struct list_head rqbd_list;
1607 /** History of requests for this buffer */
1608 struct list_head rqbd_reqs;
1609 /** Back pointer to service for which this buffer is registered */
1610 struct ptlrpc_service_part *rqbd_svcpt;
1611 /** LNet descriptor */
1612 struct lnet_handle_md rqbd_md_h;
1614 /** The buffer itself */
1616 struct ptlrpc_cb_id rqbd_cbid;
1618 * This "embedded" request structure is only used for the
1619 * last request to fit into the buffer
1621 struct ptlrpc_request rqbd_req;
1624 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1626 struct ptlrpc_service_ops {
1628 * if non-NULL called during thread creation (ptlrpc_start_thread())
1629 * to initialize service specific per-thread state.
1631 int (*so_thr_init)(struct ptlrpc_thread *thr);
1633 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1634 * destruct state created by ->srv_init().
1636 void (*so_thr_done)(struct ptlrpc_thread *thr);
1638 * Handler function for incoming requests for this service
1640 int (*so_req_handler)(struct ptlrpc_request *req);
1642 * function to determine priority of the request, it's called
1643 * on every new request
1645 int (*so_hpreq_handler)(struct ptlrpc_request *);
1647 * service-specific print fn
1649 void (*so_req_printer)(void *, struct ptlrpc_request *);
1652 #ifndef __cfs_cacheline_aligned
1653 /* NB: put it here for reducing patche dependence */
1654 # define __cfs_cacheline_aligned
1658 * How many high priority requests to serve before serving one normal
1661 #define PTLRPC_SVC_HP_RATIO 10
1664 * Definition of PortalRPC service.
1665 * The service is listening on a particular portal (like tcp port)
1666 * and perform actions for a specific server like IO service for OST
1667 * or general metadata service for MDS.
1669 struct ptlrpc_service {
1670 /** serialize /proc operations */
1671 spinlock_t srv_lock;
1672 /** most often accessed fields */
1673 /** chain thru all services */
1674 struct list_head srv_list;
1675 /** service operations table */
1676 struct ptlrpc_service_ops srv_ops;
1677 /** only statically allocated strings here; we don't clean them */
1679 /** only statically allocated strings here; we don't clean them */
1680 char *srv_thread_name;
1681 /** service thread list */
1682 struct list_head srv_threads;
1683 /** threads # should be created for each partition on initializing */
1684 int srv_nthrs_cpt_init;
1685 /** limit of threads number for each partition */
1686 int srv_nthrs_cpt_limit;
1687 /** Root of debugfs dir tree for this service */
1688 struct dentry *srv_debugfs_entry;
1689 /** Pointer to statistic data for this service */
1690 struct lprocfs_stats *srv_stats;
1691 /** # hp per lp reqs to handle */
1692 int srv_hpreq_ratio;
1693 /** biggest request to receive */
1694 int srv_max_req_size;
1695 /** biggest reply to send */
1696 int srv_max_reply_size;
1697 /** size of individual buffers */
1699 /** # buffers to allocate in 1 group */
1700 int srv_nbuf_per_group;
1701 /** Local portal on which to receive requests */
1702 __u32 srv_req_portal;
1703 /** Portal on the client to send replies to */
1704 __u32 srv_rep_portal;
1706 * Tags for lu_context associated with this thread, see struct
1710 /** soft watchdog timeout multiplier */
1711 int srv_watchdog_factor;
1712 /** under unregister_service */
1713 unsigned srv_is_stopping:1;
1714 /** Whether or not to restrict service threads to CPUs in this CPT */
1715 unsigned srv_cpt_bind:1;
1717 /** max # request buffers */
1719 /** max # request buffers in history per partition */
1720 int srv_hist_nrqbds_cpt_max;
1721 /** number of CPTs this service associated with */
1723 /** CPTs array this service associated with */
1725 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1727 /** CPT table this service is running over */
1728 struct cfs_cpt_table *srv_cptable;
1731 struct kobject srv_kobj;
1732 struct completion srv_kobj_unregister;
1734 * partition data for ptlrpc service
1736 struct ptlrpc_service_part *srv_parts[0];
1740 * Definition of PortalRPC service partition data.
1741 * Although a service only has one instance of it right now, but we
1742 * will have multiple instances very soon (instance per CPT).
1744 * it has four locks:
1746 * serialize operations on rqbd and requests waiting for preprocess
1748 * serialize operations active requests sent to this portal
1750 * serialize adaptive timeout stuff
1752 * serialize operations on RS list (reply states)
1754 * We don't have any use-case to take two or more locks at the same time
1755 * for now, so there is no lock order issue.
1757 struct ptlrpc_service_part {
1758 /** back reference to owner */
1759 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1760 /* CPT id, reserved */
1762 /** always increasing number */
1764 /** # of starting threads */
1765 int scp_nthrs_starting;
1766 /** # of stopping threads, reserved for shrinking threads */
1767 int scp_nthrs_stopping;
1768 /** # running threads */
1769 int scp_nthrs_running;
1770 /** service threads list */
1771 struct list_head scp_threads;
1774 * serialize the following fields, used for protecting
1775 * rqbd list and incoming requests waiting for preprocess,
1776 * threads starting & stopping are also protected by this lock.
1778 spinlock_t scp_lock __cfs_cacheline_aligned;
1779 /** userland serialization */
1780 struct mutex scp_mutex;
1781 /** total # req buffer descs allocated */
1782 int scp_nrqbds_total;
1783 /** # posted request buffers for receiving */
1784 int scp_nrqbds_posted;
1785 /** in progress of allocating rqbd */
1786 int scp_rqbd_allocating;
1787 /** # incoming reqs */
1788 int scp_nreqs_incoming;
1789 /** request buffers to be reposted */
1790 struct list_head scp_rqbd_idle;
1791 /** req buffers receiving */
1792 struct list_head scp_rqbd_posted;
1793 /** incoming reqs */
1794 struct list_head scp_req_incoming;
1795 /** timeout before re-posting reqs, in jiffies */
1796 long scp_rqbd_timeout;
1798 * all threads sleep on this. This wait-queue is signalled when new
1799 * incoming request arrives and when difficult reply has to be handled.
1801 wait_queue_head_t scp_waitq;
1803 /** request history */
1804 struct list_head scp_hist_reqs;
1805 /** request buffer history */
1806 struct list_head scp_hist_rqbds;
1807 /** # request buffers in history */
1808 int scp_hist_nrqbds;
1809 /** sequence number for request */
1811 /** highest seq culled from history */
1812 __u64 scp_hist_seq_culled;
1815 * serialize the following fields, used for processing requests
1816 * sent to this portal
1818 spinlock_t scp_req_lock __cfs_cacheline_aligned;
1819 /** # reqs in either of the NRS heads below */
1820 /** # reqs being served */
1821 int scp_nreqs_active;
1822 /** # HPreqs being served */
1823 int scp_nhreqs_active;
1824 /** # hp requests handled */
1827 /** NRS head for regular requests */
1828 struct ptlrpc_nrs scp_nrs_reg;
1829 /** NRS head for HP requests; this is only valid for services that can
1830 * handle HP requests */
1831 struct ptlrpc_nrs *scp_nrs_hp;
1836 * serialize the following fields, used for changes on
1839 spinlock_t scp_at_lock __cfs_cacheline_aligned;
1840 /** estimated rpc service time */
1841 struct adaptive_timeout scp_at_estimate;
1842 /** reqs waiting for replies */
1843 struct ptlrpc_at_array scp_at_array;
1844 /** early reply timer */
1845 struct timer_list scp_at_timer;
1847 ktime_t scp_at_checktime;
1848 /** check early replies */
1849 unsigned scp_at_check;
1853 * serialize the following fields, used for processing
1854 * replies for this portal
1856 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1857 /** all the active replies */
1858 struct list_head scp_rep_active;
1859 /** List of free reply_states */
1860 struct list_head scp_rep_idle;
1861 /** waitq to run, when adding stuff to srv_free_rs_list */
1862 wait_queue_head_t scp_rep_waitq;
1863 /** # 'difficult' replies */
1864 atomic_t scp_nreps_difficult;
1867 #define ptlrpc_service_for_each_part(part, i, svc) \
1869 i < (svc)->srv_ncpts && \
1870 (svc)->srv_parts != NULL && \
1871 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1874 * Declaration of ptlrpcd control structure
1876 struct ptlrpcd_ctl {
1878 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1880 unsigned long pc_flags;
1882 * Thread lock protecting structure fields.
1888 struct completion pc_starting;
1892 struct completion pc_finishing;
1894 * Thread requests set.
1896 struct ptlrpc_request_set *pc_set;
1898 * Thread name used in kthread_run()
1902 * CPT the thread is bound on.
1906 * Index of ptlrpcd thread in the array.
1910 * Pointer to the array of partners' ptlrpcd_ctl structure.
1912 struct ptlrpcd_ctl **pc_partners;
1914 * Number of the ptlrpcd's partners.
1918 * Record the partner index to be processed next.
1922 * Error code if the thread failed to fully start.
1927 /* Bits for pc_flags */
1928 enum ptlrpcd_ctl_flags {
1930 * Ptlrpc thread start flag.
1932 LIOD_START = 1 << 0,
1934 * Ptlrpc thread stop flag.
1938 * Ptlrpc thread force flag (only stop force so far).
1939 * This will cause aborting any inflight rpcs handled
1940 * by thread if LIOD_STOP is specified.
1942 LIOD_FORCE = 1 << 2,
1944 * This is a recovery ptlrpc thread.
1946 LIOD_RECOVERY = 1 << 3,
1953 * Service compatibility function; the policy is compatible with all services.
1955 * \param[in] svc The service the policy is attempting to register with.
1956 * \param[in] desc The policy descriptor
1958 * \retval true The policy is compatible with the service
1960 * \see ptlrpc_nrs_pol_desc::pd_compat()
1962 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
1963 const struct ptlrpc_nrs_pol_desc *desc)
1969 * Service compatibility function; the policy is compatible with only a specific
1970 * service which is identified by its human-readable name at
1971 * ptlrpc_service::srv_name.
1973 * \param[in] svc The service the policy is attempting to register with.
1974 * \param[in] desc The policy descriptor
1976 * \retval false The policy is not compatible with the service
1977 * \retval true The policy is compatible with the service
1979 * \see ptlrpc_nrs_pol_desc::pd_compat()
1981 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
1982 const struct ptlrpc_nrs_pol_desc *desc)
1984 LASSERT(desc->pd_compat_svc_name != NULL);
1985 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
1990 /* ptlrpc/events.c */
1991 extern struct lnet_handle_eq ptlrpc_eq_h;
1992 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1993 struct lnet_process_id *peer, lnet_nid_t *self);
1995 * These callbacks are invoked by LNet when something happened to
1999 extern void request_out_callback(struct lnet_event *ev);
2000 extern void reply_in_callback(struct lnet_event *ev);
2001 extern void client_bulk_callback(struct lnet_event *ev);
2002 extern void request_in_callback(struct lnet_event *ev);
2003 extern void reply_out_callback(struct lnet_event *ev);
2004 #ifdef HAVE_SERVER_SUPPORT
2005 extern void server_bulk_callback(struct lnet_event *ev);
2009 /* ptlrpc/connection.c */
2010 struct ptlrpc_connection *ptlrpc_connection_get(struct lnet_process_id peer,
2012 struct obd_uuid *uuid);
2013 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2014 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2015 int ptlrpc_connection_init(void);
2016 void ptlrpc_connection_fini(void);
2017 extern lnet_pid_t ptl_get_pid(void);
2020 * Check if the peer connection is on the local node. We need to use GFP_NOFS
2021 * for requests from a local client to avoid recursing into the filesystem
2022 * as we might end up waiting on a page sent in the request we're serving.
2024 * Use __GFP_HIGHMEM so that the pages can use all of the available memory
2025 * on 32-bit machines. Use more aggressive GFP_HIGHUSER flags from non-local
2026 * clients to be able to generate more memory pressure on the OSS and allow
2027 * inactive pages to be reclaimed, since it doesn't have any other processes
2028 * or allocations that generate memory reclaim pressure.
2030 * See b=17576 (bdf50dc9) and b=19529 (3dcf18d3) for details.
2032 static inline bool ptlrpc_connection_is_local(struct ptlrpc_connection *conn)
2037 if (conn->c_peer.nid == conn->c_self)
2040 RETURN(LNetIsPeerLocal(conn->c_peer.nid));
2043 /* ptlrpc/niobuf.c */
2045 * Actual interfacing with LNet to put/get/register/unregister stuff
2048 #ifdef HAVE_SERVER_SUPPORT
2049 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2050 unsigned nfrags, unsigned max_brw,
2053 const struct ptlrpc_bulk_frag_ops
2055 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2056 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2058 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2062 LASSERT(desc != NULL);
2064 spin_lock(&desc->bd_lock);
2065 rc = desc->bd_md_count;
2066 spin_unlock(&desc->bd_lock);
2071 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2072 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2074 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2076 struct ptlrpc_bulk_desc *desc;
2079 LASSERT(req != NULL);
2080 desc = req->rq_bulk;
2085 if (req->rq_bulk_deadline > ktime_get_real_seconds())
2089 spin_lock(&desc->bd_lock);
2090 rc = desc->bd_md_count;
2091 spin_unlock(&desc->bd_lock);
2095 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2096 #define PTLRPC_REPLY_EARLY 0x02
2097 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2098 int ptlrpc_reply(struct ptlrpc_request *req);
2099 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2100 int ptlrpc_error(struct ptlrpc_request *req);
2101 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2102 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2103 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2106 /* ptlrpc/client.c */
2108 * Client-side portals API. Everything to send requests, receive replies,
2109 * request queues, request management, etc.
2112 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2114 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2115 struct ptlrpc_client *);
2116 void ptlrpc_cleanup_client(struct obd_import *imp);
2117 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
2118 lnet_nid_t nid4refnet);
2120 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2121 int ptlrpc_replay_req(struct ptlrpc_request *req);
2122 void ptlrpc_restart_req(struct ptlrpc_request *req);
2123 void ptlrpc_abort_inflight(struct obd_import *imp);
2124 void ptlrpc_cleanup_imp(struct obd_import *imp);
2125 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2127 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2128 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2130 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2131 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *);
2132 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2133 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2134 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2136 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2137 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2139 struct ptlrpc_request_pool *
2140 ptlrpc_init_rq_pool(int, int,
2141 int (*populate_pool)(struct ptlrpc_request_pool *, int));
2143 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2144 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2145 const struct req_format *format);
2146 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2147 struct ptlrpc_request_pool *,
2148 const struct req_format *format);
2149 void ptlrpc_request_free(struct ptlrpc_request *request);
2150 int ptlrpc_request_pack(struct ptlrpc_request *request,
2151 __u32 version, int opcode);
2152 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2153 const struct req_format *format,
2154 __u32 version, int opcode);
2155 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2156 __u32 version, int opcode, char **bufs,
2157 struct ptlrpc_cli_ctx *ctx);
2158 void ptlrpc_req_finished(struct ptlrpc_request *request);
2159 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2160 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2161 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2162 unsigned nfrags, unsigned max_brw,
2165 const struct ptlrpc_bulk_frag_ops
2168 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
2169 void *frag, int len);
2170 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2171 struct page *page, int pageoffset, int len,
2173 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2174 struct page *page, int pageoffset,
2177 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2180 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2181 struct page *page, int pageoffset,
2184 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2187 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
2189 static inline void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
2193 for (i = 0; i < desc->bd_iov_count ; i++)
2194 put_page(BD_GET_KIOV(desc, i).kiov_page);
2197 static inline void ptlrpc_release_bulk_noop(struct ptlrpc_bulk_desc *desc)
2201 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2202 struct obd_import *imp);
2203 __u64 ptlrpc_next_xid(void);
2204 __u64 ptlrpc_sample_next_xid(void);
2205 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2207 /* Set of routines to run a function in ptlrpcd context */
2208 void *ptlrpcd_alloc_work(struct obd_import *imp,
2209 int (*cb)(const struct lu_env *, void *), void *data);
2210 void ptlrpcd_destroy_work(void *handler);
2211 int ptlrpcd_queue_work(void *handler);
2214 struct ptlrpc_service_buf_conf {
2215 /* nbufs is buffers # to allocate when growing the pool */
2216 unsigned int bc_nbufs;
2217 /* buffer size to post */
2218 unsigned int bc_buf_size;
2219 /* portal to listed for requests on */
2220 unsigned int bc_req_portal;
2221 /* portal of where to send replies to */
2222 unsigned int bc_rep_portal;
2223 /* maximum request size to be accepted for this service */
2224 unsigned int bc_req_max_size;
2225 /* maximum reply size this service can ever send */
2226 unsigned int bc_rep_max_size;
2229 struct ptlrpc_service_thr_conf {
2230 /* threadname should be 8 characters or less - 6 will be added on */
2232 /* threads increasing factor for each CPU */
2233 unsigned int tc_thr_factor;
2234 /* service threads # to start on each partition while initializing */
2235 unsigned int tc_nthrs_init;
2237 * low water of threads # upper-limit on each partition while running,
2238 * service availability may be impacted if threads number is lower
2239 * than this value. It can be ZERO if the service doesn't require
2240 * CPU affinity or there is only one partition.
2242 unsigned int tc_nthrs_base;
2243 /* "soft" limit for total threads number */
2244 unsigned int tc_nthrs_max;
2245 /* user specified threads number, it will be validated due to
2246 * other members of this structure. */
2247 unsigned int tc_nthrs_user;
2248 /* bind service threads to only CPUs in their associated CPT */
2249 unsigned int tc_cpu_bind;
2250 /* Tags for lu_context associated with service thread */
2254 struct ptlrpc_service_cpt_conf {
2255 struct cfs_cpt_table *cc_cptable;
2256 /* string pattern to describe CPTs for a service */
2258 /* whether or not to have per-CPT service partitions */
2262 struct ptlrpc_service_conf {
2265 /* soft watchdog timeout multiplifier to print stuck service traces */
2266 unsigned int psc_watchdog_factor;
2267 /* buffer information */
2268 struct ptlrpc_service_buf_conf psc_buf;
2269 /* thread information */
2270 struct ptlrpc_service_thr_conf psc_thr;
2271 /* CPU partition information */
2272 struct ptlrpc_service_cpt_conf psc_cpt;
2273 /* function table */
2274 struct ptlrpc_service_ops psc_ops;
2277 /* ptlrpc/service.c */
2279 * Server-side services API. Register/unregister service, request state
2280 * management, service thread management
2284 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
2285 int mode, bool no_ack, bool convert_lock);
2286 void ptlrpc_commit_replies(struct obd_export *exp);
2287 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2288 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2289 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2290 struct ptlrpc_service *ptlrpc_register_service(
2291 struct ptlrpc_service_conf *conf,
2292 struct kset *parent,
2293 struct dentry *debugfs_entry);
2294 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2296 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2297 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2298 int ptlrpc_service_health_check(struct ptlrpc_service *);
2299 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2300 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2301 struct obd_export *export);
2302 void ptlrpc_update_export_timer(struct obd_export *exp,
2303 time64_t extra_delay);
2305 int ptlrpc_hr_init(void);
2306 void ptlrpc_hr_fini(void);
2310 /* ptlrpc/import.c */
2315 int ptlrpc_connect_import(struct obd_import *imp);
2316 int ptlrpc_init_import(struct obd_import *imp);
2317 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2318 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp);
2319 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2320 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2322 void ptlrpc_import_enter_resend(struct obd_import *imp);
2323 /* ptlrpc/pack_generic.c */
2324 int ptlrpc_reconnect_import(struct obd_import *imp);
2328 * ptlrpc msg buffer and swab interface
2332 bool ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2334 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2336 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2337 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2339 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2340 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2342 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2343 __u32 *lens, char **bufs);
2344 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2346 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2347 __u32 *lens, char **bufs, int flags);
2348 #define LPRFL_EARLY_REPLY 1
2349 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2350 char **bufs, int flags);
2351 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2352 unsigned int newlen, int move_data);
2353 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2354 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2355 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
2356 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2357 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
2358 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
2359 __u32 lustre_msg_early_size(void);
2360 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
2361 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
2362 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
2363 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
2364 __u32 lustre_msg_bufcount(struct lustre_msg *m);
2365 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
2366 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2367 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2368 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2369 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
2370 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
2371 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
2372 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2373 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
2374 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2375 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2376 enum lustre_msg_version lustre_msg_get_version(struct lustre_msg *msg);
2377 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
2378 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2379 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2380 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
2381 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2382 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2383 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2384 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2385 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2386 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2387 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2388 int lustre_msg_get_status(struct lustre_msg *msg);
2389 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2390 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2391 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2392 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2393 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2394 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2395 __u64 lustre_msg_get_mbits(struct lustre_msg *msg);
2396 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2397 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2398 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2399 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2400 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2401 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
2402 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2403 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2404 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2405 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2406 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2407 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2408 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2409 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2410 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2411 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2412 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2413 void lustre_msg_set_mbits(struct lustre_msg *msg, __u64 mbits);
2416 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2417 unsigned int newlen, int move_data)
2419 LASSERT(req->rq_reply_state);
2420 LASSERT(req->rq_repmsg);
2421 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2425 #ifdef LUSTRE_TRANSLATE_ERRNOS
2427 static inline int ptlrpc_status_hton(int h)
2430 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2431 * ELDLM_LOCK_ABORTED, etc.
2434 return -lustre_errno_hton(-h);
2439 static inline int ptlrpc_status_ntoh(int n)
2442 * See the comment in ptlrpc_status_hton().
2445 return -lustre_errno_ntoh(-n);
2452 #define ptlrpc_status_hton(h) (h)
2453 #define ptlrpc_status_ntoh(n) (n)
2458 /** Change request phase of \a req to \a new_phase */
2460 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2462 if (req->rq_phase == new_phase)
2465 if (new_phase == RQ_PHASE_UNREG_RPC ||
2466 new_phase == RQ_PHASE_UNREG_BULK) {
2467 /* No embedded unregistering phases */
2468 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2469 req->rq_phase == RQ_PHASE_UNREG_BULK)
2472 req->rq_next_phase = req->rq_phase;
2474 atomic_inc(&req->rq_import->imp_unregistering);
2477 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2478 req->rq_phase == RQ_PHASE_UNREG_BULK) {
2480 atomic_dec(&req->rq_import->imp_unregistering);
2483 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2484 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2486 req->rq_phase = new_phase;
2490 * Returns true if request \a req got early reply and hard deadline is not met
2493 ptlrpc_client_early(struct ptlrpc_request *req)
2495 return req->rq_early;
2499 * Returns true if we got real reply from server for this request
2502 ptlrpc_client_replied(struct ptlrpc_request *req)
2504 if (req->rq_reply_deadline > ktime_get_real_seconds())
2506 return req->rq_replied;
2509 /** Returns true if request \a req is in process of receiving server reply */
2511 ptlrpc_client_recv(struct ptlrpc_request *req)
2513 if (req->rq_reply_deadline > ktime_get_real_seconds())
2515 return req->rq_receiving_reply;
2519 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2523 spin_lock(&req->rq_lock);
2524 if (req->rq_reply_deadline > ktime_get_real_seconds()) {
2525 spin_unlock(&req->rq_lock);
2528 if (req->rq_req_deadline > ktime_get_real_seconds()) {
2529 spin_unlock(&req->rq_lock);
2533 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
2534 req->rq_receiving_reply;
2535 spin_unlock(&req->rq_lock);
2540 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2543 if (req->rq_set == NULL)
2544 wake_up(&req->rq_reply_waitq);
2546 wake_up(&req->rq_set->set_waitq);
2550 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2552 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2553 atomic_inc(&rs->rs_refcount);
2557 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2559 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2560 if (atomic_dec_and_test(&rs->rs_refcount))
2561 lustre_free_reply_state(rs);
2564 /* Should only be called once per req */
2565 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2567 if (req->rq_reply_state == NULL)
2568 return; /* shouldn't occur */
2569 ptlrpc_rs_decref(req->rq_reply_state);
2570 req->rq_reply_state = NULL;
2571 req->rq_repmsg = NULL;
2574 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2576 return lustre_msg_get_magic(req->rq_reqmsg);
2579 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2581 switch (req->rq_reqmsg->lm_magic) {
2582 case LUSTRE_MSG_MAGIC_V2:
2583 return req->rq_reqmsg->lm_repsize;
2585 LASSERTF(0, "incorrect message magic: %08x\n",
2586 req->rq_reqmsg->lm_magic);
2591 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2593 if (req->rq_delay_limit != 0 &&
2594 req->rq_queued_time + req->rq_delay_limit < ktime_get_seconds())
2599 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2601 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2602 spin_lock(&req->rq_lock);
2603 req->rq_no_resend = 1;
2604 spin_unlock(&req->rq_lock);
2606 return req->rq_no_resend;
2610 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2612 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2614 return svcpt->scp_service->srv_watchdog_factor *
2615 max_t(int, at, obd_timeout);
2618 static inline struct ptlrpc_service *
2619 ptlrpc_req2svc(struct ptlrpc_request *req)
2621 LASSERT(req->rq_rqbd != NULL);
2622 return req->rq_rqbd->rqbd_svcpt->scp_service;
2625 /* ldlm/ldlm_lib.c */
2627 * Target client logic
2630 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2631 int client_obd_cleanup(struct obd_device *obddev);
2632 int client_connect_import(const struct lu_env *env,
2633 struct obd_export **exp, struct obd_device *obd,
2634 struct obd_uuid *cluuid, struct obd_connect_data *,
2636 int client_disconnect_export(struct obd_export *exp);
2637 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2639 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2640 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2641 struct obd_uuid *uuid);
2642 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2643 void client_destroy_import(struct obd_import *imp);
2646 #ifdef HAVE_SERVER_SUPPORT
2647 int server_disconnect_export(struct obd_export *exp);
2650 /* ptlrpc/pinger.c */
2652 * Pinger API (client side only)
2655 enum timeout_event {
2658 struct timeout_item;
2659 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2660 int ptlrpc_pinger_add_import(struct obd_import *imp);
2661 int ptlrpc_pinger_del_import(struct obd_import *imp);
2662 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
2663 timeout_cb_t cb, void *data,
2664 struct list_head *obd_list);
2665 int ptlrpc_del_timeout_client(struct list_head *obd_list,
2666 enum timeout_event event);
2667 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2668 int ptlrpc_obd_ping(struct obd_device *obd);
2669 void ping_evictor_start(void);
2670 void ping_evictor_stop(void);
2671 void ptlrpc_pinger_ir_up(void);
2672 void ptlrpc_pinger_ir_down(void);
2674 int ptlrpc_pinger_suppress_pings(void);
2676 /* ptlrpc/ptlrpcd.c */
2677 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2678 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2679 void ptlrpcd_wake(struct ptlrpc_request *req);
2680 void ptlrpcd_add_req(struct ptlrpc_request *req);
2681 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2682 int ptlrpcd_addref(void);
2683 void ptlrpcd_decref(void);
2685 /* ptlrpc/lproc_ptlrpc.c */
2687 * procfs output related functions
2690 const char* ll_opcode2str(__u32 opcode);
2691 const int ll_str2opcode(const char *ops);
2692 #ifdef CONFIG_PROC_FS
2693 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2694 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2695 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2697 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2698 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2699 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2703 /* ptlrpc/llog_server.c */
2704 int llog_origin_handle_open(struct ptlrpc_request *req);
2705 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2706 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2707 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2709 /* ptlrpc/llog_client.c */
2710 extern struct llog_operations llog_client_ops;