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/rhashtable.h>
55 #include <linux/uio.h>
56 #include <libcfs/libcfs.h>
58 #include <lnet/lib-types.h>
59 #include <uapi/linux/lnet/nidstr.h>
60 #include <uapi/linux/lustre/lustre_idl.h>
61 #include <lustre_ha.h>
62 #include <lustre_sec.h>
63 #include <lustre_import.h>
64 #include <lprocfs_status.h>
65 #include <lu_object.h>
66 #include <lustre_req_layout.h>
67 #include <obd_support.h>
68 #include <uapi/linux/lustre/lustre_ver.h>
70 /* MD flags we _always_ use */
71 #define PTLRPC_MD_OPTIONS 0
74 * log2 max # of bulk operations in one request: 2=4MB/RPC, 5=32MB/RPC, ...
75 * In order for the client and server to properly negotiate the maximum
76 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
77 * value. The client is free to limit the actual RPC size for any bulk
78 * transfer via cl_max_pages_per_rpc to some non-power-of-two value.
79 * NOTE: This is limited to 16 (=64GB RPCs) by IOOBJ_MAX_BRW_BITS. */
80 #define PTLRPC_BULK_OPS_BITS 6
81 #if PTLRPC_BULK_OPS_BITS > 16
82 #error "More than 65536 BRW RPCs not allowed by IOOBJ_MAX_BRW_BITS."
84 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
86 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
87 * should not be used on the server at all. Otherwise, it imposes a
88 * protocol limitation on the maximum RPC size that can be used by any
89 * RPC sent to that server in the future. Instead, the server should
90 * use the negotiated per-client ocd_brw_size to determine the bulk
92 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
95 * Define maxima for bulk I/O.
97 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
98 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
99 * currently supported maximum between peers at connect via ocd_brw_size.
101 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
102 #define PTLRPC_MAX_BRW_SIZE (1U << PTLRPC_MAX_BRW_BITS)
103 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_SHIFT)
105 #define ONE_MB_BRW_SIZE (1U << LNET_MTU_BITS)
106 #define MD_MAX_BRW_SIZE (1U << LNET_MTU_BITS)
107 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_SHIFT)
108 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
109 #define DT_DEF_BRW_SIZE (4 * ONE_MB_BRW_SIZE)
110 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_SHIFT)
111 #define OFD_MAX_BRW_SIZE (1U << LNET_MTU_BITS)
113 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
114 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
115 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
117 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_SIZE))
118 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_SIZE"
120 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
121 # error "PTLRPC_MAX_BRW_SIZE too big"
123 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
124 # error "PTLRPC_MAX_BRW_PAGES too big"
127 #define PTLRPC_NTHRS_INIT 2
132 * Constants determine how memory is used to buffer incoming service requests.
134 * ?_NBUFS # buffers to allocate when growing the pool
135 * ?_BUFSIZE # bytes in a single request buffer
136 * ?_MAXREQSIZE # maximum request service will receive
138 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
139 * of ?_NBUFS is added to the pool.
141 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
142 * considered full when less than ?_MAXREQSIZE is left in them.
147 * Constants determine how threads are created for ptlrpc service.
149 * ?_NTHRS_INIT # threads to create for each service partition on
150 * initializing. If it's non-affinity service and
151 * there is only one partition, it's the overall #
152 * threads for the service while initializing.
153 * ?_NTHRS_BASE # threads should be created at least for each
154 * ptlrpc partition to keep the service healthy.
155 * It's the low-water mark of threads upper-limit
156 * for each partition.
157 * ?_THR_FACTOR # threads can be added on threads upper-limit for
158 * each CPU core. This factor is only for reference,
159 * we might decrease value of factor if number of cores
160 * per CPT is above a limit.
161 * ?_NTHRS_MAX # overall threads can be created for a service,
162 * it's a soft limit because if service is running
163 * on machine with hundreds of cores and tens of
164 * CPU partitions, we need to guarantee each partition
165 * has ?_NTHRS_BASE threads, which means total threads
166 * will be ?_NTHRS_BASE * number_of_cpts which can
167 * exceed ?_NTHRS_MAX.
171 * #define MDS_NTHRS_INIT 2
172 * #define MDS_NTHRS_BASE 64
173 * #define MDS_NTHRS_FACTOR 8
174 * #define MDS_NTHRS_MAX 1024
177 * ---------------------------------------------------------------------
178 * Server(A) has 16 cores, user configured it to 4 partitions so each
179 * partition has 4 cores, then actual number of service threads on each
181 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
183 * Total number of threads for the service is:
184 * 96 * partitions(4) = 384
187 * ---------------------------------------------------------------------
188 * Server(B) has 32 cores, user configured it to 4 partitions so each
189 * partition has 8 cores, then actual number of service threads on each
191 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
193 * Total number of threads for the service is:
194 * 128 * partitions(4) = 512
197 * ---------------------------------------------------------------------
198 * Server(B) has 96 cores, user configured it to 8 partitions so each
199 * partition has 12 cores, then actual number of service threads on each
201 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
203 * Total number of threads for the service is:
204 * 160 * partitions(8) = 1280
206 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
207 * as upper limit of threads number for each partition:
208 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
211 * ---------------------------------------------------------------------
212 * Server(C) have a thousand of cores and user configured it to 32 partitions
213 * MDS_NTHRS_BASE(64) * 32 = 2048
215 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
216 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
217 * to keep service healthy, so total number of threads will just be 2048.
219 * NB: we don't suggest to choose server with that many cores because backend
220 * filesystem itself, buffer cache, or underlying network stack might
221 * have some SMP scalability issues at that large scale.
223 * If user already has a fat machine with hundreds or thousands of cores,
224 * there are two choices for configuration:
225 * a) create CPU table from subset of all CPUs and run Lustre on
227 * b) bind service threads on a few partitions, see modparameters of
228 * MDS and OSS for details
230 * NB: these calculations (and examples below) are simplified to help
231 * understanding, the real implementation is a little more complex,
232 * please see ptlrpc_server_nthreads_check() for details.
237 * LDLM threads constants:
239 * Given 8 as factor and 24 as base threads number
242 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
245 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
246 * threads for each partition and total threads number will be 112.
249 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
250 * threads for each partition to keep service healthy, so total threads
251 * number should be 24 * 8 = 192.
253 * So with these constants, threads number will be at the similar level
254 * of old versions, unless target machine has over a hundred cores
256 #define LDLM_THR_FACTOR 8
257 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
258 #define LDLM_NTHRS_BASE 24
259 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
261 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
262 #define LDLM_CLIENT_NBUFS 1
263 #define LDLM_SERVER_NBUFS 64
264 #define LDLM_BUFSIZE (8 * 1024)
265 #define LDLM_MAXREQSIZE (5 * 1024)
266 #define LDLM_MAXREPSIZE (1024)
269 * MDS threads constants:
271 * Please see examples in "Thread Constants", MDS threads number will be at
272 * the comparable level of old versions, unless the server has many cores.
274 #ifndef MDS_MAX_THREADS
275 #define MDS_MAX_THREADS 1024
276 #define MDS_MAX_OTHR_THREADS 256
278 #else /* MDS_MAX_THREADS */
279 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
280 #undef MDS_MAX_THREADS
281 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
283 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
286 /* default service */
287 #define MDS_THR_FACTOR 8
288 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
289 #define MDS_NTHRS_MAX MDS_MAX_THREADS
290 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
292 /* read-page service */
293 #define MDS_RDPG_THR_FACTOR 4
294 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
295 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
296 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
298 /* these should be removed when we remove setattr service in the future */
299 #define MDS_SETA_THR_FACTOR 4
300 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
301 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
302 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
304 /* non-affinity threads */
305 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
306 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
311 * Assume file name length = FNAME_MAX = 256 (true for ext3).
312 * path name length = PATH_MAX = 4096
313 * LOV MD size max = EA_MAX = 24 * 2000
314 * (NB: 24 is size of lov_ost_data)
315 * LOV LOGCOOKIE size max = 32 * 2000
316 * (NB: 32 is size of llog_cookie)
317 * symlink: FNAME_MAX + PATH_MAX <- largest
318 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
319 * rename: FNAME_MAX + FNAME_MAX
320 * open: FNAME_MAX + EA_MAX
322 * MDS_MAXREQSIZE ~= 4736 bytes =
323 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
324 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
326 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
327 * except in the open case where there are a large number of OSTs in a LOV.
329 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
330 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
333 * MDS incoming request with LOV EA
334 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
336 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
337 362 + LOV_MAX_STRIPE_COUNT * 24)
339 * MDS outgoing reply with LOV EA
341 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
342 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
344 * but 2.4 or later MDS will never send reply with llog_cookie to any
345 * version client. This macro is defined for server side reply buffer size.
347 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
350 * This is the size of a maximum REINT_SETXATTR request:
352 * lustre_msg 56 (32 + 4 x 5 + 4)
354 * mdt_rec_setxattr 136
356 * name 256 (XATTR_NAME_MAX)
357 * value 65536 (XATTR_SIZE_MAX)
359 #define MDS_EA_MAXREQSIZE 66288
362 * These are the maximum request and reply sizes (rounded up to 1 KB
363 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
365 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
366 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
367 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
370 * The update request includes all of updates from the create, which might
371 * include linkea (4K maxim), together with other updates, we set it to 1000K:
372 * lustre_msg + ptlrpc_body + OUT_UPDATE_BUFFER_SIZE_MAX
374 #define OUT_MAXREQSIZE (1000 * 1024)
375 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
377 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
378 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
382 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
383 * However, we need to allocate a much larger buffer for it because LNet
384 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
385 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
386 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
387 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
388 * utilization is very low.
390 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
391 * reused until all requests fit in it have been processed and released,
392 * which means one long blocked request can prevent the rqbd be reused.
393 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
394 * from LNet with unused 65 KB, buffer utilization will be about 59%.
395 * Please check LU-2432 for details.
397 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
401 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
402 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
403 * extra bytes to each request buffer to improve buffer utilization rate.
405 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
408 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
409 #define FLD_MAXREQSIZE (160)
411 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
412 #define FLD_MAXREPSIZE (152)
413 #define FLD_BUFSIZE (1 << 12)
416 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
418 #define SEQ_MAXREQSIZE (160)
420 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
421 #define SEQ_MAXREPSIZE (152)
422 #define SEQ_BUFSIZE (1 << 12)
424 /** MGS threads must be >= 3, see bug 22458 comment #28 */
425 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
426 #define MGS_NTHRS_MAX 32
429 #define MGS_BUFSIZE (8 * 1024)
430 #define MGS_MAXREQSIZE (7 * 1024)
431 #define MGS_MAXREPSIZE (9 * 1024)
434 * OSS threads constants:
436 * Given 8 as factor and 64 as base threads number
439 * On 8-core server configured to 2 partitions, we will have
440 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
443 * On 32-core machine configured to 4 partitions, we will have
444 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
445 * will be 112 * 4 = 448.
448 * On 64-core machine configured to 4 partitions, we will have
449 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
450 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
451 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
452 * for each partition.
454 * So we can see that with these constants, threads number wil be at the
455 * similar level of old versions, unless the server has many cores.
457 /* depress threads factor for VM with small memory size */
458 #define OSS_THR_FACTOR min_t(int, 8, \
459 NUM_CACHEPAGES >> (28 - PAGE_SHIFT))
460 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
461 #define OSS_NTHRS_BASE 64
463 /* threads for handling "create" request */
464 #define OSS_CR_THR_FACTOR 1
465 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
466 #define OSS_CR_NTHRS_BASE 8
467 #define OSS_CR_NTHRS_MAX 64
470 * OST_IO_MAXREQSIZE ~=
471 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
472 * DT_MAX_BRW_PAGES * niobuf_remote
474 * - single object with 16 pages is 512 bytes
475 * - OST_IO_MAXREQSIZE must be at least 1 niobuf per page of data
476 * - Must be a multiple of 1024
477 * - should allow a reasonably large SHORT_IO_BYTES size (64KB)
479 #define _OST_MAXREQSIZE_BASE ((unsigned long)(sizeof(struct lustre_msg) + \
480 /* lm_buflens */ sizeof(__u32) * 4 + \
481 sizeof(struct ptlrpc_body) + \
482 sizeof(struct obdo) + \
483 sizeof(struct obd_ioobj) + \
484 sizeof(struct niobuf_remote)))
485 #define _OST_MAXREQSIZE_SUM ((unsigned long)(_OST_MAXREQSIZE_BASE + \
486 sizeof(struct niobuf_remote) * \
489 * FIEMAP request can be 4K+ for now
491 #define OST_MAXREQSIZE (16UL * 1024UL)
492 #define OST_IO_MAXREQSIZE max(OST_MAXREQSIZE, \
493 ((_OST_MAXREQSIZE_SUM - 1) | \
495 /* Safe estimate of free space in standard RPC, provides upper limit for # of
496 * bytes of i/o to pack in RPC (skipping bulk transfer). */
497 #define OST_MAX_SHORT_IO_BYTES ((OST_IO_MAXREQSIZE - _OST_MAXREQSIZE_BASE) & \
500 /* Actual size used for short i/o buffer. Calculation means this:
501 * At least one page (for large PAGE_SIZE), or 16 KiB, but not more
502 * than the available space aligned to a page boundary. */
503 #define OBD_DEF_SHORT_IO_BYTES min(max(PAGE_SIZE, 16UL * 1024UL), \
504 OST_MAX_SHORT_IO_BYTES)
506 #define OST_MAXREPSIZE (9 * 1024)
507 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
510 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
511 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 32 * 1024)
513 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
514 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
516 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
518 /* Macro to hide a typecast and BUILD_BUG. */
519 #define ptlrpc_req_async_args(_var, req) ({ \
520 BUILD_BUG_ON(sizeof(*_var) > sizeof(req->rq_async_args)); \
521 (typeof(_var))&req->rq_async_args; \
524 struct ptlrpc_replay_async_args {
530 * Structure to single define portal connection.
532 struct ptlrpc_connection {
533 /** linkage for connections hash table */
534 struct rhash_head c_hash;
535 /** Our own lnet nid for this connection */
537 /** Remote side nid for this connection */
538 struct lnet_process_id c_peer;
539 /** UUID of the other side */
540 struct obd_uuid c_remote_uuid;
541 /** reference counter for this connection */
545 /** Client definition for PortalRPC */
546 struct ptlrpc_client {
547 /** What lnet portal does this client send messages to by default */
548 __u32 cli_request_portal;
549 /** What portal do we expect replies on */
550 __u32 cli_reply_portal;
551 /** Name of the client */
552 const char *cli_name;
555 /** state flags of requests */
556 /* XXX only ones left are those used by the bulk descs as well! */
557 #define PTL_RPC_FL_INTR BIT(0) /* reply wait was interrupted by user */
558 #define PTL_RPC_FL_TIMEOUT BIT(7) /* request timed out waiting for reply */
560 #define REQ_MAX_ACK_LOCKS 8
562 union ptlrpc_async_args {
564 * Scratchpad for passing args to completion interpreter. Users
565 * cast to the struct of their choosing, and BUILD_BUG_ON that this is
566 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
567 * a pointer to it here. The pointer_arg ensures this struct is at
568 * least big enough for that.
570 void *pointer_arg[11];
574 struct ptlrpc_request_set;
575 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
578 * Definition of request set structure.
579 * Request set is a list of requests (not necessary to the same target) that
580 * once populated with RPCs could be sent in parallel.
581 * There are two kinds of request sets. General purpose and with dedicated
582 * serving thread. Example of the latter is ptlrpcd set.
583 * For general purpose sets once request set started sending it is impossible
584 * to add new requests to such set.
585 * Provides a way to call "completion callbacks" when all requests in the set
588 struct ptlrpc_request_set {
589 atomic_t set_refcount;
590 /** number of in queue requests */
591 atomic_t set_new_count;
592 /** number of uncompleted requests */
593 atomic_t set_remaining;
594 /** wait queue to wait on for request events */
595 wait_queue_head_t set_waitq;
596 /** List of requests in the set */
597 struct list_head set_requests;
599 * Lock for \a set_new_requests manipulations
600 * locked so that any old caller can communicate requests to
601 * the set holder who can then fold them into the lock-free set
603 spinlock_t set_new_req_lock;
604 /** List of new yet unsent requests. Only used with ptlrpcd now. */
605 struct list_head set_new_requests;
607 /** rq_status of requests that have been freed already */
609 /** Additional fields used by the flow control extension */
610 /** Maximum number of RPCs in flight */
611 int set_max_inflight;
612 /** Callback function used to generate RPCs */
613 set_producer_func set_producer;
614 /** opaq argument passed to the producer callback */
615 void *set_producer_arg;
616 unsigned int set_allow_intr:1;
619 struct ptlrpc_bulk_desc;
620 struct ptlrpc_service_part;
621 struct ptlrpc_service;
624 * ptlrpc callback & work item stuff
626 struct ptlrpc_cb_id {
627 void (*cbid_fn)(struct lnet_event *ev); /* specific callback fn */
628 void *cbid_arg; /* additional arg */
631 /** Maximum number of locks to fit into reply state */
632 #define RS_MAX_LOCKS 8
636 * Structure to define reply state on the server
637 * Reply state holds various reply message information. Also for "difficult"
638 * replies (rep-ack case) we store the state after sending reply and wait
639 * for the client to acknowledge the reception. In these cases locks could be
640 * added to the state for replay/failover consistency guarantees.
642 struct ptlrpc_reply_state {
643 /** Callback description */
644 struct ptlrpc_cb_id rs_cb_id;
645 /** Linkage for list of all reply states in a system */
646 struct list_head rs_list;
647 /** Linkage for list of all reply states on same export */
648 struct list_head rs_exp_list;
649 /** Linkage for list of all reply states for same obd */
650 struct list_head rs_obd_list;
652 struct list_head rs_debug_list;
654 /** A spinlock to protect the reply state flags */
656 /** Reply state flags */
657 unsigned long rs_difficult:1; /* ACK/commit stuff */
658 unsigned long rs_no_ack:1; /* no ACK, even for
659 difficult requests */
660 unsigned long rs_scheduled:1; /* being handled? */
661 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
662 unsigned long rs_handled:1; /* been handled yet? */
663 unsigned long rs_on_net:1; /* reply_out_callback pending? */
664 unsigned long rs_prealloc:1; /* rs from prealloc list */
665 unsigned long rs_committed:1;/* the transaction was committed
666 and the rs was dispatched
667 by ptlrpc_commit_replies */
668 unsigned long rs_convert_lock:1; /* need to convert saved
669 * locks to COS mode */
670 atomic_t rs_refcount; /* number of users */
671 /** Number of locks awaiting client ACK */
674 /** Size of the state */
678 /** Transaction number */
682 struct obd_export *rs_export;
683 struct ptlrpc_service_part *rs_svcpt;
684 /** Lnet metadata handle for the reply */
685 struct lnet_handle_md rs_md_h;
687 /** Context for the sevice thread */
688 struct ptlrpc_svc_ctx *rs_svc_ctx;
689 /** Reply buffer (actually sent to the client), encoded if needed */
690 struct lustre_msg *rs_repbuf; /* wrapper */
691 /** Size of the reply buffer */
692 int rs_repbuf_len; /* wrapper buf length */
693 /** Size of the reply message */
694 int rs_repdata_len; /* wrapper msg length */
696 * Actual reply message. Its content is encrupted (if needed) to
697 * produce reply buffer for actual sending. In simple case
698 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
700 struct lustre_msg *rs_msg; /* reply message */
702 /** Handles of locks awaiting client reply ACK */
703 struct lustre_handle rs_locks[RS_MAX_LOCKS];
704 /** Lock modes of locks in \a rs_locks */
705 enum ldlm_mode rs_modes[RS_MAX_LOCKS];
708 struct ptlrpc_thread;
712 RQ_PHASE_NEW = 0xebc0de00,
713 RQ_PHASE_RPC = 0xebc0de01,
714 RQ_PHASE_BULK = 0xebc0de02,
715 RQ_PHASE_INTERPRET = 0xebc0de03,
716 RQ_PHASE_COMPLETE = 0xebc0de04,
717 RQ_PHASE_UNREG_RPC = 0xebc0de05,
718 RQ_PHASE_UNREG_BULK = 0xebc0de06,
719 RQ_PHASE_UNDEFINED = 0xebc0de07
722 /** Type of request interpreter call-back */
723 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
724 struct ptlrpc_request *req,
726 /** Type of request resend call-back */
727 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
731 * Definition of request pool structure.
732 * The pool is used to store empty preallocated requests for the case
733 * when we would actually need to send something without performing
734 * any allocations (to avoid e.g. OOM).
736 struct ptlrpc_request_pool {
737 /** Locks the list */
739 /** list of ptlrpc_request structs */
740 struct list_head prp_req_list;
741 /** Maximum message size that would fit into a rquest from this pool */
743 /** Function to allocate more requests for this pool */
744 int (*prp_populate)(struct ptlrpc_request_pool *, int);
752 #include <lustre_nrs.h>
755 * Basic request prioritization operations structure.
756 * The whole idea is centered around locks and RPCs that might affect locks.
757 * When a lock is contended we try to give priority to RPCs that might lead
758 * to fastest release of that lock.
759 * Currently only implemented for OSTs only in a way that makes all
760 * IO and truncate RPCs that are coming from a locked region where a lock is
761 * contended a priority over other requests.
763 struct ptlrpc_hpreq_ops {
765 * Check if the lock handle of the given lock is the same as
766 * taken from the request.
768 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
770 * Check if the request is a high priority one.
772 int (*hpreq_check)(struct ptlrpc_request *);
774 * Called after the request has been handled.
776 void (*hpreq_fini)(struct ptlrpc_request *);
779 struct ptlrpc_cli_req {
780 /** For bulk requests on client only: bulk descriptor */
781 struct ptlrpc_bulk_desc *cr_bulk;
782 /** optional time limit for send attempts. This is a timeout
783 * not a timestamp so timeout_t (s32) is used instead of time64_t
785 timeout_t cr_delay_limit;
786 /** time request was first queued */
787 time64_t cr_queued_time;
788 /** request sent in nanoseconds */
790 /** time for request really sent out */
791 time64_t cr_sent_out;
792 /** when req reply unlink must finish. */
793 time64_t cr_reply_deadline;
794 /** when req bulk unlink must finish. */
795 time64_t cr_bulk_deadline;
796 /** when req unlink must finish. */
797 time64_t cr_req_deadline;
798 /** Portal to which this request would be sent */
800 /** Portal where to wait for reply and where reply would be sent */
802 /** request resending number */
803 unsigned int cr_resend_nr;
804 /** What was import generation when this request was sent */
806 enum lustre_imp_state cr_send_state;
807 /** Per-request waitq introduced by bug 21938 for recovery waiting */
808 wait_queue_head_t cr_set_waitq;
809 /** Link item for request set lists */
810 struct list_head cr_set_chain;
811 /** link to waited ctx */
812 struct list_head cr_ctx_chain;
814 /** client's half ctx */
815 struct ptlrpc_cli_ctx *cr_cli_ctx;
816 /** Link back to the request set */
817 struct ptlrpc_request_set *cr_set;
818 /** outgoing request MD handle */
819 struct lnet_handle_md cr_req_md_h;
820 /** request-out callback parameter */
821 struct ptlrpc_cb_id cr_req_cbid;
822 /** incoming reply MD handle */
823 struct lnet_handle_md cr_reply_md_h;
824 wait_queue_head_t cr_reply_waitq;
825 /** reply callback parameter */
826 struct ptlrpc_cb_id cr_reply_cbid;
827 /** Async completion handler, called when reply is received */
828 ptlrpc_interpterer_t cr_reply_interp;
829 /** Resend handler, called when request is resend to update RPC data */
830 ptlrpc_resend_cb_t cr_resend_cb;
831 /** Async completion context */
832 union ptlrpc_async_args cr_async_args;
833 /** Opaq data for replay and commit callbacks. */
835 /** Link to the imp->imp_unreplied_list */
836 struct list_head cr_unreplied_list;
838 * Commit callback, called when request is committed and about to be
841 void (*cr_commit_cb)(struct ptlrpc_request *);
842 /** Replay callback, called after request is replayed at recovery */
843 void (*cr_replay_cb)(struct ptlrpc_request *);
846 /** client request member alias */
847 /* NB: these alias should NOT be used by any new code, instead they should
848 * be removed step by step to avoid potential abuse */
849 #define rq_bulk rq_cli.cr_bulk
850 #define rq_delay_limit rq_cli.cr_delay_limit
851 #define rq_queued_time rq_cli.cr_queued_time
852 #define rq_sent_ns rq_cli.cr_sent_ns
853 #define rq_real_sent rq_cli.cr_sent_out
854 #define rq_reply_deadline rq_cli.cr_reply_deadline
855 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
856 #define rq_req_deadline rq_cli.cr_req_deadline
857 #define rq_nr_resend rq_cli.cr_resend_nr
858 #define rq_request_portal rq_cli.cr_req_ptl
859 #define rq_reply_portal rq_cli.cr_rep_ptl
860 #define rq_import_generation rq_cli.cr_imp_gen
861 #define rq_send_state rq_cli.cr_send_state
862 #define rq_set_chain rq_cli.cr_set_chain
863 #define rq_ctx_chain rq_cli.cr_ctx_chain
864 #define rq_set rq_cli.cr_set
865 #define rq_set_waitq rq_cli.cr_set_waitq
866 #define rq_cli_ctx rq_cli.cr_cli_ctx
867 #define rq_req_md_h rq_cli.cr_req_md_h
868 #define rq_req_cbid rq_cli.cr_req_cbid
869 #define rq_reply_md_h rq_cli.cr_reply_md_h
870 #define rq_reply_waitq rq_cli.cr_reply_waitq
871 #define rq_reply_cbid rq_cli.cr_reply_cbid
872 #define rq_interpret_reply rq_cli.cr_reply_interp
873 #define rq_resend_cb rq_cli.cr_resend_cb
874 #define rq_async_args rq_cli.cr_async_args
875 #define rq_cb_data rq_cli.cr_cb_data
876 #define rq_unreplied_list rq_cli.cr_unreplied_list
877 #define rq_commit_cb rq_cli.cr_commit_cb
878 #define rq_replay_cb rq_cli.cr_replay_cb
880 struct ptlrpc_srv_req {
881 /** initial thread servicing this request */
882 struct ptlrpc_thread *sr_svc_thread;
884 * Server side list of incoming unserved requests sorted by arrival
885 * time. Traversed from time to time to notice about to expire
886 * requests and sent back "early replies" to clients to let them
887 * know server is alive and well, just very busy to service their
890 struct list_head sr_timed_list;
891 /** server-side per-export list */
892 struct list_head sr_exp_list;
893 /** server-side history, used for debuging purposes. */
894 struct list_head sr_hist_list;
895 /** history sequence # */
897 /** the index of service's srv_at_array into which request is linked */
901 /** authed uid mapped to */
902 uid_t sr_auth_mapped_uid;
903 /** RPC is generated from what part of Lustre */
904 enum lustre_sec_part sr_sp_from;
905 /** request session context */
906 struct lu_context sr_ses;
910 /** stub for NRS request */
911 struct ptlrpc_nrs_request sr_nrq;
913 /** request arrival time */
914 struct timespec64 sr_arrival_time;
915 /** server's half ctx */
916 struct ptlrpc_svc_ctx *sr_svc_ctx;
917 /** (server side), pointed directly into req buffer */
918 struct ptlrpc_user_desc *sr_user_desc;
919 /** separated reply state, may be vmalloc'd */
920 struct ptlrpc_reply_state *sr_reply_state;
921 /** server-side hp handlers */
922 struct ptlrpc_hpreq_ops *sr_ops;
923 /** incoming request buffer */
924 struct ptlrpc_request_buffer_desc *sr_rqbd;
927 /** server request member alias */
928 /* NB: these alias should NOT be used by any new code, instead they should
929 * be removed step by step to avoid potential abuse */
930 #define rq_svc_thread rq_srv.sr_svc_thread
931 #define rq_timed_list rq_srv.sr_timed_list
932 #define rq_exp_list rq_srv.sr_exp_list
933 #define rq_history_list rq_srv.sr_hist_list
934 #define rq_history_seq rq_srv.sr_hist_seq
935 #define rq_at_index rq_srv.sr_at_index
936 #define rq_auth_uid rq_srv.sr_auth_uid
937 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
938 #define rq_sp_from rq_srv.sr_sp_from
939 #define rq_session rq_srv.sr_ses
940 #define rq_nrq rq_srv.sr_nrq
941 #define rq_arrival_time rq_srv.sr_arrival_time
942 #define rq_reply_state rq_srv.sr_reply_state
943 #define rq_svc_ctx rq_srv.sr_svc_ctx
944 #define rq_user_desc rq_srv.sr_user_desc
945 #define rq_ops rq_srv.sr_ops
946 #define rq_rqbd rq_srv.sr_rqbd
949 * Represents remote procedure call.
951 * This is a staple structure used by everybody wanting to send a request
954 struct ptlrpc_request {
955 /* Request type: one of PTL_RPC_MSG_* */
957 /** Result of request processing */
960 * Linkage item through which this request is included into
961 * sending/delayed lists on client and into rqbd list on server
963 struct list_head rq_list;
964 /** Lock to protect request flags and some other important bits, like
968 spinlock_t rq_early_free_lock;
969 /** client-side flags are serialized by rq_lock @{ */
970 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
971 rq_timedout:1, rq_resend:1, rq_restart:1,
973 * when ->rq_replay is set, request is kept by the client even
974 * after server commits corresponding transaction. This is
975 * used for operations that require sequence of multiple
976 * requests to be replayed. The only example currently is file
977 * open/close. When last request in such a sequence is
978 * committed, ->rq_replay is cleared on all requests in the
982 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
983 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
985 rq_req_unlinked:1, /* unlinked request buffer from lnet */
986 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
987 rq_memalloc:1, /* req originated from "kswapd" */
989 rq_reply_truncated:1,
990 /** whether the "rq_set" is a valid one */
993 /** do not resend request on -EINPROGRESS */
994 rq_no_retry_einprogress:1,
995 /* allow the req to be sent if the import is in recovery
998 /* bulk request, sent to server, but uncommitted */
1000 rq_early_free_repbuf:1, /* free reply buffer in advance */
1004 /** server-side flags @{ */
1006 rq_hp:1, /**< high priority RPC */
1007 rq_at_linked:1, /**< link into service's srv_at_array */
1008 rq_packed_final:1, /**< packed final reply */
1009 rq_obsolete:1; /* aborted by a signal on a client */
1012 /** one of RQ_PHASE_* */
1013 enum rq_phase rq_phase;
1014 /** one of RQ_PHASE_* to be used next */
1015 enum rq_phase rq_next_phase;
1017 * client-side refcount for SENT race, server-side refcounf
1018 * for multiple replies
1020 atomic_t rq_refcount;
1023 * !rq_truncate : # reply bytes actually received,
1024 * rq_truncate : required repbuf_len for resend
1026 int rq_nob_received;
1027 /** Request length */
1031 /** Pool if request is from preallocated list */
1032 struct ptlrpc_request_pool *rq_pool;
1033 /** Request message - what client sent */
1034 struct lustre_msg *rq_reqmsg;
1035 /** Reply message - server response */
1036 struct lustre_msg *rq_repmsg;
1037 /** Transaction number */
1041 /** bulk match bits */
1044 * List item to for replay list. Not yet committed requests get linked
1046 * Also see \a rq_replay comment above.
1047 * It's also link chain on obd_export::exp_req_replay_queue
1049 struct list_head rq_replay_list;
1050 /** non-shared members for client & server request*/
1052 struct ptlrpc_cli_req rq_cli;
1053 struct ptlrpc_srv_req rq_srv;
1056 * security and encryption data
1058 /** description of flavors for client & server */
1059 struct sptlrpc_flavor rq_flvr;
1062 * SELinux policy info at the time of the request
1063 * sepol string format is:
1064 * <mode>:<policy name>:<policy version>:<policy hash>
1066 char rq_sepol[LUSTRE_NODEMAP_SEPOL_LENGTH + 1];
1068 /* client/server security flags */
1070 rq_ctx_init:1, /* context initiation */
1071 rq_ctx_fini:1, /* context destroy */
1072 rq_bulk_read:1, /* request bulk read */
1073 rq_bulk_write:1, /* request bulk write */
1074 /* server authentication flags */
1075 rq_auth_gss:1, /* authenticated by gss */
1076 rq_auth_usr_root:1, /* authed as root */
1077 rq_auth_usr_mdt:1, /* authed as mdt */
1078 rq_auth_usr_ost:1, /* authed as ost */
1079 /* security tfm flags */
1082 /* doesn't expect reply FIXME */
1084 rq_pill_init:1, /* pill initialized */
1085 rq_srv_req:1; /* server request */
1088 /** various buffer pointers */
1089 struct lustre_msg *rq_reqbuf; /**< req wrapper, vmalloc*/
1090 char *rq_repbuf; /**< rep buffer, vmalloc */
1091 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
1092 /** only in priv mode */
1093 struct lustre_msg *rq_clrbuf;
1094 int rq_reqbuf_len; /* req wrapper buf len */
1095 int rq_reqdata_len; /* req wrapper msg len */
1096 int rq_repbuf_len; /* rep buffer len */
1097 int rq_repdata_len; /* rep wrapper msg len */
1098 int rq_clrbuf_len; /* only in priv mode */
1099 int rq_clrdata_len; /* only in priv mode */
1101 /** early replies go to offset 0, regular replies go after that */
1102 unsigned int rq_reply_off;
1105 /** Fields that help to see if request and reply were swabbed or not */
1106 __u32 rq_req_swab_mask;
1107 __u32 rq_rep_swab_mask;
1109 /** how many early replies (for stats) */
1111 /** Server-side, export on which request was received */
1112 struct obd_export *rq_export;
1113 /** import where request is being sent */
1114 struct obd_import *rq_import;
1117 /** Peer description (the other side) */
1118 struct lnet_process_id rq_peer;
1119 /** Descriptor for the NID from which the peer sent the request. */
1120 struct lnet_process_id rq_source;
1122 * service time estimate (secs)
1123 * If the request is not served by this time, it is marked as timed out.
1124 * Do not change to time64_t since this is transmitted over the wire.
1126 * The linux kernel handles timestamps with time64_t and timeouts
1127 * are normally done with jiffies. Lustre shares the rq_timeout between
1128 * nodes. Since jiffies can vary from node to node Lustre instead
1129 * will express the timeout value in seconds. To avoid confusion with
1130 * timestamps (time64_t) and jiffy timeouts (long) Lustre timeouts
1131 * are expressed in s32 (timeout_t). Also what is transmitted over
1132 * the wire is 32 bits.
1134 timeout_t rq_timeout;
1136 * when request/reply sent (secs), or time when request should be sent
1139 /** when request must finish. */
1140 time64_t rq_deadline;
1141 /** request format description */
1142 struct req_capsule rq_pill;
1146 * Call completion handler for rpc if any, return it's status or original
1147 * rc if there was no handler defined for this request.
1149 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1150 struct ptlrpc_request *req, int rc)
1152 if (req->rq_interpret_reply != NULL) {
1153 req->rq_status = req->rq_interpret_reply(env, req,
1154 &req->rq_async_args,
1156 return req->rq_status;
1165 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
1166 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
1167 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1168 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1169 struct ptlrpc_nrs_pol_info *info);
1172 * Can the request be moved from the regular NRS head to the high-priority NRS
1173 * head (of the same PTLRPC service partition), if any?
1175 * For a reliable result, this should be checked under svcpt->scp_req lock.
1177 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1179 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1182 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1183 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1184 * to make sure it has not been scheduled yet (analogous to previous
1185 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1187 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1192 * Returns true if request buffer at offset \a index was already swabbed
1194 static inline bool lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
1196 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1197 return req->rq_req_swab_mask & BIT(index);
1201 * Returns true if request reply buffer at offset \a index was already swabbed
1203 static inline bool lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
1205 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1206 return req->rq_rep_swab_mask & BIT(index);
1210 * Returns true if request needs to be swabbed into local cpu byteorder
1212 static inline bool ptlrpc_req_need_swab(struct ptlrpc_request *req)
1214 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1218 * Returns true if request reply needs to be swabbed into local cpu byteorder
1220 static inline bool ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1222 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1226 * Mark request buffer at offset \a index that it was already swabbed
1228 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
1231 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1232 LASSERT((req->rq_req_swab_mask & BIT(index)) == 0);
1233 req->rq_req_swab_mask |= BIT(index);
1237 * Mark request reply buffer at offset \a index that it was already swabbed
1239 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
1242 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1243 LASSERT((req->rq_rep_swab_mask & BIT(index)) == 0);
1244 req->rq_rep_swab_mask |= BIT(index);
1248 * Convert numerical request phase value \a phase into text string description
1250 static inline const char *
1251 ptlrpc_phase2str(enum rq_phase phase)
1260 case RQ_PHASE_INTERPRET:
1262 case RQ_PHASE_COMPLETE:
1264 case RQ_PHASE_UNREG_RPC:
1266 case RQ_PHASE_UNREG_BULK:
1274 * Convert numerical request phase of the request \a req into text stringi
1277 static inline const char *
1278 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1280 return ptlrpc_phase2str(req->rq_phase);
1284 * Debugging functions and helpers to print request structure into debug log
1287 /* Spare the preprocessor, spoil the bugs. */
1288 #define FLAG(field, str) (field ? str : "")
1290 /** Convert bit flags into a string */
1291 #define DEBUG_REQ_FLAGS(req) \
1292 ptlrpc_rqphase2str(req), \
1293 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1294 FLAG(req->rq_err, "E"), FLAG(req->rq_net_err, "e"), \
1295 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1296 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1297 FLAG(req->rq_no_resend, "N"), FLAG(req->rq_no_reply, "n"), \
1298 FLAG(req->rq_waiting, "W"), \
1299 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1300 FLAG(req->rq_committed, "M"), \
1301 FLAG(req->rq_req_unlinked, "Q"), \
1302 FLAG(req->rq_reply_unlinked, "U"), \
1303 FLAG(req->rq_receiving_reply, "r")
1305 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s"
1307 void _debug_req(struct ptlrpc_request *req,
1308 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1309 __attribute__ ((format (printf, 3, 4)));
1312 * Helper that decides if we need to print request accordig to current debug
1315 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1317 CFS_CHECK_STACK(msgdata, mask, cdls); \
1319 if (((mask) & D_CANTMASK) != 0 || \
1320 ((libcfs_debug & (mask)) != 0 && \
1321 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1322 _debug_req((req), msgdata, fmt, ##a); \
1326 * This is the debug print function you need to use to print request sturucture
1327 * content into lustre debug log.
1328 * for most callers (level is a constant) this is resolved at compile time */
1329 #define DEBUG_REQ(level, req, fmt, args...) \
1331 if ((level) & (D_ERROR | D_WARNING)) { \
1332 static struct cfs_debug_limit_state cdls; \
1333 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1334 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1336 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1337 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1342 enum ptlrpc_bulk_op_type {
1343 PTLRPC_BULK_OP_ACTIVE = 0x00000001,
1344 PTLRPC_BULK_OP_PASSIVE = 0x00000002,
1345 PTLRPC_BULK_OP_PUT = 0x00000004,
1346 PTLRPC_BULK_OP_GET = 0x00000008,
1347 PTLRPC_BULK_GET_SOURCE = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_GET,
1348 PTLRPC_BULK_PUT_SINK = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_PUT,
1349 PTLRPC_BULK_GET_SINK = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_GET,
1350 PTLRPC_BULK_PUT_SOURCE = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_PUT,
1353 static inline bool ptlrpc_is_bulk_op_get(enum ptlrpc_bulk_op_type type)
1355 return (type & PTLRPC_BULK_OP_GET) == PTLRPC_BULK_OP_GET;
1358 static inline bool ptlrpc_is_bulk_get_source(enum ptlrpc_bulk_op_type type)
1360 return (type & PTLRPC_BULK_GET_SOURCE) == PTLRPC_BULK_GET_SOURCE;
1363 static inline bool ptlrpc_is_bulk_put_sink(enum ptlrpc_bulk_op_type type)
1365 return (type & PTLRPC_BULK_PUT_SINK) == PTLRPC_BULK_PUT_SINK;
1368 static inline bool ptlrpc_is_bulk_get_sink(enum ptlrpc_bulk_op_type type)
1370 return (type & PTLRPC_BULK_GET_SINK) == PTLRPC_BULK_GET_SINK;
1373 static inline bool ptlrpc_is_bulk_put_source(enum ptlrpc_bulk_op_type type)
1375 return (type & PTLRPC_BULK_PUT_SOURCE) == PTLRPC_BULK_PUT_SOURCE;
1378 static inline bool ptlrpc_is_bulk_op_active(enum ptlrpc_bulk_op_type type)
1380 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1381 (type & PTLRPC_BULK_OP_PASSIVE))
1382 == PTLRPC_BULK_OP_ACTIVE;
1385 static inline bool ptlrpc_is_bulk_op_passive(enum ptlrpc_bulk_op_type type)
1387 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1388 (type & PTLRPC_BULK_OP_PASSIVE))
1389 == PTLRPC_BULK_OP_PASSIVE;
1392 struct ptlrpc_bulk_frag_ops {
1394 * Add a page \a page to the bulk descriptor \a desc
1395 * Data to transfer in the page starts at offset \a pageoffset and
1396 * amount of data to transfer from the page is \a len
1398 void (*add_kiov_frag)(struct ptlrpc_bulk_desc *desc,
1399 struct page *page, int pageoffset, int len);
1402 * Add a \a fragment to the bulk descriptor \a desc.
1403 * Data to transfer in the fragment is pointed to by \a frag
1404 * The size of the fragment is \a len
1406 int (*add_iov_frag)(struct ptlrpc_bulk_desc *desc, void *frag, int len);
1409 * Uninitialize and free bulk descriptor \a desc.
1410 * Works on bulk descriptors both from server and client side.
1412 void (*release_frags)(struct ptlrpc_bulk_desc *desc);
1415 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops;
1416 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops;
1419 * Definition of bulk descriptor.
1420 * Bulks are special "Two phase" RPCs where initial request message
1421 * is sent first and it is followed bt a transfer (o receiving) of a large
1422 * amount of data to be settled into pages referenced from the bulk descriptors.
1423 * Bulks transfers (the actual data following the small requests) are done
1424 * on separate LNet portals.
1425 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1426 * Another user is readpage for MDT.
1428 struct ptlrpc_bulk_desc {
1429 unsigned int bd_refs; /* number MD's assigned including zero-sends */
1430 /** completed with failure */
1431 unsigned long bd_failure:1;
1433 unsigned long bd_registered:1;
1434 /** For serialization with callback */
1436 /** {put,get}{source,sink}{kvec,kiov} */
1437 enum ptlrpc_bulk_op_type bd_type;
1438 /** LNet portal for this bulk */
1440 /** Server side - export this bulk created for */
1441 struct obd_export *bd_export;
1442 /** Client side - import this bulk was sent on */
1443 struct obd_import *bd_import;
1444 /** Back pointer to the request */
1445 struct ptlrpc_request *bd_req;
1446 const struct ptlrpc_bulk_frag_ops *bd_frag_ops;
1447 wait_queue_head_t bd_waitq; /* server side only WQ */
1448 int bd_iov_count; /* # entries in bd_iov */
1449 int bd_max_iov; /* allocated size of bd_iov */
1450 int bd_nob; /* # bytes covered */
1451 int bd_nob_transferred; /* # bytes GOT/PUT */
1452 unsigned int bd_nob_last; /* # bytes in last MD */
1454 __u64 bd_last_mbits;
1456 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1457 lnet_nid_t bd_sender; /* stash event::sender */
1458 int bd_md_count; /* # valid entries in bd_mds */
1459 int bd_md_max_brw; /* max entries in bd_mds */
1461 /** array of offsets for each MD */
1462 unsigned int bd_mds_off[PTLRPC_BULK_OPS_COUNT];
1463 /** array of associated MDs */
1464 struct lnet_handle_md bd_mds[PTLRPC_BULK_OPS_COUNT];
1466 /* encrypted iov, size is either 0 or bd_iov_count. */
1467 struct bio_vec *bd_enc_vec;
1468 struct bio_vec *bd_vec;
1473 SVC_STOPPED = BIT(0),
1474 SVC_STOPPING = BIT(1),
1475 SVC_STARTING = BIT(2),
1476 SVC_RUNNING = BIT(3),
1479 #define PTLRPC_THR_NAME_LEN 32
1481 * Definition of server service thread structure
1483 struct ptlrpc_thread {
1485 * List of active threads in svcpt->scp_threads
1487 struct list_head t_link;
1489 * thread-private data (preallocated vmalloc'd memory)
1494 * service thread index, from ptlrpc_start_threads
1500 struct task_struct *t_task;
1504 * put watchdog in the structure per thread b=14840
1506 struct delayed_work t_watchdog;
1508 * the svc this thread belonged to b=18582
1510 struct ptlrpc_service_part *t_svcpt;
1511 wait_queue_head_t t_ctl_waitq;
1512 struct lu_env *t_env;
1513 char t_name[PTLRPC_THR_NAME_LEN];
1516 static inline int thread_is_init(struct ptlrpc_thread *thread)
1518 return thread->t_flags == 0;
1521 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1523 return !!(thread->t_flags & SVC_STOPPED);
1526 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1528 return !!(thread->t_flags & SVC_STOPPING);
1531 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1533 return !!(thread->t_flags & SVC_STARTING);
1536 static inline int thread_is_running(struct ptlrpc_thread *thread)
1538 return !!(thread->t_flags & SVC_RUNNING);
1541 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1543 thread->t_flags &= ~flags;
1546 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1548 thread->t_flags = flags;
1551 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1553 thread->t_flags |= flags;
1556 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1559 if (thread->t_flags & flags) {
1560 thread->t_flags &= ~flags;
1567 * Request buffer descriptor structure.
1568 * This is a structure that contains one posted request buffer for service.
1569 * Once data land into a buffer, event callback creates actual request and
1570 * notifies wakes one of the service threads to process new incoming request.
1571 * More than one request can fit into the buffer.
1573 struct ptlrpc_request_buffer_desc {
1574 /** Link item for rqbds on a service */
1575 struct list_head rqbd_list;
1576 /** History of requests for this buffer */
1577 struct list_head rqbd_reqs;
1578 /** Back pointer to service for which this buffer is registered */
1579 struct ptlrpc_service_part *rqbd_svcpt;
1580 /** LNet descriptor */
1581 struct lnet_handle_md rqbd_md_h;
1583 /** The buffer itself */
1585 struct ptlrpc_cb_id rqbd_cbid;
1587 * This "embedded" request structure is only used for the
1588 * last request to fit into the buffer
1590 struct ptlrpc_request rqbd_req;
1593 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1595 struct ptlrpc_service_ops {
1597 * if non-NULL called during thread creation (ptlrpc_start_thread())
1598 * to initialize service specific per-thread state.
1600 int (*so_thr_init)(struct ptlrpc_thread *thr);
1602 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1603 * destruct state created by ->srv_init().
1605 void (*so_thr_done)(struct ptlrpc_thread *thr);
1607 * Handler function for incoming requests for this service
1609 int (*so_req_handler)(struct ptlrpc_request *req);
1611 * function to determine priority of the request, it's called
1612 * on every new request
1614 int (*so_hpreq_handler)(struct ptlrpc_request *);
1616 * service-specific print fn
1618 void (*so_req_printer)(void *, struct ptlrpc_request *);
1621 #ifndef __cfs_cacheline_aligned
1622 /* NB: put it here for reducing patche dependence */
1623 # define __cfs_cacheline_aligned
1627 * How many high priority requests to serve before serving one normal
1630 #define PTLRPC_SVC_HP_RATIO 10
1633 * Definition of PortalRPC service.
1634 * The service is listening on a particular portal (like tcp port)
1635 * and perform actions for a specific server like IO service for OST
1636 * or general metadata service for MDS.
1638 struct ptlrpc_service {
1639 /** serialize /proc operations */
1640 spinlock_t srv_lock;
1641 /** most often accessed fields */
1642 /** chain thru all services */
1643 struct list_head srv_list;
1644 /** service operations table */
1645 struct ptlrpc_service_ops srv_ops;
1646 /** only statically allocated strings here; we don't clean them */
1648 /** only statically allocated strings here; we don't clean them */
1649 char *srv_thread_name;
1650 /** threads # should be created for each partition on initializing */
1651 int srv_nthrs_cpt_init;
1652 /** limit of threads number for each partition */
1653 int srv_nthrs_cpt_limit;
1654 /** Root of debugfs dir tree for this service */
1655 struct dentry *srv_debugfs_entry;
1656 /** Pointer to statistic data for this service */
1657 struct lprocfs_stats *srv_stats;
1658 /** # hp per lp reqs to handle */
1659 int srv_hpreq_ratio;
1660 /** biggest request to receive */
1661 int srv_max_req_size;
1662 /** biggest reply to send */
1663 int srv_max_reply_size;
1664 /** size of individual buffers */
1666 /** # buffers to allocate in 1 group */
1667 int srv_nbuf_per_group;
1668 /** Local portal on which to receive requests */
1669 __u32 srv_req_portal;
1670 /** Portal on the client to send replies to */
1671 __u32 srv_rep_portal;
1673 * Tags for lu_context associated with this thread, see struct
1677 /** soft watchdog timeout multiplier */
1678 int srv_watchdog_factor;
1679 /** under unregister_service */
1680 unsigned srv_is_stopping:1;
1681 /** Whether or not to restrict service threads to CPUs in this CPT */
1682 unsigned srv_cpt_bind:1;
1684 /** max # request buffers */
1686 /** max # request buffers in history per partition */
1687 int srv_hist_nrqbds_cpt_max;
1688 /** number of CPTs this service associated with */
1690 /** CPTs array this service associated with */
1692 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1694 /** CPT table this service is running over */
1695 struct cfs_cpt_table *srv_cptable;
1698 struct kobject srv_kobj;
1699 struct completion srv_kobj_unregister;
1701 * partition data for ptlrpc service
1703 struct ptlrpc_service_part *srv_parts[0];
1707 * Definition of PortalRPC service partition data.
1708 * Although a service only has one instance of it right now, but we
1709 * will have multiple instances very soon (instance per CPT).
1711 * it has four locks:
1713 * serialize operations on rqbd and requests waiting for preprocess
1715 * serialize operations active requests sent to this portal
1717 * serialize adaptive timeout stuff
1719 * serialize operations on RS list (reply states)
1721 * We don't have any use-case to take two or more locks at the same time
1722 * for now, so there is no lock order issue.
1724 struct ptlrpc_service_part {
1725 /** back reference to owner */
1726 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1727 /* CPT id, reserved */
1729 /** always increasing number */
1731 /** # of starting threads */
1732 int scp_nthrs_starting;
1733 /** # running threads */
1734 int scp_nthrs_running;
1735 /** service threads list */
1736 struct list_head scp_threads;
1739 * serialize the following fields, used for protecting
1740 * rqbd list and incoming requests waiting for preprocess,
1741 * threads starting & stopping are also protected by this lock.
1743 spinlock_t scp_lock __cfs_cacheline_aligned;
1744 /** userland serialization */
1745 struct mutex scp_mutex;
1746 /** total # req buffer descs allocated */
1747 int scp_nrqbds_total;
1748 /** # posted request buffers for receiving */
1749 int scp_nrqbds_posted;
1750 /** in progress of allocating rqbd */
1751 int scp_rqbd_allocating;
1752 /** # incoming reqs */
1753 int scp_nreqs_incoming;
1754 /** request buffers to be reposted */
1755 struct list_head scp_rqbd_idle;
1756 /** req buffers receiving */
1757 struct list_head scp_rqbd_posted;
1758 /** incoming reqs */
1759 struct list_head scp_req_incoming;
1760 /** timeout before re-posting reqs, in jiffies */
1761 long scp_rqbd_timeout;
1763 * all threads sleep on this. This wait-queue is signalled when new
1764 * incoming request arrives and when difficult reply has to be handled.
1766 wait_queue_head_t scp_waitq;
1768 /** request history */
1769 struct list_head scp_hist_reqs;
1770 /** request buffer history */
1771 struct list_head scp_hist_rqbds;
1772 /** # request buffers in history */
1773 int scp_hist_nrqbds;
1774 /** sequence number for request */
1776 /** highest seq culled from history */
1777 __u64 scp_hist_seq_culled;
1780 * serialize the following fields, used for processing requests
1781 * sent to this portal
1783 spinlock_t scp_req_lock __cfs_cacheline_aligned;
1784 /** # reqs in either of the NRS heads below */
1785 /** # reqs being served */
1786 int scp_nreqs_active;
1787 /** # HPreqs being served */
1788 int scp_nhreqs_active;
1789 /** # hp requests handled */
1792 /** NRS head for regular requests */
1793 struct ptlrpc_nrs scp_nrs_reg;
1794 /** NRS head for HP requests; this is only valid for services that can
1795 * handle HP requests */
1796 struct ptlrpc_nrs *scp_nrs_hp;
1801 * serialize the following fields, used for changes on
1804 spinlock_t scp_at_lock __cfs_cacheline_aligned;
1805 /** estimated rpc service time */
1806 struct adaptive_timeout scp_at_estimate;
1807 /** reqs waiting for replies */
1808 struct ptlrpc_at_array scp_at_array;
1809 /** early reply timer */
1810 struct timer_list scp_at_timer;
1812 ktime_t scp_at_checktime;
1813 /** check early replies */
1814 unsigned scp_at_check;
1818 * serialize the following fields, used for processing
1819 * replies for this portal
1821 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1822 /** all the active replies */
1823 struct list_head scp_rep_active;
1824 /** List of free reply_states */
1825 struct list_head scp_rep_idle;
1826 /** waitq to run, when adding stuff to srv_free_rs_list */
1827 wait_queue_head_t scp_rep_waitq;
1828 /** # 'difficult' replies */
1829 atomic_t scp_nreps_difficult;
1832 #define ptlrpc_service_for_each_part(part, i, svc) \
1834 i < (svc)->srv_ncpts && \
1835 (svc)->srv_parts != NULL && \
1836 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1839 * Declaration of ptlrpcd control structure
1841 struct ptlrpcd_ctl {
1843 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1845 unsigned long pc_flags;
1847 * Thread lock protecting structure fields.
1853 struct completion pc_starting;
1857 struct completion pc_finishing;
1859 * Thread requests set.
1861 struct ptlrpc_request_set *pc_set;
1863 * Thread name used in kthread_run()
1867 * CPT the thread is bound on.
1871 * Index of ptlrpcd thread in the array.
1875 * Pointer to the array of partners' ptlrpcd_ctl structure.
1877 struct ptlrpcd_ctl **pc_partners;
1879 * Number of the ptlrpcd's partners.
1883 * Record the partner index to be processed next.
1887 * Error code if the thread failed to fully start.
1892 /* Bits for pc_flags */
1893 enum ptlrpcd_ctl_flags {
1895 * Ptlrpc thread start flag.
1897 LIOD_START = BIT(0),
1899 * Ptlrpc thread stop flag.
1903 * Ptlrpc thread force flag (only stop force so far).
1904 * This will cause aborting any inflight rpcs handled
1905 * by thread if LIOD_STOP is specified.
1907 LIOD_FORCE = BIT(2),
1909 * This is a recovery ptlrpc thread.
1911 LIOD_RECOVERY = BIT(3),
1918 * Service compatibility function; the policy is compatible with all services.
1920 * \param[in] svc The service the policy is attempting to register with.
1921 * \param[in] desc The policy descriptor
1923 * \retval true The policy is compatible with the service
1925 * \see ptlrpc_nrs_pol_desc::pd_compat()
1927 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
1928 const struct ptlrpc_nrs_pol_desc *desc)
1934 * Service compatibility function; the policy is compatible with only a specific
1935 * service which is identified by its human-readable name at
1936 * ptlrpc_service::srv_name.
1938 * \param[in] svc The service the policy is attempting to register with.
1939 * \param[in] desc The policy descriptor
1941 * \retval false The policy is not compatible with the service
1942 * \retval true The policy is compatible with the service
1944 * \see ptlrpc_nrs_pol_desc::pd_compat()
1946 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
1947 const struct ptlrpc_nrs_pol_desc *desc)
1949 LASSERT(desc->pd_compat_svc_name != NULL);
1950 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
1955 /* ptlrpc/events.c */
1956 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1957 struct lnet_process_id *peer, lnet_nid_t *self);
1959 * These callbacks are invoked by LNet when something happened to
1963 extern void request_out_callback(struct lnet_event *ev);
1964 extern void reply_in_callback(struct lnet_event *ev);
1965 extern void client_bulk_callback(struct lnet_event *ev);
1966 extern void request_in_callback(struct lnet_event *ev);
1967 extern void reply_out_callback(struct lnet_event *ev);
1968 #ifdef HAVE_SERVER_SUPPORT
1969 extern void server_bulk_callback(struct lnet_event *ev);
1973 /* ptlrpc/connection.c */
1974 struct ptlrpc_connection *ptlrpc_connection_get(struct lnet_process_id peer,
1976 struct obd_uuid *uuid);
1978 static inline void ptlrpc_connection_put(struct ptlrpc_connection *conn)
1983 LASSERT(atomic_read(&conn->c_refcount) > 0);
1986 * We do not remove connection from hashtable and
1987 * do not free it even if last caller released ref,
1988 * as we want to have it cached for the case it is
1991 * Deallocating it and later creating new connection
1992 * again would be wastful. This way we also avoid
1993 * expensive locking to protect things from get/put
1994 * race when found cached connection is freed by
1995 * ptlrpc_connection_put().
1997 * It will be freed later in module unload time,
1998 * when ptlrpc_connection_fini()->lh_exit->conn_exit()
2001 atomic_dec(&conn->c_refcount);
2003 CDEBUG(D_INFO, "PUT conn=%p refcount %d to %s\n",
2004 conn, atomic_read(&conn->c_refcount),
2005 libcfs_nid2str(conn->c_peer.nid));
2008 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2009 int ptlrpc_connection_init(void);
2010 void ptlrpc_connection_fini(void);
2011 extern lnet_pid_t ptl_get_pid(void);
2014 * Check if the peer connection is on the local node. We need to use GFP_NOFS
2015 * for requests from a local client to avoid recursing into the filesystem
2016 * as we might end up waiting on a page sent in the request we're serving.
2018 * Use __GFP_HIGHMEM so that the pages can use all of the available memory
2019 * on 32-bit machines. Use more aggressive GFP_HIGHUSER flags from non-local
2020 * clients to be able to generate more memory pressure on the OSS and allow
2021 * inactive pages to be reclaimed, since it doesn't have any other processes
2022 * or allocations that generate memory reclaim pressure.
2024 * See b=17576 (bdf50dc9) and b=19529 (3dcf18d3) for details.
2026 static inline bool ptlrpc_connection_is_local(struct ptlrpc_connection *conn)
2031 if (conn->c_peer.nid == conn->c_self)
2034 RETURN(LNetIsPeerLocal(conn->c_peer.nid));
2037 /* ptlrpc/niobuf.c */
2039 * Actual interfacing with LNet to put/get/register/unregister stuff
2042 #ifdef HAVE_SERVER_SUPPORT
2043 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2044 unsigned nfrags, unsigned max_brw,
2047 const struct ptlrpc_bulk_frag_ops
2049 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2050 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2052 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2056 LASSERT(desc != NULL);
2058 spin_lock(&desc->bd_lock);
2060 spin_unlock(&desc->bd_lock);
2065 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2066 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2068 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2070 struct ptlrpc_bulk_desc *desc;
2073 LASSERT(req != NULL);
2074 desc = req->rq_bulk;
2079 if (req->rq_bulk_deadline > ktime_get_real_seconds())
2083 spin_lock(&desc->bd_lock);
2085 spin_unlock(&desc->bd_lock);
2089 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2090 #define PTLRPC_REPLY_EARLY 0x02
2091 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2092 int ptlrpc_reply(struct ptlrpc_request *req);
2093 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2094 int ptlrpc_error(struct ptlrpc_request *req);
2095 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2096 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2097 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2100 /* ptlrpc/client.c */
2102 * Client-side portals API. Everything to send requests, receive replies,
2103 * request queues, request management, etc.
2106 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2108 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
2109 struct ptlrpc_client *);
2110 void ptlrpc_cleanup_client(struct obd_import *imp);
2111 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
2112 lnet_nid_t nid4refnet);
2114 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2115 int ptlrpc_replay_req(struct ptlrpc_request *req);
2116 void ptlrpc_restart_req(struct ptlrpc_request *req);
2117 void ptlrpc_abort_inflight(struct obd_import *imp);
2118 void ptlrpc_cleanup_imp(struct obd_import *imp);
2119 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2121 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2122 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2124 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2125 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *);
2126 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2127 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2128 #define PTLRPCD_SET ((struct ptlrpc_request_set *)1)
2130 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2131 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2133 struct ptlrpc_request_pool *
2134 ptlrpc_init_rq_pool(int, int,
2135 int (*populate_pool)(struct ptlrpc_request_pool *, int));
2137 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2138 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2139 const struct req_format *format);
2140 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2141 struct ptlrpc_request_pool *,
2142 const struct req_format *format);
2143 void ptlrpc_request_free(struct ptlrpc_request *request);
2144 int ptlrpc_request_pack(struct ptlrpc_request *request,
2145 __u32 version, int opcode);
2146 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2147 const struct req_format *format,
2148 __u32 version, int opcode);
2149 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2150 __u32 version, int opcode, char **bufs,
2151 struct ptlrpc_cli_ctx *ctx);
2152 void ptlrpc_req_finished(struct ptlrpc_request *request);
2153 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2154 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2155 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2156 unsigned nfrags, unsigned max_brw,
2159 const struct ptlrpc_bulk_frag_ops
2162 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2163 struct page *page, int pageoffset, int len,
2166 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
2168 static inline void ptlrpc_release_bulk_noop(struct ptlrpc_bulk_desc *desc)
2172 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2173 struct obd_import *imp);
2174 __u64 ptlrpc_next_xid(void);
2175 __u64 ptlrpc_sample_next_xid(void);
2176 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2177 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req);
2178 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req);
2180 /* Set of routines to run a function in ptlrpcd context */
2181 void *ptlrpcd_alloc_work(struct obd_import *imp,
2182 int (*cb)(const struct lu_env *, void *), void *data);
2183 void ptlrpcd_destroy_work(void *handler);
2184 int ptlrpcd_queue_work(void *handler);
2187 struct ptlrpc_service_buf_conf {
2188 /* nbufs is buffers # to allocate when growing the pool */
2189 unsigned int bc_nbufs;
2190 /* buffer size to post */
2191 unsigned int bc_buf_size;
2192 /* portal to listed for requests on */
2193 unsigned int bc_req_portal;
2194 /* portal of where to send replies to */
2195 unsigned int bc_rep_portal;
2196 /* maximum request size to be accepted for this service */
2197 unsigned int bc_req_max_size;
2198 /* maximum reply size this service can ever send */
2199 unsigned int bc_rep_max_size;
2202 struct ptlrpc_service_thr_conf {
2203 /* threadname should be 8 characters or less - 6 will be added on */
2205 /* threads increasing factor for each CPU */
2206 unsigned int tc_thr_factor;
2207 /* service threads # to start on each partition while initializing */
2208 unsigned int tc_nthrs_init;
2210 * low water of threads # upper-limit on each partition while running,
2211 * service availability may be impacted if threads number is lower
2212 * than this value. It can be ZERO if the service doesn't require
2213 * CPU affinity or there is only one partition.
2215 unsigned int tc_nthrs_base;
2216 /* "soft" limit for total threads number */
2217 unsigned int tc_nthrs_max;
2218 /* user specified threads number, it will be validated due to
2219 * other members of this structure. */
2220 unsigned int tc_nthrs_user;
2221 /* bind service threads to only CPUs in their associated CPT */
2222 unsigned int tc_cpu_bind;
2223 /* Tags for lu_context associated with service thread */
2227 struct ptlrpc_service_cpt_conf {
2228 struct cfs_cpt_table *cc_cptable;
2229 /* string pattern to describe CPTs for a service */
2231 /* whether or not to have per-CPT service partitions */
2235 struct ptlrpc_service_conf {
2238 /* soft watchdog timeout multiplifier to print stuck service traces */
2239 unsigned int psc_watchdog_factor;
2240 /* buffer information */
2241 struct ptlrpc_service_buf_conf psc_buf;
2242 /* thread information */
2243 struct ptlrpc_service_thr_conf psc_thr;
2244 /* CPU partition information */
2245 struct ptlrpc_service_cpt_conf psc_cpt;
2246 /* function table */
2247 struct ptlrpc_service_ops psc_ops;
2250 /* ptlrpc/service.c */
2252 * Server-side services API. Register/unregister service, request state
2253 * management, service thread management
2257 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
2258 int mode, bool no_ack, bool convert_lock);
2259 void ptlrpc_commit_replies(struct obd_export *exp);
2260 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2261 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2262 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2263 struct ptlrpc_service *ptlrpc_register_service(
2264 struct ptlrpc_service_conf *conf,
2265 struct kset *parent,
2266 struct dentry *debugfs_entry);
2267 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2269 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2270 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2271 int ptlrpc_service_health_check(struct ptlrpc_service *);
2272 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2273 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2274 struct obd_export *export);
2275 void ptlrpc_update_export_timer(struct obd_export *exp,
2276 time64_t extra_delay);
2278 int ptlrpc_hr_init(void);
2279 void ptlrpc_hr_fini(void);
2281 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout);
2282 void ptlrpc_watchdog_disable(struct delayed_work *work);
2283 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout);
2287 /* ptlrpc/import.c */
2292 int ptlrpc_connect_import(struct obd_import *imp);
2293 int ptlrpc_connect_import_locked(struct obd_import *imp);
2294 int ptlrpc_init_import(struct obd_import *imp);
2295 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2296 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp);
2297 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2298 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2300 void ptlrpc_import_enter_resend(struct obd_import *imp);
2301 /* ptlrpc/pack_generic.c */
2302 int ptlrpc_reconnect_import(struct obd_import *imp);
2306 * ptlrpc msg buffer and swab interface
2310 #define PTLRPC_MAX_BUFCOUNT \
2311 (sizeof(((struct ptlrpc_request *)0)->rq_req_swab_mask) * 8)
2312 #define MD_MAX_BUFLEN (MDS_REG_MAXREQSIZE > OUT_MAXREQSIZE ? \
2313 MDS_REG_MAXREQSIZE : OUT_MAXREQSIZE)
2314 #define PTLRPC_MAX_BUFLEN (OST_IO_MAXREQSIZE > MD_MAX_BUFLEN ? \
2315 OST_IO_MAXREQSIZE : MD_MAX_BUFLEN)
2316 bool ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2318 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2320 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2321 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2323 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2324 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2326 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2327 __u32 *lens, char **bufs);
2328 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2330 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2331 __u32 *lens, char **bufs, int flags);
2332 #define LPRFL_EARLY_REPLY 1
2333 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2334 char **bufs, int flags);
2335 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2336 unsigned int newlen, int move_data);
2337 int lustre_grow_msg(struct lustre_msg *msg, int segment, unsigned int newlen);
2338 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2339 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2340 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
2341 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2342 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
2343 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
2344 __u32 lustre_msg_early_size(void);
2345 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
2346 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
2347 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
2348 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
2349 __u32 lustre_msg_bufcount(struct lustre_msg *m);
2350 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
2351 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2352 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2353 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2354 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
2355 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
2356 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
2357 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2358 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
2359 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2360 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2361 enum lustre_msg_version lustre_msg_get_version(struct lustre_msg *msg);
2362 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
2363 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2364 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2365 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
2366 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2367 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2368 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2369 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2370 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2371 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2372 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2373 int lustre_msg_get_status(struct lustre_msg *msg);
2374 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2375 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2376 timeout_t lustre_msg_get_timeout(struct lustre_msg *msg);
2377 timeout_t lustre_msg_get_service_timeout(struct lustre_msg *msg);
2378 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2379 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2380 __u64 lustre_msg_get_mbits(struct lustre_msg *msg);
2381 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, __u32 buf);
2382 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2383 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2384 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2385 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2386 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
2387 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2388 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2389 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2390 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2391 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2392 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2393 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2394 void lustre_msg_set_timeout(struct lustre_msg *msg, timeout_t timeout);
2395 void lustre_msg_set_service_timeout(struct lustre_msg *msg,
2396 timeout_t service_timeout);
2397 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2398 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2399 void lustre_msg_set_mbits(struct lustre_msg *msg, __u64 mbits);
2402 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2403 unsigned int newlen, int move_data)
2405 LASSERT(req->rq_reply_state);
2406 LASSERT(req->rq_repmsg);
2407 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2411 #ifdef LUSTRE_TRANSLATE_ERRNOS
2413 static inline int ptlrpc_status_hton(int h)
2416 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2417 * ELDLM_LOCK_ABORTED, etc.
2420 return -lustre_errno_hton(-h);
2425 static inline int ptlrpc_status_ntoh(int n)
2428 * See the comment in ptlrpc_status_hton().
2431 return -lustre_errno_ntoh(-n);
2438 #define ptlrpc_status_hton(h) (h)
2439 #define ptlrpc_status_ntoh(n) (n)
2444 /** Change request phase of \a req to \a new_phase */
2446 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2448 if (req->rq_phase == new_phase)
2451 if (new_phase == RQ_PHASE_UNREG_RPC ||
2452 new_phase == RQ_PHASE_UNREG_BULK) {
2453 /* No embedded unregistering phases */
2454 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2455 req->rq_phase == RQ_PHASE_UNREG_BULK)
2458 req->rq_next_phase = req->rq_phase;
2460 atomic_inc(&req->rq_import->imp_unregistering);
2463 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2464 req->rq_phase == RQ_PHASE_UNREG_BULK) {
2466 atomic_dec(&req->rq_import->imp_unregistering);
2469 DEBUG_REQ(D_INFO, req, "move request phase from %s to %s",
2470 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2472 req->rq_phase = new_phase;
2476 * Returns true if request \a req got early reply and hard deadline is not met
2479 ptlrpc_client_early(struct ptlrpc_request *req)
2481 return req->rq_early;
2485 * Returns true if we got real reply from server for this request
2488 ptlrpc_client_replied(struct ptlrpc_request *req)
2490 if (req->rq_reply_deadline > ktime_get_real_seconds())
2492 return req->rq_replied;
2495 /** Returns true if request \a req is in process of receiving server reply */
2497 ptlrpc_client_recv(struct ptlrpc_request *req)
2499 if (req->rq_reply_deadline > ktime_get_real_seconds())
2501 return req->rq_receiving_reply;
2505 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2509 spin_lock(&req->rq_lock);
2510 if (req->rq_reply_deadline > ktime_get_real_seconds()) {
2511 spin_unlock(&req->rq_lock);
2514 if (req->rq_req_deadline > ktime_get_real_seconds()) {
2515 spin_unlock(&req->rq_lock);
2519 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
2520 req->rq_receiving_reply;
2521 spin_unlock(&req->rq_lock);
2526 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2529 if (req->rq_set == NULL)
2530 wake_up(&req->rq_reply_waitq);
2532 wake_up(&req->rq_set->set_waitq);
2536 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2538 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2539 atomic_inc(&rs->rs_refcount);
2543 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2545 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2546 if (atomic_dec_and_test(&rs->rs_refcount))
2547 lustre_free_reply_state(rs);
2550 /* Should only be called once per req */
2551 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2553 if (req->rq_reply_state == NULL)
2554 return; /* shouldn't occur */
2555 ptlrpc_rs_decref(req->rq_reply_state);
2556 req->rq_reply_state = NULL;
2557 req->rq_repmsg = NULL;
2560 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2562 return lustre_msg_get_magic(req->rq_reqmsg);
2565 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2567 switch (req->rq_reqmsg->lm_magic) {
2568 case LUSTRE_MSG_MAGIC_V2:
2569 return req->rq_reqmsg->lm_repsize;
2571 LASSERTF(0, "incorrect message magic: %08x\n",
2572 req->rq_reqmsg->lm_magic);
2577 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2579 if (req->rq_delay_limit != 0 &&
2580 req->rq_queued_time + req->rq_delay_limit < ktime_get_seconds())
2585 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2587 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2588 spin_lock(&req->rq_lock);
2589 req->rq_no_resend = 1;
2590 spin_unlock(&req->rq_lock);
2592 return req->rq_no_resend;
2596 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2598 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2600 return svcpt->scp_service->srv_watchdog_factor *
2601 max_t(int, at, obd_timeout);
2605 * Calculate the amount of time for lock prolongation.
2607 * This is helper function to get the timeout extra time.
2609 * @req current request
2611 * Return: amount of time to extend the timeout with
2613 static inline timeout_t prolong_timeout(struct ptlrpc_request *req)
2615 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
2616 timeout_t req_timeout = 0;
2619 return obd_timeout / 2;
2621 if (req->rq_deadline > req->rq_arrival_time.tv_sec)
2622 req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec;
2624 return max(req_timeout,
2625 at_est2timeout(at_get(&svcpt->scp_at_estimate)));
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 *obd, struct lustre_cfg *lcfg);
2641 int client_obd_cleanup(struct obd_device *obd);
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_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2650 lnet_nid_t prim_nid, int priority);
2651 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
2652 int nid_count, struct obd_uuid *uuid);
2653 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2654 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2655 struct obd_uuid *uuid);
2656 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2657 void client_destroy_import(struct obd_import *imp);
2660 #ifdef HAVE_SERVER_SUPPORT
2661 int server_disconnect_export(struct obd_export *exp);
2664 /* ptlrpc/pinger.c */
2666 * Pinger API (client side only)
2669 enum timeout_event {
2672 struct timeout_item;
2673 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2674 int ptlrpc_pinger_add_import(struct obd_import *imp);
2675 int ptlrpc_pinger_del_import(struct obd_import *imp);
2676 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
2677 timeout_cb_t cb, void *data,
2678 struct list_head *obd_list);
2679 int ptlrpc_del_timeout_client(struct list_head *obd_list,
2680 enum timeout_event event);
2681 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2682 int ptlrpc_obd_ping(struct obd_device *obd);
2683 void ping_evictor_start(void);
2684 void ping_evictor_stop(void);
2685 void ptlrpc_pinger_ir_up(void);
2686 void ptlrpc_pinger_ir_down(void);
2688 int ptlrpc_pinger_suppress_pings(void);
2690 /* ptlrpc/ptlrpcd.c */
2691 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2692 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2693 void ptlrpcd_wake(struct ptlrpc_request *req);
2694 void ptlrpcd_add_req(struct ptlrpc_request *req);
2695 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2696 int ptlrpcd_addref(void);
2697 void ptlrpcd_decref(void);
2699 /* ptlrpc/lproc_ptlrpc.c */
2701 * procfs output related functions
2704 const char* ll_opcode2str(__u32 opcode);
2705 const int ll_str2opcode(const char *ops);
2706 #ifdef CONFIG_PROC_FS
2707 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2708 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2709 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2711 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2712 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2713 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2717 /* ptlrpc/llog_server.c */
2718 int llog_origin_handle_open(struct ptlrpc_request *req);
2719 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2720 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2721 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2723 /* ptlrpc/llog_client.c */
2724 extern struct llog_operations llog_client_ops;