4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 /** \defgroup PtlRPC Portal RPC and networking module.
33 * PortalRPC is the layer used by rest of lustre code to achieve network
34 * communications: establish connections with corresponding export and import
35 * states, listen for a service, send and receive RPCs.
36 * PortalRPC also includes base recovery framework: packet resending and
37 * replaying, reconnections, pinger.
39 * PortalRPC utilizes LNet as its transport layer.
52 #include <linux/kobject.h>
53 #include <linux/rhashtable.h>
54 #include <linux/uio.h>
55 #include <libcfs/libcfs.h>
57 #include <lnet/lib-types.h>
58 #include <uapi/linux/lnet/nidstr.h>
59 #include <uapi/linux/lustre/lustre_idl.h>
60 #include <lustre_ha.h>
61 #include <lustre_sec.h>
62 #include <lustre_import.h>
63 #include <lprocfs_status.h>
64 #include <lu_object.h>
65 #include <lustre_req_layout.h>
66 #include <obd_support.h>
67 #include <uapi/linux/lustre/lustre_ver.h>
69 /* MD flags we _always_ use */
70 #define PTLRPC_MD_OPTIONS 0
73 * log2 max # of bulk operations in one request: 2=4MB/RPC, 5=32MB/RPC, ...
74 * In order for the client and server to properly negotiate the maximum
75 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
76 * value. The client is free to limit the actual RPC size for any bulk
77 * transfer via cl_max_pages_per_rpc to some non-power-of-two value.
78 * NOTE: This is limited to 16 (=64GB RPCs) by IOOBJ_MAX_BRW_BITS.
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."
85 * PTLRPC_BULK_OPS_COUNT is a protocol maximum and must be a power of 2
86 * However PTLRPC_BULK_OPS_LIMIT (ops_count/2 +1) is enforced as the
87 * 64G with alignment interop limit.
89 #define PTLRPC_BULK_OPS_COUNT (1U << (PTLRPC_BULK_OPS_BITS + 1))
91 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
92 * should not be used on the server at all. Otherwise, it imposes a
93 * protocol limitation on the maximum RPC size that can be used by any
94 * RPC sent to that server in the future. Instead, the server should
95 * use the negotiated per-client ocd_brw_size to determine the bulk
98 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
100 * Unaligned DIO adjust MD size for alignment to the interop page size
101 * Enable page alignmen interop range:
102 * MD_MAX_INTEROP_PAGE_SIZE(64k) <-> MD_MIN_INTEROP_PAGE_SIZE(4k)
104 #define MD_MIN_INTEROP_PAGE_SHIFT 12
105 #define MD_MIN_INTEROP_PAGE_SIZE (1u << MD_MIN_INTEROP_PAGE_SHIFT)
106 #define MD_MAX_INTEROP_PAGE_SIZE (1u << 16)
108 * Define maxima for bulk I/O.
110 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
111 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
112 * currently supported maximum between peers at connect via ocd_brw_size.
114 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
115 #define PTLRPC_MAX_BRW_SIZE (1U << PTLRPC_MAX_BRW_BITS)
116 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_SHIFT)
118 #define ONE_MB_BRW_SIZE (1U << LNET_MTU_BITS)
119 #define MD_MAX_BRW_SIZE (1U << LNET_MTU_BITS)
120 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_SHIFT)
121 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
122 #define DT_DEF_BRW_SIZE (4 * ONE_MB_BRW_SIZE)
123 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_SHIFT)
124 #define OFD_MAX_BRW_SIZE (1U << LNET_MTU_BITS)
125 /* unaligned dio needs an extra md vector 65 instead of 64 */
126 #define PTLRPC_BULK_OPS_LIMIT ((1U << PTLRPC_BULK_OPS_BITS) + 1)
128 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
129 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
130 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
132 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_SIZE))
133 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_SIZE"
135 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
136 # error "PTLRPC_MAX_BRW_SIZE too big"
138 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
139 # error "PTLRPC_MAX_BRW_PAGES too big"
141 #if (PTLRPC_BULK_OPS_LIMIT > PTLRPC_BULK_OPS_COUNT)
142 # error "PTLRPC_BULK_OPS_LIMIT too big"
145 #define PTLRPC_NTHRS_INIT 2
150 * Constants determine how memory is used to buffer incoming service requests.
152 * ?_NBUFS # buffers to allocate when growing the pool
153 * ?_BUFSIZE # bytes in a single request buffer
154 * ?_MAXREQSIZE # maximum request service will receive
156 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
157 * of ?_NBUFS is added to the pool.
159 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
160 * considered full when less than ?_MAXREQSIZE is left in them.
165 * Constants determine how threads are created for ptlrpc service.
167 * ?_NTHRS_INIT # threads to create for each service partition on
168 * initializing. If it's non-affinity service and
169 * there is only one partition, it's the overall #
170 * threads for the service while initializing.
171 * ?_NTHRS_BASE # threads should be created at least for each
172 * ptlrpc partition to keep the service healthy.
173 * It's the low-water mark of threads upper-limit
174 * for each partition.
175 * ?_THR_FACTOR # threads can be added on threads upper-limit for
176 * each CPU core. This factor is only for reference,
177 * we might decrease value of factor if number of cores
178 * per CPT is above a limit.
179 * ?_NTHRS_MAX # overall threads can be created for a service,
180 * it's a soft limit because if service is running
181 * on machine with hundreds of cores and tens of
182 * CPU partitions, we need to guarantee each partition
183 * has ?_NTHRS_BASE threads, which means total threads
184 * will be ?_NTHRS_BASE * number_of_cpts which can
185 * exceed ?_NTHRS_MAX.
189 * #define MDS_NTHRS_INIT 2
190 * #define MDS_NTHRS_BASE 64
191 * #define MDS_NTHRS_FACTOR 8
192 * #define MDS_NTHRS_MAX 1024
195 * ---------------------------------------------------------------------
196 * Server(A) has 16 cores, user configured it to 4 partitions so each
197 * partition has 4 cores, then actual number of service threads on each
199 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
201 * Total number of threads for the service is:
202 * 96 * partitions(4) = 384
205 * ---------------------------------------------------------------------
206 * Server(B) has 32 cores, user configured it to 4 partitions so each
207 * partition has 8 cores, then actual number of service threads on each
209 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
211 * Total number of threads for the service is:
212 * 128 * partitions(4) = 512
215 * ---------------------------------------------------------------------
216 * Server(B) has 96 cores, user configured it to 8 partitions so each
217 * partition has 12 cores, then actual number of service threads on each
219 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
221 * Total number of threads for the service is:
222 * 160 * partitions(8) = 1280
224 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
225 * as upper limit of threads number for each partition:
226 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
229 * ---------------------------------------------------------------------
230 * Server(C) have a thousand of cores and user configured it to 32 partitions
231 * MDS_NTHRS_BASE(64) * 32 = 2048
233 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
234 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
235 * to keep service healthy, so total number of threads will just be 2048.
237 * NB: we don't suggest to choose server with that many cores because backend
238 * filesystem itself, buffer cache, or underlying network stack might
239 * have some SMP scalability issues at that large scale.
241 * If user already has a fat machine with hundreds or thousands of cores,
242 * there are two choices for configuration:
243 * a) create CPU table from subset of all CPUs and run Lustre on
245 * b) bind service threads on a few partitions, see modparameters of
246 * MDS and OSS for details
248 * NB: these calculations (and examples below) are simplified to help
249 * understanding, the real implementation is a little more complex,
250 * please see ptlrpc_server_nthreads_check() for details.
255 * LDLM threads constants:
257 * Given 8 as factor and 24 as base threads number
260 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
263 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
264 * threads for each partition and total threads number will be 112.
267 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
268 * threads for each partition to keep service healthy, so total threads
269 * number should be 24 * 8 = 192.
271 * So with these constants, threads number will be at the similar level
272 * of old versions, unless target machine has over a hundred cores
274 #define LDLM_THR_FACTOR 8
275 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
276 #define LDLM_NTHRS_BASE 24
277 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
279 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
280 #define LDLM_CLIENT_NBUFS 1
281 #define LDLM_SERVER_NBUFS 64
282 #define LDLM_BUFSIZE (8 * 1024)
283 #define LDLM_MAXREQSIZE (5 * 1024)
284 #define LDLM_MAXREPSIZE (1024)
287 * MDS threads constants:
289 * Please see examples in "Thread Constants", MDS threads number will be at
290 * the comparable level of old versions, unless the server has many cores.
292 #ifndef MDS_MAX_THREADS
293 #define MDS_MAX_THREADS 1024
294 #define MDS_MAX_OTHR_THREADS 256
296 #else /* MDS_MAX_THREADS */
297 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
298 #undef MDS_MAX_THREADS
299 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
301 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
304 /* default service */
305 #define MDS_THR_FACTOR 8
306 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
307 #define MDS_NTHRS_MAX MDS_MAX_THREADS
308 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
310 /* read-page service */
311 #define MDS_RDPG_THR_FACTOR 4
312 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
313 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
314 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
316 /* these should be removed when we remove setattr service in the future */
317 #define MDS_SETA_THR_FACTOR 4
318 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
319 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
320 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
322 /* non-affinity threads */
323 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
324 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
329 * Assume file name length = FNAME_MAX = 256 (true for ext3).
330 * path name length = PATH_MAX = 4096
331 * LOV MD size max = EA_MAX = 24 * 2000
332 * (NB: 24 is size of lov_ost_data)
333 * LOV LOGCOOKIE size max = 32 * 2000
334 * (NB: 32 is size of llog_cookie)
335 * symlink: FNAME_MAX + PATH_MAX <- largest
336 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
337 * rename: FNAME_MAX + FNAME_MAX
338 * open: FNAME_MAX + EA_MAX
340 * MDS_MAXREQSIZE ~= 4736 bytes =
341 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
342 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
344 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
345 * except in the open case where there are a large number of OSTs in a LOV.
347 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
348 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
351 * MDS incoming request with LOV EA
352 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
354 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
355 362 + LOV_MAX_STRIPE_COUNT * 24)
357 * MDS outgoing reply with LOV EA
359 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
360 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
362 * but 2.4 or later MDS will never send reply with llog_cookie to any
363 * version client. This macro is defined for server side reply buffer size.
365 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
368 * This is the size of a maximum REINT_SETXATTR request:
370 * lustre_msg 56 (32 + 4 x 5 + 4)
372 * mdt_rec_setxattr 136
374 * name 256 (XATTR_NAME_MAX)
375 * value 65536 (XATTR_SIZE_MAX)
377 #define MDS_EA_MAXREQSIZE 66288
380 * These are the maximum request and reply sizes (rounded up to 1 KB
381 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
383 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
384 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
385 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
388 * The update request includes all of updates from the create, which might
389 * include linkea (4K maxim), together with other updates, we set it to 1000K:
390 * lustre_msg + ptlrpc_body + OUT_UPDATE_BUFFER_SIZE_MAX
392 #define OUT_MAXREQSIZE (1000 * 1024)
393 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
395 #define BUT_MAXREQSIZE OUT_MAXREQSIZE
396 #define BUT_MAXREPSIZE BUT_MAXREQSIZE
398 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
399 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
403 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
404 * However, we need to allocate a much larger buffer for it because LNet
405 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
406 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
407 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
408 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
409 * utilization is very low.
411 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
412 * reused until all requests fit in it have been processed and released,
413 * which means one long blocked request can prevent the rqbd be reused.
414 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
415 * from LNet with unused 65 KB, buffer utilization will be about 59%.
416 * Please check LU-2432 for details.
418 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
422 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
423 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
424 * extra bytes to each request buffer to improve buffer utilization rate.
426 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
429 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
430 #define FLD_MAXREQSIZE (160)
432 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
433 #define FLD_MAXREPSIZE (152)
434 #define FLD_BUFSIZE (1 << 12)
437 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
440 #define SEQ_MAXREQSIZE (160)
442 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
443 #define SEQ_MAXREPSIZE (152)
444 #define SEQ_BUFSIZE (1 << 12)
446 /** MGS threads must be >= 3, see bug 22458 comment #28 */
447 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
448 #define MGS_NTHRS_MAX 32
451 #define MGS_BUFSIZE (8 * 1024)
452 #define MGS_MAXREQSIZE (7 * 1024)
453 #define MGS_MAXREPSIZE (9 * 1024)
456 * OSS threads constants:
458 * Given 8 as factor and 64 as base threads number
461 * On 8-core server configured to 2 partitions, we will have
462 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
465 * On 32-core machine configured to 4 partitions, we will have
466 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
467 * will be 112 * 4 = 448.
470 * On 64-core machine configured to 4 partitions, we will have
471 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
472 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
473 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
474 * for each partition.
476 * So we can see that with these constants, threads number wil be at the
477 * similar level of old versions, unless the server has many cores.
479 /* depress threads factor for VM with small memory size */
480 #define OSS_THR_FACTOR min_t(int, 8, \
481 NUM_CACHEPAGES >> (28 - PAGE_SHIFT))
482 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
483 #define OSS_NTHRS_BASE 64
485 /* threads for handling "create" request */
486 #define OSS_CR_THR_FACTOR 1
487 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
488 #define OSS_CR_NTHRS_BASE 8
489 #define OSS_CR_NTHRS_MAX 64
492 * OST_IO_MAXREQSIZE ~=
493 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
494 * DT_MAX_BRW_PAGES * niobuf_remote
496 * - single object with 16 pages is 512 bytes
497 * - OST_IO_MAXREQSIZE must be at least 1 niobuf per page of data
498 * - Must be a multiple of 1024
499 * - should allow a reasonably large SHORT_IO_BYTES size (64KB)
501 #define _OST_MAXREQSIZE_BASE ((unsigned long)(sizeof(struct lustre_msg) + \
502 /* lm_buflens */ sizeof(__u32) * 4 + \
503 sizeof(struct ptlrpc_body) + \
504 sizeof(struct obdo) + \
505 sizeof(struct obd_ioobj) + \
506 sizeof(struct niobuf_remote)))
507 #define _OST_MAXREQSIZE_SUM ((unsigned long)(_OST_MAXREQSIZE_BASE + \
508 sizeof(struct niobuf_remote) * \
511 * FIEMAP request can be 4K+ for now
513 #define OST_MAXREQSIZE (16UL * 1024UL)
514 #define OST_IO_MAXREQSIZE max(OST_MAXREQSIZE, \
515 ((_OST_MAXREQSIZE_SUM - 1) | \
517 /* Safe estimate of free space in standard RPC, provides upper limit for # of
518 * bytes of i/o to pack in RPC (skipping bulk transfer).
520 #define OST_MAX_SHORT_IO_BYTES ((OST_IO_MAXREQSIZE - _OST_MAXREQSIZE_BASE) & \
523 /* Actual size used for short i/o buffer. Calculation means this:
524 * At least one page (for large PAGE_SIZE), or 16 KiB, but not more
525 * than the available space aligned to a page boundary.
527 #define OBD_DEF_SHORT_IO_BYTES min(max(PAGE_SIZE, 16UL * 1024UL), \
528 OST_MAX_SHORT_IO_BYTES)
530 #define OST_MAXREPSIZE (9 * 1024)
531 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
534 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
535 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 32 * 1024)
537 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
538 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
540 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
542 /* Macro to hide a typecast and BUILD_BUG. */
543 #define ptlrpc_req_async_args(_var, req) ({ \
544 BUILD_BUG_ON(sizeof(*_var) > sizeof(req->rq_async_args)); \
545 (typeof(_var))&req->rq_async_args; \
548 struct ptlrpc_replay_async_args {
554 * Structure to single define portal connection.
556 struct ptlrpc_connection {
557 /** linkage for connections hash table */
558 struct rhash_head c_hash;
559 /** Our own lnet nid for this connection */
560 struct lnet_nid c_self;
561 /** Remote side nid for this connection */
562 struct lnet_processid c_peer;
563 /** UUID of the other side */
564 struct obd_uuid c_remote_uuid;
565 /** reference counter for this connection */
569 /** Client definition for PortalRPC */
570 struct ptlrpc_client {
571 /** What lnet portal does this client send messages to by default */
572 __u32 cli_request_portal;
573 /** What portal do we expect replies on */
574 __u32 cli_reply_portal;
575 /** Name of the client */
576 const char *cli_name;
579 /** state flags of requests */
580 /* XXX only ones left are those used by the bulk descs as well! */
581 #define PTL_RPC_FL_INTR BIT(0) /* reply wait was interrupted by user */
582 #define PTL_RPC_FL_TIMEOUT BIT(7) /* req timed out waiting for reply */
584 #define REQ_MAX_ACK_LOCKS 8
586 union ptlrpc_async_args {
588 * Scratchpad for passing args to completion interpreter. Users
589 * cast to the struct of their choosing, and BUILD_BUG_ON that this is
590 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
591 * a pointer to it here. The pointer_arg ensures this struct is at
592 * least big enough for that.
594 void *pointer_arg[11];
598 struct ptlrpc_request_set;
599 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
602 * Definition of request set structure.
603 * Request set is a list of requests (not necessary to the same target) that
604 * once populated with RPCs could be sent in parallel.
605 * There are two kinds of request sets. General purpose and with dedicated
606 * serving thread. Example of the latter is ptlrpcd set.
607 * For general purpose sets once request set started sending it is impossible
608 * to add new requests to such set.
609 * Provides a way to call "completion callbacks" when all requests in the set
612 struct ptlrpc_request_set {
613 atomic_t set_refcount;
614 /** number of in queue requests */
615 atomic_t set_new_count;
616 /** number of uncompleted requests */
617 atomic_t set_remaining;
618 /** wait queue to wait on for request events */
619 wait_queue_head_t set_waitq;
620 /** List of requests in the set */
621 struct list_head set_requests;
623 * Lock for \a set_new_requests manipulations
624 * locked so that any old caller can communicate requests to
625 * the set holder who can then fold them into the lock-free set
627 spinlock_t set_new_req_lock;
628 /** List of new yet unsent requests. Only used with ptlrpcd now. */
629 struct list_head set_new_requests;
631 /** rq_status of requests that have been freed already */
633 /** Additional fields used by the flow control extension */
634 /** Maximum number of RPCs in flight */
635 int set_max_inflight;
636 /** Callback function used to generate RPCs */
637 set_producer_func set_producer;
638 /** opaq argument passed to the producer callback */
639 void *set_producer_arg;
640 unsigned int set_allow_intr:1;
643 struct ptlrpc_bulk_desc;
644 struct ptlrpc_service_part;
645 struct ptlrpc_service;
648 * ptlrpc callback & work item stuff
650 struct ptlrpc_cb_id {
651 void (*cbid_fn)(struct lnet_event *ev); /* specific callback fn */
652 void *cbid_arg; /* additional arg */
655 /** Maximum number of locks to fit into reply state, migrating directory max
656 * stripe count is 2 * LMV_MAX_STRIPES_PER_MDT, plus source parent, target
657 * parent, source and target master object:
658 * 2 * LMV_MAX_STRIPES_PER_MDT + 4
660 #define RS_MAX_LOCKS 14
664 * Structure to define reply state on the server
665 * Reply state holds various reply message information. Also for "difficult"
666 * replies (rep-ack case) we store the state after sending reply and wait
667 * for the client to acknowledge the reception. In these cases locks could be
668 * added to the state for replay/failover consistency guarantees.
670 struct ptlrpc_reply_state {
671 /** Callback description */
672 struct ptlrpc_cb_id rs_cb_id;
673 /** Linkage for list of all reply states in a system */
674 struct list_head rs_list;
675 /** Linkage for list of all reply states on same export */
676 struct list_head rs_exp_list;
677 /** Linkage for list of all reply states for same obd */
678 struct list_head rs_obd_list;
680 struct list_head rs_debug_list;
682 /** A spinlock to protect the reply state flags */
684 /** Reply state flags */
685 unsigned long rs_difficult:1; /* ACK/commit stuff */
686 unsigned long rs_no_ack:1; /* no ACK, (incl difficult reqs) */
687 unsigned long rs_scheduled:1; /* being handled? */
688 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
689 unsigned long rs_handled:1; /* been handled yet? */
690 unsigned long rs_sent:1; /* Got LNET_EVENT_SEND? */
691 unsigned long rs_unlinked:1; /* Reply MD unlinked? */
692 unsigned long rs_prealloc:1; /* rs from prealloc list */
693 /* transaction committed and rs dispatched by ptlrpc_commit_replies */
694 unsigned long rs_committed:1;
695 atomic_t rs_refcount; /* number of users */
696 /** Number of locks awaiting client ACK */
699 /** Size of the state */
703 /** Transaction number */
707 struct obd_export *rs_export;
708 struct ptlrpc_service_part *rs_svcpt;
709 /** Lnet metadata handle for the reply */
710 struct lnet_handle_md rs_md_h;
712 /** Context for the sevice thread */
713 struct ptlrpc_svc_ctx *rs_svc_ctx;
714 /** Reply buffer (actually sent to the client), encoded if needed */
715 struct lustre_msg *rs_repbuf; /* wrapper */
716 /** Size of the reply buffer */
717 int rs_repbuf_len; /* wrapper buf length */
718 /** Size of the reply message */
719 int rs_repdata_len; /* wrapper msg length */
721 * Actual reply message. Its content is encrupted (if needed) to
722 * produce reply buffer for actual sending. In simple case
723 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
725 struct lustre_msg *rs_msg; /* reply message */
727 /** Handles of locks awaiting client reply ACK */
728 struct lustre_handle rs_locks[RS_MAX_LOCKS];
731 struct ptlrpc_thread;
735 RQ_PHASE_NEW = 0xebc0de00,
736 RQ_PHASE_RPC = 0xebc0de01,
737 RQ_PHASE_BULK = 0xebc0de02,
738 RQ_PHASE_INTERPRET = 0xebc0de03,
739 RQ_PHASE_COMPLETE = 0xebc0de04,
740 RQ_PHASE_UNREG_RPC = 0xebc0de05,
741 RQ_PHASE_UNREG_BULK = 0xebc0de06,
742 RQ_PHASE_UNDEFINED = 0xebc0de07
745 /** Type of request interpreter call-back */
746 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
747 struct ptlrpc_request *req,
749 /** Type of request resend call-back */
750 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
754 * Definition of request pool structure.
755 * The pool is used to store empty preallocated requests for the case
756 * when we would actually need to send something without performing
757 * any allocations (to avoid e.g. OOM).
759 struct ptlrpc_request_pool {
760 /** Locks the list */
762 /** list of ptlrpc_request structs */
763 struct list_head prp_req_list;
764 /** Maximum message size that would fit into a rquest from this pool */
766 /** Function to allocate more requests for this pool */
767 int (*prp_populate)(struct ptlrpc_request_pool *, int);
775 #include <lustre_nrs.h>
778 * Basic request prioritization operations structure.
779 * The whole idea is centered around locks and RPCs that might affect locks.
780 * When a lock is contended we try to give priority to RPCs that might lead
781 * to fastest release of that lock.
782 * Currently only implemented for OSTs only in a way that makes all
783 * IO and truncate RPCs that are coming from a locked region where a lock is
784 * contended a priority over other requests.
786 struct ptlrpc_hpreq_ops {
788 * Check if the lock handle of the given lock is the same as
789 * taken from the request.
791 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
793 * Check if the request is a high priority one.
795 int (*hpreq_check)(struct ptlrpc_request *);
797 * Called after the request has been handled.
799 void (*hpreq_fini)(struct ptlrpc_request *);
802 struct ptlrpc_cli_req {
803 /** For bulk requests on client only: bulk descriptor */
804 struct ptlrpc_bulk_desc *cr_bulk;
805 /** optional time limit for send attempts. This is a timeout
806 * not a timestamp so timeout_t (s32) is used instead of time64_t
808 timeout_t cr_delay_limit;
809 /** time request was first queued */
810 time64_t cr_queued_time;
811 /** request sent in nanoseconds */
813 /** time for request really sent out */
814 time64_t cr_sent_out;
815 /** when req reply unlink must finish. */
816 time64_t cr_reply_deadline;
817 /** when req bulk unlink must finish. */
818 time64_t cr_bulk_deadline;
819 /** when req unlink must finish. */
820 time64_t cr_req_deadline;
821 /** Portal to which this request would be sent */
823 /** Portal where to wait for reply and where reply would be sent */
825 /** request resending number */
826 unsigned int cr_resend_nr;
827 /** What was import generation when this request was sent */
829 enum lustre_imp_state cr_send_state;
830 /** Per-request waitq introduced by bug 21938 for recovery waiting */
831 wait_queue_head_t cr_set_waitq;
832 /** Link item for request set lists */
833 struct list_head cr_set_chain;
834 /** link to waited ctx */
835 struct list_head cr_ctx_chain;
837 /** client's half ctx */
838 struct ptlrpc_cli_ctx *cr_cli_ctx;
839 /** Link back to the request set */
840 struct ptlrpc_request_set *cr_set;
841 /** outgoing request MD handle */
842 struct lnet_handle_md cr_req_md_h;
843 /** request-out callback parameter */
844 struct ptlrpc_cb_id cr_req_cbid;
845 /** incoming reply MD handle */
846 struct lnet_handle_md cr_reply_md_h;
847 wait_queue_head_t cr_reply_waitq;
848 /** reply callback parameter */
849 struct ptlrpc_cb_id cr_reply_cbid;
850 /** Async completion handler, called when reply is received */
851 ptlrpc_interpterer_t cr_reply_interp;
852 /** Resend handler, called when request is resend to update RPC data */
853 ptlrpc_resend_cb_t cr_resend_cb;
854 /** Async completion context */
855 union ptlrpc_async_args cr_async_args;
856 /** Opaq data for replay and commit callbacks. */
858 /** Link to the imp->imp_unreplied_list */
859 struct list_head cr_unreplied_list;
861 * Commit callback, called when request is committed and about to be
864 void (*cr_commit_cb)(struct ptlrpc_request *);
865 /** Replay callback, called after request is replayed at recovery */
866 void (*cr_replay_cb)(struct ptlrpc_request *);
869 /** client request member alias */
870 /* NB: these alias should NOT be used by any new code, instead they should
871 * be removed step by step to avoid potential abuse
873 #define rq_bulk rq_cli.cr_bulk
874 #define rq_delay_limit rq_cli.cr_delay_limit
875 #define rq_queued_time rq_cli.cr_queued_time
876 #define rq_sent_ns rq_cli.cr_sent_ns
877 #define rq_real_sent rq_cli.cr_sent_out
878 #define rq_reply_deadline rq_cli.cr_reply_deadline
879 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
880 #define rq_req_deadline rq_cli.cr_req_deadline
881 #define rq_nr_resend rq_cli.cr_resend_nr
882 #define rq_request_portal rq_cli.cr_req_ptl
883 #define rq_reply_portal rq_cli.cr_rep_ptl
884 #define rq_import_generation rq_cli.cr_imp_gen
885 #define rq_send_state rq_cli.cr_send_state
886 #define rq_set_chain rq_cli.cr_set_chain
887 #define rq_ctx_chain rq_cli.cr_ctx_chain
888 #define rq_set rq_cli.cr_set
889 #define rq_set_waitq rq_cli.cr_set_waitq
890 #define rq_cli_ctx rq_cli.cr_cli_ctx
891 #define rq_req_md_h rq_cli.cr_req_md_h
892 #define rq_req_cbid rq_cli.cr_req_cbid
893 #define rq_reply_md_h rq_cli.cr_reply_md_h
894 #define rq_reply_waitq rq_cli.cr_reply_waitq
895 #define rq_reply_cbid rq_cli.cr_reply_cbid
896 #define rq_interpret_reply rq_cli.cr_reply_interp
897 #define rq_resend_cb rq_cli.cr_resend_cb
898 #define rq_async_args rq_cli.cr_async_args
899 #define rq_cb_data rq_cli.cr_cb_data
900 #define rq_unreplied_list rq_cli.cr_unreplied_list
901 #define rq_commit_cb rq_cli.cr_commit_cb
902 #define rq_replay_cb rq_cli.cr_replay_cb
904 struct ptlrpc_srv_req {
905 /** initial thread servicing this request */
906 struct ptlrpc_thread *sr_svc_thread;
908 * Server side list of incoming unserved requests sorted by arrival
909 * time. Traversed from time to time to notice about to expire
910 * requests and sent back "early replies" to clients to let them
911 * know server is alive and well, just very busy to service their
914 struct list_head sr_timed_list;
915 /** server-side per-export list */
916 struct list_head sr_exp_list;
917 /** server-side history, used for debuging purposes. */
918 struct list_head sr_hist_list;
919 /** history sequence # */
921 /** the index of service's srv_at_array into which request is linked */
925 /** authed uid mapped to */
926 uid_t sr_auth_mapped_uid;
927 /** RPC is generated from what part of Lustre */
928 enum lustre_sec_part sr_sp_from;
929 /** request session context */
930 struct lu_context sr_ses;
934 /** stub for NRS request */
935 struct ptlrpc_nrs_request sr_nrq;
937 /** request arrival time */
938 struct timespec64 sr_arrival_time;
939 /** server's half ctx */
940 struct ptlrpc_svc_ctx *sr_svc_ctx;
941 /** (server side), pointed directly into req buffer */
942 struct ptlrpc_user_desc *sr_user_desc;
943 /** separated reply state, may be vmalloc'd */
944 struct ptlrpc_reply_state *sr_reply_state;
945 /** server-side hp handlers */
946 struct ptlrpc_hpreq_ops *sr_ops;
947 /** incoming request buffer */
948 struct ptlrpc_request_buffer_desc *sr_rqbd;
951 /** server request member alias */
952 /* NB: these alias should NOT be used by any new code, instead they should
953 * be removed step by step to avoid potential abuse
955 #define rq_svc_thread rq_srv.sr_svc_thread
956 #define rq_timed_list rq_srv.sr_timed_list
957 #define rq_exp_list rq_srv.sr_exp_list
958 #define rq_history_list rq_srv.sr_hist_list
959 #define rq_history_seq rq_srv.sr_hist_seq
960 #define rq_at_index rq_srv.sr_at_index
961 #define rq_auth_uid rq_srv.sr_auth_uid
962 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
963 #define rq_sp_from rq_srv.sr_sp_from
964 #define rq_session rq_srv.sr_ses
965 #define rq_nrq rq_srv.sr_nrq
966 #define rq_arrival_time rq_srv.sr_arrival_time
967 #define rq_reply_state rq_srv.sr_reply_state
968 #define rq_svc_ctx rq_srv.sr_svc_ctx
969 #define rq_user_desc rq_srv.sr_user_desc
970 #define rq_ops rq_srv.sr_ops
971 #define rq_rqbd rq_srv.sr_rqbd
972 #define rq_reqmsg rq_pill.rc_reqmsg
973 #define rq_repmsg rq_pill.rc_repmsg
974 #define rq_req_swab_mask rq_pill.rc_req_swab_mask
975 #define rq_rep_swab_mask rq_pill.rc_rep_swab_mask
978 * Represents remote procedure call.
980 * This is a staple structure used by everybody wanting to send a request
983 struct ptlrpc_request {
984 /* Request type: one of PTL_RPC_MSG_* */
986 /** Result of request processing */
989 * Linkage item through which this request is included into
990 * sending/delayed lists on client and into rqbd list on server
992 struct list_head rq_list;
993 /** Lock to protect request flags and some other important bits, like
997 spinlock_t rq_early_free_lock;
998 /** client-side flags are serialized by rq_lock @{ */
999 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1000 rq_timedout:1, rq_resend:1, rq_restart:1,
1002 * when ->rq_replay is set, request is kept by the client even
1003 * after server commits corresponding transaction. This is
1004 * used for operations that require sequence of multiple
1005 * requests to be replayed. The only example currently is file
1006 * open/close. When last request in such a sequence is
1007 * committed, ->rq_replay is cleared on all requests in the
1011 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1012 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1014 rq_req_unlinked:1, /* unlinked request buffer from lnet */
1015 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
1016 rq_memalloc:1, /* req originated from "kswapd" */
1018 rq_reply_truncated:1,
1019 /** whether the "rq_set" is a valid one */
1021 rq_generation_set:1,
1022 /** do not resend request on -EINPROGRESS */
1023 rq_no_retry_einprogress:1,
1024 /* allow req to be sent if the import is in recovery status */
1026 /* bulk request, sent to server, but uncommitted */
1028 rq_early_free_repbuf:1, /* free reply buffer in advance */
1032 /** server-side flags are serialized by rq_lock @{ */
1034 rq_hp:1, /**< high priority RPC */
1035 rq_at_linked:1, /**< link into service's srv_at_array */
1036 rq_packed_final:1, /**< packed final reply */
1037 rq_obsolete:1; /* aborted by a signal on a client */
1040 /** one of RQ_PHASE_* */
1041 enum rq_phase rq_phase;
1042 /** one of RQ_PHASE_* to be used next */
1043 enum rq_phase rq_next_phase;
1045 * client-side refcount for SENT race, server-side refcounf
1046 * for multiple replies
1048 atomic_t rq_refcount;
1051 * !rq_truncate : # reply bytes actually received,
1052 * rq_truncate : required repbuf_len for resend
1054 int rq_nob_received;
1055 /** Request length */
1059 /** Pool if request is from preallocated list */
1060 struct ptlrpc_request_pool *rq_pool;
1061 /** Transaction number */
1065 /** bulk match bits */
1067 /** reply match bits */
1070 * List item to for replay list. Not yet committed requests get linked
1072 * Also see \a rq_replay comment above.
1073 * It's also link chain on obd_export::exp_req_replay_queue
1075 struct list_head rq_replay_list;
1076 /** non-shared members for client & server request*/
1078 struct ptlrpc_cli_req rq_cli;
1079 struct ptlrpc_srv_req rq_srv;
1081 /* security and encryption data */
1082 /* description of flavors for client & server */
1083 struct sptlrpc_flavor rq_flvr;
1085 /* client/server security flags */
1087 rq_ctx_init:1, /* context initiation */
1088 rq_ctx_fini:1, /* context destroy */
1089 rq_bulk_read:1, /* request bulk read */
1090 rq_bulk_write:1, /* request bulk write */
1091 /* server authentication flags */
1092 rq_auth_gss:1, /* authenticated by gss */
1093 rq_auth_usr_root:1, /* authed as root */
1094 rq_auth_usr_mdt:1, /* authed as mdt */
1095 rq_auth_usr_ost:1, /* authed as ost */
1096 /* security tfm flags */
1099 /* doesn't expect reply FIXME */
1101 rq_pill_init:1, /* pill initialized */
1102 rq_srv_req:1; /* server request */
1104 /** various buffer pointers */
1105 struct lustre_msg *rq_reqbuf; /**< req wrapper, vmalloc*/
1106 char *rq_repbuf; /**< rep buffer, vmalloc */
1107 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
1108 /** only in priv mode */
1109 struct lustre_msg *rq_clrbuf;
1110 int rq_reqbuf_len; /* req wrapper buf len */
1111 int rq_reqdata_len; /* req wrapper msg len */
1112 int rq_repbuf_len; /* rep buffer len */
1113 int rq_repdata_len; /* rep wrapper msg len */
1114 int rq_clrbuf_len; /* only in priv mode */
1115 int rq_clrdata_len; /* only in priv mode */
1117 /** early replies go to offset 0, regular replies go after that */
1118 unsigned int rq_reply_off;
1121 /** how many early replies (for stats) */
1123 /** Server-side, export on which request was received */
1124 struct obd_export *rq_export;
1125 /** import where request is being sent */
1126 struct obd_import *rq_import;
1128 struct lnet_nid rq_self;
1129 /** Peer description (the other side) */
1130 struct lnet_processid rq_peer;
1131 /** Descriptor for the NID from which the peer sent the request. */
1132 struct lnet_processid rq_source;
1134 * service time estimate (secs)
1135 * If the request is not served by this time, it is marked as timed out.
1136 * Do not change to time64_t since this is transmitted over the wire.
1138 * The linux kernel handles timestamps with time64_t and timeouts
1139 * are normally done with jiffies. Lustre shares the rq_timeout between
1140 * nodes. Since jiffies can vary from node to node Lustre instead
1141 * will express the timeout value in seconds. To avoid confusion with
1142 * timestamps (time64_t) and jiffy timeouts (long) Lustre timeouts
1143 * are expressed in s32 (timeout_t). Also what is transmitted over
1144 * the wire is 32 bits.
1146 timeout_t rq_timeout;
1148 * when request/reply sent (secs), or time when request should be sent
1151 /** when request must finish. */
1152 time64_t rq_deadline;
1153 /** request format description */
1154 struct req_capsule rq_pill;
1158 * Call completion handler for rpc if any, return it's status or original
1159 * rc if there was no handler defined for this request.
1161 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1162 struct ptlrpc_request *req, int rc)
1164 if (req->rq_interpret_reply != NULL) {
1165 req->rq_status = req->rq_interpret_reply(env, req,
1166 &req->rq_async_args,
1168 return req->rq_status;
1177 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1180 * Can the request be moved from the regular NRS head to the high-priority NRS
1181 * head (of the same PTLRPC service partition), if any?
1183 * For a reliable result, this should be checked under svcpt->scp_req lock.
1185 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1187 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1190 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1191 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1192 * to make sure it has not been scheduled yet (analogous to previous
1193 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1195 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1199 static inline bool req_capsule_ptlreq(struct req_capsule *pill)
1201 struct ptlrpc_request *req = pill->rc_req;
1203 return req != NULL && pill == &req->rq_pill;
1206 static inline bool req_capsule_subreq(struct req_capsule *pill)
1208 struct ptlrpc_request *req = pill->rc_req;
1210 return req == NULL || pill != &req->rq_pill;
1214 * Returns true if request needs to be swabbed into local cpu byteorder
1216 static inline bool req_capsule_req_need_swab(struct req_capsule *pill)
1218 struct ptlrpc_request *req = pill->rc_req;
1220 return req && req_capsule_req_swabbed(&req->rq_pill,
1221 MSG_PTLRPC_HEADER_OFF);
1225 * Returns true if request reply needs to be swabbed into local cpu byteorder
1227 static inline bool req_capsule_rep_need_swab(struct req_capsule *pill)
1229 struct ptlrpc_request *req = pill->rc_req;
1231 return req && req_capsule_rep_swabbed(&req->rq_pill,
1232 MSG_PTLRPC_HEADER_OFF);
1236 * Convert numerical request phase value \a phase into text string description
1238 static inline const char *
1239 ptlrpc_phase2str(enum rq_phase phase)
1248 case RQ_PHASE_INTERPRET:
1250 case RQ_PHASE_COMPLETE:
1252 case RQ_PHASE_UNREG_RPC:
1254 case RQ_PHASE_UNREG_BULK:
1262 * Convert numerical request phase of the request \a req into text stringi
1265 static inline const char *
1266 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1268 return ptlrpc_phase2str(req->rq_phase);
1272 * Debugging functions and helpers to print request structure into debug log
1275 /* Spare the preprocessor, spoil the bugs. */
1276 #define FLAG(field, str) (field ? str : "")
1278 /** Convert bit flags into a string */
1279 #define DEBUG_REQ_FLAGS(req) \
1280 ptlrpc_rqphase2str(req), \
1281 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1282 FLAG(req->rq_err, "E"), FLAG(req->rq_net_err, "e"), \
1283 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1284 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1285 FLAG(req->rq_no_resend, "N"), FLAG(req->rq_no_reply, "n"), \
1286 FLAG(req->rq_waiting, "W"), \
1287 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1288 FLAG(req->rq_committed, "M"), \
1289 FLAG(req->rq_req_unlinked, "Q"), \
1290 FLAG(req->rq_reply_unlinked, "U"), \
1291 FLAG(req->rq_receiving_reply, "r")
1293 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s"
1295 __printf(3, 4) /* __attribute__ */
1296 void _debug_req(struct ptlrpc_request *req,
1297 struct libcfs_debug_msg_data *data, const char *fmt, ...);
1300 * Helper that decides if we need to print request accordig to current debug
1303 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1305 if (((mask) & D_CANTMASK) != 0 || \
1306 ((libcfs_debug & (mask)) != 0 && \
1307 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1308 _debug_req((req), msgdata, fmt, ##a); \
1312 * This is the debug print function you need to use to print request sturucture
1313 * content into lustre debug log.
1314 * for most callers (level is a constant) this is resolved at compile time
1316 #define DEBUG_REQ(level, req, fmt, args...) \
1318 if ((level) & (D_ERROR | D_WARNING)) { \
1319 static struct cfs_debug_limit_state cdls; \
1320 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1321 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1323 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1324 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1329 enum ptlrpc_bulk_op_type {
1330 PTLRPC_BULK_OP_ACTIVE = 0x00000001,
1331 PTLRPC_BULK_OP_PASSIVE = 0x00000002,
1332 PTLRPC_BULK_OP_PUT = 0x00000004,
1333 PTLRPC_BULK_OP_GET = 0x00000008,
1334 PTLRPC_BULK_GET_SOURCE = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_GET,
1335 PTLRPC_BULK_PUT_SINK = PTLRPC_BULK_OP_PASSIVE | PTLRPC_BULK_OP_PUT,
1336 PTLRPC_BULK_GET_SINK = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_GET,
1337 PTLRPC_BULK_PUT_SOURCE = PTLRPC_BULK_OP_ACTIVE | PTLRPC_BULK_OP_PUT,
1340 static inline bool ptlrpc_is_bulk_op_get(enum ptlrpc_bulk_op_type type)
1342 return (type & PTLRPC_BULK_OP_GET) == PTLRPC_BULK_OP_GET;
1345 static inline bool ptlrpc_is_bulk_get_source(enum ptlrpc_bulk_op_type type)
1347 return (type & PTLRPC_BULK_GET_SOURCE) == PTLRPC_BULK_GET_SOURCE;
1350 static inline bool ptlrpc_is_bulk_put_sink(enum ptlrpc_bulk_op_type type)
1352 return (type & PTLRPC_BULK_PUT_SINK) == PTLRPC_BULK_PUT_SINK;
1355 static inline bool ptlrpc_is_bulk_get_sink(enum ptlrpc_bulk_op_type type)
1357 return (type & PTLRPC_BULK_GET_SINK) == PTLRPC_BULK_GET_SINK;
1360 static inline bool ptlrpc_is_bulk_put_source(enum ptlrpc_bulk_op_type type)
1362 return (type & PTLRPC_BULK_PUT_SOURCE) == PTLRPC_BULK_PUT_SOURCE;
1365 static inline bool ptlrpc_is_bulk_op_active(enum ptlrpc_bulk_op_type type)
1367 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1368 (type & PTLRPC_BULK_OP_PASSIVE))
1369 == PTLRPC_BULK_OP_ACTIVE;
1372 static inline bool ptlrpc_is_bulk_op_passive(enum ptlrpc_bulk_op_type type)
1374 return ((type & PTLRPC_BULK_OP_ACTIVE) |
1375 (type & PTLRPC_BULK_OP_PASSIVE))
1376 == PTLRPC_BULK_OP_PASSIVE;
1379 struct ptlrpc_bulk_frag_ops {
1381 * Add a page \a page to the bulk descriptor \a desc
1382 * Data to transfer in the page starts at offset \a pageoffset and
1383 * amount of data to transfer from the page is \a len
1385 void (*add_kiov_frag)(struct ptlrpc_bulk_desc *desc,
1386 struct page *page, int pageoffset, int len);
1389 * Add a \a fragment to the bulk descriptor \a desc.
1390 * Data to transfer in the fragment is pointed to by \a frag
1391 * The size of the fragment is \a len
1393 int (*add_iov_frag)(struct ptlrpc_bulk_desc *desc, void *frag, int len);
1396 * Uninitialize and free bulk descriptor \a desc.
1397 * Works on bulk descriptors both from server and client side.
1399 void (*release_frags)(struct ptlrpc_bulk_desc *desc);
1402 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops;
1403 extern const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops;
1406 * Definition of bulk descriptor.
1407 * Bulks are special "Two phase" RPCs where initial request message
1408 * is sent first and it is followed bt a transfer (o receiving) of a large
1409 * amount of data to be settled into pages referenced from the bulk descriptors.
1410 * Bulks transfers (the actual data following the small requests) are done
1411 * on separate LNet portals.
1412 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1413 * Another user is readpage for MDT.
1415 struct ptlrpc_bulk_desc {
1416 unsigned int bd_refs; /* number MD's assigned including zero-sends */
1417 /** completed with failure */
1418 unsigned long bd_failure:1;
1420 unsigned short bd_md_offset; /* offset in 4k pages ranged [0, 15] */
1421 unsigned int bd_registered:1,
1422 /* bulk request is RDMA transfer, use page->host as real address */
1424 /** For serialization with callback */
1426 /** {put,get}{source,sink}{kvec,kiov} */
1427 enum ptlrpc_bulk_op_type bd_type;
1428 /** LNet portal for this bulk */
1430 /** Server side - export this bulk created for */
1431 struct obd_export *bd_export;
1432 /** Client side - import this bulk was sent on */
1433 struct obd_import *bd_import;
1434 /** Back pointer to the request */
1435 struct ptlrpc_request *bd_req;
1436 const struct ptlrpc_bulk_frag_ops *bd_frag_ops;
1437 wait_queue_head_t bd_waitq; /* server side only WQ */
1438 int bd_iov_count; /* # entries in bd_iov */
1439 int bd_max_iov; /* allocated size of bd_iov */
1440 int bd_nob; /* # bytes covered */
1441 int bd_nob_transferred; /* # bytes GOT/PUT */
1442 unsigned int bd_iop_len; /* md iop bytes */
1444 __u64 bd_last_mbits;
1446 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1447 struct lnet_nid bd_sender; /* stash event::sender */
1448 int bd_md_count; /* # valid entries in bd_mds */
1449 int bd_md_max_brw; /* max entries in bd_mds */
1451 /** array of offsets for each MD */
1452 unsigned int bd_mds_off[PTLRPC_BULK_OPS_LIMIT];
1453 /** array of associated MDs */
1454 struct lnet_handle_md bd_mds[PTLRPC_BULK_OPS_LIMIT];
1456 /* encrypted iov, size is either 0 or bd_iov_count. */
1457 struct bio_vec *bd_enc_vec;
1458 struct bio_vec *bd_vec;
1461 enum ptlrpc_thread_state {
1463 SVC_STOPPED = BIT(0),
1464 SVC_STOPPING = BIT(1),
1465 SVC_STARTING = BIT(2),
1466 SVC_RUNNING = BIT(3),
1467 SVC_WATCHDOG = BIT(4),
1470 #define PTLRPC_THR_NAME_LEN 32
1472 * Definition of server service thread structure
1474 struct ptlrpc_thread {
1475 struct list_head t_link; /* active threads svcpt->scp_threads */
1476 void *t_data; /* thread-private memory (vmalloc) */
1477 enum ptlrpc_thread_state t_flags;
1478 unsigned int t_id; /* index, from ptlrpc_start_threads */
1479 struct task_struct *t_task; /* service thread */
1480 pid_t t_pid; /* kernel process ID */
1481 ktime_t t_touched; /* last active time */
1482 struct delayed_work t_watchdog; /* inactivity watchdog timer */
1483 struct ptlrpc_service_part *t_svcpt; /* service the thread belongs to */
1484 wait_queue_head_t t_ctl_waitq;/* waiter for thread starter */
1485 struct lu_env *t_env;
1486 char t_name[PTLRPC_THR_NAME_LEN];
1489 static inline int thread_is_init(struct ptlrpc_thread *thread)
1491 return thread->t_flags == 0;
1494 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1496 return !!(thread->t_flags & SVC_STOPPED);
1499 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1501 return !!(thread->t_flags & SVC_STOPPING);
1504 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1506 return !!(thread->t_flags & SVC_STARTING);
1509 static inline int thread_is_running(struct ptlrpc_thread *thread)
1511 return !!(thread->t_flags & SVC_RUNNING);
1514 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1516 thread->t_flags &= ~flags;
1519 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1521 thread->t_flags = flags;
1524 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1526 thread->t_flags |= flags;
1529 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1532 if (thread->t_flags & flags) {
1533 thread->t_flags &= ~flags;
1540 * Request buffer descriptor structure.
1541 * This is a structure that contains one posted request buffer for service.
1542 * Once data land into a buffer, event callback creates actual request and
1543 * notifies wakes one of the service threads to process new incoming request.
1544 * More than one request can fit into the buffer.
1546 struct ptlrpc_request_buffer_desc {
1547 /** Link item for rqbds on a service */
1548 struct list_head rqbd_list;
1549 /** History of requests for this buffer */
1550 struct list_head rqbd_reqs;
1551 /** Back pointer to service for which this buffer is registered */
1552 struct ptlrpc_service_part *rqbd_svcpt;
1553 /** LNet descriptor */
1554 struct lnet_handle_md rqbd_md_h;
1556 /** The buffer itself */
1558 struct ptlrpc_cb_id rqbd_cbid;
1560 * This "embedded" request structure is only used for the
1561 * last request to fit into the buffer
1563 struct ptlrpc_request rqbd_req;
1566 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1568 struct ptlrpc_service_ops {
1570 * if non-NULL called during thread creation (ptlrpc_start_thread())
1571 * to initialize service specific per-thread state.
1573 int (*so_thr_init)(struct ptlrpc_thread *thr);
1575 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1576 * destruct state created by ->srv_init().
1578 void (*so_thr_done)(struct ptlrpc_thread *thr);
1580 * Handler function for incoming requests for this service
1582 int (*so_req_handler)(struct ptlrpc_request *req);
1584 * function to determine priority of the request, it's called
1585 * on every new request
1587 int (*so_hpreq_handler)(struct ptlrpc_request *);
1589 * service-specific print fn
1591 void (*so_req_printer)(void *, struct ptlrpc_request *);
1594 #ifndef __cfs_cacheline_aligned
1595 /* NB: put it here for reducing patche dependence */
1596 # define __cfs_cacheline_aligned
1600 * How many high priority requests to serve before serving one normal
1603 #define PTLRPC_SVC_HP_RATIO 10
1606 * Definition of PortalRPC service.
1607 * The service is listening on a particular portal (like tcp port)
1608 * and perform actions for a specific server like IO service for OST
1609 * or general metadata service for MDS.
1611 struct ptlrpc_service {
1612 /** serialize /proc operations */
1613 spinlock_t srv_lock;
1614 /** most often accessed fields */
1615 /** chain thru all services */
1616 struct list_head srv_list;
1617 /** service operations table */
1618 struct ptlrpc_service_ops srv_ops;
1619 /** only statically allocated strings here; we don't clean them */
1621 /** only statically allocated strings here; we don't clean them */
1622 char *srv_thread_name;
1623 /** threads # should be created for each partition on initializing */
1624 int srv_nthrs_cpt_init;
1625 /** limit of threads number for each partition */
1626 int srv_nthrs_cpt_limit;
1627 /** Root of debugfs dir tree for this service */
1628 struct dentry *srv_debugfs_entry;
1629 /** Pointer to statistic data for this service */
1630 struct lprocfs_stats *srv_stats;
1631 /** # hp per lp reqs to handle */
1632 int srv_hpreq_ratio;
1633 /** biggest request to receive */
1634 int srv_max_req_size;
1635 /** biggest reply to send */
1636 int srv_max_reply_size;
1637 /** size of individual buffers */
1639 /** # buffers to allocate in 1 group */
1640 int srv_nbuf_per_group;
1641 /** Local portal on which to receive requests */
1642 __u32 srv_req_portal;
1643 /** Portal on the client to send replies to */
1644 __u32 srv_rep_portal;
1646 * Tags for lu_context associated with this thread, see struct
1650 /** soft watchdog timeout multiplier */
1651 int srv_watchdog_factor;
1652 /** under unregister_service */
1653 unsigned srv_is_stopping:1;
1654 /** Whether or not to restrict service threads to CPUs in this CPT */
1655 unsigned srv_cpt_bind:1;
1657 /** max # request buffers */
1659 /** max # request buffers in history per partition */
1660 int srv_hist_nrqbds_cpt_max;
1661 /** number of CPTs this service associated with */
1663 /** CPTs array this service associated with */
1665 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1667 /** CPT table this service is running over */
1668 struct cfs_cpt_table *srv_cptable;
1671 struct kobject srv_kobj;
1672 struct completion srv_kobj_unregister;
1674 * partition data for ptlrpc service
1676 struct ptlrpc_service_part *srv_parts[0];
1680 * Definition of PortalRPC service partition data.
1681 * Although a service only has one instance of it right now, but we
1682 * will have multiple instances very soon (instance per CPT).
1684 * it has four locks:
1686 * serialize operations on rqbd and requests waiting for preprocess
1688 * serialize operations active requests sent to this portal
1690 * serialize adaptive timeout stuff
1692 * serialize operations on RS list (reply states)
1694 * We don't have any use-case to take two or more locks at the same time
1695 * for now, so there is no lock order issue.
1697 struct ptlrpc_service_part {
1698 /** back reference to owner */
1699 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1700 /* CPT id, reserved */
1702 /** always increasing number */
1704 /** # of starting threads */
1705 int scp_nthrs_starting;
1706 /** # running threads */
1707 int scp_nthrs_running;
1708 /** service threads list */
1709 struct list_head scp_threads;
1712 * serialize the following fields, used for protecting
1713 * rqbd list and incoming requests waiting for preprocess,
1714 * threads starting & stopping are also protected by this lock.
1716 spinlock_t scp_lock __cfs_cacheline_aligned;
1717 /** userland serialization */
1718 struct mutex scp_mutex;
1719 /** total # req buffer descs allocated */
1720 int scp_nrqbds_total;
1721 /** # posted request buffers for receiving */
1722 int scp_nrqbds_posted;
1723 /** in progress of allocating rqbd */
1724 int scp_rqbd_allocating;
1725 /** # incoming reqs */
1726 int scp_nreqs_incoming;
1727 /** request buffers to be reposted */
1728 struct list_head scp_rqbd_idle;
1729 /** req buffers receiving */
1730 struct list_head scp_rqbd_posted;
1731 /** incoming reqs */
1732 struct list_head scp_req_incoming;
1733 /** timeout before re-posting reqs, in jiffies */
1734 long scp_rqbd_timeout;
1736 * all threads sleep on this. This wait-queue is signalled when new
1737 * incoming request arrives and when difficult reply has to be handled.
1739 wait_queue_head_t scp_waitq;
1741 /** request history */
1742 struct list_head scp_hist_reqs;
1743 /** request buffer history */
1744 struct list_head scp_hist_rqbds;
1745 /** # request buffers in history */
1746 int scp_hist_nrqbds;
1747 /** sequence number for request */
1749 /** highest seq culled from history */
1750 __u64 scp_hist_seq_culled;
1753 * serialize the following fields, used for processing requests
1754 * sent to this portal
1756 spinlock_t scp_req_lock __cfs_cacheline_aligned;
1757 /** # reqs in either of the NRS heads below */
1758 /** # reqs being served */
1759 int scp_nreqs_active;
1760 /** # HPreqs being served */
1761 int scp_nhreqs_active;
1762 /** # hp requests handled */
1765 /** NRS head for regular requests */
1766 struct ptlrpc_nrs scp_nrs_reg;
1767 /** NRS head for HP reqs; valid for services that can handle HP reqs */
1768 struct ptlrpc_nrs *scp_nrs_hp;
1773 * serialize the following fields, used for changes on
1776 spinlock_t scp_at_lock __cfs_cacheline_aligned;
1777 /** estimated rpc service time */
1778 struct adaptive_timeout scp_at_estimate;
1779 /** reqs waiting for replies */
1780 struct ptlrpc_at_array scp_at_array;
1781 /** early reply timer */
1782 struct timer_list scp_at_timer;
1784 ktime_t scp_at_checktime;
1785 /** check early replies */
1786 unsigned int scp_at_check;
1790 * serialize the following fields, used for processing
1791 * replies for this portal
1793 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1794 /** all the active replies */
1795 struct list_head scp_rep_active;
1796 /** List of free reply_states */
1797 struct list_head scp_rep_idle;
1798 /** waitq to run, when adding stuff to srv_free_rs_list */
1799 wait_queue_head_t scp_rep_waitq;
1800 /** # 'difficult' replies */
1801 atomic_t scp_nreps_difficult;
1804 #define ptlrpc_service_for_each_part(part, i, svc) \
1806 i < (svc)->srv_ncpts && \
1807 (svc)->srv_parts != NULL && \
1808 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1811 * Declaration of ptlrpcd control structure
1813 struct ptlrpcd_ctl {
1815 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1817 unsigned long pc_flags;
1819 * Thread lock protecting structure fields.
1825 struct completion pc_starting;
1829 struct completion pc_finishing;
1831 * Thread requests set.
1833 struct ptlrpc_request_set *pc_set;
1835 * Thread name used in kthread_run()
1839 * CPT the thread is bound on.
1843 * Index of ptlrpcd thread in the array.
1847 * Pointer to the array of partners' ptlrpcd_ctl structure.
1849 struct ptlrpcd_ctl **pc_partners;
1851 * Number of the ptlrpcd's partners.
1855 * Record the partner index to be processed next.
1859 * Error code if the thread failed to fully start.
1864 /* Bits for pc_flags */
1865 enum ptlrpcd_ctl_flags {
1867 * Ptlrpc thread start flag.
1869 LIOD_START = BIT(0),
1871 * Ptlrpc thread stop flag.
1875 * Ptlrpc thread force flag (only stop force so far).
1876 * This will cause aborting any inflight rpcs handled
1877 * by thread if LIOD_STOP is specified.
1879 LIOD_FORCE = BIT(2),
1881 * This is a recovery ptlrpc thread.
1883 LIOD_RECOVERY = BIT(3),
1890 * Service compatibility function; the policy is compatible with all services.
1892 * \param[in] svc The service the policy is attempting to register with.
1893 * \param[in] desc The policy descriptor
1895 * \retval true The policy is compatible with the service
1897 * \see ptlrpc_nrs_pol_desc::pd_compat()
1899 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
1900 const struct ptlrpc_nrs_pol_desc *desc)
1906 * Service compatibility function; the policy is compatible with only a specific
1907 * service which is identified by its human-readable name at
1908 * ptlrpc_service::srv_name.
1910 * \param[in] svc The service the policy is attempting to register with.
1911 * \param[in] desc The policy descriptor
1913 * \retval false The policy is not compatible with the service
1914 * \retval true The policy is compatible with the service
1916 * \see ptlrpc_nrs_pol_desc::pd_compat()
1918 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
1919 const struct ptlrpc_nrs_pol_desc *desc)
1921 LASSERT(desc->pd_compat_svc_name != NULL);
1922 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
1927 /* ptlrpc/events.c */
1928 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1929 struct lnet_processid *peer,
1930 struct lnet_nid *self,
1933 * These callbacks are invoked by LNet when something happened to
1937 extern void request_out_callback(struct lnet_event *ev);
1938 extern void reply_in_callback(struct lnet_event *ev);
1939 extern void client_bulk_callback(struct lnet_event *ev);
1940 extern void request_in_callback(struct lnet_event *ev);
1941 extern void reply_out_callback(struct lnet_event *ev);
1942 #ifdef HAVE_SERVER_SUPPORT
1943 extern void server_bulk_callback(struct lnet_event *ev);
1947 /* ptlrpc/connection.c */
1948 struct ptlrpc_connection *ptlrpc_connection_get(struct lnet_processid *peer,
1949 struct lnet_nid *self,
1950 struct obd_uuid *uuid);
1952 static inline void ptlrpc_connection_put(struct ptlrpc_connection *conn)
1957 LASSERT(atomic_read(&conn->c_refcount) > 0);
1960 * We do not remove connection from hashtable and
1961 * do not free it even if last caller released ref,
1962 * as we want to have it cached for the case it is
1965 * Deallocating it and later creating new connection
1966 * again would be wastful. This way we also avoid
1967 * expensive locking to protect things from get/put
1968 * race when found cached connection is freed by
1969 * ptlrpc_connection_put().
1971 * It will be freed later in module unload time,
1972 * when ptlrpc_connection_fini()->lh_exit->conn_exit()
1975 atomic_dec(&conn->c_refcount);
1977 CDEBUG(D_INFO, "PUT conn=%p refcount %d to %s\n",
1978 conn, atomic_read(&conn->c_refcount),
1979 libcfs_nidstr(&conn->c_peer.nid));
1982 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1983 int ptlrpc_connection_init(void);
1984 void ptlrpc_connection_fini(void);
1985 extern lnet_pid_t ptl_get_pid(void);
1988 * Check if the peer connection is on the local node. We need to use GFP_NOFS
1989 * for requests from a local client to avoid recursing into the filesystem
1990 * as we might end up waiting on a page sent in the request we're serving.
1992 * Use __GFP_HIGHMEM so that the pages can use all of the available memory
1993 * on 32-bit machines. Use more aggressive GFP_HIGHUSER flags from non-local
1994 * clients to be able to generate more memory pressure on the OSS and allow
1995 * inactive pages to be reclaimed, since it doesn't have any other processes
1996 * or allocations that generate memory reclaim pressure.
1998 * See b=17576 (bdf50dc9) and b=19529 (3dcf18d3) for details.
2000 static inline bool ptlrpc_connection_is_local(struct ptlrpc_connection *conn)
2005 if (nid_same(&conn->c_peer.nid, &conn->c_self))
2008 RETURN(LNetIsPeerLocal(&conn->c_peer.nid));
2011 /* ptlrpc/niobuf.c */
2013 * Actual interfacing with LNet to put/get/register/unregister stuff
2016 #ifdef HAVE_SERVER_SUPPORT
2017 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2018 unsigned int nfrags,
2019 unsigned int max_brw,
2021 unsigned int portal,
2022 const struct ptlrpc_bulk_frag_ops
2024 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2025 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2027 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2031 LASSERT(desc != NULL);
2033 spin_lock(&desc->bd_lock);
2035 spin_unlock(&desc->bd_lock);
2040 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2041 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2043 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2045 struct ptlrpc_bulk_desc *desc;
2048 LASSERT(req != NULL);
2049 desc = req->rq_bulk;
2054 if (req->rq_bulk_deadline > ktime_get_real_seconds())
2058 spin_lock(&desc->bd_lock);
2060 spin_unlock(&desc->bd_lock);
2064 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2065 #define PTLRPC_REPLY_EARLY 0x02
2066 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2067 int ptlrpc_reply(struct ptlrpc_request *req);
2068 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2069 int ptlrpc_error(struct ptlrpc_request *req);
2070 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2071 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2072 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2075 /* ptlrpc/client.c */
2077 * Client-side portals API. Everything to send requests, receive replies,
2078 * request queues, request management, etc.
2081 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2083 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
2084 struct ptlrpc_client *pc);
2085 void ptlrpc_cleanup_client(struct obd_import *imp);
2086 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
2089 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2090 int ptlrpc_replay_req(struct ptlrpc_request *req);
2091 void ptlrpc_restart_req(struct ptlrpc_request *req);
2092 void ptlrpc_abort_inflight(struct obd_import *imp);
2093 void ptlrpc_cleanup_imp(struct obd_import *imp);
2094 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2096 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2097 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2099 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2100 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set);
2101 void ptlrpc_set_destroy(struct ptlrpc_request_set *set);
2102 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
2103 struct ptlrpc_request *req);
2104 #define PTLRPCD_SET ((struct ptlrpc_request_set *)1)
2106 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2107 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2109 struct ptlrpc_request_pool *
2110 ptlrpc_init_rq_pool(int num_rq, int msgsiz,
2111 int (*populate_pool)(struct ptlrpc_request_pool *, int));
2113 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2114 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2115 const struct req_format *format);
2116 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2117 struct ptlrpc_request_pool *pr,
2118 const struct req_format *fmat);
2119 void ptlrpc_request_free(struct ptlrpc_request *request);
2120 int ptlrpc_request_pack(struct ptlrpc_request *request, __u32 version,
2122 struct ptlrpc_request *
2123 ptlrpc_request_alloc_pack(struct obd_import *imp,
2124 const struct req_format *format,
2125 __u32 version, int opcode);
2126 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2127 __u32 version, int opcode, char **bufs,
2128 struct ptlrpc_cli_ctx *ctx);
2129 #define ptlrpc_req_finished(rq) ptlrpc_req_put(rq)
2130 void ptlrpc_req_put(struct ptlrpc_request *request);
2131 void ptlrpc_req_put_with_imp_lock(struct ptlrpc_request *request);
2132 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2133 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2134 unsigned int nfrags,
2135 unsigned int max_brw, unsigned int type,
2136 unsigned int portal,
2137 const struct ptlrpc_bulk_frag_ops *ops);
2139 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2140 struct page *page, int pageoffset, int len,
2143 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk);
2145 static inline void ptlrpc_release_bulk_noop(struct ptlrpc_bulk_desc *desc)
2149 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2150 struct obd_import *imp);
2151 __u64 ptlrpc_next_xid(void);
2152 __u64 ptlrpc_sample_next_xid(void);
2153 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2154 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req);
2155 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req);
2157 /* Set of routines to run a function in ptlrpcd context */
2158 void *ptlrpcd_alloc_work(struct obd_import *imp,
2159 int (*cb)(const struct lu_env *, void *), void *data);
2160 void ptlrpcd_destroy_work(void *handler);
2161 int ptlrpcd_queue_work(void *handler);
2164 struct ptlrpc_service_buf_conf {
2165 /* nbufs is buffers # to allocate when growing the pool */
2166 unsigned int bc_nbufs;
2167 /* buffer size to post */
2168 unsigned int bc_buf_size;
2169 /* portal to listed for requests on */
2170 unsigned int bc_req_portal;
2171 /* portal of where to send replies to */
2172 unsigned int bc_rep_portal;
2173 /* maximum request size to be accepted for this service */
2174 unsigned int bc_req_max_size;
2175 /* maximum reply size this service can ever send */
2176 unsigned int bc_rep_max_size;
2179 struct ptlrpc_service_thr_conf {
2180 /* threadname should be 8 characters or less - 6 will be added on */
2182 /* threads increasing factor for each CPU */
2183 unsigned int tc_thr_factor;
2184 /* service threads # to start on each partition while initializing */
2185 unsigned int tc_nthrs_init;
2187 * low water of threads # upper-limit on each partition while running,
2188 * service availability may be impacted if threads number is lower
2189 * than this value. It can be ZERO if the service doesn't require
2190 * CPU affinity or there is only one partition.
2192 unsigned int tc_nthrs_base;
2193 /* "soft" limit for total threads number */
2194 unsigned int tc_nthrs_max;
2195 /* user specified threads number, it will be validated due to
2196 * other members of this structure.
2198 unsigned int tc_nthrs_user;
2199 /* bind service threads to only CPUs in their associated CPT */
2200 unsigned int tc_cpu_bind;
2201 /* Tags for lu_context associated with service thread */
2205 struct ptlrpc_service_cpt_conf {
2206 struct cfs_cpt_table *cc_cptable;
2207 /* string pattern to describe CPTs for a service */
2209 /* whether or not to have per-CPT service partitions */
2213 struct ptlrpc_service_conf {
2216 /* soft watchdog timeout multiplifier to print stuck service traces */
2217 unsigned int psc_watchdog_factor;
2218 /* buffer information */
2219 struct ptlrpc_service_buf_conf psc_buf;
2220 /* thread information */
2221 struct ptlrpc_service_thr_conf psc_thr;
2222 /* CPU partition information */
2223 struct ptlrpc_service_cpt_conf psc_cpt;
2224 /* function table */
2225 struct ptlrpc_service_ops psc_ops;
2228 /* ptlrpc/service.c */
2230 * Server-side services API. Register/unregister service, request state
2231 * management, service thread management
2235 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
2237 void ptlrpc_commit_replies(struct obd_export *exp);
2238 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2239 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2240 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2241 struct ptlrpc_service *ptlrpc_register_service(
2242 struct ptlrpc_service_conf *conf,
2243 struct kset *parent,
2244 struct dentry *debugfs_entry);
2246 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2247 int ptlrpc_service_health_check(struct ptlrpc_service *service);
2248 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2249 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2250 struct obd_export *export);
2251 void ptlrpc_update_export_timer(struct obd_export *exp,
2252 time64_t extra_delay);
2254 int ptlrpc_hr_init(void);
2255 void ptlrpc_hr_fini(void);
2257 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout);
2258 void ptlrpc_watchdog_disable(struct delayed_work *work);
2259 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout);
2260 void ptlrpc_watchdog_delete(struct delayed_work *work);
2264 /* ptlrpc/import.c */
2269 int ptlrpc_connect_import(struct obd_import *imp);
2270 int ptlrpc_connect_import_locked(struct obd_import *imp);
2271 int ptlrpc_init_import(struct obd_import *imp);
2272 int ptlrpc_disconnect_import_async(struct obd_import *imp, int noclose,
2273 struct completion *a, int *r);
2274 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2275 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp);
2276 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2277 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2279 void ptlrpc_import_enter_resend(struct obd_import *imp);
2280 /* ptlrpc/pack_generic.c */
2281 int ptlrpc_reconnect_import(struct obd_import *imp);
2285 * ptlrpc msg buffer and swab interface
2289 #define PTLRPC_MAX_BUFCOUNT \
2290 (sizeof(((struct ptlrpc_request *)0)->rq_req_swab_mask) * 8)
2291 #define MD_MAX_BUFLEN (MDS_REG_MAXREQSIZE > OUT_MAXREQSIZE ? \
2292 MDS_REG_MAXREQSIZE : OUT_MAXREQSIZE)
2293 #define PTLRPC_MAX_BUFLEN (OST_IO_MAXREQSIZE > MD_MAX_BUFLEN ? \
2294 OST_IO_MAXREQSIZE : MD_MAX_BUFLEN)
2295 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2296 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2298 int lustre_msg_check_version(struct lustre_msg *msg,
2299 enum lustre_msg_version version);
2300 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2302 int lustre_pack_request(struct ptlrpc_request *req, __u32 magic, int count,
2303 __u32 *lens, char **bufs);
2304 int lustre_pack_reply(struct ptlrpc_request *req, int count, __u32 *lens,
2306 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2307 __u32 *lens, char **bufs, int flags);
2308 #define LPRFL_EARLY_REPLY 1
2309 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2310 char **bufs, int flags);
2311 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2312 unsigned int newlen, int move_data);
2313 int lustre_grow_msg(struct lustre_msg *msg, int segment, unsigned int newlen);
2314 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2315 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2316 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
2317 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2318 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
2319 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
2320 extern __u32 lustre_msg_early_size;
2321 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
2322 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
2323 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
2324 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
2325 __u32 lustre_msg_bufcount(struct lustre_msg *m);
2326 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
2327 enum lustre_msghdr lustre_msghdr_get_flags(struct lustre_msg *msg);
2328 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2329 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2330 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
2331 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
2332 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
2333 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2334 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
2335 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2336 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2337 enum lustre_msg_version lustre_msg_get_version(struct lustre_msg *msg);
2338 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
2339 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2340 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2341 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
2342 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2343 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2344 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2345 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2346 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2347 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2348 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2349 int lustre_msg_get_status(struct lustre_msg *msg);
2350 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2351 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2352 timeout_t lustre_msg_get_timeout(struct lustre_msg *msg);
2353 timeout_t lustre_msg_get_service_timeout(struct lustre_msg *msg);
2354 int lustre_msg_get_uid_gid(struct lustre_msg *msg, __u32 *uid, __u32 *gid);
2355 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2356 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2357 __u64 lustre_msg_get_mbits(struct lustre_msg *msg);
2358 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, __u32 buf);
2359 void lustre_msg_set_handle(struct lustre_msg *msg,
2360 struct lustre_handle *handle);
2361 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2362 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2363 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2364 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
2365 void lustre_msg_set_last_committed(struct lustre_msg *msg,
2366 __u64 last_committed);
2367 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2368 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2369 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2370 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2371 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2372 void lustre_msg_set_timeout(struct lustre_msg *msg, timeout_t timeout);
2373 void lustre_msg_set_service_timeout(struct lustre_msg *msg,
2374 timeout_t service_timeout);
2376 /* jobid/uid/gid process information to pack */
2378 char ji_jobid[LUSTRE_JOBID_SIZE];
2383 void lustre_msg_set_jobinfo(struct lustre_msg *msg, const struct job_info *ji);
2384 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2385 void lustre_msg_set_mbits(struct lustre_msg *msg, __u64 mbits);
2388 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2389 unsigned int newlen, int move_data)
2391 LASSERT(req->rq_reply_state);
2392 LASSERT(req->rq_repmsg);
2393 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2397 #ifdef LUSTRE_TRANSLATE_ERRNOS
2399 static inline int ptlrpc_status_hton(int h)
2402 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2403 * ELDLM_LOCK_ABORTED, etc.
2406 return -lustre_errno_hton(-h);
2411 static inline int ptlrpc_status_ntoh(int n)
2414 * See the comment in ptlrpc_status_hton().
2417 return -lustre_errno_ntoh(-n);
2424 #define ptlrpc_status_hton(h) (h)
2425 #define ptlrpc_status_ntoh(n) (n)
2430 /** Change request phase of \a req to \a new_phase */
2432 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2434 if (req->rq_phase == new_phase)
2437 if (new_phase == RQ_PHASE_UNREG_RPC ||
2438 new_phase == RQ_PHASE_UNREG_BULK) {
2439 /* No embedded unregistering phases */
2440 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2441 req->rq_phase == RQ_PHASE_UNREG_BULK)
2444 req->rq_next_phase = req->rq_phase;
2446 atomic_inc(&req->rq_import->imp_unregistering);
2449 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
2450 req->rq_phase == RQ_PHASE_UNREG_BULK) {
2452 atomic_dec(&req->rq_import->imp_unregistering);
2455 DEBUG_REQ(D_INFO, req, "move request phase from %s to %s",
2456 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2458 req->rq_phase = new_phase;
2462 * Returns true if request \a req got early reply and hard deadline is not met
2465 ptlrpc_client_early(struct ptlrpc_request *req)
2467 return req->rq_early;
2471 * Returns true if we got real reply from server for this request
2474 ptlrpc_client_replied(struct ptlrpc_request *req)
2476 if (req->rq_reply_deadline > ktime_get_real_seconds())
2478 return req->rq_replied;
2481 /** Returns true if request \a req is in process of receiving server reply */
2483 ptlrpc_client_recv(struct ptlrpc_request *req)
2485 if (req->rq_reply_deadline > ktime_get_real_seconds())
2487 return req->rq_receiving_reply;
2490 #define ptlrpc_cli_wait_unlink(req) __ptlrpc_cli_wait_unlink(req, NULL)
2493 __ptlrpc_cli_wait_unlink(struct ptlrpc_request *req, bool *discard)
2497 spin_lock(&req->rq_lock);
2498 if (req->rq_reply_deadline > ktime_get_real_seconds()) {
2499 spin_unlock(&req->rq_lock);
2502 if (req->rq_req_deadline > ktime_get_real_seconds()) {
2503 spin_unlock(&req->rq_lock);
2509 if (req->rq_reply_unlinked && req->rq_req_unlinked == 0) {
2511 spin_unlock(&req->rq_lock);
2512 return 1; /* Should call again after LNetMDUnlink */
2516 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
2517 req->rq_receiving_reply;
2518 spin_unlock(&req->rq_lock);
2523 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2525 smp_mb(); /* ensure client context is deleted before wakeing up */
2526 if (req->rq_set == NULL)
2527 wake_up(&req->rq_reply_waitq);
2529 wake_up(&req->rq_set->set_waitq);
2533 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2535 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2536 atomic_inc(&rs->rs_refcount);
2540 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2542 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2543 if (atomic_dec_and_test(&rs->rs_refcount))
2544 lustre_free_reply_state(rs);
2547 /* Should only be called once per req */
2548 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2550 if (req->rq_reply_state == NULL)
2551 return; /* shouldn't occur */
2553 /* req_repmsg equals rq_reply_state->rs_msg,
2554 * so set it to NULL before rq_reply_state is possibly freed
2556 spin_lock(&req->rq_early_free_lock);
2557 req->rq_repmsg = NULL;
2558 spin_unlock(&req->rq_early_free_lock);
2560 ptlrpc_rs_decref(req->rq_reply_state);
2561 req->rq_reply_state = NULL;
2564 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2566 return lustre_msg_get_magic(req->rq_reqmsg);
2569 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2571 if (req->rq_delay_limit != 0 &&
2572 req->rq_queued_time + req->rq_delay_limit < ktime_get_seconds())
2577 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2579 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2580 spin_lock(&req->rq_lock);
2581 req->rq_no_resend = 1;
2582 spin_unlock(&req->rq_lock);
2584 return req->rq_no_resend;
2587 int ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt);
2589 static inline struct ptlrpc_service *
2590 ptlrpc_req2svc(struct ptlrpc_request *req)
2592 LASSERT(req->rq_rqbd != NULL);
2593 return req->rq_rqbd->rqbd_svcpt->scp_service;
2596 /* ldlm/ldlm_lib.c */
2598 * Target client logic
2601 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg);
2602 int client_obd_cleanup(struct obd_device *obd);
2603 int client_connect_import(const struct lu_env *env,
2604 struct obd_export **exp, struct obd_device *obd,
2605 struct obd_uuid *cluuid, struct obd_connect_data *ocd,
2607 int client_disconnect_export(struct obd_export *exp);
2608 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2610 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2611 struct lnet_nid *prim_nid, int priority);
2612 int client_import_add_nids_to_conn(struct obd_import *imp,
2613 struct lnet_nid *nidlist,
2614 int nid_count, int nid_size,
2615 struct obd_uuid *uuid);
2616 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2617 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2618 void client_destroy_import(struct obd_import *imp);
2621 #ifdef HAVE_SERVER_SUPPORT
2622 int server_disconnect_export(struct obd_export *exp);
2625 /* ptlrpc/pinger.c */
2627 * Pinger API (client side only)
2630 enum timeout_event {
2633 struct timeout_item;
2634 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2635 int ptlrpc_pinger_add_import(struct obd_import *imp);
2636 int ptlrpc_pinger_del_import(struct obd_import *imp);
2637 struct ptlrpc_request *ptlrpc_prep_ping(struct obd_import *imp);
2638 int ptlrpc_obd_ping(struct obd_device *obd);
2639 void ping_evictor_start(void);
2640 void ping_evictor_stop(void);
2641 void ptlrpc_pinger_ir_up(void);
2642 void ptlrpc_pinger_ir_down(void);
2644 int ptlrpc_pinger_suppress_pings(void);
2646 /* ptlrpc/ptlrpcd.c */
2647 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2648 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2649 void ptlrpcd_wake(struct ptlrpc_request *req);
2650 void ptlrpcd_add_req(struct ptlrpc_request *req);
2651 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2652 int ptlrpcd_addref(void);
2653 void ptlrpcd_decref(void);
2655 /* ptlrpc/lproc_ptlrpc.c */
2657 * procfs output related functions
2660 const char *ll_opcode2str(__u32 opcode);
2661 const int ll_str2opcode(const char *ops);
2662 #ifdef CONFIG_PROC_FS
2663 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2664 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2665 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2667 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2668 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2669 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2673 /* ptlrpc/llog_server.c */
2674 int llog_origin_handle_open(struct ptlrpc_request *req);
2675 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2676 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2677 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2679 /* ptlrpc/llog_client.c */
2680 extern const struct llog_operations llog_client_ops;